改进YOLOv5s-CSC和Deep-Sort的远端小目标车辆检测跟踪算法

胡欣; 邵良彬; 周运强; 肖剑; 程鸿亮

doi:10.3969/j.issn.0258-2724.20240528

改进YOLOv5s-CSC和Deep-Sort的远端小目标车辆检测跟踪算法

doi: 10.3969/j.issn.0258-2724.20240528

胡欣^1,,
邵良彬¹,
周运强²,
肖剑^3, ,,
程鸿亮³

1.
长安大学能源与电气工程学院，陕西西安 710064
2.
重庆长安汽车股份有限公司，重庆 400023
3.
长安大学电子与控制工程学院，陕西西安 710064

基金项目: 陕西省秦创原“科学家 + 工程师”队伍建设项目（2024QCY-KXJ-161）；咸阳市重点研发计划项目（L2024-ZDYF-ZDYF-GY-0004）

详细信息

作者简介:
胡欣（1975—），女，教授，博士，研究方向为计算机视觉、机器学习及深度学习、能源管理等，E-mail：huxin@chd.edu.cn

通讯作者:
肖剑（1975—），男，副教授，博士，研究方向为人工智能应用、模式识别及计算机视觉等，E-mail：xiaojian@chd.edu.cn

中图分类号: TP391
计量
- 文章访问数: 22
- HTML全文浏览量: 13
- PDF下载量: 12
- 被引次数: 0
出版历程
- 收稿日期: 2024-10-22
- 修回日期: 2025-03-07
- 网络出版日期: 2026-07-27

Vehicle Detection and Tracking Algorithm for Distant Small Targets Based on Improved YOLOv5s-CSC and Deep-Sort

1.
School of Energy and Electrical Engineering, Chang’an University, Xi’an 710064, China
2.
Chongqing Chang’an Automobile Company Limited, Chongqing 400023, China
3.
School of Electronics and Control Engineering, Chang’an University, Xi’an 710064, China

摘要

摘要:
针对交通监控场景中远端小目标车辆易漏检、跟踪过程中ID切换与跳变等问题，提出一种改进YOLOv5s （You Only Look Once version 5 small）与改进Deep-Sort （deep simple online and realtime tracking）相结合的小目标检测跟踪算法. 在检测方面，以YOLOv5s为基础网络，引入CBAM （convolutional block attention module）注意力机制增强对小目标特征的捕捉，将Swin Transformer Block与主干网络中的CSP （cross stage partial）融合，构建CSPSTB模块，强化全局信息关注；并采用CARAFE （content-aware reassembly of features）上采样算子减少特征损失，形成YOLOv5s-CSC检测网络. 在跟踪方面，设计含Res_SE （residual squeeze-and-excitation）结构的深度表观特征提取网络，在数据关联过程中更能适应车辆特征. 在UA-DETRAC数据集上的实验结果表明，相较于原YOLOv5s模型，改进后的检测算法在平均精度均值m_AP、交并比（IoU）阈值为0.50的平均精度均值m_AP50及IoU阈值为0.75的平均精度均值m_AP75分别提高1.5%、1.5%、1.9%，小目标检测精度提升4.2%. 在VERI-Wild数据集上的跟踪实验显示，结合改进检测器与跟踪网络后，多目标跟踪精度、高阶跟踪精度、IDF1分数（identity F1 score）分别提高9.6%、3.2%、7.7%. YOLOv5s-CSC与改进的Deep-Sort结合后，有效克服了远端小目标车辆跟踪失效及跟踪过程中ID切换、跳变等难题，验证了改进后模型在跟踪精度与鲁棒性方面的优越性.
- 目标检测 /
- 目标跟踪 /
- YOLOv5s /
- Deep-Sort /
- Swin Transformer
Abstract:
To address the problems of missed detection of distant small vehicle targets, as well as ID switching and jumping during tracking in traffic surveillance scenarios, a small-target detection and tracking algorithm combining improved You Only Look Once version 5 small (YOLOv5s) and improved deep simple online and realtime tracking (Deep-Sort) was proposed. For detection, based on the YOLOv5s baseline network, the convolutional block attention module (CBAM) mechanism was introduced to enhance the feature capture of small targets. The Swin Transformer Block was fused with the cross stage partial (CSP) structure in the backbone network to construct the CSPSTB module and strengthen global information attention. The content-aware reassembly of features (CARAFE) upsampling operator was adopted to reduce feature loss, forming the YOLOv5s-CSC detection network. For tracking, a deep appearance feature extraction network containing a residual squeeze-and-excitation (Res_SE) structure was designed, which was more adaptable to vehicle features during data association. Experimental results on the UA-DETRAC dataset indicate that, compared with the original YOLOv5s model, the improved detection algorithm increases the mean average precision (m_AP)_, mean average precision at an intersection over union (IoU) threshold of 0.50 (m_AP50), and mean average precision at an IoU threshold of 0.75 (mAP75) by 1.5%, 1.5%, and 1.9%, respectively, and the small-target detection accuracy improves by 4.2%. Tracking experiments on the VERI-Wild dataset show that, after combining the improved detector and tracking network, the multi-target tracking accuracy, high-order tracking accuracy, and IDF1 (identity F1) score increase by 9.6%, 3.2%, and 7.7%, respectively. The combination of YOLOv5s-CSC and the improved Deep-Sort effectively overcomes difficulties of tracking failure of distant small vehicle targets, as well as ID switching and jumping during tracking, verifying the superiority of the improved model in tracking accuracy and robustness.
- target detection /
- target tracking /
- YOLOv5s /
- Deep-Sort /
- Swin Transformer

HTML全文

图 1 YOLOv5s-CSC网络结构

Figure 1. YOLOv5s-CSC network structure

下载: 全尺寸图片幻灯片

图 2 CBAM网络结构

Figure 2. CBAM network structure

下载: 全尺寸图片幻灯片

图 3 STB网络结构

Figure 3. STB network structure

下载: 全尺寸图片幻灯片

图 4 CSPSTB网络结构

Figure 4. CSPSTB network structure

下载: 全尺寸图片幻灯片

图 5 CARAFE网络结构

Figure 5. CARAFE network structure

下载: 全尺寸图片幻灯片

图 6 改进Deep-Sort深度学习表观特征提取网络结构

Figure 6. Structure of improved Deep-Sort deep learning appearance feature extraction network

下载: 全尺寸图片幻灯片

图 7 改进前后残差块对比

Figure 7. Comparison of residual blocks before and after improvement

下载: 全尺寸图片幻灯片

图 8 目标检测与跟踪算法结合流程图

Figure 8. Flowchart of combination of target detection and tracking algorithms

下载: 全尺寸图片幻灯片

图 9 UA-DETRAC部分数据集

Figure 9. Partial UA-DETRAC dataset

下载: 全尺寸图片幻灯片

图 10 VERI-Wild部分数据集

Figure 10. Partial VERI-Wild dataset

下载: 全尺寸图片幻灯片

图 11 ID跳变跟踪效果

Figure 11. Tracking effect of ID jumping

下载: 全尺寸图片幻灯片

图 12 ID切换跟踪效果

Figure 12. Tracking effect of ID switching

下载: 全尺寸图片幻灯片

表 1 实验参数配置

Table 1. Experimental parameter configuration

名称	目标检测算法	目标跟踪算法
数据集	UA-DETRAC	VERI-Wild
输入尺寸	640 × 640	64 × 128
优化器（Optimizer）	SGD	SGD
训练批次（Batch size）	8	64
训练轮次（Epoch）	100	200
学习率	0.01	0.10
最小学习率	0.0001	0.0100

下载: 导出CSV

表 2 YOLOv5s-CSC算法的检测结果

Table 2. Detection results of YOLOv5s-CSC algorithm

类型	m_AP/%	m_AP50/%	m_AP75/%
car	65.2	85.4	74.4
bus	66.1	89.4	76.4
van	60.5	80.5	73.5
others	50.4	70.1	58.1
全部	58.3	84.0	69.2

下载: 导出CSV

表 3 YOLOv5s添加不同模块检测算法实验对比

Table 3. Experimental comparison of detection algorithms with different modules added to YOLOv5s %

模型	m_AP	m_AP50	m_AP75	A_PS	A_PM	A_PL
YOLOv5s	56.8	82.5	67.3	11.0	54.9	78.0
YOLOv5s + CSPSTB	58.2	83.5	69.0	11.7	56.2	78.9
YOLOv5s + CBAM	57.9	83.9	68.5	13.6	55.8	78.3
YOLOv5s + CARAFE	57.4	83.6	67.9	11.5	55.8	78.3
YOLOv5s + CSPSTB + CBAM + CARAFE	58.3	84.0	69.2	15.2	56.3	78.4

下载: 导出CSV

表 4 不同天气下算法改进前后对比

Table 4. Comparison of algorithm before and after improvement under different weather conditions

时段	原图	YOLOv5s	YOLOv5s-CSC
白天
雨天
傍晚
夜间

下载: 导出CSV

表 5 YOLOv5s加入不同注意力

Table 5. YOLOv5s with different attention mechanisms %

模型	m_AP	m_AP50	m_AP75
YOLOv5s + SE	57.1	82.6	67.8
YOLOv5s + ECA	57.5	83.7	68.0
YOLOv5s + CA	57.8	83.7	68.6
YOLOv5s + CBAM	57.9	83.9	68.5

下载: 导出CSV

表 6 不同网络性能对比

Table 6. Performance comparison of different networks

模型	主干网络	m_AP/%	m_AP50/%	m_AP75/%	参数量/MB	F_PS/（帧·s⁻¹）
SSD	VGG	45.1	72.0	50.8	26.29	88.7
Faster R-CNN	ResNet-50	43.8	77.8	44.4	137.10	13.4
YOLOv3	Darknet-53	46.6	78.6	51.3	61.95	56.4
YOLOv4	CSPDarknet-53	52.5	81.8	60.3	64.36	45.3
YOLOv5s	CSPDarknet-53	56.8	82.5	67.3	7.27	78.5
YOLOX	Darknet-53	58.1	83.5	68.8	8.97	77.2
CenterNet	ResNet-50	53.4	81.3	62.2	32.67	70.3
YOLOv7-tiny	ELAN	55.4	80.3	65.2	6.23	85.6
YOLOv8s	CSPDarknet-53	56.9	82.4	66.9	5.64	95.4
YOLOv9s	CSPDarknet-53	57.2	82.8	67.2	6.42	90.8
YOLOv10s	CSPDarknet-53	57.4	83.0	67.0	8.69	86.4
YOLO11s	CSPDarknet-53	58.0	83.6	68.4	7.53	87.7
YOLOv5s-CSC	CSPDarknet-53	58.3	84.0	69.2	8.42	63.7

下载: 导出CSV

表 7 不同改进方法对比实验

Table 7. Comparison experiments of different improvement methods %

模型	A_MOTA	A_HOTA	F_1,ID
YOLOv5s + WRN	69.2	63.3	82.5
YOLOv5s-CSC + WRN	78.1	65.2	87.9
YOLOv5s-CSC + Impoved Net	78.5	66.3	89.4
YOLOv5s-CSC + Res_SE Net	78.8	66.5	90.2

下载: 导出CSV

参考文献(31)

[1]	赵雪亭, 胡立伟. 城市交通拥塞源参数反演性能综合对比[J]. 西南交通大学学报, 2025, 60(1): 243-252. doi: 10.3969/j.issn.0258-2724.20230008 Zhao Xueting, Hu Liwei. Comprehensive comparison of inversion performance of urban traffic congestion source parameters[J]. Journal of Southwest Jiaotong University, 2025, 60(1): 243-252. doi: 10.3969/j.issn.0258-2724.20230008
[2]	夏英, 刘敏. 基于时空注意力卷积神经网络的交通流量预测[J]. 西南交通大学学报, 2023, 58(2): 340-347. doi: 10.3969/j.issn.0258-2724.20210526 Xia Ying, Liu Min. Traffic flow prediction based on spatial-temporal attention convolutional neural network[J]. Journal of Southwest Jiaotong University, 2023, 58(2): 340-347. doi: 10.3969/j.issn.0258-2724.20210526
[3]	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 (CVPR). Piscataway: IEEE, 2016: 779-788.
[4]	Redmon J, Farhadi A. YOLO9000: better, faster, stronger[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2017: 6517-6525.
[5]	Redmon J, Farhadi A. YOLOv3: an incremental improvement[PP/OL]. V1. arXiv (2018-04-08)[2024-10-12]. https://doi.org/10.48550/arXiv.1804.02767.
[6]	Bochkovskiy A, Wang C Y, Liao H M. YOLOv4: optimal speed and accuracy of object detection[PP/OL]. V1. arXiv (2020-04-23) [2024-10-12]. https://doi.org/10.48550/arXiv.2004.10934.
[7]	Li C Y, Li L L, Jiang H L, et al. YOLOv6: a single-stage object detection framework for industrial applications[PP/OL]. V1. arXiv (2022-09-07) [2024-10-14]. https://doi.org/10.48550/arXiv.2209.02976.
[8]	Wang C Y, Bochkovskiy A, Liao H M. YOLOv7: trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2023: 7464-7475.
[9]	Tian Z, Shen C H, Chen H, et al. FCOS: fully convolutional one-stage object detection[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2019: 9626-9635.
[10]	Girshick R, Donahue J, Darrell T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//2014 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2014: 580-587.
[11]	Ren S Q, He K M, 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. doi: 10.1109/TPAMI.2016.2577031
[12]	蒲玲玲, 杨柳. 改进YOLOv5的多车辆目标实时检测及跟踪算法[J]. 科学技术与工程, 2023, 23(28): 12159-12167. doi: 10.3969/j.issn.1671-1815.2023.28.031 Pu Lingling, Yang Liu. Improved real-time detection and tracking algorithm for multi vehicle targets in YOLOv5[J]. Science Technology and Engineering, 2023, 23(28): 12159-12167. doi: 10.3969/j.issn.1671-1815.2023.28.031
[13]	Abdelali H A, Derrouz H, Zennayi Y, et al. Multiple hypothesis detection and tracking using deep learning for video traffic surveillance[J]. IEEE Access, 2021, 9: 164282-164291. doi: 10.1109/ACCESS.2021.3133529
[14]	Hassaballah M, Kenk M A, Muhammad K, et al. Vehicle detection and tracking in adverse weather using a deep learning framework[J]. IEEE Transactions on Intelligent Transportation Systems, 2021, 22(7): 4230-4242. doi: 10.1109/TITS.2020.3014013
[15]	Zhang X X, Story B, Rajan D. Night time vehicle detection and tracking by fusing vehicle parts from multiple cameras[J]. IEEE Transactions on Intelligent Transportation Systems, 2022, 23(7): 8136-8156. doi: 10.1109/TITS.2021.3076406
[16]	Jin J N, Wu Y F, Gao J P. Research on target detection and tracking algorithm for campus driverless vehicle based on improved DeepSORT[C]//2022 China Automation Congress (CAC). Piscataway: IEEE, 2022: 1304-1308.
[17]	Tian Z L, Wahab M N A, Akbar M F, et al. SFFSORT multi-object tracking by shallow feature fusion for vehicle counting[J]. IEEE Access, 2023, 11: 76827-76841. doi: 10.1109/ACCESS.2023.3297190
[18]	Choi J, Ha S, Lee Y, et al. Vehicle tracking system in drone imagery with YOLOv5 and histogram[C]//2023 62nd Annual Conference of the Society of Instrument and Control Engineers (SICE). Piscataway: IEEE, 2023: 778-783.
[19]	华泽玺, 施会斌, 罗彦, 等. 基于轻量级YOLO-v4模型的变电站数字仪表检测识别[J]. 西南交通大学学报, 2024, 59(1): 70-80. doi: 10.3969/j.issn.0258-2724.20210544 Hua Zexi, Shi Huibin, Luo Yan, et al. Detection and recognition of digital instruments based on lightweight YOLO-v4 model at substations[J]. Journal of Southwest Jiaotong University, 2024, 59(1): 70-80. doi: 10.3969/j.issn.0258-2724.20210544
[20]	Woo S, Park J, Lee J Y, et al. CBAM: convolutional block attention module[C]//Computer Vision– ECCV 2018. Cham: Springer International Publishing, 2018: 3-19.
[21]	Liu Z, Lin Y T, Cao Y, et al. Swin transformer: hierarchical vision transformer using shifted windows[C]//2021 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2021: 9992-10002.
[22]	Wang J Q, Chen K, Xu R, et al. CARAFE: content-aware reassembly of features[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2019: 3007-3016.
[23]	胡欣, 胡帅, 马丽军, 等. 基于融合MBAM与YOLOv5的PCB缺陷检测方法[J]. 图学学报, 2024, 45(1): 47-55. doi: 10.11996/JG.j.2095-302X.2024010047 Hu Xin, Hu Shuai, Ma Lijun, et al. PCB defect detection method based on fusion of MBAM and YOLOv5[J]. Journal of Graphics, 2024, 45(1): 47-55. doi: 10.11996/JG.j.2095-302X.2024010047
[24]	Wojke N, Bewley A, Paulus D. Simple online and realtime tracking with a deep association metric[C]//2017 IEEE International Conference on Image Processing (ICIP). Piscataway: IEEE, 2017: 3645-3649.
[25]	Bewley A, Ge Z Y, Ott L, et al. Simple online and realtime tracking[C]//2016 IEEE International Conference on Image Processing (ICIP). Piscataway: IEEE, 2016: 3464-3468.
[26]	Zagoruyko S, Komodakis N. Wide residual networks[PP/OL]. V4. arXiv (2017-06-14)[2024-09-20]. https://doi.org/10.48550/arXiv.1605.07146.
[27]	Hu J, Shen L, Sun G. Squeeze-and-excitation networks[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2018: 7132-7141.
[28]	Wen L Y, Du D W, Cai Z W, et al. UA-DETRAC: a new benchmark and protocol for multi-object detection and tracking[J]. Computer Vision and Image Understanding, 2020, 193: 102907. doi: 10.1016/j.cviu.2020.102907
[29]	Lou Y H, Bai Y, Liu J, et al. VERI-wild: a large dataset and a new method for vehicle re-identification in the wild[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2019: 3230-3238.
[30]	Wang Q L, Wu B G, Zhu P F, et al. ECA-net: efficient channel attention for deep convolutional neural networks[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2020: 11531-11539.
[31]	Hou Q B, Zhou D Q, Feng J S. Coordinate attention for efficient mobile network design[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2021: 13708-13717.