Dynamic blur association rule extraction method in real-time visual tracking

Abstract

A dynamic blur-aware association rule extraction framework is presented for real-time visual tracking in environments characterized by severe and varying motion blur, frequent occlusions, and rapid background changes. The approach innovatively models blur as a spatially and temporally variant process, embedding local blur statistics into compact feature descriptors that are central to the formation of robust association rules. This technique enables the tracking system to selectively suppress unreliable information caused by transient blur. This allows them to distinguish between target dynamics and visual artifacts. The blur-adaptive association module is used to optimize computational efficiency and spatiotemporal accuracy. The module mainly focuses on flow-guided marginal loss and sparse optical flow strategy. The architecture adopts a hybrid CPU/FPGA acceleration scheme to meet the demanding real-time requirements. In the experimental study of UAV urban navigation and robot operation platform, the framework shows high accuracy, strong anti-drift and stability in complex high-speed environment. Comparative analysis and component ablation experiments show that motion-guided optimization, adaptive association and blur modeling have a significant impact on system robustness. These results establish the framework as a significant step forward for reliable, real-time visual tracking in challenging, dynamic settings.