Scale-Adaptive Aerial Object Tracking Network via Location Estimation

The objective in aerial object tracking with the goal of accurately capturing and tracking the dynamic and variable characteristics of objects in various complex environments. However, existing tracking methods often encounter performance bottlenecks when dealing with rapid object movement, occlusions, and changes in appearance. Summarizing the aforementioned issues, we introduce a scale adaptive aerial object tracking network (SATNet), which can not only effectively handles scale changes and occlusions of tracking objects from an aerial perspective, but also accurately estimates the positions of tracking objects against complex backgrounds, ensuring the stability and precision for the tracking model. Specifically, SATNet incorporates a scale-adaptive feature fusion enhancement module (SFFEM) to integrate multiscale detailed and semantic features, strengthening object feature representation while mitigating interference from similar objects, thereby improving robustness to occlusions and enabling accurate detection of small or distant objects. In addition, an object search strategy based on location estimation module (LEM) is designed to analyze classification and regression information, achieving precise object localization in complex environments and significantly enhancing the performance of the SATNet tracking model in processing video sequences for reliable and effective object tracking. Widespread experiments proving the efficacy and supremacy of the proposed SATNet against numerous cutting-edge competitors across three public datasets.