AMV-tracker: an Adaptive Multi-View Complementary Information Tracking Algorithm with a Dynamic Activation Mechanism

In response to the challenges of frequent viewpoint changes, appearance variations, and occlusions in UAV tracking scenarios, as well as the issue of excessive computational redundancy in simple scenes, we propose an Adaptive Multi-View Complementary Tracker (AMV-Tracker) with a dynamic activation mechanism. First, we introduce a dual-template tracking strategy, where a static template preserves the original target information while a dynamic template provides short-term variation information, thereby enhancing robustness against appearance changes and occlusions. Second, we design a dynamic activation module to reformulate the Transformer-based backbone network, enabling a feature extraction network that dynamically activates relevant modules according to different backgrounds. Third, we employ a multi-view image mutual information module to model the target, ensuring stable feature representation under viewpoint variations. Additionally, to mitigate localization errors caused by scale variations, we introduce a target boundary regression branch and propose a more precise head prediction network. The proposed method is validated on the LaSOT and UAV123 datasets, demonstrating its effectiveness and superiority. Furthermore, the algorithm is deployed on an embedded device, Jetson Orin NX, for performance evaluation, achieving a processing speed of 46 FPS. Compared to classical tracking algorithms, the proposed method exhibits higher accuracy across various challenging scenarios and effectively addresses issues related to viewpoint changes, rapid motion, and scale variations.