Motion detection based on RGB-D data and scene flow clustering

A 3D motion detection method based on RGB-D data and scene flow clustering is proposed in this paper. The 3D extension of the optical flow called scene flow which describes the 3D motion of every point in a scene between two consecutive frames. Firstly, we estimate the scene flow using color maps and the aligned depth maps in two consecutive frames, then extract the feature vector of each point from the estimated scene flow, including the direction information and the amplitude of the scene flow. Finally, an unsupervised classification algorithm: Iterative Self Organizing Data Analysis Techniques Algorithm(ISODATA) is applied to cluster the feature vectors and acquire the motion region. We also get the 3D information of the motion which consists of the location and distance relative to the camera. Experiment shows satisfied results of the method.