A Novel Human Activity Recognition Method Using Joint Points Information

In order to increase the activeness and intelligence of robot services, Kinect is used to collect the joint point information for the recognition of human activities. Firstly, coordinates of human joint points are acquired by Kinect, and the three-dimensional structure vector is constructed to represent the human body structure. Then, the angles and the ratios of structure vector lengths between the space vectors are computed to represent the human postures. Meanwhile, the behavior is represented by continuous posture sequences within a limited timespan. At last, dynamic time warping (DTW) algorithm is applied to implementing the activity recognition by calculating the similarity between test behavior sequence and reference activity template. The experimental results show that the behavior representation scheme is invariant to rotation and translation. Moreover, experiments are conducted to recognize 6 kinds of human activities in daily living, and the results show that better recognition results can be acquired by the proposed algorithm.