Answers are grounded in this paper and cite their sources.
Remote sensing satellite video object tracking plays a critical role in dynamic Earth observation, yet faces unique challenges, including small targets, low resolution, platform motion interference, and sparse scene information. In the real world, how to keep the stability of the tracker in various scenes is one of the key challenges of visual tracking. Traditional methods often lack adaptability when dealing with complex backgrounds. Therefore, this article proposes an adaptive tracking framework enhanced by positive incentive noise. The framework is based on motion state transitions to dynamically respond to complex backgrounds. It fully utilizes the hierarchical features from a Siamese network. During tracking, by utilizing both semantic and texture information, the tracker selects the most appropriate features based on the object's current motion state. To further enhance robustness, we have introduced two key innovations. First, introduce structured noise during the feature extraction phase to enable the model to learn robust feature representations that are resilient to appearance variations. Second, motion bias noise is introduced into the decision-making module based on reinforcement learning. This noise integrates physical priors to guide the generation of more stable and rational tracking actions. By the integration of multilevel features and the proposed noise strategies, our method demonstrates exceptional adaptability in complex scenarios. Extensive experiments conducted on public datasets have substantiated the effectiveness and robustness of the proposed framework.
No takes yet. Share an insight, caveat, or question.
Lu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: