E2TRACK: Overcoming Occlusion in Aerial Tracking with Trajectory Estimation

This paper introduces E2Track, a novel pipeline designed to overcome challenges in aerial tracking, particularly focusing on target occlusion scenarios. While template matching-based aerial tracking methods have shown progress in handling appearance changes, they often struggle in scenes with significant occlusion. We identify a fundamental issue leading to the limitations of existing methods in occlusion scenarios: the lack of temporal trajectory modeling. To address this, we propose to leverage the track information as a key solution to the occlusion problem. E2Track comprises two essential modules: the trajectory prediction module and the trajectory decoding module. The trajectory prediction module utilizes the entire video’s temporal contextual features to predict target trajectories. On the other hand, the trajectory decoding module performs feature queries on the predicted trajectories to determine the target’s position in each frame. This dual-module architecture enhances the model’s ability to handle occlusion scenarios by incorporating temporal context and leveraging trajectory information. To validate the efficacy of our proposed approach, we conducted experiments comparing E2Track with state-of-the-art trackers on two publicly available datasets. The results demonstrate the superiority of E2Track over competing methods, emphasizing its effectiveness in addressing the challenges posed by occluded scenes.