改进YOLOv5的多车辆目标实时检测及跟踪算法

多车辆目标跟踪时间主要花费在车辆检测模块和对每个车辆表观特征提取模块，一般情况下，车辆检测和车辆表观特征提取是在不同的神经网络中进行的，且一张图中的车辆目标越多，对车辆表观特征提取耗费时间的也越多，推理时间也相应变长。针对这一问题，基于经典的Tracking-By-Detection模式，提出一种改进的YOLO模型：在YOLO网络中添加ReID特征识别模块，使YOLO在输出目标位置信息的同时输出目标特征信息，以提高算法的跟踪速度。针对车辆间彼此覆盖的情况，提出一种基于动态IOU阈值的非极大抑制算法，以提高算法的跟踪精度。最后将YOLO输出的信息进行数据匹配，从而实现多目标跟踪。在UA-DETRAC数据集上验证改进模型的有效性，实验结果表明，将YOLOv5网络进行改进后运用在目标跟踪算法中，相对于经典的YOLO+DeepSORT跟踪模型，在车辆密集的情景下平均推理时间减少了17%；在改进后的网络上添加动态IOU阈值非极大抑制，跟踪精度提高了3.9个百分点。改进后的模型有较好的实时性与跟踪准确率。

Improved real-time detection and tracking algorithm for multi vehicle targets in YOLOv5

Multi vehicle target tracking time is mainly spent on vehicle detection module and each vehicle apparent feature extrac-tion module. Generally, vehicle detection and vehicle apparent feature extraction are carried out in different neural net-works, and the more vehicle targets in a graph, the more time is spent on vehicle apparent feature extraction, and the reasoning time is correspondingly longer. To solve this problem, based on the classic tracking by detection mode, an im-proved YOLO model is proposed: add a ReID feature recognition module to the YOLO network, so that YOLO can out-put target feature information while outputting target location information, so as to improve the tracking speed of the algorithm. Aiming at the situation that vehicles cover each other, a non maximum suppression algorithm based on dynamic IOU threshold is proposed to improve the tracking accuracy of the algorithm. Finally, the information output by YOLO is matched with data to realize multi-target tracking. The effectiveness of the improved model is veri-fied on the UA-DETRAC data set. The experimental results show that compared with the classic YOLO+DeepSORT tracking model, the average reasoning time in vehicle intensive scenarios is reduced by 17% when the YOLOv5 network is improved and applied to the target tracking algorithm; Adding dynamic IOU threshold non maximum suppression to the improved network, the tracking accuracy is improved by 3.9 percentage points. The improved model has better real-time performance and tracking accuracy.