Vehicle Detection and Classification based on

2023 Third International Conference on Advances in Electrical, Computing, Communication and Sustainable Technologies (ICAECT) | 978-1-6654-9400-7/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICAECT57570.2023.10117923

C-DSO Dataset using YOLO v3 with SRBD Method

for Intelligent Transportation Applications

R.Anitha P.Prabakaran
Department of Computer Science Department of Computer Science
Vels University, V I S T A S Vels University, V I S T A S
Chennai, India Chennai, India
rs.anitha@ gmail.com [email protected]

Abstract— The tremendous evolution of Computer vision and by Faster-RCNN to efficiently choose object candidates. After
technology, the most challenging application of vehicle category receiving any size image as input, an RPN outputs a series of
detection is an important role in our Indian Traffic system. The most rectangular item proposals for the purpose of vehicle object
of the vehicle detection algorithms facing some struggles with vehicle categorization.
object detection and recognition rate. Using Deep Learning with
Conventional Neural networks to undercover the problem of vehicle Modern state-of-the-art models are primarily used to
detection and to understand the right way of vehicle classification, identify different types of vehicles, rather than being used to
detection, and vehicle recognition. This paper proposes our model of identify same types of vehicle objects. It has some restrictions
designing a dataset (C-DSO) and combines the method YOLO(You on the each and every model applications. In this paper,
Only Look Once)V3 algorithm with SRBD to introduce the SRBD-C- proposes our methodology has to finding a appropriate
DSO algorithm, it detects the smart way of integrating the pre- network, selecting and creating a new datasets, and leveraging
training of the image, labeling of vehicle image, image smoothing, knowledge between the datasets to conventional network
and Annotation. Furthermore, renovating the dataset, adding the application of vehicle classification and vehicle detection. It is
various categories of vehicle images, and the experimental results mostly used in our intelligent transportation systems.
show a better detection rate and reduction of loss of validation in the
dataset.
II. RELATED WORK
Keywords— Vehicle Detection, Label smoothing, YOLO-v3, (A. Krizhevsky, 2012) [1] created by AlexNet, it took first
SRBD-C-DSO Algorithm, Hyper Parameter. place in the ImageNet competition. (Girshick R, 2014) [2]
Proposed object detection and performing 200 class
I. INTRODUCTION ILSVRC2013 detection dataset. (Redmon J F. A., 2018)
Object detection has many applications in various areas, et.al[3] proposed detection metric inYOLOv3, it achieves
including surveillance, automatic emergency braking systems, 57.9 mAP@50 ms on TitanX. (IOFFE S, 2015)[4] proposed
and other applications. In modern computer vision, the to improve the imageNet classification: reaching 4.9% top-5
research area has to plays a vital role in Vehicle detection of validation error. (Wei Liu D. A.-Y., 2016) et.al[5] proposed
Intelligent Traffic surveillance systems. Despite the popularity method on our VehicleID dataset which as recently-released
and significance of object detection research, it is also one of vehicle model classification dataset. (Yang, 2015) et.al[6]
the most difficult tasks in computer vision. There is a growing implemented the experiment results for fine-grained
demand for not just detecting things that differ significantly classification on the surveillance data of CompCars. (Krause,
from one another, but also identifying the type of comparable 2013) et.al[7] proposed the 2D object representations to 3D,
objects. ID recognition, facial recognition, warehouse and estimate fine-grained categorization of both local feature
categorization, and other applications are trying to achieve the appearance and location. (Wei Liu D. A.-Y., 2016) et.al[8]
detection target. Deep learning technology outperforms all proposed the new tools for evaluation and metrics for MOT
other machine learning techniques in sectors linked to systems that consider both object tracking and object detection
computer vision. An automatic feature extraction is a powerful for comprehensive analysis.
tool that deep learning technology can use to improve
performance [1]. III. YOLO-V3 VEHICLE DETECTION STAGES
There are more Region-based vehicle object detection Deep learning neural networks, which are successful in a
architectures that has introduced afterward included the most one-stage method like YOLO-v3[3], are the foundation of the
well-known designs for Region-based vehicle object majority of vehicle object detection techniques.
identification, R-CNN[2], Fast R- CNN, and Faster R-CNN.
The Region Proposal Net (RPN) layer is highly recommended

Authorized licensed use limited to: b-on: Instituto Politecnico do Porto. Downloaded on May 19,2023 at 15:53:59 UTC from IEEE Xplore. Restrictions apply.
 Stage 1: The entire image is divided into S ∗ S grid, and The various image data is collected from various sources
every cell occupies a image prediction with indicates boundary of the dataset. The dataset contains various vehicle images
box on frame B, and the Hyper parameter such as x, w, h, y collected from different sources and focuses on the Chennai
and object_conf. Every single box is an image predication area vehicles moving on Chennai Metropolitan city roadways.
frame.The (x,y) coordinates represents for the predicted So, similar vehicles like autos, bikes, bicycles, cars,
border, centre, and offset value of the grid boundary. The motorcycle, trucks, vans, etc.
object_ conf provides the object confidence score of the related
image, and the frame reports a confidence value. The image
width (w,h) represents the height-to-width ratio of the frame in
the whole image. As shown in equation (1).

object _ conf = pr ( object ) * IOU pred

truth
(1)

The prediction result shows object Pr(object) indicates the Fig. 1. Model of SRBD-C-DSO Vehicle classification and labeling
corresponding cell value. And it is the prediction border it has In our custom C-DSO dataset shows the input new raw
object , the value of inclusion=1 and non-inclusion=0. The image data entries in (a), in (b) the vehicles are classified in
intersection over Union of the predicted image frame border terms of small, medium, large categories. In our proposed
represents as IOUpred and IOUtruth . system illustrated in (c) type of vehicle such as auto, bike,
bicycle, car, motorcycle, truck, van etc and various pose
object _ conf = pr ( class | object ) * pr ( class ) * IOU pred
truth

detection and grouping the similarity images for future

prediction.
= pr ( class ) * IOU truth
pred
(2)
b) Process of Label Smoothing:
Stage 2: The Convolutional neural network plays a major In pursuit of real-time detection, the YOLO-v3 model used
role of Feature Extraction, which helps to make prediction. to label and multi-label classification for category prediction,
The conditional probability of the objects in the grid is which somewhat decreased the accuracy. Our method used the
represented in [Pr(class | object)], the class of probability
Label Smoothing constraint model by which targets classified,
obtaining in various prediction Pr(classi). And the networks
reducing the degree of overfitting from the model and then
calculates total number of object in the network it is
represented as C. As shown in equation(2), a specific class is increasing detection rate accuracy.
determined by the product of the probabilities, and the class's The network works typically employs the soft-max function
confidence value corresponds to the intersection ratio. to compute the probability value of the all classes for each
distribution of object.
Stage 3: The Non-maximum suppression, has filtering the exp zi
predicted image border and remove the duplicate border, and pi = (3)
detect the final output. First, it has to filtering the predicted i e xp z i

border by NMS (non-maximum suppression), secondly,

remove the replicate border, and predicted the final output. The Where, Zi represents non-standardized logical format it is
loss function calculates mean square value and dimension- directly connected from the end layer of network, used for
vector [S∗S∗[B∗5+C]]. It shows value of image whether it is image classification prediction. At the time of training, the
correct or incorrect value and evaluate to execute the truth end-to-end detection, it can be modifies only the image
value of the current image. In last steps, it succeeds in the classification. Where P represents loss of distribution by
CNN structure model optimizing its goal. comparing the final output distribution. And q represents the
However, the loss function has to set different coefficient real-value distribution of cross-entropy.
of prediction box and without the object in boundary box L =  qi * log pi (4)
respectively, its balances the difference when there is no i

object in many grids [4]. The result is sum of the bounding The q represents a mono-thermal distribution, where as
box's loss coefficients and the judgments category, the objects class probability is 1, and others value is 0. Since, value 1 is
in the bounding box's loss coefficients may be given more considerable. Otherwise, pointing to the model is overestimate
weight. of prediction in that category, rendering the model to analyse
the generalization ability and easily to find overlapping.
IV. METHODOLOGY
c) Vehicle Detection SRBD-C-DSO strategy
a) Collecting, Classifying and labeling data:
First, our proposed model uses the existing dataset of the
In this paper, our proposed dataset C-DSO is evaluating Comprehensive Cars (CompCars) [6] dataset. In the
with our proposed SRBD vehicle Detection model levels. The environment appears a mostly large vehicle. And also try to
evaluation of current state-of-the-art of models is conducted use Standford Cars[7] dataset, but it has to missing the similar
on various famous vehicle datasets, VehicleID[5], Visual QA, vehicles that are used in our system. The Proposed algorithm
compcars, CIFAR10,MS COCO and ImageNet. model was tried on the UA-DETRAC [8] dataset, it has a lot
of vehicles is composed, but it does not exactly match our

Authorized licensed use limited to: b-on: Instituto Politecnico do Porto. Downloaded on May 19,2023 at 15:53:59 UTC from IEEE Xplore. Restrictions apply.
traffic system. Furthermore, to set up the custom dataset of C- Source_classif= “Ready to classification of vehicle(auto,
DSO and develop a Vehicle detection goal. bicycle, bus, car,Motorcycle truck,van).
The proposed SRBD-C-DSO model employs the
categorization, of Softmax classifiers which is classifies The SRBD-C-DSO Algorithm proposes a way of training
small, Medium, and large vehicles with the proposed SRBD data in the C-DSO dataset that combines both raw data and
with Yolov3 and checks similarity learning and to generate metadata. Then the vehicle images are trained in the YOLO v3
highly discriminating features. Fig 2, shows Our model is model architecture with our custom dataset C-DSO. The
framed of these main activities such as customizing the resulting model of the C-DSO dataset captures the vehicle
dataset, vehicle classification, and similarity checking. For the images and stores them in form of hyperdata, which represents
vehicle, categorization learning is deploy with Softmax hyper parameters such as Learning rate, number of training
classifier has to categorize the different vehicles based on their images, and calculating loss_function in YOLO-v3 with
vehicle identification number in dataset. This helps to build SRBD architecture. Finally, the detection process succeeds in
the model works more efficiently features for each vehicle vehicle classification, prediction of ground truth value and
identification from various perspectives. detection of learning rate.

V. DATASET SETUP
To solve vehicle detection problems, finding the solution
by reconstructing the dataset. Datasets is a very vital role that
plays in vehicle detection when it is using machine learning
techniques. To collecting a vehicle images in Chennai
Fig. 2. Method of SRBD using C-DSO Dataset Vehicle Detection
Roadways from various datasets and gather from the Internet
d) SRBD-C-DSO Algorithm and capture images through the camera. And the processed
vehicle images are made to convert json format.
The paper proposes to create its custom Dataset our main
The proposed algorithm model is training on the C-DSO
goal is to measures vehicles of visual similarity, and to
dataset. In the custom dataset is bounded with 170 videos, and
generate features extraction from the images and then evaluate
it is recorded by Cannon camera in 12 various locations in
some similarity measures between different vehicle images
Chennai, India. To collect the open source vehicle images and
using aforesaid features. In addition Conventional Neural
store in C-DSO (Chennai-Dataset Object) datasets, it is
Networks working in C-DSO dataset which carries the pre-
captured by the VHS hospital bridge of 120° angle. The
trained various vehicle images. The Process of the SRBD-C-
captured images are uploaded in the C-DSO dataset. The
DSO algorithm was originally trained to classify the images of
Chennai roadways vehicles dataset is named IITM-HrTr_v2
the Deep Learning Network among more than 10000 labels
which is taken from an elevated top view surveillance camera
(eg. Bus, truck, Motorcycle, bicycle, Van, Lorry, etc.,). This
to collect the images and store into dataset, to better
layer is fully-connected with YOLO-V3 which does the final
discriminate all vehicle types with a depth-angle view. The
detection and mapping between vehicle image features and
vehicle classification in the dataset is marked as rectangular
class labels and uses the final layer of SRBD.
boxes and defined as various colors of detection (Figure 3).
The initially setup the Five stages of process in this SRBD-
First, the vehicle images in the database are arranged in
C-DSO Algorithm:
ascending order and exported into YOLO format. Secondly,
Step 1: To create a Dataset and to upload the vehicle
the vehicle images in the data set are randomly split into pair
images from video frames.
of portions one is the training set and another one is the testing
Step 2: into the dataset for both training raw data and
set. In the third part, generate the vehicle object location
metadata = source object
information file in .json format. And to create a Python
Given, source_name= CV-image and setup
program to detecting the location vehicle and marked in
source_url=http://custom.dataset/C-DSO, then source_ data=
bounding boxes, and convert it to TXT format as the vehicle
“path”.
datasets labeled.
Step 3. If <To Setup training hyperparameter> then
learning_rate=0.01, training_epochs:=2100; TABLE I. VEHICLES IMAGES IN THE PROPOSED SRBD-C-DSO DATASET
Loss_function:=BCE “GOTO NEXT STEP”
Step 4: To reference to parameter, Given, training the Dataset Number of Images
Metadata from source_path, C-DSO-1 (depth angle view) 3946
Then training_epochs:=2100, C-DSO-2 (videos 1080P) 170
learning_rate=0.01,optimizer:=YOLOv3
Save, trained image format:= model.pt and CV.json
Step 5: Classify the new image data Table I shows the total number of vehicles stores in
Path_Running:= SRBD+v3 concates Training data datasets used to train the proposed model. The Dataset
Repeat the step until reach training_epochs=0 and C-DSO-1 is the collection of image with a depth angle view.
Execute Both step 3 and step 4 The dataset C-DSO-2 has different view angles and videos

Authorized licensed use limited to: b-on: Instituto Politecnico do Porto. Downloaded on May 19,2023 at 15:53:59 UTC from IEEE Xplore. Restrictions apply.
 captured as 1080p quality. The main use of the C-DSO-1 The experiment result is shown in Fig.4a~Fig.4d the
dataset is to store the different category of vehicle, which proposed method detects the different kinds of vehicle in
helps to detecting small, medium, and large vehicles in Chennai on the roadway. In this paper, as the input taking
running on-road applications. The additional datasets and multi-resolution images as the training set of input, which
well-tuned outcomes allowed for accurate vehicle type includes image sizes such as 127x890, 239x320, 320x240,
predictions. 640x401, 675x340, 700x466, 1280x720, and 1200x801.
Based on the proposed algorithm the SRBD-C-DSO has fine-
tuned vehicle detection and the model can achieve better
performance results. It can be detected in various classes of
vehicles which include, auto, bicycle, car, motorcycle, bus,
vans, and trucks. Fig:4 shows only a sample vehicle detection
image.

B) Results and Discussion

The experiment results are shown Fig.4a~Fig.4d the
proposed method detects the vehicles using the C-DSO
dataset; it has achieved a better detection rate. The SRBD
method strictly works in the CNN layer and reduces the Loss
Fig.3. Labeling representation of Vehicle
of training validation value.
The initial stage of vehicle detection is to classify the
different types of objects, including large-size cars, small–size
cars, trucks, autos, motorcycles, and bicycles, to establish the
all branded and new vehicles images are stored in the
datasets.
As the list shown in Table II, C-DSO Dataset vehicles are
classified into small, medium, and large a category, which is
directly related to the nature of quality vehicle classification
and detection. According to the list of vehicle classification
Table II, the sample vehicles are only given, if their needs for
future evaluation can be increased as many as possible.

TABLE : II C-DSO-DATASET VEHICLE CLASSIFICATIONS

Vehicle Type C-DSO Dataset-1 C-DSO Dataset-2
Truck 167 12
Sedan/SUV/Car 129 42 Fig.4. Sample result of vehicle detection in SRBD-C-DSO Algorithm
Bus 140 25
Van 205 17 The detection outcomes are shown in Fig.4(a)~Fig. 4(d) The
Motorcycle 4832 52 different colors listed in each and every vehicle. The violet
Cycle 1205 0
rectangle border represents sedan/SUV/car, the Blue rectangle
Auto 268 31
border represents truck, the orange rectangle border represents
VI. EXPERIMENTS bus, the purple rectangle border represents van, the yellow
rectangle border represents auto, the Green colored rectangle
A) Image Training and System Configuration: border represents cycle and the Cream-pink border represents
This paper focuses experiment is taken out under the Scooter, motorcycle and Bike. Finally, our model achieves the
Windows 11 version system. Using Intel (R) Core (TM) i7 better mAP and least value loss function in the image
CPU processor, which is to improve the 8.5m/s computing validation set. Furthermore, the evaluation of this model is
speed and reduces the training time, using the Nvidia RTX improved in the future and increased the number of vehicle
3050 graphics, CUDA10.0, and tools to be used in python 3.8, images and uploads of the C-DSO dataset. And calculate the
OpenCV. During the process of SRBD-C-DSO algorithm to validation loss function which is compared with another
evaluate the training set, validation set and testing set. To detection model.
setting the proper way of Hyper Parameter in this model. The
main parameter such as Model name, Training time, model VII. CONCLUSION
version, output type, classes, the backbone of CNN, fixing the In this paper, the YOLO-V3 with SRBD-C-DSO Algorithm is
learning rate, batch size=16, and resizing the image is 800 employed for vehicle detection on the custom vehicle dataset
pixels. These are used to the indication of the algorithm to C-DSO. This YOLO-V3 model effectively works in vehicle
performs and produce a better detection result, along with classification but it sometimes struggles to detect some small
increasing the number of images, then the prediction value is vehicles. Our proposed system trying to solve the issues and
changed. designed the custom Dataset C-DSO, which as working in

Authorized licensed use limited to: b-on: Instituto Politecnico do Porto. Downloaded on May 19,2023 at 15:53:59 UTC from IEEE Xplore. Restrictions apply.
label smoothing, training, testing, validating data, and setting a [3] J Redmon, A. Farhadi, ”YOLOv3: An Incremental Improvement,”
arXiv 2018.
Hyper parameter of the dataset. Through the experiment, our
[4] S. Ioffe, and C. Szegedy, “Batch Normalization: Accelerating Deep
SRBD-C-DSO algorithm performs a smart way of detection Network Training by Reducing Internal Covariate Shift,” arXiv, Mar
that can achieve optimal vehicle detection results. The 2015.
enhanced detection rate was achieved by the experimental [5] H. Liu, Y. Tian, Y. Wang, L. Pang, and T. Huang, ‘‘Deep relative
results of our YOLO- v3 model with SRBD and it reduced the distance learning: Tell the difference between similar vehicles,’’ Proc.
loss of validation value. In the future, the training images are IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 2167–2175,
increased in the C-DSO dataset, and improved the validation Jun. 2016.
rate and learning rate. [6] L. Yang, P. Luo, C. Change Loy, and X. Tang, “A large-scale car dataset
for fine-grained categorization and verification,” IEEE CVPR, pp. 3973-
3981, 2015.
REFERENCES
[7] J. Krause, M. Stark, J. Deng, and L. Fei-Fei, “3d object representations
[1] A. Krizhevsky, I. Sutskever, and G. Hinton, “Imagenet classification for fine-grained categorization,” ICCVW, pp. 554-561, 2013
with deep convolutional neural networks,” in NIPS, 2012.
[8] L. Wen, D. Du, Z. Cai, Z. Lei, M. Chang, H. Qi, J. Lim, M. Yang, and S.
[2] R. Girshick, J. Donahue, T. Darrell and J. Malik, "Rich Feature Lyu, "UA-DETRAC: A New Benchmark and Protocol for Multi-Object
Hierarchies for Accurate Object Detection and Semantic Detection and Tracking,” arXiv, 2015.
Segmentation," IEEE Conference on Computer Vision and Pattern
Recognition, pp. 580-587, 2014.

Authorized licensed use limited to: b-on: Instituto Politecnico do Porto. Downloaded on May 19,2023 at 15:53:59 UTC from IEEE Xplore. Restrictions apply.