YINGJIANG Li et al: HUMAN ACTION RECOGNITION BASED ON DYNAMIC TIME WARPING

Computers can recognize human gestures by man-machine interactive systems such as the Xbox Kinect depth camera. Although they can identify easy body movements these systems lack flexible and robust mechanism to perform high-level gesture recognition. This paper proposes a robust method to recognize body actions, it relies on Dynamic Time Warping (DTW), which is a technique used to recognize actions by finding an optimal alignment between two sequences. We propose an improved DTW algorithm with first and second derivatives to reduce singular mappings. We get the skeleton joints through Kinect camera, store the 3D coordinates of the joints into joint-model, then translate and normalize the joint-models. We select the hand and foot joints as the key joints, and exercise the coordinate time sequence by the improved DTW. During the process of exercises, we get a threshold for each action recognition class. Finally, we demonstrate the performance of our method and compare the result with other two methods by experiments.