Quality-Aware Motion Compensation and Adaptive Filtering for UAV Multi-Object Tracking

Abstract

Uncrewed aerial vehicle videos provide a flexible platform for observing dynamic scenes from aerial viewpoints, making multi-object tracking increasingly important in various applications. However, platform ego-motion and rapid viewpoint changes can destabilize inter-frame global motion estimation and increase geometric alignment errors, thereby complicating cross-frame matching and causing identity switches, biased state estimation, and fragmented tracks. To address these issues, we propose a quality-aware motion compensation and adaptive filtering framework for multi-object tracking in aerial videos. The framework employs an object-masked global motion compensation module to suppress foreground-induced interference, stabilize inter-frame alignment, and estimate motion-compensation quality. It further introduces a dual-branch Kalman filter measurement update strategy guided by detection confidence and the normalized innovation squared statistic, enabling efficient updates for reliable measurements and conservative updates for degraded ones. In addition, a quality-to-noise normalization strategy adaptively adjusts observation noise according to motion-compensation quality. Experiments on two widely used aerial video tracking benchmarks demonstrate consistent improvements over competitive methods.