Real-time Human Action Recognition From Motion Capture Data

Recognition of human actions is one of the important tasks in various computer vision applications including video surveillance, human computer interaction etc. Traditionally RGB or depth cameras are utilized for this task. In this work we propose an approach that utilizes motion capture data for recognizing actions. Motion capture provides accurate motion information of joints of body in 3D space. The 3D skeleton joint co-ordinates of the user provided by motion capture system are used to analyze the dynamics of the action being performed. The temporal variance of each joint of the skeleton and its time weighted variance serve as the features for classification. The time weighted variance feature embeds temporal information in the feature and helps in discriminating confusing actions such as sit-down and stand-up. These features can be extracted rapidly and suitable for real-time recognition. We demonstrate the performance of the proposed approach using correlation based metric and support vector machines (SVM) on the Multimodal Human Action Detection dataset. The recognition accuracy of above 95% has been achieved.