Efficient Recognition of Human Actions by Limiting the Search Space in Deep Learning Methods

The efficiency of human action recognition systems depends on extracting appropriate representations from the video data. In recent years, deep learning methods have been proposed to extract efficient spatial-temporal representations. Deep learning methods, on the other hand, have a high computational complexity for development over temporal domain. Challenges such as the sparsity and limitation of discriminative data, and highly noise factors increase the computational complexity of representing human actions. Therefore, creating a high accurate representation requires a very high computational cost. In this paper, spatial and temporal deep learning networks have been enhanced by adding appropriate feature selection mechanisms to reduce the search space. In this regard, non-online and online feature selection mechanisms have been studied to identify human actions with less computational complexity and higher accuracy. The results showed that the non-linear feature selection mechanism leads to a significant reduction in computational complexity and the online feature selection mechanism increases the accuracy while controlling the computational complexity.