融合全尺度特征与轨迹修正的遮挡车辆跟踪

郭俊锋; 张子华

doi:10.13700/j.bh.1001-5965.2023.0288

融合全尺度特征与轨迹修正的遮挡车辆跟踪

doi: 10.13700/j.bh.1001-5965.2023.0288

郭俊锋,
张子华^,

兰州理工大学机电工程学院，兰州 730050

基金项目:

国家自然科学基金(51465034)

详细信息

通讯作者:
E-mail：zzh15536308429@163.com

中图分类号: TP391
计量
- 文章访问数: 254
- HTML全文浏览量: 39
- PDF下载量: 10
- 被引次数: 0
出版历程
- 收稿日期: 2023-05-29
- 录用日期: 2023-07-28
- 网络出版日期: 2023-09-01
- 整期出版日期: 2025-05-31

Track obscured vehicles by fusing full-scale features with trajectory correction

GUO Junfeng,
ZHANG Zihua^,

School of Electrical and Mechanical Engineering，Lanzhou University of Technology，Lanzhou 730050，China

Funds:

National Natural Science Foundation of China (51465034)

More Information

Corresponding author: E-mail：zzh15536308429@163.com

摘要

摘要:
为改善车辆跟踪过程中因遮挡造成的跟踪漂移及身份切换（IDS）问题，在Deep SORT算法基础上提出一种融合全尺度特征与轨迹修正的遮挡车辆跟踪方法。引入全尺度特征提取网络提取目标不同尺度特征并实现自适应动态融合，增强目标表观特征。提出一种轨迹修正算法对遮挡过程中的跟踪轨迹进行修复，重新更新Kalman滤波参数以减小遮挡过程中累积的线性误差，优化目标运动特征。结合外观特征和运动特征实现遮挡车辆跟踪。通过消融实验与可视化分析验证所提方法的可行性，在KITTI数据集上的实验结果表明：与现有典型方法相比，所提方法取得最高综合得分为78.13%和最低的IDS次数为192，有效改善遮挡车辆跟踪中的IDS问题，提高车辆跟踪鲁棒性。
- 计算机应用 /
- 多目标跟踪 /
- 车辆跟踪 /
- 全尺度 /
- 轨迹修正
Abstract:
An occluded vehicle tracking method that fuses full-scale features with trajectory correction based on the Deep SORT algorithm is proposed to improve the tracking drift and identity switching (IDS) problems caused by occlusion in vehicle tracking. First, a full-scale feature extraction network is introduced to extract features of different scales of the target and to achieve adaptive dynamic fusion to enhance the apparent features of the target. Then, a trajectory correction approach is suggested to fix the tracking trajectory during the occlusion process, and the Kalman filter parameters are adjusted in order to minimize the cumulative linear errors during the occlusion process and optimize the target motion features. Finally, occluded vehicle tracking is achieved by combining the appearance features and the motion features. The feasibility of the proposed method is verified by ablation experiments and visualized analysis. The suggested approach successfully resolves the IDS issue in obscured vehicle tracking and increases the robustness of vehicle tracking, as demonstrated by experimental results on the KITTI dataset, which yield the highest overall score of 78.13% and the fewest number of IDS (192), when compared to typical existing methods.
- computer application /
- multiple object tracking /
- vehicle tracking /
- full scale /
- trajectory correction

HTML全文

图 1 Kalman滤波示例

Figure 1. Kalman filter example

下载: 全尺寸图片幻灯片

图 2 不同车辆外观特征

Figure 2. Different vehicle appearance features

下载: 全尺寸图片幻灯片

图 3 FSNet示意图

Figure 3. Schematic diagram of FSNet

下载: 全尺寸图片幻灯片

图 4 全尺度特征提取网络结构

Figure 4. FSNet structure

下载: 全尺寸图片幻灯片

图 5 目标不同时刻参数更新

Figure 5. Target parameter update at different moments

下载: 全尺寸图片幻灯片

图 6 轨迹修正示意图

Figure 6. Schematic diagram of trajectory correction

下载: 全尺寸图片幻灯片

图 7 遮挡车辆跟踪流程

Figure 7. Obscured vehicle tracking process

下载: 全尺寸图片幻灯片

图 8 不同算法的成功率曲线

Figure 8. Success rate curves of different algorithms

下载: 全尺寸图片幻灯片

图 9 不同输入下车辆特征注意力热图

Figure 9. Heat map of vehicle feature attention with different inputs

下载: 全尺寸图片幻灯片

图 10 KITTI数据集场景车辆特征热力图

Figure 10. Heat map of vehicle characteristics for dataset scenarios of the KITTI

下载: 全尺寸图片幻灯片

图 11 车辆跟踪轨迹

Figure 11. Vehicle tracking trajectory

下载: 全尺寸图片幻灯片

图 12 不同算法跟踪结果对比

Figure 12. Comparison of tracking results of different algorithms

下载: 全尺寸图片幻灯片

图 13 真实道路场景跟踪结果

Figure 13. Real road scene tracking results

下载: 全尺寸图片幻灯片

表 1 全尺度特征提取网络架构

Table 1. FSNet architecture

层	输出（h×w×c）	操作
conv1	128×64×64 64×32×64	7×7 conv，stride 2 3×3 max pooling，stride 2
conv2	64×32×256	bottleneck×2
transition	64×32×256 32×16×256	1×1 conv 2×2 average pooling，stride 2
conv3	32×16×384	bottleneck×2
transition	32×16×384 16×8×384	1×1 conv 2×2 average pooling，stride 2
conv4	16×8×512	bottleneck×2
conv5	16×8×512	1×1 conv
gap	1×1×512	global average pooling

下载: 导出CSV

表 2 全尺度特征提取网络训练参数

Table 2. FSNet training parameters

训练轮次	批处理大小	初始学习率	权重衰减系数
350	64	0.065	0.0005

下载: 导出CSV

表 3 消融实验

Table 3. Ablation experimental

模型	FSNet	TCA	MOTA/%	ASSA/%	HOTA/%
A			87.37	73.59	72.18
B	√		89.33	73.77	75.42
C		√	87.31	77.89	75.11
D	√	√	90.81	78.27	78.13

下载: 导出CSV

表 4 不同算法在KITTI数据集测试结果

Table 4. Test results of different algorithms on KITTI dataset

算法	MOTA/%	ASSA/%	HOTA/%	IDS次数	FPS
Deep SORT	87.37	73.59	74.02	334	34.51
PermaTrack	90.85	77.66	77.42	275	11.86
OC SORT	90.28	76.39	76.54	250	29.04
StrongSORT	90.35	78.20	77.75	240	7.09
ByteTrack	90.43	76.14	76.60	361	29.63
BOT-SORT	90.72	78.19	77.95	248	6.58
本文方法	90.81	78.27	78.13	192	17.84

下载: 导出CSV

参考文献(16)

[1]	PARK Y, DANG L M, LEE S, et al. Multiple object tracking in deep learning approaches: a survey[J]. Electronics, 2021, 10(19): 2406. doi: 10.3390/electronics10192406
[2]	LUO W H, XING J L, MILAN A, et al. Multiple object tracking: a literature review[J]. Artificial Intelligence, 2021, 293: 103448. doi: 10.1016/j.artint.2020.103448
[3]	周雪, 梁超, 何均洋, 等. 一体化多目标跟踪算法研究综述[J]. 电子科技大学学报, 2022, 51(5): 728-736. doi: 10.12178/1001-0548.2021349 ZHOU X, LIANG C, HE J Y, et al. A survey on one-shot multi-object tracking algorithm[J]. Journal of University of Electronic Science and Technology of China, 2022, 51(5): 728-736(in Chinese). doi: 10.12178/1001-0548.2021349
[4]	韩瑞泽, 冯伟, 郭青, 等. 视频单目标跟踪研究进展综述[J]. 计算机学报, 2022, 45(9): 1877-1907. doi: 10.11897/SP.J.1016.2022.01877 HAN R Z, FENG W, GUO Q, et al. Single object tracking research: a survey[J]. Chinese Journal of Computers, 2022, 45(9): 1877-1907(in Chinese). doi: 10.11897/SP.J.1016.2022.01877
[5]	DENDORFER P, OŠEP A, MILAN A, et al. MOTChallenge: a benchmark for single-camera multiple target tracking[J]. International Journal of Computer Vision, 2021, 129(4): 845-881. doi: 10.1007/s11263-020-01393-0
[6]	CIAPARRONE G, LUQUE SÁNCHEZ F, TABIK S, et al. Deep learning in video multi-object tracking: a survey[J]. Neurocomputing, 2020, 381: 61-88. doi: 10.1016/j.neucom.2019.11.023
[7]	EMAMI P, PARDALOS P M, ELEFTERIADOU L, et al. Machine learning methods for data association in multi-object tracking[J]. ACM Computing Surveys, 2020, 53(4): 1-34.
[8]	BEWLEY A, GE Z Y, OTT L, et al. Simple online and realtime tracking[C]//Proceedings of the IEEE International Conference on Image Processing. Piscataway: IEEE Press, 2016: 3464-3468.
[9]	WOJKE N, BEWLEY A, PAULUS D. Simple online and realtime tracking with a deep association metric[C]//Proceedings of the IEEE International Conference on Image Processing. Piscataway: IEEE Press, 2017: 3645-3649.
[10]	DU Y H, ZHAO Z C, SONG Y, et al. StrongSORT: make DeepSORT great again[J]. IEEE Transactions on Multimedia, 2023, 25: 8725-8737. doi: 10.1109/TMM.2023.3240881
[11]	CAO J K, PANG J M, WENG X S, et al. Observation-centric SORT: rethinking SORT for robust multi-object tracking[EB/OL]. (2023-03-16)[2023-03-20]. https://arxiv.org/abs/2203.14360v3.
[12]	MAGGIOLINO G, AHMAD A, CAO J K, et al. Deep OC-sort: multi-pedestrian tracking by adaptive re-identification[C]//Proceedings of the IEEE International Conference on Image Processing. Piscataway: IEEE Press, 2023: 3025-3029.
[13]	AHARON N, ORFAIG R, BOBROVSKY B Z. BoT-SORT: robust associations multi-pedestrian tracking[EB/OL]. (2022-07-07)[2023-03-20]. https://arxiv.org/abs/2206.14651v2.
[14]	SEIDENSCHWARZ J, BRASÓ G, SERRANO V C, et al. Simple cues lead to a strong multi-object tracker[EB/OL]. (2022-06-09)[2023-03-20]. https://arxiv.org/abs/2206.04656v7.
[15]	YANG F, ODASHIMA S, MASUI S, et al. Hard to track objects with irregular motions and similar appearances? make it easier by buffering the matching space[C]//Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. Piscataway: IEEE Press, 2023: 4788-4797.
[16]	马永杰, 马芸婷, 程时升, 等. 基于改进YOLOv3模型与Deep-SORT算法的道路车辆检测方法[J]. 交通运输工程学报, 2021, 21(2): 222-231. MA Y J, MA Y T, CHENG S S, et al. Road vehicle detection method based on improved YOLOv3 model and deep-SORT algorithm[J]. Journal of Traffic and Transportation Engineering, 2021, 21(2): 222-231(in Chinese).