SuperMOT: Decoupling Motion and Fusing Temporal Pyramid Features for UAV Multiobject Tracking

Multi-object tracking in unmanned aerial vehicle (UAV) videos is a critical research topic with applications in various fields. However, traditional multi-object trackers demonstrate limited generalization in UAV videos due to challenges such as nonlinear object motion caused by UAV movement and the multi-scale appearance of small objects caused by the oblique imaging angles of UAVs. Therefore, this article proposes a novel method called SuperMOT, specifically designed for multi-object tracking in UAV videos. SuperMOT proposes a Pyramid Deformable Alignment (PDA) module, which aligns and fuses temporal semantic features across multiple levels to enhance object recognition. It can be observed that the input sizes commonly selected by mainstream UAV multi-object trackers to reduce computational load are suboptimal for capturing the semantic features of small objects. To address this limitation, a lightweight Video Super-Resolution (VSR) module is integrated into the training pipeline to incorporate high-resolution information. Furthermore, to address complex motions in UAV view, a Motion Decoupling (MD) module is developed to separately process object motion and platform motion. Experimental results on the VisDrone2019 and UAVDT datasets indicate the effectiveness and real-time capability of the SuperMOT, achieving state-of-the-art performance.