Real-Time Human Action Recognition by using R(2+1)D Convolutional Neural Network

In Computer Vision applications, automatically recognizing human actions from videos is a big challenge. Numerous industries, including intelligent video surveillance, object detection, robot-human interaction, etc., depend significantly on this technology. In Deep learning (DL) technology, Convolutional Neural Networks (CNN) act as the basic building block of DL, it can handle complex image-oriented tasks and also enhance the performance efficiency of the model due to its simplicity. The main application of 3D convolution for video analysis is the extraction of spatial and temporal data. However, it needs more parameter layers to train the network model, which leads to creating a vanishing gradient problem, more power consumption, and computational complexity. This paper proposed an R(2+1)D CNN architecture, which uses factorized CNN to reduce the number of parameter layers by using an optimized residual model. An action dataset from UCF-101 was used in this design to confirm the effectiveness of the suggested model. Compared with the most advanced models, our model achieved an accuracy of up to 82%