Frontiers in Computing and Intelligent Systems 
ISSN: 2832-6024 | Vol. 3, No. 1, 2023 

 

134 

Road Traffic Small Target Detection Algorithm Based on 
Lightweight YOLO V5 

Chanyi Liu 1, 2 *, Dan Huang 1, 2, Tao Wang 1, 2, Tao Zhu 1, 2 

1 School of Automation and Information Engineering, Sichuan University of Science & Engineering, Yibin, 644002, China 
2Artificial Intelligence Key Laboratory of Sichuan Province, Sichuan University of Science & Engineering, Yibin, 644002, China; 

* Corresponding author: Chanyi Liu (Email: 2720806204@qq.com) 

 

Abstract: For the current problem of large number of parameters and large computation of the target detection network, the 

lightness improvement is carried out on the basis of yolov5s, and the dynamic convolution of OD is introduced to compensate 

for the accuracy degradation caused by the lightness, and the Map0.5 of the improved Yolov5s- EfficientNetV2-OD network 

reaches 82.8 through comparison experiments, and the FPS is 59 frames per second. The experiment proves that the network 

maintains the detection accuracy and detection speed of the network for small targets on the basis of light weighting. It is suitable 

for deployment on vehicle-mounted devices. 

Keywords: Deep learning; Lightweight; Small target detection. 

 

1. Introduction 

With the rise in per capita household income in China, 

motor vehicle ownership has increased dramatically, and the 

increase in vehicle ownership has brought great pressure on 

the roads and the ensuing traffic accidents. In addition to the 

weather and the vehicle itself mechanical failure and other 

objective factors, traditional car driving requires the driver to 

always observe the surrounding road information, if fatigue, 

distracted easily caused by traffic accidents. With the 

development of artificial intelligence, the technology of 

automatic driving is becoming more and more mature. There 

is a big difference between traditional manual driving 

vehicles and self-driving vehicles, and automatic driving can 

be achieved without human intervention and precise 

automatic control, and when there is an unexpected situation, 

the automatic driving system will intervene in advance as well 

as emergency braking when necessary. 

As a basic component of an autonomous driving system, 

the detection capability of the target detection is one of the 

reference indicators for judging the level of autonomous 

driving. Common detection networks are divided into two-

stage detection networks such as RCNN, Fast-Rcnn,Faster-

Rcnn,single-stage detection networks SSD, and Yolo[5-9] 

series. The two-stage network has high detection accuracy but 

slow detection speed, and the single-stage network has fast 

detection speed but slightly lower detection accuracy. For 

target detection, many scholars have been improving the 

capability of target detection networks to achieve good 

detection results, Ren et al improved the SSD algorithm to 

enhance the detection capability of the network for obscured 

objects or small objects. Li et al proposed Quality Focal Loss 

to improve the error suppression that occurs during NMS 

operation to improve the detection accuracy of the network. 

Huang et al designed a variable residual convolutional 

branching module with RetinaNet as the base framework to 

improve the detection accuracy of various types of targets. 

However, with the improvement of detection capability, the 

network structure becomes more and more complex, the 

number of parameters, and the model size also become larger, 

leading to the decrease of detection speed. 

Based on the above problems, this paper performs light 

weighting on the basis of yolov5s, and introduces OD 

dynamic convolution to improve the feature extraction 

capability of the network by in order to compensate for the 

loss of model accuracy after light weighting. The improved 

network can guarantee the detection capability on the basis of 

lightweighting. It is beneficial to be deployed on in-vehicle 

devices. 

2. Related Work 

YOLOv5 algorithm is the latest version of the current 

YOLO series detection algorithm, proposed by Ultralytics 

Yolov5 main structure is shown in Figure 1, the model 

consists of four parts: input side (Input), backbone network 

(Backbone), multi-scale feature fusion network (Neck), and 

detection network (YOLOHead). 

 

Figure 1. Yolov5 Network 

Yolov5 input consists of SPP structure, FOUCS structure, 

and C3 structure, which are used to extract the image network 

features.Neck side consists of multi-scale pyramid FPN+PAN 

structure, through such a pyramid structure to increase the 

resolution of the image, so that the image can be divided into 

more size pyramids. The more layers of the pyramid, the more 



 

135 

details can be extracted. In addition, a batch normalization 

layer is used, which can effectively prevent the occurrence of 

overfitting. The prediction side consists of three detection 

heads, corresponding to large, medium and small size targets. 

the Yolov5 structure is shown in Figure 1. 

3. Method 

In this paper, we optimize YOLOv5s from two aspects: 

lightweight and attention. Firstly, the yolov5s backbone 

network is replaced with EfficientNetV2 network to reduce 

the computational complexity and make the detection model 

more lightweight, and secondly, the conventional convolution 

is replaced with ODConv dynamic convolution in the output 

layer, which uses a multidimensional attention mechanism to 

learn complementary attention along four dimensions of the 

kernel space through a parallel strategy. 

3.1. Lightweight 

EfficientNetv2 is a lightweight network that combines 

training perceptual neural architecture search and scaling, 

which better balances the relationship between accuracy, 

training speed and number of parameters.EfficientNetV2-S 

structure table is shown in Table X. 

Table 1. EfficientNetV2- S structure table 

Stage Operator Stride Channel Layers 

0 

1 

2 

3 

4 

5 

6 

7 

Conv3x3 

Fused-MBConv1, K3x3 

Fused-MBConv4, K3x3 

Fused-MBConv4, K3x3 

MBConv4,K3x3,SE0.25 

MBConv6,K3x3,SE0.25 

MBConv6,K3x3,SE0.25 

Conv1x1&Pooling&FC 

2 

1 

2 

2 

2 

1 

2 

- 

24 

24 

48 

64 

128 

160 

272 

1792 

1 

2 

4 

4 

6 

9 

15 

1 

To address the problem of slow speed due to DW 

convolution in the MBConv module in EfficientNet, the 

efficientNetv2 network addresses the problem of accuracy 

degradation and slower training by using the Fused-MBConv 

structure instead of the MBConv structure in the efficientNet 

network.The Fused- MBConv replaces the 1x1 normal 

convolution and 3x3 deep separable convolution in MBConv 

by 3x3 normal convolution, and then uses the NAS technique 

to search for the best combined structure, the network 

structures of MBConv and Fused-MBConv are shown in 

Figure 2, Figure 3 

 

Figure 2. MBConv Module 

 

Figure 3. Fused-MBConv Module 

The structure of the Yolov5s backbone network after 

replacing it with efficientNetv2 is shown in Figure4. 

 

Figure 4. Yolov5s- EfficientNetV2 structure 

3.2. ODConv 

ODConv convolution takes into account the dynamics of 

the null field, input channel, output channel and other 

dimensions, so it is called full-dimensional dynamic 

convolution. oDConv adopts multidimensional attention 

mechanism to learn complementary attention in four 

dimensions of convolutional kernel: input channel, output 

channel, kernel space and number of kernels through parallel 

strategy, which enhances the feature extraction ability of 

convolution in all aspects. The structure of ODConv 

comparison is shown in Figure 5. 

 

Figure 5. DyConv and ODConv structure comparison diagram 

The above figure shows a schematic comparison of the 

structures of DyConv (a),ODConv (b). Unlike CondConv and 

Dyconv, which compute a single attention scalar, for the 

convolution kernel ODConv utilizes a new multidimensional 

attention mechanism that computes four types of attention in 

parallel along all four dimensions of the kernel space. The 

four types of attention convolution are illustrated in Figure 6 

 

Figure 6. ODConv illustration diagram 

(a), (b) (c) (d) represent the position multiplication along 

the spatial dimension, the channel multiplication along the 

input channel dimension, the filter multiplication along the 

output channel dimension, and the kernel multiplication along 

the kernel dimension of the convolution kernel space, 



 

136 

respectively. These four types of attention are complementary, 

and by progressively multiplying the convolution with 

different attentions along the position, channel, filter, and 

kernel dimensions will make the convolution operation differ 

for each dimension of the input, providing better performance 

to capture rich contextual information. Therefore, ODCOnv 

can significantly improve the feature extraction capability of 

convolution. The structure of the network after adding 

dynamic OD convolution is shown in Figure 7. 

 

Figure 7. Yolov5s- EfficientNetV2-OD structure diagram 

3.3. Dataset 

The KITTI dataset, co-founded by Karlsruhe Institute of 

Technology and Toyota Institute of Technology, is the largest 

international dataset for evaluating computer vision 

algorithms in autonomous driving scenarios. The KITTI 

includes real image data from urban, rural, and highway 

scenes, with up to 15 vehicles and 30 pedestrians per image, 

and various levels of occlusion and truncation. The original 

dataset is classified and subdivided into car, van, truck, 

pedestrian, pedestrian(sitting), cyclist, tram, and misc. In this 

paper, the original KITTI dataset consists of car, truck, van as 

car, pedestrian, person_sitting as person, and cyclist as rider, 

and the final KITTI dataset consists of car, person, and rider. 

These three categories. 

3.4. Configuration 

The operating system of the experimental platform in this 

paper is Window10, the hardware configuration is Intel(R) 

Core (TM) i7-10875H 2.30GHz processor, GeForceRTX2060 

graphics card with 6GB of video memory size, and the 

software environment equipped with Anaconda, CUDA11.4.0, 

using Pytorch1.7.1 Deep learning and YOLOv5 target 

detection framework. 

3.5. Evaluation Indicators 

Average precision (AP): The AP value is the area enclosed 

by the P-R curve (consisting of the values of precision P and 

recall R at different confidence thresholds) and the coordinate 

axes, and the larger the area enclosed by the P-R curve, the 

better the detection performance of the network model.  

𝐴𝑃 = ∫ 𝑃(𝑅)𝑑𝑅
1

0
         (1) 

Mean Average Precison (mAP): indicates that the AP of 

multiple categories is averaged and calculated by the 

following formula: 

𝑚𝑎𝑝 =
∑ 𝐴𝑃𝑖
𝑛
𝑖=1

𝑛
           (2) 

n denotes the number of categories. 

Detection speed (FPS): reflects the real-time nature of the 

algorithm, and being able to detect obstacle targets in real-

time is a requirement for self-driving cars. The common 

metric for detection speed is Frame Per Second (FPS), which 

indicates the number of images that can be detected per 

second, and the larger the FPS, the faster the model detects 

and the better the real-time performance. 

3.6. Training Strategies 

This experiment was conducted on the dataset KITTI, and 

5000 images in KITTI were selected and divided into training 

set, validation set, and test set in the ratio of 8:1:1. The divided 

training set is 4000 images, and the validation set and test set 

are 500 images each. The input image size is 640×640, the 

batchsize is 16, the test batchsize is 1, the initial learning rate 

is 0.01, the learning rate momentum is 0.937, and the training 

epoch is 200 rounds. Adaptive anchor is used to set the initial 

anchor box size, adaptive moment estimation (Adam) 

optimizer is used to optimize the model, and mosaic data 

augmentation is used to augment the dataset, which increases 

the diversity of the dataset and the number of targets, so that 

the algorithm has better generalization ability. The other 

parameters are set according to the official default parameters 

of YOLOv5s. 

4. Experiments 

4.1. Lightweight Improvement Experiment 

In order to verify the Effectiveness of lightweight networks, 

the EfficientNetv2 lightweight network was replaced with the 

original yolov5s and Yolov5-nano and Mobilenetv3 were 

replaced with the original yolov5s network. Comparative 

experiments were conducted using the same dataset under the 

same configuration environment. The comparison results are 

shown in Table 2 

Table 2. Lightweight experimental results 

Models Backbone Size Par Map0.5 Fps 

Yolov5s CSP1_X 14.4 7.07 83.5 60 

Ynano CSP1_X 3.7 1.76 78.2 93 

YMbv3 Mobv3 7.3 3.55 76.1 73 

YEffV2 EffV2 11.2 5.43 81.4 66 

Yolov5s-nano is short as Ynano, Yolov5s-MobileNetV3 is 

short as YMbv3 Yolov5s-EfficientNetv2 is short as YEffv2. 

From the results of the lightweight experiments in Table 2, we 

can see that the performance of all networks tested in the data 

set, YOLOv5s has the highest indexes, with Map@0.5 

reaching 83.5; among the lightweight networks, Yolov5s-

Mobilenetv3 has the lowest accuracy rate, Map@0.5 value 

among all networks, and The Map@0.5 of Yolov5s-

EfficientNetV2 reaches 81.4, which is higher than the 76.1 of 

Yolov5s-Mobilenetv3 network and 78.2 of Yolov5s-nano, and 

is the best performance among all lightweight networks. 

22.2%, 23.2%, And the FPS is higher than the 60 frames per 

second of the original yolov5s. It is slightly lower than the 73 

frames per second of Yolov5s-Mobilenetv3. 

4.2. Add ODConv Moudle 

In order to compensate for the loss of accuracy after light 

weighting, the OD dynamic convolution is fused in the light 

weighting network and the comparison experiments are 

conducted as follows. 

Yolov5s-EfficientNetv2-OD is abbreviated as YEffV2-OD. 

From Table 3, we can see that after adding ODConv 

convolution to the lightweight network Yolov5s- 

EfficientNetV2, the Map@0.5 value improved by 1.4% with 



 

137 

better detection capability for small targets, although the FPS 

decreased slightly by 7 frames per second. The original image 

and the detection effect are shown in Figure 8 and Figure 9 

Table3. Add ODConv experimental results 

Models ODConv Map0.5 FPS 

Yolov5s  83.5 60 

Ynano  78.2 93 

YMbv3  76.1 73 

YEffV2  81.4 66 

YEffV2-OD √ 82.8 59 

 

Figure 8. Original Image 

 

Figure 9. Detection effect picture 

5. Conclusion  

In this paper, firstly, the EfficientNetV2 lightweight 

network replaces the yolov5s backbone network, which 

reduces the number of network model parameters and 

computation, and secondly, the ODconv dynamic convolution 

is added to make the model feature extraction ability stronger, 

after experimental comparison, the improved Yolov5s- 

EfficientNetV2-OD network model Map@0.5 reaches 82.8 

and the PFS is 59, which reduces the requirements for the 

performance of on-board devices while ensuring the 

lightweight and detection accuracy. 

References 

[1] GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich Feature 
Hierarchies for Accurate Object Detection and Semantic 
Segmentation [J]. 2013. 

[2] GIRSHICK R. 2015 IEEE International Conference on 
Computer Vision (ICCV); proceedings of the IEEE 
International Conference on Computer Vision, F, 2015 [C]. 

[3] REN S, HE K, GIRSHICK R, et al. Faster R-CNN: Towards 
Real-Time Object Detection with Region Proposal Networks; 
proceedings of the NIPS, F, 2016 [C]. 

[4] WEI L, DRAGOMIR A, DUMITRU E, et al. SSD: Single Shot 
MultiBox Detector [J]. 2016. 

[5] REDMON J, DIVVALA S, GIRSHICK R, et al. You Only 
Look Once: Unified, Real-Time Object Detection; proceedings 
of the Computer Vision & Pattern Recognition, F, 2016 [C]. 

[6] REDMON J, FARHADI A. YOLO9000: Better, Faster, 
Stronger; proceedings of the IEEE Conference on Computer 
Vision & Pattern Recognition, F, 2017 [C]. 

[7] REDMON J, FARHADI A J A E-P. YOLOv3: An Incremental 
Improvement [J]. 2018. 

[8] BOCHKOVSKIY A, WANG C Y, LIAO H. YOLOv4: 
Optimal Speed and Accuracy of Object Detection [J]. 2020. 

[9] GE Z, LIU S, WANG F, et al. YOLOX: Exceeding YOLO 
Series in 2021 [J]. 2021. 

[10] REN J C X H, LIU J B, ET AL. Accurate single stage detector 
using recurrent rolling convolution [J]. IEEE Conference on 
Computer Vision and Pattern Recognition, 2017, Honolulu, HI, 
USA. New York: IEEE Press,: 752-60. 

[11] AL L X W W W L E. Generalized focal loss： learning 
qualified and distributed bounding boxes for dense object 
detection. Advances in Neural Information Processing Systems, 
2020. 

[12] HUANG WENHAN Y G, GENG KEKE, ET AL. Target 
detection in complex driving scenes based on adaptive fusion 
of dilated convolutional features [J]. Journal of Southeast 
University Natural Science Edition, 2021, (51(06)): 1076-83. 

[13] LI CHAO Z A Y A. Omni-dimensional dynamic convolution 
[EB/OL] https://arxivorg/pdf/220907947pdf, [2022-09-11]. 

[14] LIN T Y, DOLLáR P, GIRSHICK R, et al. Feature Pyramid 
Networks for Object Detection [J]. 2016. 

[15] LIU S, QI L, QIN H, et al. Path Aggregation Network for 
Instance Segmentation [J]. 2018. 

[16] Q T M L. Efficientnetv2: Smaller models and faster training [J]. 
International Conference on Machine Learning, 2021: 10096-
106. 

[17] C T M L Q E R M S F C N N. EfficientNet:  Rethinking  
model  scaling  for  convolutional  neural networks [J]. 
Proceedings  of  the  36th  International  Conference  on  
Machine  Learning, 2019: 6105-14.r 

 

https://arxivorg/pdf/220907947pdf