Abstract
A method for generic object recognition to categorize such as desk, chair, etc., has been already proposed by authors, in which point cloud data captured by a range image sensor (Kinect, Microsoft) is used. The point data is segmented to several regions based on depth data. As for each segmented region, the point distribution data and the region size data of width, depth, and height are considered. The data is input to AdaBoost classifier. The classifier judges whether the region is object or not. However, two (or more) objects close to each other are likely segmented to one overlapping region. To address this problem, a new segmentation/recognition method is proposed in this article as follows: i) an overlapping region is divided to several small regions based on RGB data using k-means method, ii) all combinations of small regions are considered. Depth data of all combined region is input to AdaBoost classifier, and iii) combination of which AdaBoost output is the highest is adopted as the segmentation/recognition result. This method imitates a human who seems to repeat segmentation and recognition complementally. As for overlapping chair and desk, segmentation and recognition were successfully conducted by proposed method.
