Multi-object tracking in UAVs with feature fusion distribution and occlusion awareness

Multi-object tracking (MOT) in unmanned aerial vehicles (UAVs) is a crucial computer vision task with diverse applications in both military and civilian domains. However, the unique characteristics of UAVs, such as motion uncertainty and sudden changes in viewpoints, lead to objects with scale variance, occlusion, dense distribution, and frequent appearance and disappearance in the image, posing significant challenges in MOT in UAVs. In this paper, we address these issues by proposing two novel techniques: Feature Fusion Distribution Network (FFDN) and Occlusion-Aware Prediction and Association (OAPA), which are integrated into a new MOT algorithm named OATrack. The FFDN aims to improve object detection by optimizing the fusion of multi-scale features within the detection network, especially for densely distributed and different sized objects. The OAPA aims to enhance the accuracy and robustness of prediction and association for objects lost due to occlusion, thus addressing the issue of occlusion in UAV scenes. Experiments are conducted on Visdrone2019 and UAVDT, and the results clearly demonstrate the effectiveness and superiority of the proposed method.