Pose specification based online person identification

Abstract

Generally, identification methods use high quality frames that have obvious features like whole face of human being. In human identification case, multiple recognition areas have been proved to be a significant improvement over traditional face recognition methods. The main challenge of human recognition are that in some poses, the identification leads to a result of low accuracy as there are no obvious features like a whole face. In order to solve that problem, we apply the networks to detect the additional information to process the images that are hard to be used for identification. In continuous online conditions, there may still be some frames that can not be detected with those efforts. Our method uses a weight system to record changes in posture. Then the sequence of frames that belong to the same person can be grouped and the undetected frames can be identified by the detected frames. Experimental results show that our model achieves higher recognition accuracy than the existing methods in online case.