Improved autonomous driving object detection based on YOLOv8s

Table 1. Training parameters on different versions of COCO dataset
View table
View in Article
Table 1. Training parameters on different versions of COCO dataset
Model mAP（50-95）^val Speed^CPU/ms Speed^A100/ms Params/M FLOPs/B
YOLOv8n 37.3 80.4 0.99 3.2 8.7
YOLOv8s 44.9 128.4 1.20 11.2 28.6
YOLOv8m 50.2 234.7 1.83 25.9 78.9
YOLOv8l 52.9 375.2 2.39 43.7 165.2
YOLOv8x 53.9 479.1 3.53 68.2 257.8

Table 2. RepConv experiment comparison
View table
View in Article
Table 2. RepConv experiment comparison
模型 mAP50 mAP（50-95） Recall Precision Parameters
YOLOv8s 0.797 0.572 0.739 0.793 11 137 329
YOLOv8s+RepConv 0.806 0.575 0.737 0.811 11 137 329

Table 3. Comparison of EMA adding

View table

View in Article

Table 3. Comparison of EMA adding

模型	mAP50	mAP（50-95）	Recall	Precision	Parameters
YOLOv8s	0.797	0.572	0.739	0.793	11 137 329
YOLOv8s+Triple	0.804	0.573	0.742	0.804	11 318 081
YOLOv8s+ECA	0.804	0.574	0.737	0.804	11 137 338
YOLOv8s+EMA	0.804	0.574	0.757	0.804	11 140 625

Table 4. Comparison of improved Head end experiment
View table
View in Article
Table 4. Comparison of improved Head end experiment
模型 mAP50 mAP（50-95） Recall Precision Parameters
YOLOv8s 0.797 0.572 0.739 0.793 11 137 329
YOLOv8s+P2 0.808 0.580 0.765 0.808 10 862 844

Table 5. Comparison of loss function improvement experiment

View table

View in Article

Table 5. Comparison of loss function improvement experiment

模型	mAP50	mAP（50-95）	Recall	Precision	Parameters
YOLOv8s	0.797	0.572	0.739	0.793	11 137 329
RepConv_EMA_P2_Focal-CIoU	0.804	0.579	0.744	0.785	10 866 364
RepConv_EMA_P2_Focal-DIoU	0.809	0.580	0.755	0.790	10 866 364
RepConv_EMA_P2_Focal-EIoU^［15］	0.806	0.586	0.742	0.809	10 863 664
RepConv_EMA_P2_Focal-GIoU^［16］	0.800	0.578	0.734	0.810	10 866 364
Rep_EMA_P2_WIoU	0.812	0.584	0.747	0.812	10 866 364

Table 6. Experimental image configuration
View table
View in Article
Table 6. Experimental image configuration
镜像配置版本
CUDA 11.3
Python 3.8
Pytorch 1.11.0+cu113
Torchvision 0.12.0+cu113

Table 7. Dataset details

View table

View in Article

Table 7. Dataset details

镜像配置	版本
数据集名称	Car数据集
来源	改编自A2D2数据集
A2D2数据集总量	2.3 TB
摄像头帧率	30 FPS
选取帧数	6 254张图片
采集设备	车辆前置摄像头
摄像头视野	水平视野60°，垂直视野38°
图像分辨率	1 920像素×1 208像素
采集地点	德国南部3个不同城市
数据分类	语义分割、3D边界框等
采集场景	高速公路、乡村、城市

Table 8. Number of instances of each label in the dataset
View table
View in Article
Table 8. Number of instances of each label in the dataset
标签种类实例数量
Car 21 712
Van 3 731
Bus 716
Truck 4 927
Building 8 145
Person 1 306
Traffic light 3 834
Traffic signal 3 344
Lamp 5 645
Rider 659
Bike 207

Table 9. Ablation experiment

View table

View in Article

Table 9. Ablation experiment

模型	mAP50	mAP（50-95）	Recall	Precision	Parameters
YOLOv8s	0.797	0.572	0.739	0.793	11 137 329
RepConv	0.806	0.575	0.737	0.811	11 137 329
EMA	0.804	0.574	0.757	0.804	11 140 625
P2	0.808	0.580	0.765	0.808	10 862 844
Rep_EMA	0.802	0.575	0.732	0.824	11 140 625
Rep_p2	0.806	0.583	0.733	0.797	10 909 404
EMA_p2	0.811	0.582	0.746	0.780	10 686 012
Rep_EMA_p2	0.808	0.584	0.769	0.775	10 866 364
Rep_EMA_p2_WIoU	0.812	0.584	0.747	0.812	10 866 364

Table 10. YOLOv8 ablation experiments in different versions
View table
View in Article
Table 10. YOLOv8 ablation experiments in different versions
模型 mAP50 mAP（50-95） Recall Precision Layers
YOLOv8n 0.776 0.539 0.693 0.794 207
YOLOv8n_ours 0.778 0.547 0.737 0.807 287
YOLOv8s 0.797 0.572 0.739 0.793 207
YOLOv8s_ours 0.812 0.584 0.747 0.812 287

Table 11. Comparison of Total-loss functions for different losses
View table
View in Article
Table 11. Comparison of Total-loss functions for different losses
损失函数最小损失值平均损失值
CIoU 9.125 10.011
Focal-CIoU 8.740 9.555
Focal-DIoU 9.508 10.476
Focal-EIoU 9.170 10.069
Focal-GIoU 8.774 9.557
WIoU 7.068 8.683

Table 12. Comparative experiments

View table

View in Article

Table 12. Comparative experiments

模型	mAP50	mAP（50-95）	Recall	Precision	Parameters
YOLOv3	0.792	0.557	0.752	0.815	61 551 280
YOLOv5s	0.767	0.511	0.702	0.807	7 039 792
YOLOv7-tiny	0.755	0.494	0.662	0.808	6 034 656
YOLOv8s	0.797	0.572	0.739	0.793	11 137 329
SSD	0.568	0.326	0.403	0.839	25 082 528
Gold-YOLO	0.788	0.543	0.681	0.788	23 232 300
Faster RCNN	0.755	0.430	0.793	0.611	136 893 874
RetinaNet	0.528	0.307	0.411	0.815	36 537 287
Rep_EMA_p2_WIoU	0.812	0.584	0.747	0.812	10 866 364

Tools

Get Citation

Copy Citation Text

Longchun WANG, Wei FANG, Lijuan ZHANG, Dongming LI. Improved autonomous driving object detection based on YOLOv8s[J]. Chinese Journal of Liquid Crystals and Displays, 2025, 40(5): 773

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category:

Received: Sep. 20, 2024

Accepted: --

Published Online: Jun. 18, 2025

The Author Email: Dongming LI (LDM0214@163.com)

DOI:10.37188/CJLCD.2024-0290

Topics

Table 1. Training parameters on different versions of COCO dataset

Table 1. Training parameters on different versions of COCO dataset

Table 2. RepConv experiment comparison

Table 2. RepConv experiment comparison

Table 3. Comparison of EMA adding

Table 3. Comparison of EMA adding

Table 4. Comparison of improved Head end experiment

Table 4. Comparison of improved Head end experiment

Table 5. Comparison of loss function improvement experiment

Table 5. Comparison of loss function improvement experiment

Table 6. Experimental image configuration

Table 6. Experimental image configuration

Table 7. Dataset details

Table 7. Dataset details

Table 8. Number of instances of each label in the dataset

Table 8. Number of instances of each label in the dataset

Table 9. Ablation experiment

Table 9. Ablation experiment

Table 10. YOLOv8 ablation experiments in different versions

Table 10. YOLOv8 ablation experiments in different versions

Table 11. Comparison of Total-loss functions for different losses

Table 11. Comparison of Total-loss functions for different losses

Table 12. Comparative experiments

Table 12. Comparative experiments