ABSTRACT In target tracking under background motion scenarios, background movement often significantly compromises detection and tracking accuracy. This paper proposes a visual motion compensation algorithm for unmanned aerial vehicles, integrating feature point detection, matching and robust geometric estimation methods. First, the SuperPoint network is employed for end‐to‐end point detection and description, ensuring robust feature extraction in complex environments. Subsequently, the SuperGlue graph neural network is introduced. It optimises feature matching relationships through self‐attention and cross‐attention mechanisms, whereas the optimal transport layer yields high‐confidence matching pairs. Building upon this, MAGSAC++ sampling is employed for outlier rejection and homography matrix estimation of matched points. This enables modelling and compensation for global background motion, effectively isolating the true motion trajectory of the foreground target. This approach balances local precision with global robustness, maintaining high matching accuracy and stability under background interference, noise and partial occlusion. It provides reliable background motion compensation support for infrared small target detection.
Wang et al. (Thu,) studied this question.