Using CNN, GRU, and B/idirectional Multiscale Convolutional Neural Networks for Human Behavior Recognition

The main challenge in recognizing human behavior is constructing a network for the extraction and categorization of spatiotemporal features. In order to address the issue that the current channel attention mechanism simply aggregates each channel's global average information while ignoring its specific spatial information, this work suggests two enhanced channel attention modules: the depth separable convolutions section and the time-space (ST) interaction section of matrices operation. These modules are also combined with research on the recognition of human behavior. Proposing a multiple habitats convolutional neural network technique for human behavior detection, it is combined with the excellent performance using convolutional neural network (CNN) for video and image processing. First, the behavior video is divided into segments. Next, low rank learning is applied to each segment to extract the associated low rank actions information. Finally, these minimal position behavior information are linked together in the time axis to get the low are behavior data for the entire video. This allows for the efficient extraction of behavior information from the video without the need for laborious extraction processes or assumptions. Neural networks can simulate human behavior in a variety of network topologies by transferring and reusing this capacity. To lessen the distinction between features derived from various network topologies, two efficient feature difference measurement methods are presented, taking into account the various properties of data features at various network levels. The suggested strategy has a decent categorization impact, according to experiments on a number of available datasets. The experimental findings demonstrate that the method's accuracy in identifying human behavior is excellent. It has been shown that the suggested model increases recognition accuracy while simultaneously enhancing the compactness for the model structure and successfully lowering the computational cost of the output weights.