Multiscale Deep Learning Framework for Human Behaviour Classification and Analysis

Abstract

Human behaviour identification is a significant research domain in computer vision and deep learning, as it has extensive use in areas such as observation, healthcare monitoring, and intelligent systems. Nevertheless, precisely discerning human actions from visual data remains difficult due to fluctuations in mobility, diverse camera angles, and evolving settings. Numerous current systems employ conventional machine learning techniques or rudimentary convolutional neural networks, that frequently fail to adequately capture both both temporal and spatial information. This research presents a multi-scale deep learning techniques framework for the classification and analysis of human behaviour. The system initiates with dataset upload, preprocessing, and partitioning to ready the data for testing and training purposes. The basic CNN2D model is employed to extract spatial data from images, whilst a CNN3D model is implemented to collect spatial as well as temporal characteristics from video sequences. A hybrid model integrating CNN, GRU, and bidirectional layers is generated to enhance the comprehension of sequential patterns and elevate overall predictive performance. The system comprises modules for showing accuracy and loss graphs, in addition to facilitating the comparison of various models' performance. Experimental findings indicate that the suggested framework attains superior accuracy and enhanced performance relative to conventional methods. This enhances the system's reliability for real-time human behaviour identification tasks.