Human Activity Recognition from Video Sequences for Usual and Unusual Activities

Abstract

Human Activity Recognition (HAR) has become a pivotal research domain over the newlinepast decade due to its growing applications and the wide availability of data. It serves as newlinethe foundation for various computer vision applications, including home security, video newlinesurveillance, human-computer interaction, and elderly care. Advancements in sensing newlinetechnologies and the limitations of wearable sensors have driven Human Activity Recognition newline(HAR) systems to shift from traditional sensor- and ambient-based approaches to newlinevision-based solutions. These limitations include privacy concerns in surveillance applications, newlineuser discomfort, and non-compliance due to continuous wear, and limited battery newlinelife. This transition has motivated researchers to propose innovative solutions that address newlinethe challenges in vision-based HAR. newlineThe smart surveillance system that automatically detects human activities from video newlinesequences and classifies them as usual and unusual human activity is a necessity for the newlinecommercial and public sectors. Over decades, many efforts have been made by researchers newlineto accurately recognize human activity from video sequences. However, challenges like newlineview angle, camera motion, occlusion, diverse contexts, and model opacity continue to newlinemotivate researchers to further tackle this problem. newlineIn this thesis, research gaps in the HAR field and current challenges are identified. newlineThis process involves a comprehensive review of leading-edge techniques, followed by newlineanalysis and evaluation. Based on the literature review, the primarily used popular newlinemethods in HAR are implemented and tested for the targeted datasets, and enhancements newlineare introduced using optimization techniques. newlineSubsequently, the study progresses toward deep learning methodologies to capitalize newlineon their advantages and align the research with current advancements. Prominent deep newlinelearning methodologies, such as Transfer Learning (TL), Convolutional Long Short-Term newlineMemory (ConvLSTM), and 3D Convolutional Neural Networks (3DCNN), are employed newlinewith innovative arch