Multi-Motion Segmentation: Combining Geometric Model-Fitting and Optical Flow for RGB Sensors

Abstract

This paper presents a novel multi-motion segmentation framework by combining the geometric model fitting and optical flow. More precisely, we use the spatial information extracted by the multi-model fitting and the motion information provided by the optical flow to improve the practicality of the motion segmentation algorithm. The real scenes contain challenges such as multiple motions, camera translation, large field of views, and occlusions. To improve the segmentation accuracy of the real scenes, we first propose a cluster selection method to generate several clusters that can cover all motion models in the data, then a cluster merging strategy is proposed to estimate the basic motion models and the number of motions accurately for the scene. Finally, motion assignment of the remaining points is carried out to complete the motion segmentation. Detailed experiments on two challenging real-world benchmark datasets (62-clip and KT3DMoSeg) show that the proposed framework is superior to the state-of-the-art multi-motion segmentation methods.