Research on Moving Object Segmentation Based on Optical Flow and FastSAM

FastSAM is a lightweight general-purpose segmentation model built upon YOLOv8-seg as its backbone network, which demonstrates several advantages including fast inference speed, low deployment cost, and high training efficiency. However, it suffers from notable limitations in that object segmentation relies on manually defined prompts, and the model lacks the capability of moving object discovery. To achieve moving object segmentation in a fixed field of view, vector operators are employed to jointly measure the optical flow perturbation features induced by object motion. Based on this, a threshold-adaptive optical flow perturbation region extraction algorithm is designed. The bounding box of the perturbation region is subsequently computed and fed into FastSAM as a box prompt, thereby realizing moving object segmentation in a fixed field of view. The proposed approach achieves satisfactory performance in both single-object and multi-object segmentation scenarios.