Abstract: The current methods of human action recognition by computer vision are mostly based on hand-craft features and usually prior knowledge-required. They inevitably depend on specific applications and neglect the inner structure of visional information. A novel method which integrated self-learned pose features and combined posture symbol rules was proposed to achieve the recognition of human action more efficiently. The structural features of posture silhouette were extracted and a codebook of primary posture was built through the establishment of a sparse auto-encoder network. Then, in the phase of recognition, the Hidden Markov Model was employed to train the models for different action categories. Besides, a key frame extraction algorithm was developed to reduce the redundancy of long code sequence before training HMMs. Simulation experiments manifest the effectiveness of the proposed method.

Key words: pose recognition, self-learned feature, SAE, HMM, posture codebook