Video Stabilization Based on Key-Structure Semantic Constraints and Grid Motion Fusion

To address video jitter and key-structure positional drift in industrial surveillance scenarios, this paper proposes a video stabilization method based on key-structure semantic constraints and grid motion fusion. First, Segment Anything Model 3 (SAM3) is employed to extract key structures and estimate semantic motion from their positional variations. The resulting semantic motion is then fused with deep unsupervised trajectory-based (DUT) grid motion to enhance key-structure positional stability while maintaining overall video stability and preserving local geometric structures. Subsequently, a geometric valid support domain is constructed using the fused motion to identify invalid boundary regions introduced by frame reprojection. The Flow-Guided Transformer (FGT) model is further employed to complete these regions and improve the visual completeness of the stabilized videos. Experiments conducted on general video sequences and industrial surveillance scenes demonstrate that the proposed method achieves favorable performance in geometric structure preservation and overall stability while reducing key-structure positional drift. Qualitative results further illustrate that the geometric support masks identify regions requiring completion and guide FGT in generating visually plausible boundary content in the presented examples. Ablation studies and fusion-weight sensitivity analysis further demonstrate the complementary roles of semantic motion and grid motion in achieving balanced stabilization performance.