2A2-J06 Object extraction and recognition in a room using both of depth and color images obtained by Kinect(Robot Vision (2))

Abstract

A method for generic object recognition for categories, such as desk, chair, human beings, etc., is proposed, in which point cloud data captured by a range image sensor (Kinect, Microsoft) is used. Regions in which objects exist are extracted by clustering the points in the neighborhood. One region is divided to 5×5×5=125 sub-regions. The three dimensional (3D) distribution of point cloud in the region is considered by counting the point number in each sub-region. The size data of width, depth, and height is considered, making totally 125+3=128 data. The data is dealt with in the form of histogram, which is input to AdaBoost classifier. The classifier judges whether the region is object or not. As examples, "chair", "desk", and "human" are considered. Simulation and experiment were carried out and the recognition rate around 90% was successfully achieved, provided that the object regions be correctly extracted.