Research on target detection and tracking algorithm for campus driverless vehicle based on improved DeepSORT

Abstract

Compared with other structured road scenes such as highways, the campus scene has the characteristics of multi-target, multi-cluster and multi-occlusion, the detection accuracy of traditional detection algorithms is low in this scene. In order to solve this problem, this paper proposes an improved DeepSORT target detection and tracking algorithm. The improved algorithm uses the YOLOv4-tiny algorithm for target detection and obtains the detection frame, then take the detection result as input, using the Kalman filter to predict the target trajectory, the detection and the prediction frames of the target are matched by the Hungarian algorithm. It uses DIOU (Distance Intersection Over Union) instead of IOU (Intersection Over Union) to perform correlation matching between unsuccessfully matched trajectories and detection results, followed by the Kalman filter update. Finally, the experimental of the campus driverless vehicle shows that the improved algorithm MOTA (Multiple Object Tracking Accuracy) increased by 8.6% and MOTP (Multiple Object Tracking Precision) increased by 1.5%.