Lightweight skeleton-based action recognition research

Abstract

Action recognition is a cutting-edge research direction in the field of computer vision, widely used in video surveillance, video retrieval, human-computer interaction, and other fields. However, existing action recognition algorithms improve prediction accuracy at the cost of having too many model parameters and high computational complexity, which limits their practical application. In order to address this issue and improve recognition accuracy with higher computational efficiency, we propose a dense residual 3D depth-separable convolutional network model based on channel attention mechanism. By extending the depth separable convolution to 3D, the model can maintain high accuracy while greatly reducing the number of model parameters and operational complexity. The dense residual structure is introduced to enhance the training ability of the network, improve feature extraction, and further improve network performance by learning the correlation between channels through the channel attention mechanism. Compared with other action recognition algorithms on two datasets, NTU RGB-D and Fall Dataset Lei2, the proposed action recognition algorithm has better robustness under lightweight condition, with 415.9 MFLOPs of computation, 205.4k of parameters, and 90.5% of CS, 95.5% of CV and 95.97% of accuracy, respectively. The proposed algorithm has a high recognition accuracy and greatly reduces the calculation cost, which can provide better performance for practical applications.