AUTOMATED HUMAN ACTIVITY RECOGNITION FROM CONTROLLED ENVIRONMENT VIDEOS

Abstract

This thesis explores deep learning methods for Human Activity Recognition (HAR) from videos to automate the annotation of human activities in videos. The research is particularly relevant for continuous monitoring in healthcare settings such as nursing homes and hospitals. The innovative part of the approach lies in using YOLO models to first detect humans in video frames and then isolating them from the rest of the image for activity recognition which leads to an improvement in accuracy. The study employs pre-trained deep residual networks, such as ResNet50, ResNet152-V2, and Inception-ResNetV2, which were found to work better than custom CNN-based models. The methodology involved extracting frames at one-minute intervals from 12-hour-long videos of 18 subjects and using this data for training and testing the models for human activity recognition. This thesis contributes to HAR research by demonstrating the effectiveness of combining deep learning with advanced image processing, suggesting new directions for healthcare monitoring applications.