EdgeFlow: Combining Local Feature Enhancement and Sobel-Based Channel Attention for Optical Flow Estimation

High-precision optical flow estimation is critical for achieving precise self-localization and environmental mapping. Existing optical flow approaches, however, struggle with accuracy issues when dealing with regions of high gradient variation, such as edges. This paper introduces a new optical flow estimation framework named EdgeFlow, designed to enhance the detection accuracy of edge pixel motion and to coherently extract high-gradient features, all while keeping computational complexity low. By integrating multi-scale local detail feature extraction with a Sobel-Channel Attention mechanism, EdgeFlow seeks to improve performance in these challenging areas. The experimental results indicate that our model surpasses existing state-of-the-art methods on the KITTI optical flow benchmark, exhibiting particularly notable advancements in the movement estimation of edge areas. These findings affirm the method's efficacy and practical viability.