| 摘 要: 在监测交通违法行为的检测任务中,传统的人工监控与基础传感器方法因准确性和实时性存在局限而面临挑战。通过引入多目标跟踪技术,可以显著提升检测效率和系统识别的准确性。文章提出了一种针对车辆信息的多目标跟踪算法,使用MAU-DLA34(MixedAttentionUnit-DeepLayerAggregation34)作为主干网络,替换传统的DLA34主干网络。在处理高密度交通和复杂交互情况下,该算法的表现出色,AP达到73.6%,MOTA为82.9%,IDF1为79.3%,明显优于传统方法和其他深度学习方法,如JDE、DeepSort、CenterTrack等。该算法有效减少了身份切换,确保了对交通参与者的精准且连续跟踪。 |
| 关键词: 多目标跟踪;目标检测;Fairmot;交通违法行为 |
|
中图分类号: TP391.41
文献标识码: A
|
|
| Research on Traffic Violation Detecti on Algorithm Basedon Fairmot |
|
SHI Kaibin, LI Wenshu
|
(School of Computer Science and Technology, Zhejiang Sc-i Tech University, Hangzhou 310018, China)
1103991878@qq.com; charlie@zstu.edu.cn
|
| Abstract: In the detection task of monitoring traffic violations, traditional manual monitoring and basic sensor methods face challenges due to limitations in accuracy and real-time capabilities. By introducing multi-object tracking technology, the detection efficiency and the accuracy of the system identification can be significantly enhanced. This paper proposes a multi-object tracking algorithm specifically designed for vehicle information, using MAU-DLA34 (Mixed Attention Unit-Deep Layer Aggregation 34) as the backbone network, replacing the traditional DLA34 backbone. In handling high-density traffic and complex interaction scenarios, the algorithm demonstrates outstanding performance, achieving an Average Precision (AP) of 73.6%, a Multi-Object Tracking Accuracy (MOTA) of 82.9%,and an IDF1 of 79.3%. These results significantly outperform traditional methods and other deep learning approaches,such as JDE, DeepSort, and CenterTrack. The pr |
| Keywords: multi-object tracking; object detection; Fairmot; traffic violations |
