D2Track: Decoupled and Discriminative Representation Learning for Online UAV Multiple Object Tracking

Abstract

Multiple object tracking (MOT) from unmanned aerial vehicles (UAVs) presents significant challenges due to drastic viewpoint changes and ambiguous top-down appearances. The former often result in large, misleading displacements of objects in the image plane, while the latter makes it difficult to distinguish between objects with similar appearances. To address these, we propose D2Track, a novel graph neural network (GNN)-based tracker that focuses on learning Decoupled and Discriminative representations for robust association. Our main contribution consists of two specialized modules. First, we propose the adaptive feature rectification module (AFRM), which effectively decouples the camera’s ego-motion from the object’s true motion. By estimating the global projective transformation between frames, the AFRM generates a decoupled motion feature that provides a more accurate motion cue for the GNN. Second, to address the difficulty of distinguishing objects with similar appearances, we design a KFD feature using the kernel Fisher discriminant (KFD) algorithm. This method projects the initial reidentification (ReID) features into a more discriminative feature space, significantly improving the model’s ability to differentiate between such similar objects. Extensive experiments on the challenging VisDrone, UAVDT, and SportsMOT benchmark datasets demonstrate that our proposed tracker, D2Track, achieves excellent performance, proving the effectiveness of its decoupled motion and discriminative appearance strategies.