Spatiotemporal attention fusion network for multiple object tracking of unmanned aerial vehicle

The task of multiple object tracking from the perspective of Unmanned Aerial Vehicle (UAV) is becoming increasingly important and has a wide range of applications. However, conventional multiple object trackers do not fully exploit temporal and spatial information, facing challenges such as target blurring and variable trajectories due to the high-speed motion of UAV. In this paper, we propose STAF(Spatiotemporal Attention Fusion Network), which is based on spatiotemporal multi-head attention and fully integrates information from video sequence frames, enhancing the detection capability of targets. To better handle the camera shake, we develop an appearance feature update algorithm based confidence. The proposed method has demonstrated improvements on the VisDrone2019 dataset.