Human body orientation recognition with arbitrary postures

Human body orientation recognition is a fundamental preprocessing task in registering two human body surfaces, analyzing motions of human bodies, and generating data-driven animations. The existing methods for human body orientation recognition are usually pose dependent and not very robust. To solve this problem, we propose an automatic pose-independent method for human body orientation recognition, which is based on SVM and geometric features. Specifically, the geometric features are constructed by using the relations between the joints of the lower limbs and body orientation. The SVM classifier is then trained to determine the body orientations. The experimental results show the effectiveness of the features (with good linear separability) and the SVM classifier can achieve very good performance in a pose-independent manner. Moreover, it is convenient to integrate this method into the existing approaches of extracting the skeleton from human body shapes.