Abstract:

In order to achieve skeleton identification function of body feeling controller, engineers must implement human body detection. We present a real-time detection method based on rgb-d data and the effect of its implementation on body detection. First, simplify the original 3D point cloud. Second, remove the ground plane. Third, for the rest of the point cloud data, the preliminary classification is performed to get human point cloud data cluster. Again after the first classification do the secondary fine classification, and then implement multiple human body detection. This method can detect multiple human body in static, and also can detect multiple human body in walking. Experimental results show that this method has good real-time performance,and CPU real-time processing speed can reach 25 frames per second. No matter to the static human body or the walking human body, the method is more accurate and rapid in detection, and the human body detection accuracy can reach 86%.

Key words: real-time people detection, K-nearest neighbors, simplification of point cloud, histogram of oriented depths(HOD)