基于改进YOLOv5算法的道路目标检测方法

doi:10.13229/j.cnki.jdxbgxb.20221461

摘要/Abstract

摘要：

针对现有网络在道路交通场景下的远处目标识别效果欠佳、目标特征表达不充分及目标定位不准确等目标检测问题，提出一种基于改进YOLOv5算法的道路目标检测方法。首先，总结了YOLOv5算法的特征提取结构，分析出原网络结构的不足之处；其次，在原网络基础上增加小目标检测层，通过补充融合特征层及引入额外检测头，提高网络对远处目标的识别能力；再次，对原检测头进行解耦，通过将边框回归和目标分类过程改为两个分支进行，提升网络对目标特征的表达能力；然后，对先验框进行重聚类，通过K-means++算法调整先验框的高宽比例，增强网络对目标的定位能力；最后，以AP、mAP和FPS为评价指标进行消融、对比和可视化验证实验。结果表明：本文算法在 BDD100K数据集上检测速度为95.2帧/s，平均精度达到55.6%，较YOLOv5算法提高6.7%。可见，改进YOLOv5算法在满足检测实时性要求的同时，具备较高的目标检测精度，适用于复杂交通环境下的道路目标检测任务，对提升自动驾驶汽车的视觉感知能力具有指导意义。

关键词: 交通运输安全工程, 环境感知, 目标检测, YOLOv5, 多尺度检测

Abstract:

Aiming at the object detection problems of the existing network such as poor remote object recognition effect， insufficient object feature expression and inaccurate object positioning， a road object detection method based on the improved YOLOv5 algorithm was proposed. Firstly， the feature extraction structure of YOLOv5 algorithm was summarized， and the shortcomings of the original network structure were analyzed. Secondly， the small object detection layer was added to the original network， and the recognition ability of the network for distant objects was improved by supplementing the fusion feature layer and introducing additional detection heads. Thirdly， the original detection head was decoupled， and the expression ability of the network to object features was improved by changing the border regression and object classification process to two branches. Then， the prior box was reclustered， and the height to width ratio of the prior box was adjusted by the K-means++ algorithm to enhance the network's ability to locate the object. Finally， AP， mAP and FPS were used as evaluation indicators for ablation， comparison and visual verification experiments. The results show that the detection speed of the proposed algorithm on the BDD100K dataset is 95.2 frames per second， and the average accuracy reaches 55.6%， which is 6.7% higher than that of the YOLOv5 algorithm. It can be seen that the improved YOLOv5 algorithm not only meets the requirements of real-time detection， but also has good object detection accuracy， which is suitable for road object detection tasks in complex traffic environments， and has guiding significance for improving the visual perception ability of autonomous vehicles.

Key words: transporstation safety engineering, situational awareness, object detection, YOLOv5, multi-scale detection

中图分类号:

U492.8

王宏志,宋明轩,程超,解东旋. 基于改进YOLOv5算法的道路目标检测方法[J]. 吉林大学学报(工学版), 2024, 54(9): 2658-2667.

Hong-zhi WANG,Ming-xuan SONG,Chao CHENG,Dong-xuan XIE. Road object detection method based on improved YOLOv5 algorithm[J]. Journal of Jilin University(Engineering and Technology Edition), 2024, 54(9): 2658-2667.

图/表 15

图1

图2

图3

图4

图5

图6

图7

图8

表1

图9

表2

表3

表4

图10

图11

参考文献 26

1	Gao X, Wu Y, Yang K, et al. Vehicle bottom anomaly detection algorithm based on SIFT[J]. Optik International Journal for Light and Electron Optics, 2015, 126(23): 3562-3566.
2	Chacon I, Rivero A, Ramirez A. Adaptive fuzzy weighted color histogram and HOG appearance model for object tracking with a dynamic trained neural network prediction[J]. Signal, Image and Video Processing, 2021, 15(7): 1585-1592.
3	裘莉娅, 陈玮琳, 李范鸣, 等. 复杂背景下基于LBP纹理特征的运动目标快速检测算法[J]. 红外与毫米波学报, 2022, 41(3): 639-651.
	Qiu Li-ya, Chen Wei-lin, Li Fan-ming, et al. A fast moving target detection algorithm based on LBP texture feature in complex background[J]. Journal of Infrared and Millimeter Waves, 2022, 41(3): 639-651.
4	李雄飞, 王婧, 张小利, 等. 基于SVM和窗口梯度的多焦距图像融合方法[J]. 吉林大学学报: 工学版, 2020, 50(1): 227-236.
	Li Xiong-fei, Wang Jing, Zhang Xiao-li, et al. Multi focus image fusion based on support vector machines and window gradient[J]. Journal of Jilin University(Engineering and Technology Edition), 2020, 50(1): 227-236.
5	Mehmood Z, Asghar S. Customizing SVM as a base learner with AdaBoost ensemble to learn from multi-class problems: a hybrid approach AdaBoost MSVM[J]. Knowledge-based Systems, 2021, 217:No. 106845.
6	He X, Lou B, Yang H, et al. Robust lane change decision making for autonomous vehicles: an observation adversa-rial reinforcement learning approachin[J]. IEEE Transactions on Intelligent Vehicles, 2023, 8(1): 184-193.
7	曹磊, 王强, 史润佳, 等. 基于改进RPN的Faster-RCNN网络SAR图像车辆目标检测方法[J]. 东南大学学报: 自然科学版, 2021, 51(1): 87-91.
	Cao Lei, Wang Qiang, Shi Run-jia, et al. Method for vehicle target detection on SAR image based on improved RPN in Faster-RCNN[J]. Journal of Southeast University (Natural Science Edition), 2021, 51(1): 87-91.
8	Agrawal P, Girshick R, Malik J. Analyzing the performance of multilayer neural networks for object recognition[C]∥European Conference on Computer Vision (ECCV), California, USA, 2014: 329-344.
9	He K M, Zhang X Y, Ren S Q, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 37(9): 1904-1916.
10	Girshick R. Fast R-CNN[C]∥Proceedings of the IEEE International Conference on Computer Vision (ICCV), Santiago, Chile, 2015: 1440-1448.
11	He X, Lou B, Yang H, et al. Robust decision making for autonomous vehicles at highway on-ramps: a constra-ined adversarial reinforcement learning approach[J]. IEEE Transactions on Intelligent Transportation Systems, 2023, 24(4): 4103-4113.
12	Redmon J, Divvala S, Girshick R, et al. You only look once: unified, real-time object detection[C]∥2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, 2016: 779-788.
13	Redmon J, Farhadi A. YOLO9000: better, faster, stronger[C]∥IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, USA, 2017: 6517-6525.
14	Redmon J, Farhadi A. YOLOv3: an incremental improve-ment[DB/OL].[2022-10-26].
15	Bochkovskiy A, Wang C Y, Liao H. YOLOv4: optimal speed and accuracy of object detection[DB/OL].[2022-10-26].
16	侯庆山, 邢进生. 基于Grad-CAM与KL损失的SSD目标检测算法[J]. 电子学报, 2020, 48(12): 2409-2416.
	Hou Qing-shan, Xing Jin-sheng. SSD object detection algorithm based on KL loss and Grad-CAM[J]. Acta Electronica Sinica, 2020, 48(12): 2409-2416.
17	Girshick R, Donahue J, Darrell T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Lake City, USA, 2014: 580-587.
18	Zhu X, Lyu S, Wang X, et al. TPH-YOLOv5: improved YOLOv5 based on transformer prediction head for object detection on drone-captured scenarios[J]. IEEE International Conference on Computer Vision Workshops (ICCVW), Montreal,Canada,2021: 2778-2788.
19	He K M, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, USA, 2016: 770-778.
20	Lin T Y, Dollar P, Girshick R, et al. Feature pyramid networks for object detection[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, USA, 2017: 2117-2125.
21	Liu S, Qi L, Qin H, et al. Path aggregation network for instance segmentation[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recogni-tion, Salt Lake City, USA, 2018: 8759-8768.
22	Song G, Liu Y, Wang X. Revisiting the sibling head in object detector[C]∥2020 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, USA, 2020: 11560-11569.
23	Ren S, He K, Girshick R, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149.
24	Amrouche A, Bentrcia Y, Abed A, et al. Vehicle detection and tracking in real-time using YOLOv4-tiny[C]∥2022 7th International Conference on Image and Signal Processing and their Applications (ISPA), Mostaganem, Algeria, 2022: 1-5.
25	Zhang Y S, Xu W X, Yang S S, et al. Improved YOLOX detection algorithm for contraband in X-ray images[J]. Applied Optics, 2022,61(21): 6297-6310.
26	Hussain M, Al-Aqrabi H, Munawar M, et al. Domain feature mapping with YOLOv7 for automated edge-based pallet racking inspections[J]. Sensors, 2022, 22(18): 6927-6939.

相关文章 15

[1]	杨军,韩鹏飞. 采用神经网络架构搜索的高分辨率遥感影像目标检测[J]. 吉林大学学报(工学版), 2024, 54(9): 2646-2657.
[2]	朱圣杰,王宣,徐芳,彭佳琦,王远超. 机载广域遥感图像的尺度归一化目标检测方法[J]. 吉林大学学报(工学版), 2024, 54(8): 2329-2337.
[3]	游新冬,郭磊,韩晶,吕学强. 一种工件表面压印字符识别网络[J]. 吉林大学学报(工学版), 2024, 54(7): 2072-2079.
[4]	戢晓峰,徐迎豪,普永明,郝京京,覃文文. 山区双车道公路货车移动遮断小客车跟驰风险预测模型[J]. 吉林大学学报(工学版), 2024, 54(5): 1323-1331.
[5]	高云龙,任明,吴川,高文. 基于注意力机制改进的无锚框舰船检测模型[J]. 吉林大学学报(工学版), 2024, 54(5): 1407-1416.
[6]	邬岚,赵乐,李根. 基于方差异质性随机参数模型的汇合行为分析[J]. 吉林大学学报(工学版), 2024, 54(4): 883-889.
[7]	陈仁祥,胡超超,胡小林,杨黎霞,张军,何家乐. 基于改进YOLOv5的驾驶员分心驾驶检测[J]. 吉林大学学报(工学版), 2024, 54(4): 959-968.
[8]	张云佐,郭威,李文博. 遥感图像密集小目标全方位精准检测算法[J]. 吉林大学学报(工学版), 2024, 54(4): 1105-1113.
[9]	张洪,朱志伟,胡天宇,龚燕峰,周建庭. 基于改进YOLOv5s的桥梁螺栓缺陷识别方法[J]. 吉林大学学报(工学版), 2024, 54(3): 749-760.
[10]	王宏志,宋明轩,程超,解东旋. 基于改进YOLOv4-tiny算法的车距预警方法[J]. 吉林大学学报(工学版), 2024, 54(3): 741-748.
[11]	王春华,李恩泽,肖敏. 多特征融合和孪生注意力网络的高分辨率遥感图像目标检测[J]. 吉林大学学报(工学版), 2024, 54(1): 240-250.
[12]	何杰,张长健,严欣彤,王琛玮,叶云涛. 基于微观动力学参数的高速公路特征路段事故风险分析[J]. 吉林大学学报(工学版), 2024, 54(1): 162-172.
[13]	薛珊,张亚亮,吕琼莹,曹国华. 复杂背景下的反无人机系统目标检测算法[J]. 吉林大学学报(工学版), 2023, 53(3): 891-901.
[14]	陶博,颜伏伍,尹智帅,武冬梅. 基于高精度地图增强的三维目标检测算法[J]. 吉林大学学报(工学版), 2023, 53(3): 802-809.
[15]	潘恒彦,张文会,梁婷婷,彭志鹏,高维,王永岗. 基于MIMIC与机器学习的出租车驾驶员交通事故诱因分析[J]. 吉林大学学报(工学版), 2023, 53(2): 457-467.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

检测层	聚类前	聚类后
160×160	None	［6，12，8，29，14，16］
80×80	［10，13，16，30，33，23］	［15，61，19，31，33，129］
40×40	［30，61，62，45，59，119］	［34，32，38，55，65，76］
20×20	［116，90，156，198，373，326］	［98，124，137，204，213，328］

类别名称	训练数量	测试数量
Person （行人）	11 846	1 419
Rider （骑手）	581	68
Car （小汽车）	92 014	10 524
Bus （公共汽车）	1 402	195
Truck （货车）	3 796	451
Bike （自行车）	885	122
Motor （摩托车）	409	43
Traffic light （交通灯）	24 218	2 672
Traffic sign （交通标志）	31 373	3 541

状态	网络结构	mAP@0.5/%	mAP@0.5∶0.95/%
1	YOLOv5	48.9	25.9
2	YOLOv5+小目标检测层	52.9	27.1
3	YOLOv5+小目标检测层+检测头解耦	54.4	28.3
4	YOLOv5+小目标检测层+检测头解耦+先验框重聚类	55.6	28.9

算法	AP/%									mAP@0.5/%	FPS
算法	Person	Rider	Car	Bus	Truck	Bike	Motor	Traffic light	Traffic sign	mAP@0.5/%	FPS
Faster R-CNN^［23］	30.6	23.9	56.7	47.5	49.2	32.3	28.4	8.9	21.6	33.2	15.3
SSD^［16］	31.1	17.6	59.5	44.0	47.1	30.6	27.5	22.1	26.2	34.0	68.8
YOLOv5^［18］	52.6	31.5	72.4	48.3	52.2	38.6	35.8	53.0	55.9	48.9	175.4
YOLOv4-tiny^［24］	18.5	2.9	44.0	26.5	31.0	16.7	5.8	10.6	24.4	20.2	94.5
YOLOX^［25］	55.3	28.7	74.3	53.2	56.3	40.0	42.0	55.8	58.0	51.5	46.7
YOLOv7^［26］	66.3	45.9	76.4	51.9	53.7	48.8	46.1	56.8	60.7	56.3	33.2
YOLOv7-tiny^［26］	49.2	34.0	69.5	44.9	48.5	41.0	29.1	41.3	45.5	44.8	63.4
本文	60.8	38.4	78.3	49.9	56.8	45.4	42.5	64.1	64.4	55.6	95.2