An Occlusion-Aware Tracker With Local-Global Features Modeling in UAV Videos
Qiuyu Jin,Yuqi Han,Wenzheng Wang,Linbo Tang,Jianan Li,Chenwei Deng
DOI: https://doi.org/10.1109/jstars.2024.3368035
IF: 4.715
2024-03-09
IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Abstract:Recently, tracking with unmanned aerial vehicle (UAVs) platforms has played significant roles in Earth observation tasks. However, target occlusion remains a challenging factor during the continuous tracking procedure. In particular, incomplete local appearance features can mislead the tracking network to produce inaccurate size and position estimations when the target is occluded. Furthermore, the tracking network lacks sufficient occlusion supervision information, which may lead to template degradation during template updating. To address these challenges, in this article, we design an occlusion-aware tracker with local-global features modeling, which contains two key components, namely the feature intrinsic association module (FIAM) and the feature verification module (FVM). Specifically, the FIAM divides the local features into blocks and utilizes the transformer network to explore the relative relationships among each subblock, which supplements the damaged local target features and assists the modeling for global target features. In addition, the FVM establishes a correlation measurement network between the target and the template. To precisely evaluate the occlusion status, masked samples with occlusion exceeding 50% are selected as negative samples for independent training, which ensures the purity of the target template. Qualitative and quantitative experiments are conducted on publicly available datasets, including UAV20 L, UAV123, and LaSOT. Qualitative and quantitative experiments have demonstrated the effectiveness of the proposed tracking algorithm over the other state-of-the-art trackers in occlusion scenarios.
imaging science & photographic technology,remote sensing,engineering, electrical & electronic,geography, physical