INTRODUCTION

1.1 Project Overview

In today's bustling public environments, ensuring public safety through effective crowd management
has become an imperative. The rise in crowd-related incidents emphasizes the urgency of developing
an advanced crowd detection and management system. Traditional manual counting methods have
demonstrated limitations in accuracy and efficiency, particularly in complex and dynamic spaces. To
address these challenges, the "Crowd Watch" project aims to create a cutting-edge crowd detection
model using deep learning and CCTV cameras.

The central focus of the project is the critical need for maintaining public safety within crowded public
spaces. Existing manual methods for crowd management are not only time-consuming but also
susceptible to errors, especially in intricate settings. The proposed solution seeks to develop a
dependable and efficient system capable of accurately detecting and counting individuals across diverse
public areas.

The "Crowd Watch" project revolves around a set of core objectives that collectively aim to
revolutionize crowd management and enhance public safety. It begins with an in-depth exploration of
the existing literature within the domains of crowd detection, deep learning, and computer vision. This
comprehensive literature review not only identifies gaps and challenges but also sheds light on emerging
trends and best practices, effectively guiding the overarching approach of the project.

With insights gleaned from the literature review, the project focuses on a pivotal aspect: the
development of a sophisticated deep learning model tailored for crowd detection and counting using
CCTV footage. Leveraging the power of convolutional neural networks (CNNs), this model endeavors
to achieve precise density map estimation. The primary goal here is to enable accurate and efficient
crowd counting and detection within various public spaces.

Expanding the project's scope, a real-time safety alert system takes center stage. Crafted to operate in
real-time, this alert system is strategically designed to promptly notify authorities when predetermined
crowd thresholds are exceeded or breaches of established rules and regulations are detected. By

1
swiftly furnishing actionable insights during critical situations, this real-time response capability plays
a pivotal role in enhancing public safety.

Furthermore, the project undertakes the task of integrating the meticulously designed deep learning
model and the real-time safety alert system into a user-friendly software application. This application
caters to authorized personnel, furnishing them with direct access to real-time safety alerts. By
facilitating immediate and seamless communication, this application empowers swift decision-
making, thereby bolstering the efficiency of crowd management strategies.

Validation and testing constitute another pivotal phase of the project. Real-time data, harvested from
strategically placed CCTV cameras across diverse public spaces, serves as the testing ground. This
phase serves the dual purpose of assessing the deep learning model's accuracy in detecting and counting
crowds, while also evaluating the responsiveness and reliability of the alert system.

Anticipated outcomes span several critical deliverables. The literature review, elucidates existing
methodologies and techniques pertinent to crowd detection and management. The deep learning model
emerges as a pinnacle achievement, capable of accurately identifying and enumerating individuals
within crowded scenes. The operational real-time safety alert system stands as a beacon of proactive
vigilance, providing timely notifications to authorities regarding crowd thresholds and rule
infringements.

Moreover, the user-centric software application caters to authorized personnel, rendering them
instantaneous access to real-time safety alerts. Most crucially, the project's holistic validation and testing
phase culminate in the demonstration of the deep learning model's accuracy, the efficiency of the alert
system, and their pragmatic applicability.

In summation, the "Crowd Watch" project is poised to disrupt conventional crowd management
strategies by harnessing cutting-edge technologies. By seamlessly integrating deep learning and
computer vision, the project aspires to establish a robust system capable of effective crowd detection
and enhanced public safety. The real-time alert system's proactive approach further amplifies its
potential impact, ensuring swift intervention during critical scenarios. Through meticulous validation

2
and real-world testing, the project endeavors to validate its proposed solutions, thereby making a
substantial contribution to the realm of crowd management and public safety.

1.1.1 Technical terminology

Crowd Detection Model:
Utilizes deep learning techniques.
Specifically employs Convolutional Neural Networks (CNNs).
Aims for precise density map estimation.

Real-time Safety Alert System:

Operates in real-time.
Notifies authorities when predefined crowd thresholds are exceeded or rule breaches are detected.

Validation and Testing:

Involves real-time data from CCTV cameras.
Assessing accuracy of the deep learning model.
Evaluating responsiveness and reliability of the alert system.

User-Friendly Software Application:

Integrates the deep learning model and real-time safety alert system.
Provides authorized personnel with direct access to real-time safety alerts.nology

1.1.2 Problem statement

Challenge:
Inefficient traditional manual counting methods for crowd management.
Susceptibility to errors, especially in complex and dynamic public spaces.

1.1.3 Goal
Main Objective:
Develop a cutting-edge crowd detection model using deep learning and CCTV cameras.
Revolutionize crowd management and enhance public safety.

3
Specific Objectives:
Conduct an in-depth literature review on crowd detection, deep learning, and computer vision.
Develop a deep learning model for accurate crowd detection and counting.
Implement a real-time safety alert system for proactive intervention.
Integrate the model and alert system into a user-friendly software application.
Validate the proposed solutions through real-world testing.

1.1.4 Solution
Proposed Solution:
Develop a sophisticated deep learning model using CNNs for accurate crowd detection.
Implement a real-time safety alert system to notify authorities during critical situations.
Integrate the deep learning model and alert system into a user-friendly software application for
authorized personnel.

Anticipated Outcomes:
Literature review providing insights into existing methodologies.
Deep learning model for accurate crowd identification and enumeration.
Operational real-time safety alert system providing timely notifications.
User-centric software application for seamless communication.
Validation and testing demonstrating the accuracy and efficiency of proposed solutions.

1.2 Need Analysis

Crowd management and control is a critical aspect of public safety, particularly in crowded public
spaces such as airports, shopping malls, and sports stadiums. In recent years, there have been several
incidents of crowd-related accidents and crimes, highlighting the need for an efficient and effective
system for crowd detection and management. The proposed project aims to address this need by
developing a deep learning model for crowd detection and counting using CCTV cameras.The need for
such a system is evident in many real-world scenarios. For instance, in the wake of the COVID-19
pandemic, many governments have imposed restrictions on the number of people allowed in public
spaces. A system for crowd detection and counting could be instrumental in enforcing these

4
restrictions and preventing the spread of the virus. Similarly, in areas prone to riots or civil unrest, a
system for crowd detection and management could be used to prevent the gathering of large crowds and
mitigate the risk of violence. In such scenarios, the proposed system can be instrumental in monitoring
the crowd size and detecting any potential breaches of the law. If the crowd count exceeds the set limit
or violates any pre-set rules, the system can send real-time alerts to the local authorities, enabling timely
action to prevent any potential incidents.
The proposed system has significant potential to enhance public safety and security in various public
spaces. It can provide real-time alerts to authorities in case of any safety concerns, enabling a timely
response to any potential incidents. Furthermore, the system can be extended to include advanced
analytics and predictive modeling capabilities to provide insights into crowd behavior, such as peak
hours and average footfall, which can be used for better crowd management and planning.
In conclusion, there is a clear need for a reliable and efficient system for crowd detection and counting
in public spaces. The proposed project has significant relevance in the real world, and its potential
impact on public safety and security cannot be overstated. By leveraging the latest advancements in
deep learning and computer vision, the proposed system has the potential to revolutionize crowd
management and make public spaces safer and more secure for all.

1.3 Research Gaps

Real-World Validation and Generalization: While many crowd counting methods demonstrate
impressive accuracy on benchmark datasets, their performance in real-world scenarios with diverse
lighting conditions, camera angles, and crowd densities remains a challenge. Real-world validation and
generalization are crucial to assess the adaptability of crowd detection models. Addressing this gap
involves testing models on data collected from actual public spaces, as highlighted by Zhang et al.
[12] in their study on wide-area crowd counting using ground-plane density maps and multi-view fusion
CNNs.

Privacy-Preserving Crowd Analysis: The deployment of crowd detection systems raises concerns about
privacy infringement due to the potential identification of individuals within the crowd.

5
Research focusing on privacy-preserving crowd analysis methods is lacking. Techniques that can
extract crowd insights while safeguarding personal information deserve exploration. This gap is
supported by the ethical dimension of crowd management, as emphasized by Varior et al. [8] in their
study on scale-aware attention networks for crowd counting.

Robustness to Environmental Factors: Most crowd detection models focus on controlled environments
and may struggle to handle adverse weather conditions, occlusions, and other environmental factors that
affect the visibility of individuals within a crowd. Developing models that are robust to various
environmental challenges is imperative for ensuring reliable crowd detection in real-world settings. This
gap resonates with the need to address variations in crowd density and illumination conditions, as
discussed by Jiang et al. [2] in their study on trellis encoder-decoder networks for crowd counting.

Robustness to Occlusions and Crowdedness: Existing crowd detection models may struggle to
accurately count and locate individuals in highly crowded scenes with significant occlusions.
Developing methods that can handle occlusions and crowdedness while maintaining accuracy is a
research challenge. This gap resonates with the work of Liu et al. [3], who introduce AdCrowdNet with
an attention-injective deformable convolutional network to address variations in crowd density.

Multi-Camera Crowd Analysis: While many studies focus on single-camera setups, crowd analysis in
large areas often requires multiple camera viewpoints for comprehensive coverage. Research on
effectively fusing data from multiple cameras to achieve accurate crowd detection and behavior analysis
is limited. This gap aligns with the work of Zhang et al. [12], who propose multi-view fusion CNNs for
wide-area crowd counting.

1.4 Problem Definition and Scope

Problem Definition:

The "Crowd Watch" project aims to address the critical challenge of crowd detection, counting, and
management. The existing methods for crowd management often rely on manual crowd counting and
detection, which are prone to errors and inefficiencies. Moreover, these methods are time-consuming,
costly, and may not be suitable for large and complex environments. The lack of accurate and efficient

6
crowd detection and management systems poses significant risks to public safety, leading to accidents,
crimes, and potential security breaches.

The primary problem is to design and implement a reliable and efficient system for automated crowd
detection and counting using CCTV cameras. The system should be capable of identifying and counting
individuals within a given area and providing real-time alerts to authorities when crowd thresholds are
exceeded or rule violations occur. Additionally, the project aims to explore advanced analytics to gain
insights into crowd behavior, which can be used for better crowd management and planning.

Scope:

Literature Review and Gap Identification: Comprehensive literature review to identify crowd detection,
deep learning, and computer vision methods, highlighting research gaps.

Deep Learning Model Development: Focus on implementing accurate crowd detection and counting
using convolutional neural networks (CNNs) and density map estimation techniques.

Real-Time Safety Alert System: Development of a real-time alert system triggering notifications to
authorities upon surpassing crowd thresholds or rule violations for enhanced public safety.

Software Application Integration: Integration of the deep learning model and alert system into
user-friendly software, enabling authorized personnel to receive real-time alerts on mobile devices.

Validation and Testing: Rigorous validation and testing of solutions using real-world CCTV data to
assess accuracy, efficiency, and reliability across diverse conditions.
Exploration of Advanced Analytics: Investigation into advanced analytics to understand crowd behavior
patterns, peak hours, and footfall trends, contributing to effective crowd management strategies.

1.5 Assumptions and Constraints

TABLE 1: ASSUMPTIONS

7
S. No. Assumptions
1. CCTV Coverage: The project assumes that the target public spaces are equipped with
an adequate number of CCTV cameras that provide sufficient coverage of the areas of
interest. The quality and placement of these cameras are assumed to be appropriate for
capturing crowd dynamics effectively.
2. Stable Network Connectivity: It is assumed that there is a stable and reliable network
connectivity available for transmitting data from CCTV cameras to the central
processing system. Real-time alerts and data exchange between the cameras and the
system depend on uninterrupted network access.
3. Static Camera Positions: The project assumes that the CCTV cameras are static and
do not move once installed. The crowd detection and counting algorithms are designed
with this stationary camera setup in mind, and the system may require adjustments if
cameras are repositioned.
4. Predefined Crowd Thresholds: The system assumes that predefined crowd density
thresholds and rules are established based on safety regulations and standards. The
effectiveness of the real-time alert system depends on accurate threshold definitions.

5. Homogeneous Lighting Conditions: The project assumes relatively consistent lighting

conditions within the target public spaces. Extreme variations in lighting might impact
the performance of the crowd detection model and the accuracy of the generated
density maps.

TABLE 2: CONSTRAINTS
S. No. Constraints
1. Hardware Limitations: The project operates under the constraint of the hardware
available for implementation. Processing power, memory, and storage of the deployed
system may impact the complexity and real-time performance of the deep learning
model and alert system.

8
 2. Data Privacy and Security: The development of a real-time alert system involves the
collection and processing of sensitive data from CCTV cameras. The project must
adhere to strict data privacy and security regulations to prevent unauthorized access and
potential breaches.
3. Camera Quality and Calibration: The accuracy of crowd detection and counting
heavily relies on the quality and calibration of the CCTV cameras. Poor camera quality
or misalignment can lead to inaccurate results and affect the overall system
performance.
4. Environmental Conditions: Adverse environmental conditions such as heavy rain,
fog, or strong backlighting can hinder the visibility of individuals in the crowd. The
system should be able to handle such challenges to ensure reliable crowd analysis.
5. Processing Latency: Real-time alerts are subject to processing latency, which might
impact the timeliness of notifications to authorities. Minimizing latency while
maintaining accurate results is a challenge to be addressed.
6. Ethical Considerations: The project operates within the ethical constraints of crowd
monitoring and analysis. Balancing public safety with individual privacy is paramount,
and the system must adhere to ethical guidelines and regulations.
7. Legal and Regulatory Compliance: The project must operate within legal frameworks
governing surveillance, data collection, and privacy. Compliance with local, regional,
and national laws is essential during the design, deployment, and operation of the
system.

1.6 Standards

ISO 27001 and ISO 27018: These information security and cloud privacy standards are internationally
recognized and can be adopted by Indian organizations to enhance data security and privacy. Many
Indian companies already implement ISO 27001 to safeguard their information assets.

ISO 22320: This standard provides guidance on emergency management, which can be relevant in India
to ensure effective response systems during emergencies and crises.

9
IEEE Standards: These technical standards can be used as references for technology implementation in
India, but it's important to consider India-specific regulations and guidelines for wireless
communication, data protection, and safety.

NFPA 730: The recommendations in this guide can be adapted for premises security in India, but local
regulations and practices may also need to be considered.

EN 50132: European standards might not be directly applicable, but they can provide valuable insights
when designing and implementing CCTV surveillance systems in India.

NIST SP 800-53: While designed for U.S. federal systems, the security and privacy controls outlined in
this document can inform security practices in India, but local regulations should be considered as well.

ASTM E2533: This guide's principles for assessing the efficacy of emergency management can offer
insights for evaluating real-time safety alert systems in India.

1.7 Objectives

1. To study and explore the existing literature related to the field crowd detection and prediction.
2. To design and develop an efficient deep learning model for crowd detection and counting using
CCTV cameras.
3. To additionally extend the system for enhancing public safety by providing real-time alerts to
authorities in case of any safety concerns, including cases where crowds are not allowed in
certain areas.
4. To design a simple application allowing authorized personnel to receive safety alerts.
5. To verify and validate the proposed model on real-time data. Overall, the project aims to
improve public safety and security in various public spaces through efficient and reliable crowd
detection and counting.
6. To incorporate advanced analysis in our crowd monitoring system is to empower users with
dynamic and insightful visualizations of crowd data

10
1.8 Methodology Used

1) To study and explore the existing literature related to the field crowd detection and prediction.

● Conducting a literature review to identify relevant papers, articles, and research in the field of
crowd detection and prediction.
● Analyzing the literature and identifying gaps in the existing research.
● Determining the best approaches and techniques to design the proposed model.
● Summarizing the key findings in the literature review section of the report.

2) To design and develop a deep learning model for crowd detection and counting using CCTV
cameras.

● Collection and preprocessing a dataset of images or videos of crowded scenes.

● Training and fine-tuning the deep learning model using an appropriate CNNbased density map
estimation algorithm
● Optimizing the model's performance for real-time detection.
● Evaluating the model's performance using appropriate metrics such as precision, recall, MAE,
RMSE, etc.
● Implementing the model in a software application and test it on real-time data.

3) To additionally extend the system for enhancing public safety by providing real-time alerts to
authorities in case of any safety concerns, including cases where crowds are not allowed in
certain areas.
● Developing a rules-based system that triggers alerts to authorities when a predefined
threshold of people is reached or when a violation of rules and regulations is detected.
● Integrating the system with existing security systems to provide real-time alerts in case of
safety concerns such as suspicious behavior or security threats.
● Testing the system on real-time data and evaluating its performance.

11
 4) To design a simple application allowing authorized personnel to receive safety alerts directly
on their mobile devices.
● Determine target audience and needs
● Define application features and functionality
● Conduct user testing
● Use appropriate programming languages and development tools
● Test for functionality, usability, and security

5) To verify and validate the proposed model on real-time data.

● Collect real-time data from CCTV cameras in different public areas such as airports,
shopping malls, etc.
● Evaluating the model's performance on real-time data using appropriate metrics such as
precision, recall, and F1-score.
● Analyzing the model's performance under different conditions and in different environments.

1.9 Project Outcomes and Deliverables

1. A comprehensive literature review report on existing methods and techniques for crowd
detection and prediction.
2. An efficient deep learning model for crowd detection and counting using CCTV cameras.
3. An enhanced system for public safety that provides real-time alerts to authorities in
case of safety concerns, including cases where crowds are not allowed in certain areas.
4. A software system that integrates the developed models and tools for practical use.
5. A report summarizing the findings and results of the study, including the verification
and validation of the proposed models on real-time data.

1.10 Novelty of work

Real-World Validation Framework: The project proposes a comprehensive framework for real-world
validation of crowd detection models. This framework goes beyond benchmark datasets and involves
testing the developed deep learning model on actual data collected from diverse public spaces. This

12
approach accounts for factors such as lighting conditions, camera angles, and crowd dynamics, ensuring
that the model's performance is reliable and generalizable to real-world scenarios.

Privacy-Preserving Insights: Recognizing the privacy concerns associated with crowd analysis, the
project introduces techniques for privacy-preserving crowd insights. By extracting meaningful crowd
behavior information while safeguarding individual identities, the project aims to strike a balance
between effective crowd management and respecting personal privacy, thus contributing to the ethical
dimension of crowd analysis.

Robustness to Environmental Factors: Acknowledging the challenges posed by adverse weather

conditions and environmental factors, the project focuses on developing a deep learning model that is
robust to such challenges. By training the model to handle variations in illumination, occlusions, and
other real-world constraints, the project ensures accurate crowd detection in various environments,
contributing to the reliability of the system.

User-Centric Alert System: The project introduces a user-centric alert system that empowers authorized
personnel with real-time safety notifications. This system enhances communication and decision-
making during critical situations, enabling a timely response to potential incidents. The integration of
such a system adds practical value to the crowd detection and management process.

REQUIREMENT ANALYSIS
2.1 Literature Survey
2.1.1 Related Work
TABLE 3: LITERATURE SURVEY
S. Roll Name Paper Title Tools/ Findings Citation
No. Number Technology
1 Crowd Deep learning, The paper presents a crowd
Counting Convolutional counting approach that
Using Neural utilizes scale-aware attention
Scale-Aware Networks networks to achieve accurate
Attention (CNNs), crowd analysis,

13
 Networks Scale-aware demonstrating improved
attention performance in counting [1]
mechanisms individuals within crowded
scenes.
2 Wide-Area Computer The paper proposes a method
Crowd vision, Deep for crowd counting using
Shubh Counting via learning, ground-plane density maps
102003670 Mehtani Ground-Plane Convolutional and multi-view fusion CNNs,
Density Maps Neural demonstrating effective
and Networks crowd analysis in wide-area
Multi-View (CNNs), scenarios.
Fusion CNNs Multi-view [12]
fusion
techniques
3 Relational Computer The paper introduces a
Attention vision, Deep relational attention network
Network for learning, for crowd counting,
Crowd Relational highlighting the importance [11]
Counting attention of modeling relationships
network between individuals within a
crowd for improved counting
accuracy.
4 Point in, Box Computer The paper proposes a method
Out: Beyond vision, Deep that not only counts
Counting learning, individuals in crowds but also
Persons in Object identifies the most salient
Crowds detection person for improved crowd [4]
understanding and analysis
5 Almost The paper The paper introduces an
Unsupervised explores "almost unsupervised"
Learning for unsupervised approach for crowd counting,
Dense Crowd learning highlighting the potential of

14
 Counting techniques for using limited labeled data and
102017032 Satwik crowd counting abundant unlabeled data for [5]
Ghildiyal in dense accurate crowd counting in
environments. challenging scenarios.
6 Improved The paper The paper presents a novel
Crowd proposes a approach for crowd counting
Counting scale-adaptive using a scale-adaptive CNN,
Method convolutional demonstrating improved
Based on neural network accuracy in estimating crowd [6]
Scale-Adaptiv for crowd density across varying scales
e counting. in crowded scenes.
Convolutional
Neural
Network
7 Crowd The paper The paper introduces an
Counting and proposes trellis approach utilizing trellis
Density encoder-decode encoder-decoder networks for
Estimation by r networks for accurate crowd counting and
Trellis crowd counting density estimation, achieving
Encoder-Dec and density competitive performance on [2]
oder estimation. benchmark datasets.
Networks
8 Learning The paper The paper presents a method
from explores the to leverage synthetic data for
102017027 Lakshya Synthetic use of synthetic training crowd counting
Goel Data for data for crowd models, demonstrating
Crowd counting in improved performance in
Counting in real-world handling diverse and [10]
the Wild scenarios. challenging environments.
9 Adcrowdnet: The paper The paper presents the
An introduces a Adcrowdnet model, which
Attention-Inje novel utilizes attention mechanisms

15
 ctive attention-inject and deformable convolutions
Deformable ive deformable to enhance crowd
Convolutional convolutional understanding, achieving
Network for network for improved performance in [3]
Crowd crowd analysis. crowd density estimation and
Understandin counting tasks.
g
10 "W-net: The paper The W-net model is
Reinforced proposes the introduced as an approach for
U-Net for W-net model density map estimation,
Density Map for density map incorporating reinforcement
Estimation estimation, learning to enhance [7]
leveraging performance, offering
U-Net potential advancements in
architecture crowd counting accuracy.
and
reinforcement
102003668 Angad learning
Sidhu techniques.
11 Multi-scale The paper The proposed multi-scale
Attention introduces a attention network
Network for multi-scale demonstrates improved
Crowd attention performance in crowd
Counting network model counting tasks, highlighting
for crowd the potential of attention [8]
counting. mechanisms to enhance
accuracy in crowd analysis.

16
 12 NWPU-crow The paper The NWPU-crowd dataset
d: A introduces the provides a valuable resource
Large-Scale NWPU-crowd for evaluating and advancing
Benchmark dataset and crowd counting and
for Crowd benchmark for localization methods, [9]
Counting and crowd counting contributing to improved
Localization and accuracy and performance in
localization. crowd analysis.

2.1.2 Research Gaps of Existing Literature

Beyond Counting: Individual Identification: Research trends are shifting towards distinguishing
individuals within crowds, as seen in techniques like "Point in, Box Out" [4].

Synthetic Data and Domain Adaptation: Some studies explore using synthetic data to enhance crowd
counting accuracy in real-world settings [10].

Unsupervised and Semi-Supervised Approaches: Efforts to reduce reliance on labeled data are
evident, with approaches like almost unsupervised learning showing potential [5].

Privacy Preservation and Ethical Considerations: Privacy-preserving crowd analysis methods are
lacking [8], emphasizing the need for ethical techniques.

Real-World Validation and Generalization Challenges: Methods perform well on benchmarks but
struggle in real-world scenarios [2], emphasizing the need for validation.

Dynamic Crowd Behavior Modeling: Methods often neglect dynamic behaviors like surges [5],
highlighting the need for accurate temporal modeling.

Interactions and Collective Behaviors: Limited attention is given to interactions and collective
behaviors within crowds [11], warranting further exploration.

17
Robustness to Environmental Factors: Challenges with adverse conditions are acknowledged [2],
underscoring the need for robust models.

Dataset Availability and Benchmarking: Large-scale datasets like NWPU-Crowd [9] aid benchmarking
and standardization efforts.

2.1.3 Detailed Problem Analysis

In the realm of real-world crowd management systems and solutions, several critical problems have
serving crowd analysis, which is critical for balancing the need for public safety with the protection of
individuals' rights [8]. Addressing this concern is essential to ensure the ethical and responsible use of
crowd management technologies.

Additionally, existing crowd analysis methods primarily focus on static crowd scenes and may overlook
dynamic crowd behaviors, such as sudden surges, dispersals, and variations in crowd density over time
[5]. These dynamic aspects of crowd behavior can significantly impact safety and resource allocation
strategies. Incorporating temporal dynamics into crowd management models is essential to enhance the
accuracy of behavior prediction and resource allocation, ultimately leading to more effective crowd
control.

In summary, the problems identified within the domain of real-world crowd management systems and
solutions encompass challenges related to accurate crowd counting in varying conditions, limited
generalization capabilities, privacy infringement concerns, and the need to capture dynamic crowd
behaviors. Addressing these issues through advanced techniques and methodologies is crucial for
developing comprehensive and adaptable crowd detection and management solutions that can ensure
public safety, security, and efficient resource allocation in diverse public spaces.been identified that
hinder the effective monitoring, analysis, and response to crowd-related scenarios. These issues arise
from the limitations of existing methods in accurately detecting and managing crowds within public
spaces. One key challenge is the inability of conventional crowd counting techniques to handle diverse
crowd densities, dynamic behaviors, and varying camera angles [2]. These factors contribute to
inaccuracies in crowd size estimation, hindering resource allocation and emergency response planning.

18
Furthermore, many existing crowd detection systems fail to generalize well in real-world scenarios due
to their dependency on controlled environments and specific datasets. The lack of robustness to
environmental factors, such as adverse weather conditions, occlusions, and lighting variations, further
diminishes the reliability of these systems [12]. This deficiency in adapting to different conditions limits
their practical utility in a range of public spaces.

Privacy concerns also emerge as a significant obstacle in the deployment of crowd analysis systems.
The potential identification of individuals within crowds raises ethical questions about the infringement
of personal privacy. Current approaches often lack the capability to perform privacy-preserving
crowd analysis, which is critical for balancing the need for public safety with the protection of
individuals' rights [8]. Addressing this concern is essential to ensure the ethical and responsible use of
crowd management technologies.

2.1.4 Survey of Tools and Technologies Used

A comprehensive survey of tools and technologies used in the domain of crowd management and
analysis reveals a diverse landscape of approaches aimed at addressing the complexities of crowd
detection, monitoring, and response. These tools and technologies encompass a wide range of domains,
including computer vision, machine learning, data analytics, and communication systems. Here, we
explore some of the prominent tools and technologies employed in this field:

Computer Vision Frameworks:

Computer vision forms the foundation of many crowd management solutions. Open-source frameworks
like OpenCV provide a wide range of tools for image and video analysis, enabling tasks such as crowd
counting, density estimation, and object detection. These frameworks offer pre-built modules for feature
extraction, image processing, and visualization, facilitating the development of robust crowd analysis
algorithms.

Deep Learning Libraries:

Deep learning has revolutionized crowd analysis by enabling the creation of complex models for

19
crowd detection and behavior prediction. Libraries like TensorFlow and PyTorch provide a framework
for designing, training, and deploying deep neural networks. These libraries offer a variety of pre-
trained models and optimization techniques that can be tailored to specific crowd management tasks.

Convolutional Neural Networks (CNNs):

CNNs are a class of deep learning architectures specifically designed for image analysis tasks. They
have been extensively used for crowd counting and density estimation due to their ability to learn
hierarchical features from images. Researchers have explored various CNN architectures, including U-
Net, VGG, and ResNet, to achieve accurate crowd analysis results [1][2][4].

Object Detection Frameworks:

Object detection frameworks like YOLO (You Only Look Once) and Faster R-CNN (Region
Convolutional Neural Network) have been adapted for crowd management to identify and track
individuals within crowds. These frameworks enable real-time detection and localization of people,
contributing to crowd behavior analysis and resource allocation.

IoT Sensors and Data Streams:

Internet of Things (IoT) sensors and devices have been integrated into crowd management systems to
collect real-time data on crowd density, movement, and environmental conditions. This data is crucial
for informing decision-making and triggering automated safety alerts [9].

Edge Computing and Cloud Services:

Edge computing technologies process data locally on sensors or devices, reducing latency and enabling
real-time analysis. Cloud services are utilized for storing and processing large volumes of crowd-related
data, enabling advanced analytics and insights [10].

2.1.5 Summary

● Utilized SASNet model architecture: Employed a state-of-the-art SASNet model for accurate
crowd counting in images and live video, addressing challenges faced by conventional
methods.
● Custom Dataset Creation: Developed a unique custom dataset for fine-tuning the SASNet

20
 model, enhancing the model's performance on specific real-world scenarios.

● Alert System Integration: Introduced an alert system, sounding alarms and sending email
notifications when the crowd count surpasses a predefined threshold for an extended period.
This addresses the need for real-time validation and response capabilities, filling a gap
identified in the literature.

● Dashboard for Advanced Analytics: Designed and implemented a user-friendly dashboard

using Streamlit, providing advanced analytics such as real-time graphs of crowd counts at
different time scales. This addresses the limited attention given to user interfaces and real-time
analytics in existing literature.

2.1.1 Theory Associated With Problem Area

Computer Vision Fundamentals: Computer vision is the foundation of crowd detection. Concepts
such as image processing, feature extraction, object detection, and image segmentation are crucial for
identifying and delineating individuals within a crowd. Techniques like edge detection, contour
analysis, and region-based segmentation contribute to accurate crowd boundary detection.

Density Estimation: Density estimation is a core concept in crowd management. It involves estimating
the distribution of individuals within a crowd scene. Techniques like kernel density estimation and
Gaussian mixture models are commonly used to generate density maps that represent the spatial
distribution of people.

Convolutional Neural Networks (CNNs): CNNs are a key aspect of deep learning for crowd detection.
These neural networks are designed to automatically learn hierarchical features from images, making
them well-suited for object detection and classification tasks. The convolutional layers capture local
patterns, while pooling layers aggregate information, enabling robust feature extraction from crowd
images.

Attention Mechanisms: Attention mechanisms enable networks to focus on relevant regions of an

image. These mechanisms are valuable for crowd detection as they allow the network to allocate more
attention to densely populated areas and ignore less relevant regions. Techniques like self-attention and
spatial attention enhance the accuracy of crowd density estimation.

21
Transfer Learning: Transfer learning leverages pre-trained models on large datasets to improve the
performance of crowd detection models with limited data. Concepts like fine-tuning and feature
extraction aid in adapting pre-trained models to the specific task of crowd detection, saving training
time and improving accuracy.

Real-Time Processing: Real-time processing is vital for timely safety alerts. Concepts from signal
processing, parallel computing, and hardware acceleration ensure that the crowd detection and alert
system can process video streams efficiently and provide instantaneous alerts to authorities.

Privacy-Preserving Techniques: Theoretical frameworks related to privacy-preserving techniques,

such as differential privacy and homomorphic encryption, provide ways to analyze crowd data while
safeguarding the identity and personal information of individuals in the crowd.

Ethical Considerations: Ethical theories and principles guide the development of responsible crowd
detection systems. Concepts such as privacy, fairness, transparency, and accountability play a critical
role in ensuring that crowd management systems are ethically sound and aligned with societal values.

2.2 Software Requirement Specification

2.2.1 Introduction
2.2.1.1 Purpose

The primary purpose of the "Crowd Watch" project is to develop an advanced crowd detection and
management system that leverages cutting-edge technologies to enhance public safety and optimize
crowd-related decision-making. By employing deep learning models, real-time alert systems, and user-
friendly applications, the project aims to provide accurate crowd counting, behavior analysis, and safety
alerts in various public spaces. This technology-driven solution is designed to bridge the gap in current
crowd management practices by addressing challenges related to crowd monitoring, safety, and
responsiveness. Ultimately, the purpose of the project is to contribute to the creation of more

22
secure and efficiently managed public environments, promoting the well-being of individuals within
crowded spaces.

2.2.1.2 Intended Audience and Reading Suggestions

The project's target audience encompasses urban planners, security professionals, software developers,
researchers, academics, technology enthusiasts, and the general public interested in crowd management,
public safety, and technology. It aims to provide valuable insights and solutions tailored to enhance
crowd management strategies, security measures, software development practices, academic research,
technological innovations, and overall public safety awareness.

2.2.1.3 Project Scope

The scope of the "Crowd Watch" project involves the development of an advanced crowd detection and
management system using deep learning and computer vision. This includes designing a precise density
map estimation model, implementing a real-time safety alert system, creating a user-friendly software
application for authorized personnel, conducting validation and testing using real-world data, addressing
research gaps in the field, considering ethical considerations such as privacy, and providing
comprehensive documentation. The project aims to enhance crowd management, public safety, and
emergency response through innovative technologies and methodologies.

2.2.2 Overall Description

2.2.2.1 Product Perspective

FIGURE 1: BLOCK DIAGRAM OF CROWD WATCH

23
The "Crowd Watch" project approaches the challenge of crowd management from a product
perspective, focusing on the development of a comprehensive solution that leverages advanced
technologies for crowd detection, analysis, and safety. The project aims to create a user-friendly
software application that integrates a sophisticated deep learning model for crowd counting and
detection with a real-time safety alert system. This product perspective emphasizes the practical
application of cutting-edge techniques in real-world scenarios, addressing the needs of urban planners,
security personnel, and other stakeholders involved in crowd management. The software application
serves as the central deliverable, providing authorized users with instant access to accurate crowd
counts, behavior insights, and safety alerts, thereby enhancing crowd management strategies and
contributing to public safety.

2.2.2.2 Product Features

Advanced Crowd Detection: Utilises deep learning for accurate crowd counting and monitoring.
Density Map Estimation: Generates detailed density maps for precise crowd distribution analysis.
Real-time Safety Alerts: Notifies authorities of crowd thresholds and rule violations instantly.
User-friendly Interface: Intuitive interface for easy data access and behavior insights.

24
Adaptability: Works effectively in diverse real-world conditions, including lighting and crowd densities.
Privacy Preservation: Ensures crowd analysis while protecting individual identities.

2.2.3 External Interface Requirements

2.2.3.1 User Interfaces
The user interface of the "Crowd Watch" system is designed with simplicity and functionality in mind.
It offers an intuitive dashboard that provides real-time insights into crowd density, behavior patterns,
and safety alerts.
Authorized personnel can access the system on their mobile devices, receiving instant safety alerts and
actionable information. The user interface facilitates easy customization of alert thresholds and provides
historical data for trend analysis. Its user-centric design ensures efficient decision-making and effective
crowd management.

2.2.3.2 Hardware Interfaces

Closed-Circuit Television (CCTV) Cameras: The system integrates with CCTV cameras deployed in
public spaces, leveraging their video feeds for real-time crowd monitoring. The deep learning model
processes camera inputs to estimate crowd density and behavior.

Sensors: In environments where available, additional sensors like motion detectors or thermal cameras
can provide supplementary data for more accurate crowd analysis, enhancing the system's performance
and robustness.

Communication Systems: The system connects with communication systems used by authorities and
security personnel. This integration allows for the immediate transmission of safety alerts and
notifications when predefined crowd thresholds are reached.

Central Control Center: For large-scale events or centralized crowd management operations, the
system can interface with control center equipment, facilitating coordinated responses based on the
gathered insights.

Cloud Infrastructure: Cloud-based interfaces enable data storage, processing, and remote
accessibility. The system can leverage cloud computing resources for heavy-duty data analysis and deep
learning model training.

25
2.2.3.3 Software Interfaces
Graphical User Interface (GUI): The system offers an intuitive GUI that allows authorized personnel
to interact with the software application. Through the GUI, users can access real-time crowd data,
configure alert settings, visualize crowd density maps, and review historical data.

Alert Management System: The system's software interface enables users to set and customize alert
thresholds for crowd density and behavior. When these thresholds are exceeded, the interface triggers
real-time safety alerts, providing immediate notifications to relevant authorities.

Database Management System: The system interfaces with a database to store and retrieve historical
crowd data, alert logs, and other relevant information. This database-driven approach ensures efficient
data management and allows for retrospective analysis.

Communication Protocols: To transmit real-time safety alerts, the system integrates with
communication protocols such as SMS, email, or push notifications. These protocols ensure that
authorities receive timely notifications on their mobile devices.

Deep Learning Framework: The system interfaces with a deep learning framework for model training,
testing, and inference. This interface enables the deployment of sophisticated crowd detection
algorithms that accurately estimate crowd density and behavior.

Cloud Services: Leveraging cloud services, the system interfaces with cloud-based infrastructure for
data storage, scalability, and computation. This interface enhances the system's processing capabilities
and ensures remote accessibility.

Authentication and Authorization: The software interfaces with authentication and authorization
mechanisms to ensure secure access. Users with appropriate credentials can access different levels of
functionality and data based on their roles.

26
2.2.4 Other Non-functional Requirements
2.2.4.1 Performance Requirements
Real-time Processing: The system must be capable of processing video footage from multiple CCTV
cameras in real time. It should provide real-time crowd density estimation, behavior analysis, and safety
alerts to ensure timely response.

Accuracy: The crowd counting and behavior analysis algorithms must achieve a high level of accuracy,
minimizing false positives and false negatives. The system should accurately estimate crowd density
and identify anomalies in crowd behavior.

Scalability: The system should be scalable to accommodate a growing number of CCTV cameras and
public spaces. It should maintain performance even as the scale of data and processing requirements
increases.

Response Time: The system's response time for generating safety alerts and presenting crowd insights
should be within milliseconds to ensure immediate actions can be taken by authorities.

Alert Reliability: The real-time safety alert system should have a high level of reliability, ensuring that
alerts are promptly delivered to authorized personnel. The system should have mechanisms to handle
potential communication failures.

User Interface Responsiveness: The graphical user interface (GUI) should be responsive and provide
smooth interaction. Users should be able to access data, configure settings, and review alerts without
experiencing lag or delays.

Compatibility: The system should be compatible with various types of CCTV cameras, hardware
devices, and mobile platforms. It should be able to integrate with existing infrastructure and
technologies.

Security: The system should implement robust security measures to protect user data, sensitive

27
information, and communication channels. It should adhere to industry best practices for data
encryption and secure authentication.

Usability: The user interface should be intuitive and user-friendly, requiring minimal training for
authorized personnel to navigate and utilize its features effectively.

Resource Utilization: The system should optimize the utilization of computational resources to ensure
efficient processing without overloading hardware or slowing down other tasks on the system.

Adaptability: The system should be adaptable to different environmental conditions, lighting

variations, and crowd densities commonly found in real-world scenarios.

2.2.4.2 Safety Requirements

Privacy Protection: The system must adhere to strict privacy guidelines to prevent the identification and
tracking of individuals within the crowd. Personal information should not be collected or stored, and
the system should utilize anonymization techniques to ensure privacy.

Data Security: User data, system configuration, and communication channels must be secured using
encryption and authentication mechanisms. Measures should be in place to prevent unauthorized access
and data breaches.

Emergency Protocols: The system should have protocols in place to handle emergency situations, such
as mass evacuations or potential security threats. It should be capable of providing real-time alerts to
authorities and relevant stakeholders.

False Alarm Mitigation: The real-time safety alert system should be designed to minimize false alarms
and ensure that alerts are triggered only when there is a genuine cause for concern. This helps prevent
unnecessary panic and disruptions.

28
System Reliability: The system should undergo thorough testing and validation to ensure its reliability
in detecting crowd behavior and generating alerts. It should be able to operate consistently under various
conditions.

User Training: Authorized personnel using the system should receive proper training on its features,
functionalities, and protocols. This training ensures that users can effectively interpret alerts and take
appropriate actions.

Compliance with Regulations: The system must comply with local regulations, laws, and ethical
standards related to surveillance, data protection, and public safety. It should not infringe upon legal
rights and norms.

Accessibility: The user interface of the system should be designed to be accessible to users with
disabilities, ensuring inclusivity and equal access to information.

Regular Maintenance: Regular maintenance and updates should be conducted to address vulnerabilities,
bugs, and performance issues. Updates should be applied while minimizing disruption to ongoing
operations.

Ethical Considerations: The system should be designed and operated ethically, taking into account
potential biases and implications of crowd analysis. It should not contribute to discrimination or harm
based on demographic factors.

2.2.4.3 Security Requirements

Authentication and Authorization: The system should enforce strong authentication mechanisms to
ensure that only authorized personnel can access the system. User roles and permissions should be
clearly defined to control access to different functionalities.

User Data Protection: Personally identifiable information and sensitive user data should be stored
securely using encryption and proper data access controls. Data retention policies should be

29
established and followed to minimize data exposure.

Secure Communication: The system's communication channels, including data transmission and alerts,
should be secured using encryption protocols to prevent data leakage and unauthorized access.

Secure Configuration: The system's hardware and software components should be configured securely,
with default credentials changed, unnecessary services disabled, and security patches applied regularly
to address vulnerabilities.

Secure APIs: If the system interacts with external applications or services through APIs, these interfaces
should be designed with proper authentication, access controls, and rate limiting to prevent unauthorized
access and misuse.

Physical Security: If applicable, physical access to the hardware components of the system should be
restricted to authorized personnel only.

Regular Security Audits: The system should undergo regular security audits and penetration testing to
identify vulnerabilities and weaknesses. Any identified security issues should be promptly addressed
and mitigated.

User Awareness and Training: Authorized users should be educated about security best practices,
including password hygiene. Training helps prevent human errors that could compromise security.

Secure Backup and Recovery: Data backups should be performed regularly and stored securely to
ensure data recovery in case of data loss or system failure. Backup data should also be encrypted.

2.3 Cost Analysis

TABLE 4: COST ANALYSIS

S. No Details Cost

30
 1. AWS Server for Model Rs. 2000-3000

2. CCTV Module Rs. 2000-2500

3. Domain Purchase Rs. 300-500

1. The Web Application will be deployed on Heroku as a PWA. Due to its small scope of requests,
no cost is incurred here. The scope of this server can be expanded by deploying on paid AWS
EC2 instances in the future.
2. The DB is deployed on MongoDB Atlas. The free tier is used here due to the low amount of
CRUD requests which are well within the free tier. To handle a greater number of requests, paid
shared server tiers can be used in the future.
3. The ML Model is trained on Kaggle Notebook as it offers 16GB GPUs for free. This is also
enough since it is a low frequency operation.
4. ML model inference is done every time the prediction system is used. Since it is a moderately
frequent operation requiring elastic compute, an AWS T.3 Micro instance is dedicated for this
task. To handle higher frequency loads along with on server training, paid dedicated GPU
instances like AWS G4 or higher.

2.4 Risk Analysis

Data Breach and Privacy: Risk of unauthorized access leading to data breaches and privacy
violations. Mitigated by access controls, etc.
False Alerts: Potential for inaccurate alerts due to crowd counting errors. Minimized through
continuous validation and testing.
Inaccurate Counting: Variability in conditions affecting crowd counting accuracy. Mitigated by
real-world validation and adaptive algorithms.
Dependency on Data: System accuracy relies on external data sources' quality and availability.
Operational Challenges: Implementation complexities and training needs. Addressed through proper
training and support.
Ethical and Legal Concerns: Privacy issues and biases in algorithms. Ensured compliance with
regulations.

31
User Adoption: User acceptance impacting system success. Addressed through user training and
support.
Unpredictable Behavior: System challenges due to unexpected crowd behavior. Managed with
continuous monitoring and adaptation.
Integration Challenges: Compatibility issues while integrating with existing infrastructure.
Resource Limitations: Limited resources affecting real-time performance.
Regulatory Compliance: Ensuring adherence to regulations and standards.

METHODOLOGY ADOPTED
3.1 Investigative Techniques
TABLE 5: INVESTIGATIVE TECHNIQUES

S. No. Investigative Investigative Projects Examples Investigative Techniques

Projects Description
Techniques

1. Descriptive Descriptive investigations Descriptive investigations are

involve systematically essential to establish a
documenting existing scientific foundational understanding of the
phenomena, system models, existing landscape of crowd
algorithms, and concepts. In the detection methods. By
context of our project, this documenting and categorizing
technique is applied to different approaches, the project
understand and categorize the gains insights into the strengths,
various crowd detection methods limitations, and trends in the
available in the literature. It field. This understanding informs
involves summarizing the the project's decision-making
principles, architectures, and process when designing and
features of different crowd developing innovative crowd
counting algorithms and models. detection solutions.

32
 2. Comparative Comparative investigations Comparative investigations allow
involve making side-by-side the project to benchmark its
comparisons between different proposed crowd detection
objects, methods, or phenomena. solutions against existing
In our project, this technique is state-of-the-art methods. This
used to compare the performance process helps identify gaps and
of different crowd counting opportunities for improvement.
algorithms and models. By By rigorously comparing
quantitatively assessing metrics different approaches, the project
such as accuracy, robustness, and can highlight its advancements
efficiency, the project can and contribute novel insights to
identify which methods excel the field of crowd management.
under specific conditions.

3. Experimental Experimental investigations Experimental investigations

involve designing and conducting enable the project to validate
controlled studies to test hypotheses and assess the impact
hypotheses and uncover causal of specific variables on crowd
relationships. For our project, detection performance. By
experimental techniques involved systematically altering
generating synthetic crowd data preprocessing techniques, or
with varying densities and model architectures, the project
attributes to evaluate how can gain a deeper understanding
different models respond to of factors influencing accuracy
diverse scenarios. and make informed decisions
about the best approaches to
implement in real-world
scenarios.

3.2 Proposed Solution

33
1. Literature Review and Gap Identification:
The project begins with an extensive literature review to explore existing methods and technologies
related to crowd detection, deep learning, and computer vision. By identifying research gaps and
challenges in the field, this phase informs the project's approach and lays the foundation for
innovative solutions.

2. Deep Learning Model Development:

Building upon insights from the literature review, the project focuses on the design and
implementation of a sophisticated deep learning model for crowd detection and counting.
Convolutional neural networks (CNNs) are employed to accurately estimate crowd density maps,
enabling precise crowd analysis. This model utilizes both spatial and contextual information to
enhance accuracy in diverse scenarios.

3. Real-Time Safety Alert System:

The project's scope extends beyond crowd detection to include the development of a real-time safety
alert system. This system is designed to monitor crowd sizes and detect potential safety violations in
real-time. When predefined crowd thresholds are exceeded or rule violations are detected, instant
alerts are triggered and sent to authorized personnel. This capability empowers authorities to take
swift actions in critical situations, enhancing public safety.

4. Software Application Integration:

The deep learning model and real-time alert system are seamlessly integrated into a user-friendly
software application. This application provides authorized personnel with direct access to real-time
safety alerts on their mobile devices. The intuitive interface enables efficient decision-making by
presenting actionable insights in a clear and organized manner.

5. Validation and Testing:

To ensure the reliability and effectiveness of the proposed solutions, real-time data collected from
CCTV cameras deployed in various public spaces are used for validation and testing. This phase
assesses the accuracy of the deep learning model in detecting and counting crowds across diverse
conditions. It also evaluates the responsiveness and dependability of the alert system.

34
6. Exploration of Advanced Analytics:
As an added feature, the project explores the potential of advanced analytics. By analyzing crowd
behavior patterns, peak hours, and footfall trends, authorities can optimize crowd management
strategies. This information enables proactive decision-making and resource allocation.

3.3 Work Breakdown Structure

TABLE 6: WORK BREAKDOWN STRUCTURE

Sr. Activities Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
No.
1. Identification, Formulation and
planning of the Project
2. Gathering Literature, reference
materials and literature survey
3. Survey of techniques and tools
needed for project development
and literature survey.
4. Data Gathering
5. Dataset creation
6. Model Selection and Training
7. Fine Tuning the model and
optimizing for real time output
8. Build a User Interface/WebApp
9. Adding User Dashboard and
Analytics features to the WebApp
10. Integrating the model with the
WebApp and enabling alert
functionality

35
3.4 Tools and Technology
Python: Used as the primary programming language for implementing deep learning models, data
processing, and software development.
TensorFlow: Employed for building and training deep learning models, including convolutional
neural networks (CNNs) for crowd detection and counting.
Keras: Utilized as a high-level neural networks API integrated with TensorFlow for simplifying
model design and implementation.
OpenCV: Applied for image and video processing tasks, including pre-processing of CCTV footage
and extracting features relevant to crowd analysis.
GitHub: Used for version control and collaborative development, allowing multiple team members to
work on the project simultaneously.
Jupyter Notebook: Employed for prototyping and experimentation with various deep learning
architectures and algorithms.
RESTful APIs: Used to establish communication between the real-time safety alert system, the deep
learning model, and the mobile application.
Data Visualization Libraries (e.g., Matplotlib, Plotly): Utilized to create visual representations of
crowd analysis results, trends, and patterns for better understanding.
Cloud Services (e.g., AWS, Google Cloud): Explored for potential deployment of the deep learning
model and hosting the real-time safety alert system to ensure scalability and availability.
Machine Learning Libraries (e.gScikit-learn): Investigated for potential integration of additional
machine learning algorithms to enhance crowd behavior prediction and analysis.
Web Frameworks (e.g., Flask, Django): Considered for building a web-based interface for
administrators to monitor and manage the real-time safety alert system.

DESIGN SPECIFICATIONS
4.1 System Architecture
Tier Architecture Diagram:
FIGURE 2: TIER ARCHITECTURE DIAGRAM OF CROWD WATCH

36
The first tier is the CCTV camera. This is responsible for capturing images or videos of the scene.
The second tier is the cloud. This is where the images or videos are stored and processed. The third
tier is the web/app interface. This is where the user interacts with the system to view the predicted
crowd count.

Here is an explanation of each tier:

CCTV camera: This is responsible for capturing images or videos of the scene. The camera can be
either fixed or pan-tilt-zoom. The images or videos captured by the camera are then sent to the cloud.

37
Cloud: This is where the images or videos are stored and processed. The cloud has the computing
power to process the images or videos and generate the predicted crowd count.
Web/app interface: This is where the user interacts with the system to view the predicted crowd
count. The user can also use the interface to configure the system settings.

4.2 Design Level Diagrams

Data Flow Diagrams:

Level 0:
FIGURE 3: DATA FLOW DIAGRAM LEVEL 0 OF CROWD WATCH

The main components of the system are:

1) Camera: This is responsible for capturing images or videos of the scene.
2) Image processing: This is where the images or videos are processed to extract features that
can be used to count the crowd.
3) Crowd counting: This is where the crowd is counted based on the features extracted from
the images or videos.
4) Database: This is where the crowd count data is stored.
5) User interface: This is where the user interacts with the system to view the crowd count
data.

The flow of data in the system is as follows:

1) The camera captures images or videos of the scene.
2) The images or videos are sent to the image processing component.

38
 3) The image processing component extracts features from the images or videos.
4) The crowd counting component counts the crowd based on the features extracted from the
images or videos.
5) The crowd count data is stored in the database.
6) The user can view the crowd count data through the user interface.

Level 1:
FIGURE 4: DATA FLOW DIAGRAM LEVEL 1 OF CROWD WATCH

The main sub-processes of the system are:

1) Image acquisition: This is the process of capturing images or videos of the scene.

2) Image preprocessing: This is the process of preparing the images or videos for crowd counting.
This may involve tasks such as noise removal, image enhancement, and segmentation.
3) Feature extraction: This is the process of extracting features from the images or videos that can
be used to count the crowd.

39
4) Crowd counting: This is the process of counting the crowd based on the features extracted from
the images or videos.
5) Data storage: This is the process of storing the crowd count data in a database.

6) User interface: This is the process of allowing the user to interact with the system to view the
crowd count data.

The flow of data in the system is as follows:

1) The images or videos are captured by the camera.
2) The images or videos are preprocessed.
3) Features are extracted from the images or videos.
4) The crowd is counted based on the features extracted from the images or videos.
5) The crowd count data is stored in the database.
6) The user can view the crowd count data through the user interface.

Here are some additional details about the sub-processes of the system:
1) Image acquisition: The images or videos can be captured by a CCTV camera or a mobile
phone. The resolution of the images or videos will affect the accuracy of the crowd count.
2) Image preprocessing: The image preprocessing step can remove noise, enhance the image,
and segment the image into different regions. This will help to improve the accuracy of the
crowd counting algorithm.
3) Feature extraction: The feature extraction step extracts features from the images or videos
that can be used to count the crowd. Some common features include pixel intensity, edge
density, and texture information.
4) Crowd counting: There are a variety of algorithms that can be used to count the crowd. Some
common algorithms include density estimation, object tracking, and deep learning.
5) Data storage: The crowd count data can be stored in a relational database or a NoSQL
database. The database should be able to handle large amounts of data and provide fast
queries.
6) User interface: The user interface should be easy to use and provide the user with the
information they need. The user interface should also be able to handle multiple users.

40
 Class Diagram:

FIGURE 5: CLASS DIAGRAM OF CROWD WATCH

The main classes in the system are:

1) Image: This class represents an image or video. It has attributes such as the image file name,
the image dimensions, and the number of people in the image.
2) CrowdCounter: This class is responsible for counting the crowd in an image. It uses a variety
of algorithms to count the crowd, such as density estimation, object tracking, and deep
learning.
3) Database: This class represents the database that stores the crowd count data. It has methods
for storing, retrieving, and updating the crowd count data.
4) UserInterface: This class represents the user interface of the system. It allows the user to
view the crowd count data and to configure the system settings.

The relationships between the classes are as follows:

Image has a CrowdCounter. This means that each image has an associated crowd counter.
CrowdCounter uses Database. This means that the crowd counter needs to access the database to

41
store and retrieve the crowd count data.

UserInterface uses CrowdCounter and Database. This means that the user interface needs to access
the crowd counter and the database to display the crowd count data and to configure the system
settings.

4.3 User Interface Diagrams

Swimlane Diagram:
FIGURE 6: SWIMLANE DIAGRAM OF CROWD WATCH

The swimlanes in the diagram are:

1) User: This swimlane represents the user of the system. The user can login to the web app and
make requests for information, such as retrieving snapshots or viewing analysis.
2) Web App: This swimlane represents the web app. The web app receives requests from the user
and sends requests to the CCTV DVR. The web app also displays information to the user, such
as snapshots and analysis.
3) CCTV/DVR: This swimlane represents the CCTV DVR. The CCTV DVR stores the video

42
 footage and can be used to retrieve snapshots. The CCTV DVR can also be used to generate
analysis, such as the number of people in a scene.

The arrows in the diagram represent the flow of information. For example, the arrow from the user to
the web app represents the user making a request to the web app. The arrow from the web app to the
CCTV/DVR represents the web app sending a request to the CCTV DVR.

Here are some additional details about the swimlanes in the diagram:
1) User: The user can login to the web app and make requests for information, such as retrieving
snapshots or viewing analysis. The user can also configure the system settings, such as the
threshold for detecting a crowd.
2) Web App: The web app receives requests from the user and sends requests to the CCTV DVR.
The web app also displays information to the user, such as snapshots and analysis. The web app
is responsible for ensuring that the user has the correct permissions to access the information
they requested.
3) CCTV/DVR: The CCTV DVR stores the video footage and can be used to retrieve snapshots.
The CCTV DVR can also be used to generate analysis, such as the number of people in a scene.
The CCTV DVR is responsible for ensuring that the video footage is secure and that only
authorized users can access it.

Use Case Diagram:

FIGURE 7: USE CASE DIAGRAM OF CROWD WATCH

43
The actors in the system are:
1) User: This actor represents the user of the crowd counting system. The user can view the
crowd count data and configure the system settings.
2) System: This actor represents the crowd counting system itself. The system performs the tasks
of counting the crowd and storing the crowd count data.

44
The use cases in the system are:
1) View Crowd Count Data: This use case allows the user to view the crowd count data for a
particular location or time period.
2) Configure System Settings: This use case allows the user to configure the system settings, such
as the threshold for detecting a crowd.

The arrows in the diagram represent the relationships between the actors and the use cases. For example,
the arrow from the User actor to the View Crowd Count Data use case represents the User actor being
able to perform the View Crowd Count Data use case.

Here are some additional details about the use cases in the diagram:
1) View Crowd Count Data: The View Crowd Count Data use case allows the user to view the
crowd count data for a particular location or time period. The user can view the crowd count
data in a variety of formats, such as a table, a graph, or a map.
2) Configure System Settings: The Configure System Settings use case allows the user to configure
the system settings, such as the threshold for detecting a crowd. The threshold for detecting a
crowd is the minimum number of people that must be present in a scene before the system will
count the crowd.

IMPLEMENTATION AND EXPERIMENTAL RESULTS

5.1 Experimental Setup (or simulation)

Hardware Requirements:
Computer or server for running the crowd monitoring system.
Webcam or external camera for capturing live video feeds.

Software and Frameworks:

Python environment with required libraries (OpenCV, Streamlit, PyTorch, etc.).
Database management system (SQLite) for storing crowd count data.
Web browser for accessing the Streamlit-based user interface.
Email server configuration for sending notifications.

45
Data Collection and Annotation:
Annotated benchmark datasets with accurate crowd counts for training the model.
Custom dataset for fine-tuning the SASNet model, including images and videos with varying crowd
densities.

SASNet Model Integration:

Implement the SASNet model into the system, ensuring compatibility with the chosen Python
environment.
Fine-tune the model using the custom dataset and optimize hyperparameters for crowd counting.

Video Processing:
Integrate OpenCV for video processing to handle live video feeds from webcams or external cameras.
Address challenges related to video input, ensuring smooth processing and accurate crowd counting.

Web App Development:

Develop the web application using Streamlit to allow user registration, video upload, and live webcam
monitoring of crowd count.
Implement the analytics dashboard for real-time crowd count visualization.

Database Setup:
Set up an SQLite database to store user and crowd analytics related data.

Email Notification System:

Configure the email notification system to send alerts when crowd counts exceed predefined
thresholds.
Ensure the accurate generation of email notifications, including timestamps and crowd count details.

Real-Time Analytics:
Integrate real-time analytics principles to generate dynamic graphs on the analytics dashboard.
Implement functionality to process and update crowd count data in real-time.

46
Security Measures:
Implement user authentication and authorization mechanisms to ensure secure access to the web
application.

5.2 Experimental Analysis

5.2.1 Data

Data Sources:
Our experimental analysis relies on a combination of online benchmark datasets and a custom dataset
curated from images captured within our campus environment. The online benchmark datasets serve
as a reference for standard crowd scenarios, while our custom dataset is tailored to reflect the unique
characteristics of our campus crowd dynamics. The combination of these datasets enhances the
robustness and adaptability of our crowd monitoring system.

Data Cleaning:
Data cleaning procedures involves removal of irrelevant or redundant images. This process ensures
that the dataset used for fine-tuning the SASNet model is free from outliers or noise that could impact
the model's performance negatively.

Data Pruning:
Data pruning involves the systematic removal of images that do not contribute meaningfully to the
training process. Useless or irrelevant images, such as those with no discernible crowd or containing
irrelevant background noise, are carefully identified and eliminated from the dataset.

Feature Extraction Workflow:

We extract relevant features from each image, such as crowd density, spatial distribution, and
contextual information. These features serve as the input for the deep learning model, enabling it to
learn and generalize patterns related to crowd counting.

47
*?* 5.2.2 Performance Parameters (Accuracy Type
Measures/ QOS Parameters depending upon the type
of project)

Dataset Name Mean Average Error Mean Square Error

ShanghaiTech Part A 53.59 88.38

ShanghaiTech Part B 6.35 9.9

UCF_CC_50 161.4 234.46

UCF_QNRF 85.2 147.3

JHU-CROWD 61.36 274.62

Our Dataset 23.5 48.83

Performance of our model on different datasets

5.3 Working of the project

5.3.1 Procedural Workflow

1. User Authentication:
Users initiate the workflow by logging into the web application. Secure user authentication
mechanisms guarantee access control.

2. Input Options:
Upon successful login, users are presented with two distinct options:
Live Video Monitoring:
Users can opt to utilize their webcam or an external camera for live video analysis. This feature
enables real-time observation of crowd dynamics.
Image Upload:
Alternatively, users can upload images for crowd counting. This accommodates scenarios where a
historical or predefined image is available for analysis.

3. Live Feed Processing:

In the case of live video monitoring, the system processes the incoming video feed using the SASNet
model for object detection. This involves continuous analysis of the live feed to estimate crowd counts

48
in real time.

4. Data Retrieval from SASNet Model:

The SASNet model continuously processes the live feed, extracting relevant data. The web app retrieves
this data, ensuring the display of accurate and up-to-date crowd counts on the user interface.

5. Real-Time Crowd Counts:

Users receive instantaneous crowd counts displayed on the web app interface. The system continuously
updates these counts based on the live video analysis, offering an immediate assessment of the current
crowd situation.

6. Alarm and Notification System:

If the crowd count surpasses a predefined threshold for a specified duration, an alarm is triggered.
Simultaneously, an email notification is sent to the concerned person, containing crucial details such
as the crowd count, timestamp, and date.

7. Advanced Analytics Dashboard:

The web app features an analytics dashboard providing users with live crowd count data in the form of
dynamic graphs. These graphs offer insights into crowd trends at different time scales.

8. User Interaction and Monitoring:

Users have the flexibility to interact with the live feed, toggle between different views, and monitor
crowd counts dynamically.

49
5.3.2 Algorithmic Approaches Used (Mention algorithms, pseudocodes with
explanation)
Model Related Algorithms:
● Extract features from the input image using the pretrained VGG16 network, dividing them into
five stages.

● Decode the features using the decoder module, which consists of five upsampling and
concatenation operations, followed by two convolutional layers. For each stage, predict the
density map and confidence map using separate heads.

● Aggregate the confidence maps by concatenating them and using sigmoid and softmax
functions.

● Perform soft selection on the density maps based on the confidence maps, by multiplying each
density map with its corresponding confidence map, and summing them up. Return the density
map as the final output.

Real Time Predictions:

● Video Initialization:
○ If video_option is "Webcam," initialize video capture using the default webcam.
○ If video_option is "Video Upload," prompt the user to upload a video file, save it to a
designated folder, and initialize video capture from the uploaded file.
○ Display information about the uploaded video.

50
 ● Object Detection Loop:
○ Continuously capture frames from the video feed.
○ Resize each frame for faster processing and convert it to a blob for input to a
pre-trained neural network (assumed to be previously loaded as net).
○ Run a forward pass through the neural network to detect persons in the frame.
○ Draw circles around detected persons, update person count, and display it on the frame.
○ Convert the frame to RGB and update a Streamlit image placeholder.

● Graphical Display:
○ Update a line chart in a Streamlit chart placeholder with the live human count data.

● Alert Mechanism:
○ Check if the current person count exceeds a predefined threshold and has been
sustained for a specified duration.
○ If the conditions are met, play an alert sound asynchronously, display an alert message,
and log relevant data (date, hour, and count).
○ Send an email notification with details of the high person count.

● Data Logging and Release:

○ Log maximum person count data to relevant database tables (max_count_data,
dashboard_day, dashboard_hour).
○ Release the video capture when the function exits.

*?* 5.3.3 Project Deployment (Can be explained using Component and

Deployment Diagrams)

5.3.4 System Screenshots

51
Landing page of the web app

Sign up page for a new user

52
 Login page for already registered users

Users not logged in cannot access dashboard and prediction pages

Optional Feedback page

53
 A user that has logged in

Prediction page available for logged in users

54
 Dropdown available for users

Real time crowd count below threshold set

55
Real time crowd count above set threshold

Hour-Wise Graph for crowd count

56
 Day-Wise graph for crowd count

5.4 Testing Process

5.4.1 Test Plan

Objective: Validate the system's performance, functionality, and user experience. Our primary focus is
to ensure that the crowd counting accuracy is maintained, real-time analytics are displayed correctly,
the alert system is reliable, and the user interface is user-friendly.

Scope: Test the whole system, including crowd counting, real-time analytics, live video feed
monitoring, alert system, and user interface.

Testing Environment: Our testing environment includes a server running Streamlit framework, a

57
SQLite database, and a machine learning environment for running the SASNet crowd counting model.
We use Python for automating tests and Streamlit for UI testing.

Roles and Responsibilities: Testers are responsible for developing and executing test plans, tracking
and reporting bugs, and verifying bug fixes. Developers are responsible for fixing reported bugs and
verifying the fixes. Product owners are responsible for communicating testing progress and priorities
to stakeholders.

5.4.2 Features to be tested

Crowd Counting Accuracy: Testing the accuracy of the crowd counting model in various scenarios,
including different crowd densities, lighting conditions, and dynamic behaviors.

Real Time Prediction: Ensuring that the video playback is smooth and works near real time.

Alert System: Validating the effectiveness of the alert system by triggering alarms and sending email
notifications in response to predefined crowd count thresholds.

Dashboard Functionality: Testing the features and functionalities of the dashboard, ensuring that
real-time analytics, and video upload/live monitoring options work as intended.

User registration/login: Authorization and Authentication of user login and sign up details.

5.4.3 Test Strategy

Integration Testing: Verify the seamless integration of components, ensuring that the SASNet model,
alert system, dashboard, and database interact effectively.

Performance Testing: Assess the system's performance under different load conditions, evaluating its
responsiveness and resource utilization.

Usability Testing: Evaluate the user interface and overall user experience, ensuring that the web app is
intuitive and user-friendly.

5.4.4 Test Techniques

Black Box Testing: Test the system as a black box, focusing on the behavior and output of the system
without knowledge of the internal components.

Gray Box Testing: Perform gray box testing to understand the system's internal architecture, design,
and logic, which can help uncover potential bugs and issues.

Acceptance Testing: Perform acceptance testing to validate that the system meets the acceptance
criteria and user stories.

End-to-End Testing: Validate the entire system workflow, from crowd counting to alert generation and
dashboard display.

58
*?*5.4.5 Test Cases
Develop specific test cases for each feature, outlining input conditions, expected outcomes, and steps
to reproduce.

Include edge cases and scenarios identified during the literature review and project development to
ensure comprehensive testing.

Include non-functional test cases covering performance, reliability, and scalability.

?*?5.4.6 Test Results

Document the test results, including test case execution, test summary, and final verdict (pass or fail).

Document any issues, bugs, or defects found during testing.

Document the steps taken to resolve the issues and the results of retesting.

Provide a summary of system performance, functionality, and user experience.

?*? 5.5 Results and Discussions

● Successfully collected and generated a comprehensive real-world dataset, encompassing
diverse crowd scenarios within our campus environment.

● Demonstrated the SASNet model's capability to predict crowd counts fairly accurately in
real-time scenarios, validating its effectiveness for live video analysis and image uploads.

● Implemented a robust alert system triggered by predefined crowd count thresholds, ensuring
timely notifications for potential crowd-related incidents.

● Utilized advanced data analytics to present real-time graphs and visualizations, offering users
dynamic insights into crowd trends and enhancing the overall decision-making process.

add graphs of our model vs diff models

5.6 Inferences Drawn

Our crowd monitoring system has made significant strides thanks to the implemented procedural
workflow, which involves live video monitoring and the uploading of images. This process has
yielded considerable success in our endeavors. One of the key achievements of our system has been
the effective utilization of the SASNet model for accurate object detection and crowd counting during
real-time campus events. Furthermore, the integration of an alert system has demonstrated its efficacy,
promptly notifying security personnel when unexpected high-density crowds occur. The robust data
analytics, including real-time graph generation, have also empowered users with dynamic insights into
crowd behavior.

59
While there have been numerous accomplishments, some challenges still persist. These include
limited generalization to diverse environments and dependency on predefined thresholds. To address
these issues, we plan to expand our dataset to include various crowd scenarios. These findings
underscore the successes of our system while also emphasizing the ongoing commitment to
overcoming obstacles for continuous improvement of our crowd monitoring solution.

5.7 Validation of Objectives

S.No. Objectives Status

1. Studied and explored the existing literature related to the field crowd detection and Yes
prediction.

2. Designed and developed an efficient deep learning model for crowd detection and Yes
counting using CCTV cameras.

3. Extended the system for enhancing public safety by providing real-time alerts to Yes
authorities in case of any safety concerns, including cases where crowds are not
allowed in certain areas.

4. Designed a simple application allowing authorized personnel to receive safety Yes

alerts.

5. Verified and validated the proposed model on real-time data. Overall, the project Yes
aims to improve public safety and security in various public spaces through efficient
and reliable crowd detection and counting

6. Incorporated advanced analysis in our crowd monitoring system is to empower Yes

users with dynamic and insightful visualizations of crowd data

CONCLUSIONS AND FUTURE SCOPE

6.1 Conclusions

60
Enhanced Public Safety: The project introduces an innovative approach to crowd management by
leveraging deep learning techniques and real-time alert systems. This leads to improved public safety
during events and gatherings.

Accurate Crowd Analysis: The developed deep learning model accurately detects and counts individuals
within crowded scenes, providing valuable insights into crowd density and behavior.

Real-Time Alerts: The real-time safety alert system ensures timely notifications to authorities when
crowd thresholds are exceeded or rule violations occur, enabling proactive responses to potential
incidents.

User-Friendly Interface: The user-centric web application empowers authorized personnel to receive
safety alerts.

Practical Viability: The project's solutions are validated and tested using real-world data, showcasing
their accuracy, reliability, and adaptability to various environmental conditions.

Scalability and Future Potential: The use of cloud services and modern technologies positions the
project for scalability and potential integration with other smart city systems.

Ethical Considerations: The project acknowledges privacy concerns and emphasizes privacy-
preserving crowd analysis methods to ensure ethical implementation and compliance with regulations.

Cross-Disciplinary Impact: The project bridges the gap between technology, public safety, and urban
planning, offering insights and solutions that benefit urban planners, security professionals, researchers,
and the general public.

6.2 Economical/Social Benefits

Economical Benefits:

61
Cost-Efficient Crowd Management: The project's technology optimizes resource allocation by
providing real-time insights, minimizing the need for excessive security personnel and resources.

Reduced Operational Costs: Improved crowd analysis and predictive modeling enable event organizers
to optimize staffing and logistics, resulting in reduced operational expenses.
Potential Revenue Generation: Efficient crowd management attracts more visitors to events, boosting
local businesses and generating additional revenue for the host city.

Social Benefits:
Enhanced Public Safety: The real-time safety alert system and accurate crowd analysis contribute to
safer and more secure public gatherings and events.

Improved Emergency Response: Swift alerts and real-time data empower authorities to respond
effectively to emergencies, ensuring the safety of attendees and participants.

Better Event Experiences: Well-managed crowds lead to better attendee experiences, fostering positive
memories and encouraging repeat participation in events.

6.3 Reflections
In conclusion, the development and implementation of our crowd monitoring system have provided
valuable insights and reflections on the intersection of computer vision, deep learning, and real-time
analytics. The successful integration of the SASNet model, coupled with the utilization of diverse
datasets, has demonstrated the system's efficacy in accurately detecting and counting crowds in live
video feeds and uploaded images. The proactive alarm and notification system, triggered by predefined
crowd count thresholds, enhances the system's responsiveness in critical situations. Additionally, the
incorporation of an advanced analytics dashboard, featuring real-time graphs, empowers users with
dynamic insights into crowd dynamics. Overall, the project serves as a foundation for continuous
improvement and exploration within the realm of intelligent crowd monitoring.

62
6.4 Future Work Plan

1. Advanced Model Refinement

- Incorporate newer deep learning architectures and techniques that emerge in the field to
continually improve the accuracy of the crowd detection model.
- Explore the integration of other neural network architectures like Generative Adversarial
Networks (GANs) or Recurrent Neural Networks (RNNs) to better predict crowd dynamics over
time.
2. Advanced Analytics
- Integrate machine learning models to predict crowd behavior based on historical data.
- Analyze patterns to anticipate potential crowd surges, enabling proactive management.
3. Software Enhancement
- Enhance the software application with more user-centric features, feedback systems, and
intuitive visualization tools.
- Implement a cloud-based dashboard for advanced analytics, reporting, and centralized
management.
4. Scalability and Deployment
- Design strategies to scale the system across multiple cities or larger regions.
- Implement a robust deployment plan, including training for local authorities and on-ground
teams.
5. Ethical and Regulatory Updates
- Regularly review the ethical implications of the system, ensuring that it respects individuals'
rights and privacy.
- Stay updated with local and international regulations, ensuring the system remains compliant.

PROJECT METRICS
7.1 Challenges Faced
Data Collection and Annotation:
Collecting a diverse and representative dataset for fine-tuning the SASNet model.
Annotating the dataset with accurate crowd counts for training the model.

63
Model Fine-Tuning:
Fine-tuning the SASNet model on a custom dataset and addressing issues like overfitting or
underfitting.
Optimizing hyperparameters to achieve better accuracy and performance.

Real-Time Processing:
Ensuring real-time processing for live video feeds and optimizing the object detection algorithm for
speed.
Addressing latency issues to provide a smooth user experience.

Web App Development:

Integrating the SASNet model into the web app.
Designing an intuitive and user-friendly interface for uploading videos, using webcams, and accessing
analytics.

Database Management:
Setting up and managing the SQLite database, handling concurrent requests, and ensuring data
consistency.
Optimizing database queries for efficient retrieval of crowd count data.

Security and Privacy:

Implementing user authentication and authorization to ensure secure access to the web app.
Addressing privacy concerns related to storing and processing live video feeds.

Alarm System:
Implementing a reliable alarm system triggered by the crowd count exceeding a threshold for a
specified duration.

Email Notification:
Setting up an email notification system and ensuring the delivery of timely and accurate notifications.

7.2 Relevant Subjects

64
Deep Learning:
Deep learning forms the backbone of our crowd monitoring system, enabling the SASNet model to
automatically detect and count people in images and videos through the learning of complex patterns
and features.

Object Detection:
The principles of object detection are applied to crowd counting using the SASNet model. The model
is fine-tuned on a custom dataset, adapting its capabilities to accurately identify and count individuals
in various scenarios.

Computer Vision:
Computer vision techniques are instrumental in the field of crowd monitoring. Our project utilizes
computer vision to analyze and understand crowd behavior, addressing challenges in real-time video
feeds and enhancing the accuracy of people counting.

Database Management (SQLite):

SQLite plays a crucial role in storing and retrieving crowd count data.

Web Development (Streamlit):

Streamlit is chosen as the framework for our web app development, providing a user-friendly interface
for uploading videos, accessing live feeds, and viewing real-time analytics.

Real-Time Analytics:
Real-time analytics principles are applied to create an analytics dashboard. The dashboard, built using
tools like Streamlit, dynamically generates and updates graphs to provide users with real-time insights
into crowd counts at different time scales.

Web App Security:

Security measures are implemented in our web app, including user authentication.

Video Processing (OpenCV):

OpenCV is utilized for video processing, playing a key role in integrating the SASNet model with our

65
system. Challenges related to handling video input are addressed to ensure the effective analysis of
crowd dynamics.

Email Notification Systems:

An email notification system is implemented to alert concerned parties when crowd counts exceed
predefined thresholds. This feature enhances the responsiveness of our system, providing timely
information to relevant stakeholders.

User Authentication and Authorization:

User registration, login, and access control are implemented in our web app to ensure secure interactions.

Dataset Creation and Annotation:

A custom dataset is created for fine-tuning the SASNet model, capturing various crowd scenarios. The
challenges and considerations in annotating the dataset with accurate crowd counts contribute to the
model's training efficacy.

7.3 Interdisciplinary Knowledge Sharing

The integration of deep learning, particularly with the SASNet model, allowed us to automate object
detection and accurately count people in images and videos. Simultaneously, insights from web
development were instrumental in crafting a user-friendly interface and seamlessly integrating various
components within our web application. The utilization of SQLite for database management facilitated
efficient storage. Real-time analytics principles guided the design of our analytics dashboard, offering
users dynamic, real-time insights into crowd counts. Our focus on web app security incorporated
authentication. Performance optimization strategies were employed to enhance the overall system speed
and efficiency, creating a responsive crowd monitoring solution. Video processing, particularly
leveraging OpenCV, addressed challenges related to handling video input, contributing to the system's
effectiveness. The implementation of an email notification system added a timely alerting mechanism.
Interdisciplinary knowledge sharing, including considerations of ethics and privacy, ensured a cohesive
and effective crowd monitoring solution.

66
7.4 Peer Assessment Matrix
TABLE ?: Peer Assessment Matrix Here 1 represents the minimum rating and 5 represents the
maximum rating of contribution of each member)

Evaluation of

Angad Sidhu Lakshya Goel Shubh Satwik

Mehtani Ghildiyal

Evaluation Angad Sidhu - 4.5 4.5 4.5

by
Lakshya Goel 4.5 - 4.5 4.5

Shubh Mehtani 4.5 4.5 - 4.5

Satwik Ghildiyal 4.5 4.5 4.5 -

7.5 Role Playing and Work Schedule

Name Work Done

Shubh Mehtani ● Literature survey related to crowd counting.

● Creation of a labeled dataset in the required format from the
annotated images.
● Fine Tuning the model on our custom dataset.
● Evaluating performance of the final model on our custom dataset
and optimizing hyperparameters.

Angad Sidhu ● Literature survey related to crowd counting.

● Gathering relevant data to crowd counting from locations within the
campus.
● Designing Frontend of the WebApp.
● Enabling login functionality of the WebApp.

Satwik Ghildiyal ● Literature survey related to crowd counting.

● Gathering relevant data to crowd counting from locations within the
campus
● Testing performance of different models on benchmark datasets to
obtain a baseline model.
● Enabling user alert functionality for the WebApp.

Lakshya Goel ● Literature survey related to crowd counting.

● Cleaning of the collected data and annotating the images.
● Fine Tuning the model on our custom dataset.
● Building a user dashboard displaying analytics, database design and

67
 integration with WebApp for crowd count analytics.

7.6 Student Outcomes Description and Performance Indicators (A-K Mapping)

S0 Description Outcome

A1 Conducting an in-depth The team explores existing methodologies and techniques

literature review. in crowd detection, deep learning, and computer vision,
identifying gaps, challenges, and emerging trends.

A2 Developing a sophisticated Applying mathematical concepts and leveraging CNNs,

deep learning model. the model achieves precise & accurate crowd counting
and detection.

B1 Implementing a real time The team applies engineering techniques, including the
safety alert system. strategic design of an alert system to notify authorities
promptly when crowd thresholds or rule violations occur.
Constraints, such as potential irregularity in CCTV
images, are considered.

B2 Integrating a user-friendly Using appropriate methods for data collection, real-time

software application. data from CCTV cameras is collected and fed into the
deep learning model. The system is integrated into a
user-friendly application for authorized personnel.

B3 Analyzing and interpreting Multiple real-time images with different environmental

results. backgrounds are performed. Crowd count is detected and
data analysis is provided basis the real time crowd count.

C1 Validating and testing the Analyzing and interpreting results with respect to
proposed solutions. assumptions, constraints, and theory. Real-time data from
strategically placed CCTV cameras serves as the testing

68
 ground, demonstrating the accuracy of the deep learning
model and the responsiveness and reliability of the alert
system.

C2 Designing software The system is developed in Python, ensuring efficient

system to address desired crowd detection and real-time safety alerting. The entire
needs. system runs on the Chrome browser, providing a seamless
user experience.

D1 Disrupting conventional The project, by seamlessly integrating deep learning and

crowd management. computer vision, aims to revolutionize crowd
management, enhancing public safety.

D2 Utilizing interdisciplinary The project is divided among team members for effective
skills. use of shared knowledge, fostering collaboration and
utilizing diverse skill sets.

E1 Making a substantial The anticipated outcomes include a literature review, a

contribution to crowd pinnacle achievement in the form of the deep learning
management. model, and a proactive real-time safety alert system.

E2 Developing appropriate We used models like yolo v4, v8, SASNet for crowd
models. detection, compared accuracies and hypertuned the best
models for ensuring accurate identification of individuals
in crowded scenes.

F1 Ensuring swift intervention The real-time alert system's proactive approach amplifies
during critical scenarios. its potential impact, providing timely notifications to
authorities and empowering swift decision-making.

F2 Exhibiting professional Team members are punctual in meetings, regular in

responsibility in teamwork. arriving for evaluations, and showcase professional
responsibility.

G1 Establishing a robust system Through meticulous validation and real-world testing, the

69
 for effective crowd detection. project endeavors to validate its proposed solutions,
contributing to the realm of crowd management and
public safety.

G2 Delivering well-organized oral The team effectively communicates the project idea and
presentations. implementation details to the mentor and evaluation
panel.

H1 Providing authorized The user-centric software application caters to authorized

personnel with instantaneous personnel, rendering them direct access to real-time
access. safety alerts for swift decision-making.

H2 Examining societal impact and The team gauges social, ethical, economic, and health
economic tradeoffs. benefits of the project, such as persistence,
encouragement of a learning attitude, and minimal cost
with health benefits.

I1 Harnessing cutting-edge By seamlessly integrating deep learning and computer

technologies. vision, the project aspires to establish a robust system
capable of effective crowd detection and enhanced public
safety.

J1 Making a substantial The holistic validation and testing phase culminates the
contribution to the realm of demonstration of the deep learning model's accuracy, the
crowd management. efficiency of the alert system, and their pragmatic
applicability.

K1 Contributing to public safety. The "Crowd Watch" project, through its innovative
approach, aims to contribute substantially to public safety
by addressing the limitations of traditional manual
counting methods.

K3 Using software tools Software platforms include Anaconda, Python (version

necessary for computer 3.5), OpenCV (version 3.4.3), and the Chrome browser,
engineering. showcasing proficiency in relevant tools for project

70
 development.

7.7 Brief Analytical Assessment

Dataset Collection and Annotation:

Building a comprehensive and diverse dataset for fine-tuning the SASNet model could have been a
challenge. Annotation of the dataset to mark the correct number of people in each image or video frame
might have been time-consuming.

Model Fine-Tuning:
Fine-tuning a pre-existing model on a custom dataset can be challenging, and finding the right balance
to achieve accurate crowd counting without overfitting or underfitting might have required iterative
adjustments.

Real-time Processing:
Implementing real-time crowd counting on live video feeds involves handling large amounts of data in
real-time. Ensuring that the system performs efficiently without significant latency could be a technical
challenge.

Integration of Components:
Integrating SASNet, OpenCV, Streamlit, and SQLite into a cohesive system might have posed
integration challenges. Ensuring seamless communication between these components and handling
dependencies can be complex.

User Authentication and Security:

Implementing user registration, login, and securing the web app against unauthorized access is crucial.
Ensuring the security of user data, especially when dealing with a system that involves real-time
monitoring, is a significant challenge.

Threshold Determination:
Setting the appropriate threshold for triggering an alarm and sending a notification would require

71
careful consideration. Finding a balance that minimizes false alarms while capturing significant crowd-
related events can be challenging.

Email Notification System:

Implementing a robust email notification system, including handling SMTP configurations, ensuring
reliable delivery, and formatting the notification emails appropriately, may have been a challenge.

Scalability and Performance:

As the user base grows, ensuring that the system can scale effectively to handle increased load and
maintaining performance becomes essential. This includes optimizing database queries and managing
server resources efficiently.

User Interface Design:

Designing an intuitive and user-friendly interface in Streamlit to cater to both video upload and live
webcam feed options while displaying real-time analytics on a dashboard could have been a design
challenge.

Testing and Evaluation:

Comprehensive testing of the system, especially in diverse scenarios, is essential. Ensuring that the
system performs reliably under varying crowd densities, lighting conditions, and video qualities could
have been a challenge.

72
REFERENCES

[1] Hossain M, Hosseinzadeh M, Chanda O, Wang Y. Crowd counting using scaleaware attention
networks. In2019 IEEE winter conference on applications of computer vision (WACV) 2019 Jan 7
(pp. 1280-1288). IEEE.

[2] Jiang X, Xiao Z, Zhang B, Zhen X, Cao X, Doermann D, Shao L. Crowd counting
and density estimation by trellis encoder-decoder networks. In proceedings of the IEEE/CVF
conference on computer vision and pattern recognition 2019 (pp. 6133-6142).

[3] Liu N, Long Y, Zou C, Niu Q, Pan L, Wu H. Adcrowdnet: An attention-injective deformable

convolutional network for crowd understanding. InProceedings of the IEEE/CVF conference on
computer vision and pattern recognition 2019 (pp. 3225-3234).

[4] Liu Y, Shi M, Zhao Q, Wang X. Point in, box out: Beyond counting persons in crowds.
InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2019 (pp.
6469-6478).

[5] Sam DB, Sajjan NN, Maurya H, Babu RV. Almost unsupervised learning for dense crowd
counting. InProceedings of the AAAI conference on artificial intelligence 2019 Jul 17 (Vol. 33, No.
01, pp. 8868-8875).

[6] Sang J, Wu W, Luo H, Xiang H, Zhang Q, Hu H, Xia X. Improved crowd counting method based
on scale-adaptive convolutional neural network. IEEE Access. 2019 Feb 17;7:24411-9.

[7] Valloli VK, Mehta K. W-net: Reinforced u-net for density map estimation. arXiv preprint
arXiv:1903.11249. 2019 Mar 27.

[8] Varior RR, Shuai B, Tighe J, Modolo D. Multi-scale attention network for crowd counting. arXiv
preprint arXiv:1901.06026. 2019 Jan 17.

[9] Wang Q, Gao J, Lin W, Li X. NWPU-crowd: A large-scale benchmark for crowd counting and

73
localization. IEEE transactions on pattern analysis and machine intelligence. 2020 Jul
31;43(6):2141-9.

[10] Wang Q, Gao J, Lin W, Yuan Y. Learning from synthetic data for crowd counting in the wild.
InProceedings of the IEEE/CVF conference on computer vision and pattern recognition 2019 (pp.
8198-8207).

[11] Zhang A, Shen J, Xiao Z, Zhu F, Zhen X, Cao X, Shao L. Relational attention network for crowd
counting. InProceedings of the IEEE/CVF international conference on computer vision 2019 (pp.
6788-6797).

[12] Zhang Q, Chan AB. Wide-area crowd counting via ground-plane density maps and multi-view
fusion CNNs. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition 2019 (pp. 8297-8306).
[13]

74