Application of Deep Learning In Human Activity Recognition

Abstract

newline ABSTRACT newlineHuman Activity Recognition (HAR) is an essential technology with broad applications in fields such as surveillance, healthcare, and human-computer interaction. This thesis explores the application of deep learning techniques for HAR across diverse data modalities, including RGB video, infrared video, egocentric videos, and sensor data, to overcome challenges like low accuracy and real-time processing. newlineMotivated by the increasing demand for accurate and efficient HAR systems, especially during the COVID-19 pandemic, this research develops innovative frameworks to enhance recognition performance. A real-time Face Mask Detection Framework using RGB video data achieves an accuracy of 97.75%, providing an effective solution for public safety. Furthermore, an Infrared-Based HAR Framework reaches an accuracy of 81%, improving activity recognition for surveillance and healthcare scenarios. The thesis also presents quotFlu-Net,quot an AI-powered system that detects flu-like symptoms from video data, contributing to controlling the spread of infectious diseases. newlineAdditionally, the research explores cutting-edge algorithms for Egocentric Activity Recognition, with applications in healthcare, sports, and assistive technologies, and implements Sensor-Based HAR using established datasets, offering insights into multimodal systems. newlineThe thesis is structured across eight chapters, covering the introduction, literature review, methodologies, and detailed discussions of each major contribution, with the final chapter summarizing key findings and proposing directions for future research in advanced HAR systems. newlineKeywords: Human Activity Recognition, RGB Data, Infrared Data, Egocentric Vision, Sensor Data, Face Mask Detection, Flu Symptom Detection, Deep Learning, Machine Learning, Surveillance, Healthcare Monitoring. CNN, LSTM, RNN.