An Improved Multi-object Tracking Algorithm for Autonomous Driving Based on DeepSORT

Multi-object tracking (MOT) is one of the most important tasks of the vision perception module of autonomous driving, which provides crucial inputs for the downstream task modules such as motion prediction, planning and control. However, how to balance the efficiency and accuracy of MOT in autonomous driving scenario, has always been a great challenge. To address this issue, we try to improve the classical DeepSORT model which follows the Track-By-Detection paradigm from three perspectives of object detection, feature extraction and data association, so as to further improve its tracking accuracy while retaining its online, light-weight and high-efficiency advantages. First, a single-stage detection model Yolo-v5 is adopted to replace the two-stage model Faster R-CNN used in DeepSORT, which leads to significant improvement of detection efficiency and accuracy. Second, specified light-weight ShuffleNet-v2 in stead of time-consuming DenseNet is used to extract appearance feature from detected object bounding boxes to further enhance the efficiency of MOT. Finally, an efficient and accurate data association algorithm integrating cascade matching and IoU matching is proposed. Based on the cost matrix jointly constructed by appearance features, motion and shape, effective tracking of multiple objects is realized. A large number of experiments have been conducted to compare our improved MOT model proposed and the DeepSORT model in terms of many evaluation metrics, such as IDs, IDF1, MOTA, MOTP, etc. Moreover, the sensitivity of some hyper-parameters that have important influence on MOT efficiency is also analyzed. Experimental results show that the efficiency and accuracy of MOT can be effectively balanced by keeping the lightweight and efficient characteristics of the original DeepSORT model and further improving the performance of its three key components, i.e. object detection, feature extraction and data association.