Human Posture Recognition Using Skeleton and Depth Information

Abstract

We present an approach to efficiently recognize human posture with a multi-classified support vector machine (SVM). In order to get features that input to the SVM, the approach use skeleton information obtained from a two-dimensional (2D) image and then map it into three-dimensional (3D) space using depth information. A body coordinate system is established to ensure the same postures have similar features. To deal with the problem of occlusion, we generate interpolating points using interpolation algorithm. Features contain both 3D information of the interpolating points and angles information related to joints and interpolating points. A dataset of five postures is built to verify the effectiveness of the approach. The results of experiments show that the recognition accuracy reaches 97.9% by the approach. Furthermore, the average time cost by extracting features and posture recognition with SVM after obtaining skeleton information is only 0.483 ms which meets the real-time application requirements.