Abstract: Due to the algorithm for motion object segmentation had poor adaptation of complicated scene, and the time complexity was too high, we proposed a new motion object segmentation algorithm, which used adaptive manifold denoising to achieve motion segmentation between rigid and nonrigid objects. We first introduced an adaptive kernel space in which two feature trajectories were mapped to the same point if they belonged to the same rigid object. Then, we adopted an embedded manifold denoising algorithm based on the adaptive kernel to segment the motions of rigid and nonrigid objects. Finally, we did contrast experiments with several traditional algorithms on several datasets. Experimental results show that the algorithm can achieve better segmentation and tracking effects in different scenes.

Key words: kernel space, adaptive manifold denoising, video motion segmentation, computer vision