Discriminative Pose Analysis for Human Action Recognition

Xiaofeng Zhao, Yao Huang, Jianyu Yang, Chunping Liu

Abstract: Analysis of pose-level information is critical for human action recognition. For fast visual perception, we propose to encode human pose with discriminative information in the pose feature space, which needs low computational cost. The main idea is that the common poses that exist in different actions have low discrimination while the discriminative poses exist in few actions. Therefore, we employ the Gaussian Mixture Model to cluster the feature vectors of all poses in all classes of actions, and then quantify the discrimination of each pose as well as its contribution to each action class. Meanwhile, we calculate the correlation between the clusters and action classes as a rating matrix. The query action is inferred by simply calculating the Gaussion weights of poses and multiply the rating matrix. The framework runs without deep neural networks, which is low cost for application on low power devices. The proposed method is validated on three benchmark human action datasets, and the experimental results show good performances of the proposed method. The inferring time of each query is only 24ms on a 2.5 GHz CPU that indicats the efficiency of our method.