A two-stage human body detector on depth data

Detection of human body is one of the essential stages for human robot interaction. As the depth data is available with low cost recently, design of depth data based detector is of more significance. Previous works regarded the depth data as pure 3D point cloud or 2D depth images for this task. This paper proposes a two-stage detector, which makes full use of the depth data. The detector segments the scene and selects the potential human body point cluster using the statistics and classification on the 3D information. For the second stage on 2D re-projected depth image, a Haar-based AdaBoost classifier is employed for further performance promotion. The experiment showed that the proposed two-stage detector is more effective than the two stages used in separated way as well as a popular RGB image based detector Histogram of gradient (HOG).