Robust underwater fishing net video stabilization based on the special Euclidean group

In order to solve the video shaky problem of underwater fishing net video suffering from shooting equipment and water flow disturbance, a multi-stage video stabilization algorithm on <i>S</i><i>E</i>(2) is proposed. We integrate low-rank sparse decomposition, deep feature detection, and log-space trajectory optimization. First, Iterated Robust CUR is used to decompose the video sequence into foreground and background components, extracting the dynamic fishing net region as a mask. Second, the SuperPoint feature detector is adopted to find high-confidence keypoints. The keypoints are then combined with mask and Lucas-Kanade optical flow tracking to construct motion trajectory. Finally, global motion parameters are optimized by smoothing the L1 norm trajectory of <i>S</i><i>E</i>(2) in logarithmic space. The experimental results show that our method performs better in both quantitative and qualitative analyses compared to existing methods.