Frontiers in Computing and Intelligent Systems 
ISSN: 2832-6024 | Vol. 7, No. 2, 2024 

 

75 

Research	on	the	Detection	of	Traffic	Flow	based	on	
Video	Images	
Jian He, Wei Teng *, Zeyu Zhao, Binche Liu, Bing Qin and Jun Jiang 

School of Computer Science and Software Engineering, University of Science and Technology Liaoning, Anshan Liaoning, 114051, China 

* Corresponding author: Wei Teng 

 

Abstract: Based on the current level of social development, everyone's demand for cars has increased rapidly. At present, the 
total number of motor vehicles and drivers in China ranks first in the world. With the rapid development of deep learning, the 
method of vehicle flow statistics based on video can directly use the existing traffic monitoring camera to realize the detection 
of vehicles, and some traffic flow detection based on YOLOv1, YOLOv2, YOLOv3, YOLOv4 and other algorithms have 
problems such as insufficient accuracy and low efficiency. Therefore, this paper proposes to use YOLOv5 to replace the original 
algorithm to achieve object detection, tracking, and processing. I improve the efficiency of the statistics of the traffic flow. 

Keywords: Video-based Traffic Statistics; YOLOv5; Object Detection; Tracking; Processing. 
 

1. Introduction 
The total number of motor vehicles and drivers in China 

ranks first in the world, and the detection of traffic flow is a 
crucial study in urban traffic management and planning. With 
the increase of urban population and the continuous growth 
of car ownership, the problem of traffic congestion is 
becoming increasingly prominent, which brings great 
challenges to the sustainable development of cities. 
Understanding and mastering traffic flow can provide 
important data support for traffic management departments, 
helping to optimize decision-making in road planning, traffic 
light control, traffic congestion mitigation, and more. Traffic 
flow detection plays an important role in urban traffic 
management and planning, which can provide accurate data 
support for traffic management departments, help optimize 
traffic conditions, improve road traffic efficiency, and 
promote the sustainable development of cities. However, 
many algorithms have serious problems such as low detection 
accuracy and low detection efficiency. To better improve the 
accuracy and efficiency, we use the YOLOv5 algorithm to 
improve it to achieve real-time monitoring [1].  

2. Introduction of YOLOv (1,2,3,4) and 
Their Shortcomings  

The Hinton team designed AlexNe by using a 
convolutional neural network, which works better on the 
ImageNet dataset than other traditional methods. So far, 
convolutional neural networks have gradually become 
popular in the field of computer vision, and more algorithms 
for image detection and processing have been proposed.  

YOLOv1 is the first version of YOLO, presented in 2015. 
It takes an input size of 448x448 and divides the image into 
7x7 grids, each predicting two bounding boxes, each 
containing a target. Each bounding box also predicts the 
category probability of the target. YOLOv1 was released by 
Joseph et al., which not only has a high degree of 
generalization ability, but also can meet the requirements of 
real-time detection. However, because YOLOv1 does not use 
a full convolutional network, the size of the input is fixed, and 
the image often needs to be cropped, which has a worse 

prediction effect on small targets in large images, and also has 
a certain impact on comprehensive object recognition. FCN 
network structure is divided into two parts: full convolution 
part and deconvolution part. The full convolution part is some 
classic CNN networks (such as VGG, ResNet, etc.), which is 
used to extract features; the deconvolution part is to get the 
original size of the semantic segmentation image through up-
sampling. The input of FCN can be a color image of any size, 
and the output size is the same as the input size, and the 
number of channels is n (number of target categories) + 1 
(background). The FCN network structure is as follows: 

In order to improve the problem of YOLOv1, YOLOv2 was 
proposed in 2016. YOLOv2 uses the Darknet-19 network as 
the backbone network and introduces Anchor Boxes to better 
handle targets at different scales. In addition, YOLOv2 uses 
multi-scale training and testing to improve the detection 
ability of small targets. Compared with YOLOv1, YOLOv2 
has improved in terms of accuracy and speed. However, due 
to the small number of grids, the positioning accuracy for 
large-scale targets is not high. 

YOLOv3 was introduced in 2018 and is the largest version 
of the YOLO series. YOLOv3 uses a deeper Darknet-53 
backbone network and adds detection at multiple scales. 
YOLOv3 also introduces three target prediction layers of 
different sizes for detecting targets of different sizes. In 
addition, YOLOv3 also uses FPN (Feature Pyramid Network) 
to extract multi-scale features. FPN is a top-down feature 
fusion method, but FPN is a multi-scale object detection 
algorithm, that is, there is not only one feature prediction layer. 
Although some algorithms also adopt multi-scale feature 
fusion to carry out target detection, they often only use the 
features of one scale obtained after fusion. Although this 
approach can combine the semantic information of top-level 
features with the details of low-level features, it will cause 
some deviations in the process of feature deconvolution. The 
prediction using only the features obtained after fusion will 
have a bad effect on the detection accuracy. Based on the 
above problems, FPN method can make prediction on 
multiple fusion features of different scales to maximize the 
detection accuracy. 

 
 



 

76 

 

 
Figure 1. The FCN network structure 

 
YOLOv4 was proposed in 2020. YOLOv4 has made a 

series of improvements in network structure and training 
strategy, including using CSPDarknet53 as the backbone 
network, using the SPP (Spatial Pyramid Pooling) module to 
process multi-scale information, and using PANet to fuse 
different hierarchical features. The network structure of 
PANet is shown in Figure 2, which consists of five core 
modules. Among them, (a) is an FPN[2], (b) is a bottom-up 

feature fusion layer added by PAN, (c) is an adaptive feature 
pooling layer, (d) is a bounding box prediction head of PANet, 
and (e) is a fully connected fusion layer for predicting 
masks.YOLOv4 has been significantly improved in terms of 
accuracy and speed. However, there is still the problem of 
poor handling of overlapping targets, identifying them as one 
target or missing some of them. 

 

 
Figure 2. Network Structure of Panet 

 

3. Introduction of YOLOv5 
In order to make the tracking, detection, and processing of 

images better realized, and the efficiency of traffic flow 
statistics is greatly improved, we use the YOLOv5 algorithm 
to implement it, which is compared with other YOLO 
algorithms, YOLOv5 uses CSPNet as the backbone network. 
CSPNet is a new network structure, which can effectively 
reduce the redundancy of feature maps and improve the 
computational efficiency and accuracy, using feature maps of 
four scales, which are 40x40, 80x80, 160x160, and 320x320. 
(CSPNet structure by splitting the shallow feature map in two 
at the channel dimension, a portion propagates backwards via 
a feature extraction module such as a residual block, the other 
part is directly merged with the output of the feature 
extraction module through the cross-stage hierarchy. A richer 
gradient combination is achieved, and can be reduced on the 

basis of the same or improved accuracy 10% ~ 20% of 
network parameters.）This made it possible to increase the 
ability to detect small targets, and at the same time, a variable 
input size of 640x640 was used. This allows for adaptation to 
different scenarios and needs. And there are 876 million 
parameters. YOLOv5 increased the number of parameters to 
improve the expressiveness and generalization ability of the 
model, and the mAP and FPS of YOLOv5 on the COCO 
dataset were 36.2% and 140.0. YOLOv5 is superior to the rest 
in terms of accuracy and speed. Therefore, we use YOLOv5 
to improve the accuracy and efficiency of traffic flow 
measurement[2]. 

YOLOv5 is a pre-trained object detection architecture and 
model series on the COCO dataset, which is an extension of 
the YOLO series. neck and head, among them, Backbone is 
responsible for feature extraction, Neck is responsible for 
feature fusion, and Head contains three detection heads, 



 

77 

which are responsible for outputting detection information, 
yolov5 has modified the four parts of the yolov4 network, and 
has achieved great improvements, using Mosaic data 
enhancement, adaptive anchor frame calculation, and 
adaptive image scaling on the input side. CBS module 
consists of a convolution layer Conv, a batch normalization 
layer BatchNorm and SiLu activation function layer. The 
specific structure diagram is shown below. 

 

 
Figure 3. CBS module structure diagram 

	
CBS module can extract the features in the image and sort 

out the extracted feature map. Conv of the CONvolution layer 
in CBS module changes the multiple of the feature map 
reduction by modifying the step size, and the image filling 
operation is calculated automatically by the model. In 
Backbone, the step size of Conv of CONvolution layer is 2, 
and the size of convolution kernel is 3 X3. Through this 
convolution layer, the width and height of feature map will be 
halved, and feature extraction of image is carried out at the 
same time. In Neck, convolution check feature maps with a 
size of 1X1 are mostly used to raise or reduce dimension. The 
BatchNorm layer is added after the convolution layer to 
normalize the extracted features, so that the network will not 
be unstable due to large data before the activation function, 
and to prevent the gradient explosion, disappearance and 
overfitting problems. At the very end, the SiLu activation 
function is used, adding smooth and non-monotonic features 
to increase the nonlinearity of the data. 

Focus structure is used on the backbone side. (Focus 
structure is a special layer in YOLOv5, mainly used to 
improve the detection ability of the model on small targets. 
This design can make the model better focus on the features 

of small targets, thus improving the accuracy of target 
detection and recall rate. Especially for small target detection, 
the Focus structure can bring obvious advantages. In addition, 
the Focus structure can reduce the computational effort and 
memory footprint by breaking the input feature graph into 
multiple subgraphs and concatenating the channel dimensions 
instead of directly copying the spatial dimensions of the entire 
feature graph. Such operations can improve the overall 
performance of the model by reducing the number of channels 
in the feature graph while maintaining critical information) 
and CSP structures (CSP structures improve the performance 
of deep neural networks by dividing input features into two 
parts and then cross-connecting between the two parts. CSP 
structure can effectively improve the feature representation 
ability of the model, thus improving the accuracy and 
generalization ability of the model. FPN+PAN structure was 
added to the neck side, the loss function during training was 
improved on the head side, and GIOU_Loss was used to 
improve DIOU_nms for prediction box screening. [3]. 
Compared with the fixed size and proportion used in the 
traditional method, the adaptive anchor frame has many 
advantages, such as enriching the detection of the object's 
background and small target, and calculating the Batch As it 
uses Normalization, it computations data on four images at a 
time, so that the mini-batch size doesn't have to be as large as 
it uses a single GPU to get good results. 

Adaptive anchor frame calculation 
In both YOLOv3 and YOLOv4, anchor needs to be 

calculated by K-means clustering method in advance, and the 
anchor is fixed. However, in YOLOv5, although the anchor is 
set in advance, the optimal anchor in different training sets 
can be calculated adaptively during training, so as to update 
the anchor value. This feature can also be turned off manually, 
changed in train.py, set to False. 

The specific implementation structure of YOLOv5 is as 
follows [4]: 

 
Figure 4. YOLOv5 structure diagram 

 

4. Algorithm Design 
The YOLOv5 algorithm is used to realize object detection, 

image segmentation, and target tracking to realize the 
statistics of traffic flow. 

There are five versions of YOLOv5, namely YOLOv5n, 
YOLOv5s, YOLOv5m, YOLOv5l and YOLO5x. The 
network depth, width, parameter size and number of 
convolutional nuclei in different stages are different 

1.YOLOv5s: This is the smallest model in the YOLOv5 
series. "s" stands for "small". The model performs best on 

devices with limited computing resources, such as mobile 
devices or edge devices. YOLOv5s has the fastest detection, 
but the accuracy is relatively low. 

2.YOLOv5m: This is a medium sized model in the 
YOLOv5 series. "m" stands for "medium". The YOLOv5m 
offers a good balance between speed and accuracy and is 
suitable for devices with a certain amount of computing 
power. 

3.YOLOv5l: This is a larger model in the YOLOv5 series. 
"l" stands for "large". YOLOv5l is relatively accurate, but the 



 

78 

detection speed is slower. It is suitable for devices that require 
high accuracy and have strong computing power. 

4.YOLOv5x: This is the largest model in the YOLOv5 
series. "x" stands for "extra large." YOLOv5x performed best 
in terms of accuracy, but had the slowest detection speed. 
Ideal for tasks requiring extremely high accuracy and devices 
with powerful computing power, such as Gpus. 

5.YOLOv5n: This is a variant in the YOLOv5 family, 
optimized for Nano devices such as the NVIDIA Jetson Nano. 
YOLOv5n delivers accuracy for edge devices while 
maintaining high speeds. 

We used yolov 5s in this experiment. 
First, we built a vehicle model and improved it by using the 

YOLOv5 network, which effectively alleviated the problem 
of target loss and target location. YOLOv5 is used to predict 
the position and category of the target in the image, and the 
prediction results are processed by YOLOv5. In the process 
of target tracking and detection, we conduct target tracking by 
studying the process of target positioning moving target in 
video over time. In the target tracking task in the real scene, 
considering the real-time and flexibility of the algorithm, we 
intend to adopt the online tracking method based on detection. 
Multi-target tracking technology (MOT) is designed to detect 
and obtain the position of the target that is interested or wants 
to track in the video frame, and assign an ID to each target, 
and keep each target ID unchanged during the process of 
target movement. After consulting a lot of data, we were able 
to reconstruct the feature extraction network of YOLOv5 
using the improved EfficientNetv2, and the association fusion 
network solved the competition problem between multi-task 
learning in JDE algorithm, greatly reducing the number of 
parameters and calculation amount of multi-target tracking 
algorithm, and better meeting the real-time requirements of 
mobile devices. At the same time, a fast multi-object based 
algorithm based on semi-suppressed fuzzy clustering 
proposed by Zhang Junhui and others is used to greatly 
improve the real-time and reliability of multi-object data 
association in complex environments. 

For image segmentation, a given image data is treated as a 
collection of many fuzzy elements belonging to different 
clusters. The membership degree of each pixel is blurred to 
each cluster, and the image segmentation algorithm based on 
OTSU-IFCM is used. The algorithm first utilizes the OTSU 
algorithm (Otsu algorithm is an adaptive threshold 
segmentation algorithm based on image gray histogram. It 
determines an optimal threshold by maximizing the inter-
class variance and minimizing the intra-class variance, which 
realizes the purpose of automatic image segmentation. The 
basic principle of Otsu algorithm is to select the optimal 
threshold by the maximum inter-class variance. In the process 
of image segmentation, we want the difference between the 
two categories after segmentation to be as large as possible, 

that is, to maximize the inter-class variance. Otsu algorithm 
calculates the inter-class variance under different thresholds 
and selects the threshold with the largest inter-class variance 
as the best threshold. To find the optimal segmentation 
threshold of image foreground and background, an image 
segmentation algorithm based on OTSU-IFCM is used. The 
algorithm firstly uses Otsu algorithm (Otsu algorithm is an 
adaptive threshold segmentation algorithm based on image 
gray histogram. It determines an optimal threshold by 
maximizing the inter-class variance and minimizing the intra-
class variance, which realizes the purpose of automatic image 
segmentation. The basic principle of Otsu algorithm is to 
select the optimal threshold by the maximum inter-class 
variance. In the process of image segmentation, we want the 
difference between the two categories after segmentation to 
be as large as possible, that is, to maximize the inter-class 
variance. Otsu algorithm calculates the inter-class variance 
under different thresholds and selects the threshold with the 
largest inter-class variance as the best threshold. The optimal 
segmentation threshold of foreground and background of 
image is found. By studying the positioning process of 
moving targets in video over time, the detection-based online 
tracking method is adopted in the target tracking task of real 
scene, considering the real-time and flexibility of the 
algorithm. 

Then the IFCM algorithm is used, which is a variant of the 
intuitive fuzzy C-means clustering algorithm. The basic idea 
is to treat a given image data as a set of fuzzy elements 
belonging to different clusters. By blurring the membership 
degree of each pixel to each cluster, and then iteratively 
adjusting the cluster center and membership degree according 
to the update formula, the image segmentation is finally 
realized. 

Multi-target tracking technology (MOT) is designed to 
detect and obtain the position of the target that is interested or 
wants to track in the video frame, and assign an ID to each 
target, and keep each target ID unchanged during the process 
of target movement. After consulting a lot of data, we were 
able to reconstruct the feature extraction network of YOLOv5 
using the improved EfficientNetv2, and the association fusion 
network solved the competition problem between multi-task 
learning in JDE algorithm, greatly reducing the number of 
parameters and calculation amount of multi-target tracking 
algorithm, and better meeting the real-time requirements of 
mobile devices. At the same time, a fast multi-object basis 
algorithm based on semi-suppressed fuzzy clustering 
proposed by Zhang Junhui et al is used to greatly improve the 
real-time and reliability of multi-object data association in 
complex environments. The specific realization flow chart is 
as follows: 

The specific implementation flow chart is as follows: 

 
Figure 5. Implementation flowchart 

 
 
 



 

79 

 

5. Experimental Results 
In the area of target detection, there are four types of 

samples: true positive(TP), false positive(FP), true 
positive(TP), false positive(FP), in order to quantify the 
performance comparison of the proposed network with other 
networks, this paper uses typical evaluation metrics such as 
precision (P) and recall (R), the formula is as follows: 

 

 
The average precision (AP) is the area below the P-R curve, 

which is the average of the Precision values of the P-R curve. 
For P-R curve, a definite integral is used for the calculation: 

	
In the actual calculation of target detection performance 

test, P-R curve should be smoothed. For each point on the P-
R curve, the Precision value is calculated as follows: 

	
MAP is an average value of a plurality of classes mAP. 

Because the method in this paper only detects pavement 
objects, the meaning of mAP and AP in this paper is the same. 
MAP is the most important index to measure the performance 
of target detection. 

In this experiment, we collected a lot of videos on the 
network and took some videos for testing from different 
angles, and a large number of videos were recognized well 
under the model and algorithm, and the accuracy was very 
high. The accuracy and efficiency of traffic flow statistics are 
fully improved, and the algorithm mentioned in this paper can 
have a good detection effect on traffic flow statistics. The 
YOLOv5 algorithm can be applied to a large number of object 

detection projects to improve the accuracy and efficiency of 
the project.	

6. Conclusion  
In this paper, the YOLOV5 algorithm is used to improve 

the accuracy and efficiency of traffic flow statistics, and 
improve the accuracy of actual detection. In the future, we 
will consider adapting the model algorithm to the traffic 
camera to truly achieve the statistics of the actual application 
of traffic flow. 

Acknowledgments 
College Students' Innovation and Entrepreneurship 

Training Plan of University of Science and Technology 
Liaoning in 2024. 

References 
[1] Zhang Yingwei. Research on traffic flow and violation 

detection technology in intelligent transportation[D].Liaoning 
University of Science and Technology, 2023.DOI: 10. 26923/ 
d. cnki.gasgc.2023.000474. 

[2] Research and application of multi-target tracking algorithm 
based on YOLOv5 and DeepSORT[D]. WANG Jialin 
Shandong University, 2021. 

[3] Liu Bei. Research on traffic flow detection algorithm based on 
object detection and tracking[D]. Inner Mongolia: Inner 
Mongolia University,2022. 

[4] Editorial Department of Chinese Journal of Highway and 
Transportation. Review of China Automotive Engineering 
Research Progress in 2017. China soumnal of Highway and 
Transpont,2017, 30(6):1-197. Editorial Department of Chinese 
Journal of Highway Engineering, Review of Chinese 
Automotive Engineering Academic Research, 2017 [J]. 
Chinese Journal of Highway, 2017,30 (6): 1-197. 

[5] Lin Shu, Liu Mingying, Tao Zhiying, Underwater Treasure 
Detection Based on Attention Mechanism and Improved 
YOLOv5 y!.Transactions of the chinese society of Agricutural 
Engineerng, 2021, 37(18):307-314.