Automatic Motion Segmentation Using Random Walks

This paper presents a method two segment moving objects from video sequences by combining spatial and temporal information. The segmentation is performed in two steps : automatic seeds detection, and random walk segmentation. In the first step, seeds are detected by using optical flow in specific pixels so in every dxd grid there is one pixel to represent the grid, then initial seeds are chosen automatically by thresholding the distance of estimated motion vector of the sparse pixels. In the second step, we treat our motion segmentation as a graph based problem, then an energy function is defined to evaluate spatial and temporal smoothness, and we apply random walk algorithm to solve the energy minimization problem, the solution leads to affecting a label to every pixel in sequences video and get the final segmentation. The experimental results illustrate its promising performance and can be integrated in real-time application.