Two-dimensional skeleton action recognition method based on hierarchical attention mechanism

Abstract

The traditional RGB-based action recognition method is vulnerable to background factors such as illumination and occlusion, which makes the recognition accuracy and real-time performance low. In order to solve the above problems, a 2D skeleton action recognition method incorporating a hierarchical attention mechanism was proposed. Firstly, the human skeleton joints in the video were extracted using the human pose estimation algorithm OpenPose. Secondly, the missing and abnormal points in the extracted skeleton data were pre-processed by the mean-completion method and exponential smoothing method. Finally, a network model based on CNN-LSTM fused with hierarchical attention mechanism, CNN-HALSTM, was constructed to realize action classification. The results show that the recognition accuracy of this method in interactive motion and KTH data sets is 96.73% and 98.35%, and the model parameters are significantly reduced, which has the advantage of real-time performance and is better than other methods of the same type.

Keywords

• human action recognition
• pose estimation
• skeleton joint
• average filling method
• exponential smoothing method
• hierarchical attention