Neighborhood-Constrained Video Stabilization and Frame-Centering Compensation for UAV Target Tracking

Camera jitter and target drift degrade visual tracking by unmanned aerial vehicles (UAVs), especially near frame boundaries. This study proposes a modular post-processing method that combines neighborhood-constrained video stabilization with frame-centering compensation. Adjacent-frame phase correlation estimates global translation. A local median and Gaussian temporal weights, combined with residual down-weighting, construct a robust camera path that suppresses isolated motion-estimation errors. A bounded centering term then moves the target toward a safety region while preserving frame content. The method requires no inertial sensor and can operate after correlation-filter, Siamese, or Transformer trackers. A pilot was conducted on an 898-frame UAV123 subset. Target-center RMS error decreased from 164.58 to 82.58 pixels (49.82%), safety-region coverage increased from 65.26% to 100%, and frame retention remained 90.36%. Under injected noise levels of 2, 4, and 8 pixels, trajectory-recovery error decreased by 57.95%, 68.70%, and 74.26%, respectively. These results indicate geometric feasibility, while full UAV123, UAV20L, and DTB70 benchmarks remain necessary before general performance claims.