Body gesture validation using multi-dimensional dynamic time warping on Kinect data

This paper presents a system capable of identifying and validating various human body gestures. Body data is acquired from a Kinect sensor and consist in a set of bones represented by their rotation in 3D space. The main goal is to correctly identify the user performed gesture and return feedback related to its performance accuracy. A database with several samples for each gesture is built and used as ground truth in the gesture validation process. A novel approach of dynamic time warping algorithm is proposed for synchronizing the performed gesture with the corresponding ground truth dataset. If the sequences are not synchronized, feedback is returned to user as a comparison between its performance and the closest sample in the database. Experimental results show high accuracy at about 90% success rate. The system performs better for gestures in which the users´ body is fully exposed to the sensor.