Multiple object tracking method based on multi-task joint learning

In order to improve the efficiency of the multi-object tracking method, a joint detection-apparence network based on anchor aligned convolutional feature, called AAC-JDAN, was proposed. On the basis of the object detection network YOLOv3, an anchor transformation network and an anchor aligned convolutional operation was introduced, so that the network can detect the rotated objects, while alleviating the problem of the weak correlation between the apparence feature extracted by the exsiting joint network and the rotated objects;by adding an apparence feature extraction branch in the detection network, two subtasks of object detection and object apparence feature extraction were combined in a multi-task joint learning manner to realize the sharing of the low-level feature, and the apparence feature vectors can be extracted along with the corresponding detected objects, which improves the overall efficiency of the tracking algorithm. A fast online data association method was proposed to realize the efficient tracking of multiple rotated objects in the video. The similarity matrix between the incoming detections and the trajectories was calculated with the object apparence feature extracted by AAC-JDAN and the motion prediction result given by the Kalman filter, and the matching was done by the KM algorithm. When tested on two public datasets and a custom dataset, the TPR, MOTA, and IDF-1 reached 80.4%, 71.3%, and 69.5%, respectively, and the framerate reached 20 frames per second, this showed that the proposed method achieves a better balance in the speed and accuracy of tracking. © 2023 Editorial Board of Jilin University. All rights reserved.