Analyzing Role of Joint Subset Selection in Human Action Recognition

Joints of human pose play an important role in human action recognition (HAR) due to their invariant to subject appearance, low requirement on computation and storage, as well as containing informative data. Psychologists [1] showed that human actions (e.g., walking, running, dancing) could be distinguished by using only a subset of joints rather than using full-joint model. In this paper, we tackle role of joint selection in HAR performance by proposing a pre-determined configuration of joints combined with temporal derivatives of joint positions. To analyze role of joint selection, our proposed scheme is compared with full-joint as well as automatic joint selection schemes. Proposed method is evaluated on three public datasets: MSR-Action3D, UTKinect-Action and Florence3D-Action. Experiments show that our proposed method achieves very competitive results compared with state-of-the-art methods. For MSR-Action3D dataset, our proposed method achieves better result than existing methods with accuracy of up to 95.53%. For Florence3D-Action and UTKinect-Action datasets, accuracy of our proposed method ranks in the second position while computation is 10-20 times faster than top-ranking method.