| 引用本文: | 李岩,魏笑笑,陈桂珍,李昱燃,马静.面向高效车辆轨迹提取的轻量化目标追踪算法[J].哈尔滨工业大学学报,2026,58(7):54.DOI:10.11918/202507041 |
| Li Yan,Wei Xiaoxiao,Chen Guizhen,Li Yuran,Ma Jing.Lightweight object tracking algorithm for efficient vehicle trajectory extraction[J].Journal of Harbin Institute of Technology,2026,58(7):54.DOI:10.11918/202507041 |
|
| 摘要: |
| 针对路侧设备部署的算法对车辆轨迹提取的效率及精度难以同时保障问题,提出一种高效的轻量化目标跟踪算法DeepSORT-CKFaster。该算法通过Cutmix数据增强方法缓解标签类别不平衡问题,提高YOLOv5目标检测算法的泛化能力;提出一种自适应反馈式的轻量特征提取网络,动态调整特征提取策略,减少冗余计算和频繁的内存访问,实现网络轻量化的同时提升特征提取效率;在训练过程中,提出自适应卡尔曼滤波学习率调整策略,通过实时监控梯度统计量,自动触发余弦下降或线性上升机制,在稳定提取特征的基础上提升模型收敛效果;所提出算法结合YOLOv5目标检测算法,可有效提取车辆轨迹。应用西安市高清视频在两种配置平台验证所提出算法的结果表明:所提出算法的平均目标追踪准确度为92.2%,平均目标追踪精确度为1.106 m,平均目标追踪速度为26.538 帧/s和12.819 帧/s;相比于SORT算法和基准DeepSORT算法,所提出算法的平均目标追踪准确度提高了4.3%~7.1%,平均目标追踪速度提升了4.811~11.789 帧/s。所提出算法特征提取模型的训练时间为6.2 h,计算量为1.376 GFLOPs,为基准DeepSORT算法模型训练时间的72.1%、模型计算量的71.8%。 |
| 关键词: 交通信息技术 轻量化目标追踪算法 DeepSORT-CKFaster 车辆轨迹提取 YOLOv5 |
| DOI:10.11918/202507041 |
| 分类号:U495 |
| 文献标识码:A |
| 基金项目:陕西省自然科学基础研究计划(2025JC-YBMS-7,5JC-YBQN-524);福建省交通运输科技计划重点科技项目(ZD202405,YB202431) |
|
| Lightweight object tracking algorithm for efficient vehicle trajectory extraction |
|
Li Yan1,Wei Xiaoxiao1,Chen Guizhen1,Li Yuran1,Ma Jing2
|
|
(1.School of Transportation Engineering, Chang′an University, Xi′an 710064, China; 2.School of Electronics and Control Engineering, Chang′an University, Xi′an 710064, China)
|
| Abstract: |
| In order to solve the problem that algorithm deployed on roadside equipment are difficult to guarantee the efficiency and accuracy of the vehicle trajectory extraction at the same time, an efficient lightweight object tracking algorithm DeepSORT-CKFaster was proposed. The algorithm alleviates the problem of label category imbalance and improves the generalization ability of the YOLOv5 object detection algorithm through the Cutmix data augmentation method. An adaptive feedback lightweight feature extraction network was proposed, which dynamically adjusted the feature extraction strategy, reduced redundant computing and frequent memory access, and realized the lightweight network and improved the efficiency of feature extraction. In the training process, an adaptive Kalman filter learning rate adjustment strategy was proposed, which automatically triggered the cosine descent or linear rise mechanism through real-time monitoring of the gradient statistics, so as to improve the convergence effect of the model on the basis of stable feature extraction. Combined with the YOLOv5 target detection algorithm, the vehicle trajectory can be effectively extracted. The results of the proposed algorithm verified by high-resolution video in Xi′an on two configuration platforms indicate that the average target tracking accuracy and average target tracking precision of the proposed algorithm are 92.2%, 1.106 m, respectively. The average target tracking speed of the proposed algorithm are 26.538 frames/s and 12.819 frames/s. Compared with the SORT algorithm and the basic DeepSORT algorithm, the average target tracking accuracy and average target tracking speed of the proposed algorithm are increased by 4.3%7.1% and 4.811 frames/s11.789 frames/s, respectively. The training time of the feature extraction model of the proposed algorithm is 6.2 h, and the computation amount is 1.376 GFLOPs, which is 72.1% of the training time and 71.8% of the model computation of the basic DeepSORT algorithm. |
| Key words: traffic information technology lightweight target tracking algorithm DeepSORT-CKFaster vehicle trajectory extraction YOLOv5 |