A reliable UAV tracking system with online re-detection network.

Failures in long-term tracking have been frequently reported, posing significant challenges for the practical implementation of UAV tracking systems. Previous research has often employed a metric based on the current tracking state to assess reliability, coupled with a time-consuming re-detection network designed to recover the lost target. However, this approach lacks sufficient robustness and flexibility when dealing with unknown factors present in complex tracking scenarios. To address this issue, we propose a reliable UAV tracking system that incorporates a temporal consistency deviation index and an online re-detection network. The former takes into account the temporal consistency of consecutive frames and estimates tracking uncertainty using the confidence deviation caused by interference factors. The latter applies a series of linear transformations, inspired by Ghost operations, to reduce computational load and expedite inference. Additionally, a channel-spatial attention module is integrated into the re-detection component to enhance the extraction of valuable feature information. Results from the long-term dataset UAV20L demonstrate that the proposed algorithm outperforms the baseline trackers, particularly in scenarios involving full occlusion and viewpoint change situations. Furthermore, a physically constructed UAV tracking system is utilized to validate the effectiveness and real-time performance of the algorithm in handling occlusion events. Our code is released at https://anonymous.4open.science/r/Tracking-Redetection-F06F.