Multi-Object Tracking for Unmanned Aerial Vehicles Based on Multi-Frame Feature Fusion

To address the issues of tracking trajectory loss caused by small object size, frequent view angle changes and object occlusion in the multi-object tracking task of Unmanned Aerial Vehicle (UAV), in this paper, we propose a multi-object tracker for UAV based on multi-frame feature fusion. First, in order to more fully extract and utilize the interframe information, we design an attention-based adaptive multi-frame fusion module, which introduces Efficient Channel Attention (ECA) to trade-off the importance of the information in the history frames and the current frame. Second, we use a high-resolution feature extraction network as backbone network to extract features. The proposed method is evaluated on the UAV multi-object tracking datasets of Visdrone2019 and UAVDT. Compared with other mainstream multi-object tracking algorithms, our method achieves higher accuracy and fewer identity switches, which effectively improves multi-object tracking performance.