RECOGNITION OF HUMAN BEHAVIOUR UTILISING MULTISCALE CONVOLUTIONAL NEURAL NETWORKS

Abstract

In order to recognise human conduct, the most difficult thing is to construct a network that can extract and classify features based on their spatial and temporal relationships. To enhance the existing channel attention mechanism, which only takes into account the global average data from each channel and disregards its local spatial information, we suggest using the space-time (ST) interaction matrix operation module in conjunction with the depth separable convolution module. These modules are accompanied by studies on human behaviour recognition. A multi-scale CNN method for human behaviour recognition is suggested, taking advantage of CNN's superior performance in video and image processing. Low rank learning takes the behaviour video segments and uses them to derive knowledge about low rank behaviour. Without making any assumptions or enduring any tedious extraction techniques, the complete video's low-rank behaviour data can be obtained by linking this data along the time axis. Human behaviour models trained on neural networks can be reused across many network topologies. In order to reduce the disparity between features derived from different network topologies, two efficient approaches for measuring feature difference at various network levels are presented. The suggested method is effective, according to classification tests conducted on a wide variety of publicly available datasets. Experiments show that the method accurately identifies human conduct. According to our findings, the proposed model improves recognition accuracy, streamlines model structure, and makes computing output weights easier.