Track-MT3：一种基于Transformer的新型多目标跟踪算法

陈辉; 杜双燕; 连峰; 韩崇昭

doi:10.12000/JR24164

Track-MT3：一种基于Transformer的新型多目标跟踪算法

DOI: 10.12000/JR24164

陈辉^1, ,,
杜双燕¹,
连峰²,
韩崇昭²

1.
兰州理工大学电气工程与信息工程学院兰州 730050
2.
西安交通大学自动化科学与工程学院西安 710049

基金项目: 国家自然科学基金(62163023, 61873116, 62363023, 62366031)，2024年甘肃省重点人才项目资助

详细信息

作者简介:
陈　辉，教授，博士生导师，主要研究方向为数据融合、统计信号处理、机器学习和智能决策

杜双燕，硕士生，主要研究方向为深度学习和雷达目标跟踪

连　峰，教授，博士生导师，主要研究方向为多源信息融合、滤波与估计算法、气动融合算法

韩崇昭，教授，博士生导师，主要研究方向为数据融合、电子对抗、雷达目标跟踪等

通讯作者:
陈辉 chenh@lut.edu.cn

责任主编：李天成 Corresponding Editor: LI Tiancheng
中图分类号: TP389.1
计量
- 文章访问数: 32
- HTML全文浏览量: 17
- PDF下载量: 5
- 被引次数: 0
出版历程
- 收稿日期: 2024-08-15
- 修回日期: 2024-10-11

Track-MT3: A Novel Multitarget Tracking Algorithm Based on Transformer Network

1.
School of Electrical and Information Engineering, Lanzhou University of Technology, Lanzhou 730050, China
2.
School of Automation Science and Engineering, Xi’an Jiaotong University, Xi’an 710049, China

Funds: The National Natural Science Foundation of China (62163023, 61873116, 62363023, 62366031), The Key Talent Project of Gansu Province in 2024

More Information

Corresponding author: CHEN Hui, chenh@lut.edu.cn

摘要

摘要: 针对复杂环境中多目标跟踪数据关联难度大、难以实现目标长时间稳定跟踪的问题，该文创新性地提出了一种基于Transformer网络的端到端多目标跟踪模型Track-MT3。首先，引入了检测查询和跟踪查询机制，隐式地执行量测-目标的数据关联并且实现了目标的状态估计任务。然后，采用跨帧目标对齐策略增强跟踪轨迹的时间连续性。同时，设计了查询变换与时间特征编码模块强化目标运动建模能力。最后，在模型训练中采用了集体平均损失函数，实现了模型性能的全局优化。通过构造多种复杂的多目标跟踪场景，并利用多重性能指标进行评估，Track-MT3展现了优于MT3等基线方法的长时跟踪性能，与JPDA和MHT方法相比整体性能分别提高了6%和20%，能够有效挖掘时序信息，在复杂动态环境下实现稳定、鲁棒的多目标跟踪。
- 多目标跟踪 /
- 数据关联 /
- Transformer /
- 长时跟踪 /
- 注意力机制
Abstract: To address the challenges associated with the data association and stable long-term tracking of multiple targets in complex environments, this study proposes an innovative end-to-end multitarget tracking model called Track-MT3 based on a transformer network. First, a dual-query mechanism comprising detection and tracking queries is introduced to implicitly perform measurement-to-target data association and enable accurate target state estimation. Subsequently, a cross-frame target alignment strategy is employed to enhance the temporal continuity of tracking trajectories, ensuring consistent target identities across frames. In addition, a query transformation and temporal feature encoding module is designed to improve target motion pattern modeling by adaptively combining target dynamics information at different time scales. During model training, a collective average loss function is adopted to achieve the global optimization of tracking performance, considering the entire tracking process in an end-to-end manner. Finally, the performance of Track-MT3 is extensively evaluated under various complex multitarget tracking scenarios using multiple metrics. Experimental results demonstrate that Track-MT3 exhibits superior long-term tracking performance than baseline methods such as MT3. Specifically, Track-MT3 achieves overall performance improvements of 6% and 20% against JPDA and MHT, respectively. By effectively exploiting temporal information, Track-MT3 ensures stable and robust multitarget tracking in complex dynamic environments.
- Multitarget tracking /
- Data association /
- Transformer /
- Long-term tracking /
- Attention mechanism

HTML全文

图 1 Transformer编码器

Figure 1. Transformer encoder

下载: 全尺寸图片幻灯片

图 2 改进的Transformer解码器

Figure 2. Improved Transformer decoder

下载: 全尺寸图片幻灯片

图 3 Track-MT3模型架构示意图

Figure 3. Schematic diagram of Track-MT3 model architecture

下载: 全尺寸图片幻灯片

图 4 检测查询和跟踪查询示意图

Figure 4. Schematic diagram of detection query and track query

下载: 全尺寸图片幻灯片

图 5 查询变换与时间特征编码模块

Figure 5. Query transformation and temporal feature encoding module

下载: 全尺寸图片幻灯片

图 6 训练损失函数曲线

Figure 6. Training loss function curve

下载: 全尺寸图片幻灯片

图 7 一个滑动窗口下模型的输入和输出

Figure 7. Inputs and outputs of the model under a sliding window

下载: 全尺寸图片幻灯片

图 8 编码器输出数据分析可视化

Figure 8. Visualisation of the analysis of the encoder output data

下载: 全尺寸图片幻灯片

图 9 查询向量和编码器输出的注意力分数可视化

Figure 9. Attention score visualisation of query vectors and encoder outputs

下载: 全尺寸图片幻灯片

图 10 不同实验场景下的轨迹跟踪图

Figure 10. Trajectory tracking plots for different experimental scenarios

下载: 全尺寸图片幻灯片

图 11 不同实验场景下目标数量变化图

Figure 11. Variation of the number of targets in different experimental scenarios

下载: 全尺寸图片幻灯片

图 12 不同场景下评价指标对比

Figure 12. Comparison of evaluation indicators in different scenarios

下载: 全尺寸图片幻灯片

图 13 查询置信度阈值稳健性分析

Figure 13. Robustness Analysis of Query Confidence Threshold

下载: 全尺寸图片幻灯片

图 14 鲁棒性测试

Figure 14. Robustness test

下载: 全尺寸图片幻灯片

表 1 训练样本信息

Table 1. Training Sample Information

参数	数值
总的样本数(有效量测点数)	401651991
真实目标量测点数	81664937
杂波量测点数	319987054
平均每个批次样本总数	8034
平均每个时间窗口样本总数	252

下载: 导出CSV

表 2 实验环境

Table 2. Experimental environment

项目	版本
CPU	12th Gen Intel(R) Core i5-12400
GPU	NVIDIA GeForce RTX 3090 Ti
Python	3.7.4
Pytorch	1.6.0
Torchvision	0.7.0
CUDA	4.14.0

下载: 导出CSV

表 3 Track-MT3网络参数

Table 3. Track-MT3 network parameters

参数	取值
编码器层数	6
解码器层数	6
编码器输入数据维度	256
解码器输入数据层数	256
多头注意力头数	8
查询向量数量	16
前馈网络隐藏层维度	2048
神经元Dropout	0.1
预测器MLP层数	3
预测器隐藏层维度	128

下载: 导出CSV

表 4 模型训练参数

Table 4. Model training parameters

参数	取值
优化器	Adam
Epoch数	50000
Batch Size	32
初始学习率	0.0002
学习率衰减容忍度	5000
学习率衰减因子	0.5

下载: 导出CSV

表 5 不同仿真场景参数设置

Table 5. Parameter settings for different simulation scenarios

参数	场景1	场景2	场景3
目标数量(个)	6	6	10
出生率	0.04	0.08	0.12
死亡率	0.01	0.02	0.03

下载: 导出CSV

表 6 跟踪准确性对比

Table 6. Tracking accuracy comparison

跟踪方法	定位误差	漏检误差	虚警误差
JPDA	0.1629	0.6208	4.2812
MHT	0.6006	1.5921	3.8717
Track-MT3	0.0588	2.3683	2.3708

下载: 导出CSV

表 7 计算效率对比

Table 7. Computational efficiency comparison

跟踪方法	单帧运行时间(s)	平均内存占用(MB)
JPDA	0.0041	169.6641
MHT	0.1714	209.8398
Track-MT3	0.0123	253.6656

下载: 导出CSV

表 8 QTM 消融实验

Table 8. QTM ablation experiment

评价指标	Full	No-QTM
GOSPA(×10^–1 m)	3.546362	4.76092
Pro-GOSPA(×10^–1 m)	1.340019	1.925471

下载: 导出CSV

表 9 实验参数设置

Table 9. Experimental parameter settings

参数	实验1	实验2	实验3
${P_{\mathrm{D}}}$	0.95	0.9	0.85
${\sigma _{\mathrm{q}}}$	0.01	0.02	0.03
${\sigma _{\mathrm{r}}}$	0.1	0.9	2.0
${\lambda _{\mathrm{c}}}$	5	10	15

下载: 导出CSV

参考文献(34)

[1]	BAI Xianglong, LAN Hua, WANG Zengfu, et al. Robust Multitarget Tracking in Interference Environments: A Message-Passing Approach[J]. IEEE Transactions on Aerospace and Electronic Systems, 2024, 60(1): 360–386. doi: 10.1109/TAES.2023.3323629.
[2]	YANG Jialin, JIANG Defu, TAO Jin, et al. A Sector-Matching Probability Hypothesis Density Filter for Radar Multiple Target Tracking[J]. Applied Sciences, 2023, 13(5): 2834. doi: 10.3390/app13052834.
[3]	HEM A G, BAERVELDT M, and BREKKE E F. PMBM Filtering With Fusion of Target-Provided and Exteroceptive Measurements: Applications to Maritime Point and Extended Object Tracking[J]. IEEE Access, 2024, 12: 55404–55423. doi: 10.1109/ACCESS.2024.3389824.
[4]	CHEN Jiahui, GUO Shisheng, LUO Haolan, et al. Non-Line-of-Sight Multi-Target Localization Algorithm for Driver-Assistance Radar System[J]. IEEE Transactions on Vehicular Technology, 2023, 72(4): 5332–5337. doi: 10.1109/TVT.2022.3227971.
[5]	HERZOG F, CHEN Junpeng, TEEPE T, et al. Synthehicle: Multi-Vehicle Multi-Camera Tracking in Virtual Cities[C]. 2023 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops. Waikoloa, USA, 2023: 1-11. doi: 10.1109/WACVW58289.2023.00005.
[6]	RAKAI L, SONG Huansheng, SUN Shijie, et al. Data association in multiple object tracking: A survey of recent techniques[J]. Expert Systems with Applications, 2022, 192: 116300. doi: 10.1016/j.eswa.2021.116300.
[7]	LI Tiancheng, LIANG Haozhe, XIAO Bing, et al. Finite mixture modeling in time series: A survey of Bayesian filters and fusion approaches[J]. Information Fusion, 2023, 98: 101827. doi: 10.1016/j.inffus.2023.101827.
[8]	LIU Zongxiang, LUO Junwen, and ZHOU Chunmei. Multi-Hypothesis Marginal Multi-Target Bayes Filter for a Heavy-Tailed Observation Noise[J]. Remote Sensing, 2023, 15(21): 5258. doi: 10.3390/rs15215258.
[9]	QIU Changzhen, ZHANG Zhiyong, LU Huanzhang, et al. A Survey of Motion-based Multitarget tracking methods[J]. Progress In Electromagnetics Research B, 2015, 62: 195–223. doi: 10.2528/PIERB15010503.
[10]	Vo B N and MA W K. The Gaussian Mixture Probability Hypothesis Density Filter[J]. IEEE Transactions on Signal Processing, 2006, 54(11): 4091–4104. doi: 10.1109/TSP.2006.881190.
[11]	Vo B T, Vo B N, and CANTONI A. Analytic Implementations of the Cardinalized Probability Hypothesis Density Filter[J]. IEEE Transactions on Signal Processing, 2007, 55(7): 3553–3567. doi: 10.1109/TSP.2007.894241.
[12]	Vo B T, Vo B N, and CANTONI A. The Cardinality Balanced Multi-Target Multi-Bernoulli Filter and Its Implementations[J]. IEEE Transactions on Signal Processing, 2009, 57(2): 409–423. doi: 10.1109/TSP.2008.2007924.
[13]	Vo B N, Vo B T, and PHUNG D. Labeled Random Finite Sets and the Bayes Multi-Target Tracking Filter[J]. IEEE Transactions on Signal Processing, 2014, 62(24): 6554–6567. doi: 10.1109/TSP.2014.2364014.
[14]	GARCÍA-FERNÁNDEZ Á F, WILLIAMS J L, GRANSTRÖM K, et al. Poisson multi-Bernoulli mixture filter: direct derivation and implementation[J]. IEEE Transactions on Aerospace and Electronic Systems, 2018, 54(4): 1883–1901. doi: 10.1109/TAES.2018.2805153.
[15]	CHONG C Y. An Overview of Machine Learning Methods for Multiple Target Tracking[C]. 2021 IEEE 24th International Conference on Information Fusion, Sun City, South Africa, 2021: 1-9. doi: 10.23919/FUSION49465.2021.9627045.
[16]	JONDHALE S R and DESHPANDE R S. Kalman Filtering Framework-Based Real Time Target Tracking in Wireless Sensor Networks Using Generalized Regression Neural Networks[J]. IEEE Sensors Journal, 2019, 19(1): 224–233. doi: 10.1109/JSEN.2018.2873357.
[17]	LIU Huajun, ZHANG Hui, and MERTZ C. DeepDA: LSTM-based Deep Data Association Network for Multi-Targets Tracking in Clutter[C]. 22th International Conference on Information Fusion, Ottawa, Canada, 2019: 1-8. doi: 10.23919/FUSION43075.2019.9011217.
[18]	BECKER P, PANDYA H, GEBHARDT G H W, et al. Recurrent Kalman Networks: Factorized Inference in High-Dimensional Deep Feature Spaces [C]. International Conference on Machine Learning, Long Beach, USA, 2019: 544-552. doi: 10.48550/arXiv.1905.07357.
[19]	SHI Zhuangwei. Incorporating Transformer and LSTM to Kalman Filter with EM algorithm for state estimation[OL]. https://doi.org/10.48550/arXiv.2105.00250.
[20]	GAO Chang, YAN Junkun, ZHOU Shenghua, et al. Long short-term memory-based deep recurrent neural networks for target tracking[J]. Information Sciences, 2019, 502: 279–296. doi: 10.1016/j.ins.2019.06.039.
[21]	ZHANG Yongquan, SHI Zhenyun, JI Hongbing, et al. Online multi-target intelligent tracking using a deep long-short term memory network[J]. Chinese Journal of Aeronautics, 2023, 36(9): 313–329. doi: 10.1016/j.cja.2023.02.006.
[22]	LI Jing, LIANG Xinru, YUAN Shengzhi, et al. A Strong Maneuvering Target-Tracking Filtering Based on Intelligent Algorithm[J]. International Journal of Aerospace Engineering, 2024, 2024(1): 9981332. doi: 10.1155/2024/9981332.
[23]	EMAMBAKHSH M, BAY A, and VAZQUEZ E. Deep Recurrent Neural Network for Multi-target Filtering [C]. MultiMedia Modeling: 25th International Conference, Thessaloniki, Greece, 2019: 519-531. doi: 10.1007/978-3-030-05716-9_42.
[24]	LIU Jingxian, WANG Zulin, and XU Mai. DeepMTT: A deep learning maneuvering target-tracking algorithm based on bidirectional LSTM network[J]. Information Fusion, 2020, 53: 289–304. doi: 10.1016/j.inffus.2019.06.012.
[25]	VASWANI A, SHAZEER N, PARMAR N, et al. Attention is All you Need [C]. Proceedings of the 31st International Conference on Neural Information Processing Systems, Long Beach, USA, 2017: 6000–6010.
[26]	ZENG Ailing, CHEN Muxi, ZHANG Lei, et al. Are Transformers Effective for Time Series Forecasting?[C]. 37th AAAI Conference on Artificial Intelligence, Washington, USA, 2023: 11121-11128. doi: 10.1609/aaai.v37i9.26317.
[27]	PINTO J, HESS G, LJUNGBERGH W, et al. Next Generation Multitarget Trackers: Random Finite Set Methods vs Transformer-based Deep Learning[C]. 2021 IEEE 24th International Conference on Information Fusion, Sun City, South Africa, 2021: 1-8. doi: 10.23919/FUSION49465.2021.9626990.
[28]	PINTO J, HESS G, LJUNGBERGH W, et al. Can Deep Learning be Applied to Model-Based Multi-Object Tracking?[OL]. https://doi.org/10.48550/arXiv.2202.07909.
[29]	MEINHARDT T, KIRILLOV A, LEAL-TAIXÉ L, et al. TrackFormer: Multi-Object Tracking with Transformers[C]. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, USA, 2022: 8844-8854. doi: 10.1109/CVPR52688.2022.00864.
[30]	ZENG Fangao, DONG Bin, ZHANG Yuang, et al. MOTR: End-to-End Multiple-Object Tracking with Transformer [C]. 17th European Conference on Computer Vision, Tel Aviv, Israel, 2022: 659-675. doi: 10.1007/978-3-031-19812-0_38.
[31]	WANG Qiang, LI Bei, XIAO Tong, et al. Learning Deep Transformer Models for Machine Translation [C]. 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, 2019: 1810–1822. doi: 10.18653/v1/P19-1176.
[32]	CARION N, MASSA F, SYNNAEVE G, et al. End-to-End Object Detection with Transformers [C]. 16th European conference on computer vision, Glasgow, UK, 2020: 213-229. doi: 10.1007/978-3-030-58452-8_13.
[33]	BEARD M, VO B T, and VO B N. Bayesian Multi-target tracking with Merged Measurements using Labelled Random Finite Sets[J]. IEEE Transactions on Signal Processing, 2015, 63(6): 1433–1447. doi: 10.1109/TSP.2015.2393843.
[34]	RAHMATHULLAH A S, GARCÍA-FERNÁNDEZ Á F, and SVENSSON L. Generalized optimal sub-pattern assignment metric[C]. 2017 20th International Conference on Information Fusion, Xi’an, China, 2017: 1–8. doi: 10.23919/ICIF.2017.8009645.