On trajectory segmentation and description for motion recognition

With the rapid development of robotic technologies and their extensive applications in human life, human robot interaction via human motion is a fundamental topic in robotics. The critical problem in this topic is recognizing human motion in real time. In this paper, we propose a new method to represent 3D motion trajectory for fast recognizing complex motion trajectory. 3D trajectory is segmented into sequential primitives and each primitive is assigned with a label. After that, the primitives are represented by the integral invariant descriptor. The motion trajectory representation is invariant to complex geometrical transformation, including: inconsistent scaling and 3D rotation. The proposed method is also robust to partial occlusion and local noise. Motion recognition based on the proposed parsing method is explored with a new proposed hierarchical dynamic time warping algorithm. The conductive experiments validate the accuracy and efficiency of our method, as well as the invariant properties. The comparative results verify the capability of the proposed method for fast motion recognition.