Coordinate Attention-Driven Robust Multi-Object Tracking in Autonomous Vehicles: A Hybrid Framework of CA-YOLOv5 and ResNet-DeepSORT

With the increasing complexity of traffic conditions, the computational burden of multi-object tracking algorithms has grown, making it difficult to meet the requirements for tracking accuracy and real-time performance. In this paper, we proposed a road vehicle multi-object tracking method by improving and optimizing the YOLOv5 detection algorithm and the DeepSORT tracking algorithm. A Channel Attention(CA) mechanism is introduced into the existing YOLOv5 algorithm to construct the fusion algorithm CA-YOLOv5, and the feature extraction network structure of YOLOv5 is reconstructed by adding a prediction layer to improve the accuracy of vehicle detection. The ReID (Re-identification) network in DeepSORT algorithm is adopted as ResNet neural network to construct the fusion algorithm ResNet-DeepSORT. And it combined with data and feature enhancement, as well as high accuracy detection results of road vehicles. Thus, it improves the tracking accuracy and reduces the number of ID jumps to realize the multi-target tracking of road vehicles. Experimental results show that the proposed method increases mAP by 1.7%, MOTA by 0.9%, MOTP by 12.6%, and decreases ID switch by 27.7%, Average FPS by 1.69, meeting the tracking requirements for vehicles in practical autonomous driving scenarios, compared with that of the original tracking algorithm.