Imrovement of human key posture recognition

Abstract

Human pose recognition is an important problem in such fields like scene recognition, robotics, multimedia systems an etc. It plays an essential role in full body gesture recognition. The obtained poses can be then used to detect gestures. In gesture recognition it is essential to capture all poses. Because missing some key pose may significantly decrease the gesture recognition accuracy. Most pose recognition approaches use silhouettes extracted from regular cameras that may not bring a favorable pose accuracy or different kind of special costumes that uncomfortable to wear. In this paper we propose to classify poses based on their skeleton appearance. For this we use depth camera called Microsoft Kinect. The main difference of our method is that we add some additional objects that take place in the scene into skeleton provided by Kinect camera. In another words we use an advanced type of skeleton unlike camera can provide. The skeleton can be modified upon the specific pose recognition task by modifying it which in turn increase the pose recognition accuracy. In our work we adopt it for golf postures recognition problem by tracking golf club head. The background of an image is subtracted by using depth image histogram, then we calculate object coordinates using HSV(Hue Saturation Value) color information and image moment. Finally we perform classification using Support Vector Machines. The obtained results demonstrate the efficiency of the proposed method.