Evaluation of action recognition algorithms for fall detection applications

• Main Author: Pei, Yubo
• Other Authors: Alex Chichung Kot
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Fall detection; Graph convolutional neural networks; Skeleton-based action recognition; Spatio-temporal information
• Online Access: https://hdl.handle.net/10356/175958
• Institution: Nanyang Technological University

As society ages, ensuring the elderly's safety at home, especially fall prevention, has become crucial due to the serious health risks falls pose. Consequently, developing effective fall detection and prevention technologies is a key focus in smart elder care. This dissertation explores a video surveillance-based fall detection method aiming to enhance the elderly's safety and well-being through technological innovations. This dissertation conducts a comprehensive analysis of existing action detection technologies, combined with Spatial-Temporal Graph Convolutional Networks (ST-GCN) and Two-Stream Adaptive Graph Convolutional Networks (2S-AGCN) techniques from deep learning, to verify the performance of these two models in fall action recognition. Two sets of experiments were designed, namely, 8-class and 2-class classification experiments for model training experiments and evaluation experiments respectively. The experiments were trained and validated using the NTU RGB+D dataset and evaluated through comparative experiments. The experimental results show that the fall detection methods based on ST-GCN and 2S-AGCN achieved excellent performance on the dataset designed for this experiment, especially showing higher accuracy in the 2-class of fall actions. These achievements not only provide a new direction for the development of fall detection technology but also offer strong technical support for developing home safety systems for the elderly.