Towards Discriminative Feature Learning for Multi-object Tracking in UAV Captured Videos

Recently, multi-object tracking (MOT) based on unmanned aerial vehicle (UAV) platform has become an important topic. However, in aerial photography scenes, the lack of object's appearance texture remains a challenge, as trackers are prone to confuse objects with similar appearances and lead to ID switches. Nonetheless, most of the existing methods mainly model appearance features using short-time clues, and such limited information makes it difficult to distinguish similar objects. To address this issue, we propose a novel Discriminative Multi-object Tracker (DistMOT), aiming to utilize high-quality long-term templates to mine distinctive object appearance, and further leverage the richer information of historical templates to distinguish similar objects. To this end, a Selective Memory Bank (SMB) is introduced to store multi-view historical templates; meanwhile, the Uncertainty-augmented Contrastive Learning (UACL) strategy is proposed to focus more attention on hard samples in the SMB, thereby forcing the model to highlight inter-object differential features and intra-object invariant features. Finally, the historical template differences of similar objects are considered for more accurate discrimination. Extensive experiments on the VisDrone-MOT and UAVDT datasets demonstrate the superiority of our method. Code is available at https://github.com/JackWoo0831/DistMOT