基于改进YOLOv5的道路车辆跟踪算法

Road Vehicle Tracking Algorithm Based on Improved YOLOv5

针对现有多目标跟踪算法参数量和计算量大,难以满足移动设备实时性要求的问题,本文通过改进JDE跟踪算法,提出了一种道路车辆多目标跟踪算法。首先,设计关联融合网络来解决JDE算法中多任务学习存在的竞争问题,提高算法的跟踪精度,减少身份切换次数;其次,使用改进的EfficientNetv2重新构建YOLOv5的特征提取网络,降低模型复杂度,提高模型实时检测速度;最后,使用改进的YOLOv5检测算法与JDE跟踪算法结合,实现道路车辆多目标跟踪。实验结果表明,提出的方法相比原JDE跟踪算法,MOTA提高0.3个百分点、跟踪速度提高约43.2%,可以满足实际自动驾驶场景中对车辆跟踪的速度要求。

To solve the problem that it’s difficult for the large amount of network parameters and calculations for existing multi-object tracking algorithm to meet the real-time requirements of mobile devices,a road vehicle multi-object tracking algorithm is proposed by improving the JDE tracking algorithm.Firstly,in order to improve the tracking accuracy of the algorithm and reduce the number of ID switching,the association fusion network is used to solve the competition problem of multi-task learning in the JDE algorithm.Secondly,in order to reduce the complexity of the model and improve the real-time detection speed of the model,the improved EfficientNetV2 is used to rebuild the feature extraction network in YOLOv5.Finally,the improved YOLOv5 detection algorithm is combined with the JDE tracking algorithm to achieve multi-object tracking of road vehicles.The experimental results show that compared with the original JDE tracking algorithm,the proposed method improves MOTA by 0.3 percentage point and tracking speed by about 43.2%.It can meet the speed requirements for vehicle tracking in actual autonomous driving scenarios.