Dance action recognition and pose estimation based on deep convolutional neural network

Abstract

Sports action recognition helps athletes correct their action range and standardize their poses. But it is not an easy task to recognize sports actions, due to the individual difference in action execution. Besides, the difficulty of action recognition increases with the diversity of actions and the complexity of background. The previous studies have not fully considered temporal changes, and failed to determine the exact staring point of actions. To solve the problem, this paper proposes a new method to recognize dance actions and estimate poses based on deep convolutional neural network (DCNN). Firstly, the authors presented full-effect expression of global and local features of dance actions, and derived an optimal model based on DeepPose. Next, a dance pose evaluation model was established based on time sequence segmentation network, and the sparse time sampling strategy was introduced to realize efficient and effective learning of the frame sequence of the whole video. Experimental results confirm the superiority of the full-effect expression of global and local features, and the effectiveness of the proposed model. The research results provide a reference for the application of deep learning (DL) in other scenarios of action recognition and pose estimation.