Weapon Detection using Artificial Intelligence

and Deep Learning for Security Applications

Ajmeera Kiran1 P Purushotham2 D Divya Priya3
Assistant Professor Assistant Professor Assistant Professor
Department of Computer Science and Engineering Department of Computer Science and Department of Computer Science and
MLR Institute of Technology Engineering Engineering
2022 International Conference on Advancements in Smart, Secure and Intelligent Computing (ASSIC) | 978-1-6654-6109-2/22/$31.00 ©2022 IEEE | DOI: 10.1109/ASSIC55218.2022.10088403

Hyderabad, India MLR Institute of Technology MLR Institute of Technology

[email protected] Hyderabad, India Hyderabad, India
[email protected] [email protected]

Abstract- Increased crime in packed events or lonely areas has made deep learning approach to collect lessons and understand the
security a top priority in every industry. Computer Vision is used to most obvious and effective strategies or methods that are
find and fix anomalies. Increasing needs for security, privacy, and
private property protection require video surveillance systems that utilized in locations where bizarre movement and violent
can recognize and understand scene and anomalous situations.
Monitoring such activities and recognizing antisocial behavior helps behavior are observed [7]. With the use of a comprehensive
minimize crime and social offenses. Existing surveillance and supervised learning strategy, the objective is to develop a smart
control systems need human oversight. We're interested in detecting
firearms quickly through photos and surveillance data. We recast reconnaissance framework that can identify instances of
the detection problem as decreasing false positives and solve it by
building a data set guided by a deep CNN classifier and evaluating violence or weapons in a given video overview. To design and
the best classification model using the region proposal approach. implement an effective system for detecting the weapon in the
Our model uses Faster RCNN, YOLO.
surrounding area. Input is a weapon-containing video. The
Keywords –Artificial intelligence, abnormal detection, convolution
neural network (CNN), deep learning, RCNN, YOLO. processing consists of identifying the individual component
parts of the weapon using the Faster RCNN algorithm. Any
I.INTRODUCTION
Because of the rise in the number of dangerous activities and weapons discovered after identification will be detected. When

weapons, security is extremely important in every area. Earlier a weapon is detected, the output shows the typeof weapon.
reconnaissance frameworks were more dependent on the II. LITERATURE SURVEY
involvement of a human administrator [1,2]. At this time, [1] Deep Convolutional Neural Networks have demonstrated

because surveillance cameras are being installed in a variety of performance at the cutting edge on object detection metrics

public places, such as workplaces, emergency clinics, schools, such as ILSVRC-2012. While a model is able to capture the

roadways, and so on, it tends to be useful for capturing context of a whole frame and the items that surround it, it is

valuable activities and developments for the purpose of event unable to tackle the problem of several instances of the same

anticipation and web-based monitoring [3,4,5]. It will be easier object occurring throughout an image, which results in an

for those in positions of authority to build up security and make accidentally higher quantity of output for each illustration.

decisions that are appropriate if they have access to a They proposed a detection model based on spectral neural

mechanized system that is able to detect the occurrence of networks that predicts a collection of anchor boxes that are not

violent acts in recordings and provide a continuous response. class-specific and assigns a fixed score to each box based on

The use of violence is an abnormal way of behaving, and these the likelihood that it contains an object.

activities can be recognized through a brilliant reconnaissance

framework [6]. Using this framework, we can prevent further
tragic accidents from occurring. The primary goal is to use the

978-1-6654-8695-8/22/$31.00 ©2022 IEEE

Authorized licensed use limited to: VIT University. Downloaded on December 01,2024 at 06:33:10 UTC from IEEE Xplore. Restrictions apply.
[4] Wei Liu, Dumitru Erhan, Alexander CBerg, Scott Reed,
Christian Szegedy, Dragomir Anguelov, Cheng-Yang Fu, and
others proved how to recognize objects in photos using an
SSD (Single Shot Detector). This model divides the state
space of the frames into a series of preconfigured boxes with
varying property orientations and gauges for each individual
feature map. During the training process, the model assigns
scores to each inventory item that is located nearby a specified
box and modifies those boxes in order to get a better match
shape. In addition, the network automatically controls
different sized objects by combining projections from various
Figure 1: System Architecture
feature map data with varying resolutions. This allows the A. Faster R-CNN:
network to handle objects of varying sizes. This model is As a result of the enhancements made to the detection network
simpler than models that require object proposals since it design of R-object CNN, Fast R CNN was born. Although the
incorporates all of the processing into a single network. This training and recognition periods of the network have been

makes it possible to eliminate the need for proposal greatly shortened, it is not actually quick enough to be used as
development and subsequent resolution. As a result, training a real-time system because the process of generating on an
and integrating this into systems that require detection is not input image takes about (2 seconds) [12]. The selective search
difficult. method is the bottleneck in the architectural system. As a

[3] In their research work titled "Improved Anomaly Detection direct consequence of this, K He and colleagues developed the
in Surveillance Films Based on A Deep Learning Method," architecture known as the Faster R-CNN [13,14]. They offer
Professors Ali Khaleghi and Mohammad Shahram Moin Ku an alternate mechanism for producing region ideas called the
describe the development of a program that can distinguish Region Proposal Network rather than selective search. Object
between normal and abnormal videos. The backdrop is Detection Network and Region Proposal Network are both
removed as the input video is segmented on a frame-by-frame components of the architecture (RPN) [15].
basis during the data preprocessing. The reduction process can
either be done manually or automatically, and either way, it
results in the creation of a behavioral framework that can be
used for modeling and the identification of features.
III. PROPOSED SYSTEM
The proposed system is centered on the implementation of
certain realistic artificial intelligence-based weapon
identification algorithms that can enhance the several standard
approaches that are already in place [8,9,10]. It employs SSD
and Faster RCNN algorithms, both of which are based on
convolutional neural networks, for the detection of weapons
(CNNs). Because of these enhancements, the older model is
now capable of mimicking the precision of the faster R-CNN
while making use of less concentrated photos, hence Figure 2: Components of the architecture
increasing efficiency and reducing costs [11].

Authorized licensed use limited to: VIT University. Downloaded on December 01,2024 at 06:33:10 UTC from IEEE Xplore. Restrictions apply.
 given to areas by YOLO and the other CNN algorithms is
determined by how closely they resemble the target classes.

Figure 3: Fast RCNN Architecture

Figure 4: Working of YOLO

Figure 5: Flow Chart of Proposed Technique
B. YOLO
C. Proposed Algorithm Flow
YOLO is an object detection Convolutional Neural Network
Step-1: Importing the required packages
that operates in real time (CNN). CNNs are classifier-based
Step-2: Object Detection
systems that can recognize patterns in input images as
Step-3: Loading the Model
organized arrays of data. These patterns can be extracted from
Step-4: Pretrained Model
the images using various techniques [16, 17]. The speed at
Step-5: Input Video
which YOLO operates is far higher than that of other systems,
Step-6: Preprocessing the input video
but it does not sacrifice accuracy for this advantage. It
Step-7: Boxing the object
stimulates the model to investigate the full image while it is in
Step-8: Display the output
the testing phase, which provides the model with information
regarding the bigger picture of the image [17]. The "score"

Authorized licensed use limited to: VIT University. Downloaded on December 01,2024 at 06:33:10 UTC from IEEE Xplore. Restrictions apply.
 IV. RESULT
findings of certain studies, Faster RCNN also has a higher
recall than YOLO. On the other hand, YOLO is incredibly
quick for detection in real time. In addition, a relatively
significant percentage of false positives are detected even in
photographs with a high level of detail. To find a solution to
this problem, we fine-tuned our classification model by
adding new data and gradually thawing the topmost layers.
In order to cut down on the number of false positives, the
classification model is fed with the photos that were

Figure 6: Detection of Object (Input 1) incorrectly identified. The Faster RCNN algorithm was
trained on the pre-labeled video datasets to improve its
accuracy. Both approaches are efficient and produce
superior outcomes; however, putting either one into
practice in real time necessitates making a compromise
between the two values of speed and accuracy. Even though
it is faster than SSD, faster RCNN only achieves a frame
rate of 1.606 seconds per frame. Faster RCNN provides
84.6 percent accuracy. In contrast to the more efficient
RCNN, the SSD only achieves an accuracy of 73.8 percent
of the time. Greater accuracy was achieved with faster

Figure 7: Detection of Object ( Input 2) RCNN.

For training purposes utilizing GPUs, DSP, and
FPGA kits with larger datasets. Implementation in real time
using YOLO with an uniform loss metric (such as Mean
Average Precision) for all three methods. Detecting different
types of weapons and being able to tell fakes from the real
thing are both important skills. Identifying concealed
handguns through the use of pose analysis and human key
point estimation. Experimenting with RNN-based methods
for the purpose of producing automatic annotations
(DEXTR and Polygon RNN++).
REFERENCES
[1]. Erhan, D., Szegedy, C., Toshev, A., & Anguelov, D. (2014). Scalable
object detection using deep neural networks. In Proceedings of the IEEE
Figure 8: Detection Report
conference on computer vision and pattern recognition (pp. 2147-2154).
V. CONCLUSION
In the process of modeling, we simultaneously tested [2]. Glowacz, A., Kmieć, M. & Dziech, A. Visual detection of knives in

three different object identification algorithms, including security applications using Active Appearance Models. Multimed Tools
Appl 74, 4253–4267 (2015). https://doi.org/10.1007/s11042-013-1537-2
YOLO and Faster RCNN. Methods of object detection
based on sliding windows have been abandoned since [3]. Liu, W. et al. (2016). SSD: Single Shot MultiBox Detector. In: Leibe, B.,
Matas, J., Sebe, N., Welling, M. (eds) Computer Vision – ECCV 2016.
previous research demonstrated that they performed
ECCV 2016. Lecture Notes in Computer Science(), vol 9905. Springer,
poorly. According to the Cham. https://doi.org/10.1007/978-3-319-46448-0_2

[4]. Yuming Li, Jue Wang, Tengfei Xing, Tianlu Liu, Chengjun Li, and
Kuifeng Su. 2017. TAD16K: An enhanced benchmark for autonomous
Authorized licensed use limited to: VIT University. Downloaded on December 01,2024 at 06:33:10 UTC from IEEE Xplore. Restrictions apply.
 driving. In 2017 IEEE International Conference on Image Processing Detection using Artificial Intelligence and Deep Learning for Security
(ICIP). IEEE Press, 2344–2348. Applications," 2020 International Conference on Electronics and
https://doi.org/10.1109/ICIP.2017.8296701 Sustainable Communication Systems (ICESC), 2020, pp. 193-198, doi:
10.1109/ICESC48915.2020.9155832.
[5]. A. Biswas, A. P. Jana, Mohana and S. Sai Tejas, "Classification of
Objects in Video Records using Neural Network Framework," 2018 [16]. Makkena Brahmaiah, Srinivasa Rao Madala and Ch Mastan Chowdary,
International Conference on Smart Systems and Inventive Technology Artificial Intelligence and Deep Learning for Weapon Identification in
(ICSSIT), 2018, pp. 564-569, doi: 10.1109/ICSSIT.2018.8748560. Security Systems, Journal of Physics: Conference Series, Volume
2089, 1st International Conference on Applied Mathematics, Modeling
[6]. Mohana and H. V. Ravish Aradhya, "Simulation of Object Detection
and Simulation in Engineering (AMSE) 2021 15-16 September 2021,
Algorithms for Video Survillance Applications," 2018 2nd International
India
Conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud) (I-
SMAC)I-SMAC (IoT in Social, Mobile, Analytics and Cloud) (I-SMAC), [17]. B. Arthi, S. S, K. PoornaPushkala, A. Arya and D. Rajasekhar,
2018 2nd International Conference on, 2018, pp. 651-655, doi: "Wearable Sensors and Real-Time System for Detecting violence using
10.1109/I-SMAC.2018.8653665 Artificial Intelligence," 2022 International Conference on Advanced
Computing Technologies and Applications (ICACTA), 2022, pp. 1-5,
[7]. R. K. Meghana, Y. Chitkara, A. S. and Mohana, "Background-modelling
doi: 10.1109/ICACTA54488.2022.9753223
techniques for foreground detection and Tracking using Gaussian
Mixture Model," 2019 3rd International Conference on Computing
Methodologies and Communication (ICCMC), 2019, pp. 1129-1134, doi:
10.1109/ICCMC.2019.8819825.

[8]. J. Rajendra Prasad, Avantika Tiwari, O. Venkata Siva† And J.

Nageswara Rao, Ajmeera Kiran,”A Comprehensive Study On
Normalization Techniquesf Or Privacy Preservation In Data Mining”,
Journal Of Interconnection Networks ,2141040, World Scientific
Publishing Company , Doi: 10.1142/S0219265921410401,2022.

[9]. D Shanthi, N Swapna, Ajmeera Kiran, A Anoosha”,Ensemble Approach

Of GP,ACOT,PSO,And SNN For Predicting Software Reliability”,
International Journal Of Engineering Systems Modelling And Simulation
, 2022.

[10]. Ajmeera Kiran , D. Vasumathi, “Optimal Privacy Preserving

Technique Over Big Data Analytics Using Oppositional Fruitfly
Algorithm”, Recent Patents On Computer Science, Volume 13 , Issue 2
, 2020, DOI : 10.2174/2213275911666181119113913

[11]. .Biswas, A., Jain, A., Mohana (2021). Survey on Edge Computing–
Key Technology in Retail Industry. In: Smys, S., Palanisamy, R.,
Rocha, Á., Beligiannis, G.N. (eds) Computer Networks and Inventive
Communication Technologies. Lecture Notes on Data Engineering and
Communications Technologies, vol 58. Springer, Singapore.
https://doi.org/10.1007/978-981-15-9647-6_7.

[12]. R. J. Franklin, Mohana and V. Dabbagol, "Anomaly Detection in

Videos for Video Surveillance Applications using Neural
Networks," 2020 Fourth International Conference on Inventive
Systems and Control (ICISC), 2020, pp. 632-637, doi:
10.1109/ICISC47916.2020.9171212

[13]. Karatzas, P., Melagraki, G., Ellis, L. J. A., Lynch, I., Varsou, D. D.,
Afantitis, A., ... & Sarimveis, H. (2020). Development of deep learning
models for predicting the effects of exposure to engineered
nanomaterials on Daphnia Magna. Small, 16(36), 2001080.

[14]. Zhou, Q., Li, X. Visual Positioning of Distant Wall-Climbing Robots

Using Convolutional Neural Networks. J Intell Robot Syst 98, 603–613
(2020). https://doi.org/10.1007/s10846-019-01096-w

[15]. H. Jain, A. Vikram, Mohana, A. Kashyap and A. Jain, "Weapon

Authorized licensed use limited to: VIT University. Downloaded on December 01,2024 at 06:33:10 UTC from IEEE Xplore. Restrictions apply.