A Constraints-based Dynamic Time Warping Method for Gesture Recognition with Kinect

Abstract

Gesture recognition is still actively researched in non-contact human-computer interaction (HCI), where the Dynamic Time Warping (DTW) algorithm is commonly employed. However, the computation load of traditional DTW algorithms during the matching phase presents a challenge for gesture recognition, especially when the number of reference gestures in the template library increases. In order to solve this problem, a Constraints-based Dynamic Time Warping (CDTW) method is proposed in this paper, including Global-Path Constraint, First-Frame Constraint, and Feature-Vector Constraint. These three constraints are expected to limit the region of the warping path, exclude several reference gestures from the template library, and reduce computation load directly with reduced elements of the feature vector. To verify the proposed CDTW method, Microsoft Kinect-based comparative experiments with the traditional DTW method have been carried out. Experimental results show that our CDTW method boosts the efficiency of gesture recognition with a 17% decrease in recognition time and a 3% increase in average recognition accuracy, compared to the traditional DTW algorithm.