Real-time upper-body detection and orientation estimation via depth cues for assistive technology

Automatic and efficient human pose estimation has great practical value in video surveillance. In this paper, we explore how a consumer depth sensor can assist with upper-body detection and pose estimation more precisely in the field of assistive technology for people with disabilities, and a novel real-time upper-body pose (orientation) estimation method is presented. At first, the Haar cascade based upper-body detection is conducted, and the depth information in a fixed subregion is extracted as the input feature vector. Then, support vector machine (SVM) and naive Bayes classifier are compared for estimating the upper-body orientation. Further, in order to acquire the continuous estimation data during a long time for behavioral analysis, we also adopt the support vector regression (SVR) to train a regression model. The experimental results show the effectiveness of the proposed method.