Open Collections

Shojaei Hashemi, Anahita

University of British Columbia

Remote health monitoring is evolving from a luxury to a necessity in residential buildings and smart homes. As the early detection of fall incidents can prevent mortality and alleviate morbidity, automatic fall detection systems are crucial for remote health monitoring. Providing rich content at an affordable cost, video cameras have become prevalent for fall detection. Most video-based fall detection models use classic machine learning, relying on handcrafted that limit their performance and generalization capability. In this thesis, we propose three novel video-based fall detection methods to solve the mentioned weaknesses, while considering privacy concerns associated with video data. Using deep learning, our models automatically learn the most discriminative features without resorting to definitions and assumptions. The first method is among the early video-based fall detection algorithms that employ deep learning. Consisting of a long short-term memory (LSTM) neural network, it works with depth cameras and uses the three-dimensional (3D) locations of major body joints to address privacy concerns. The second method is designed to operate with RGB cameras, targeting applications that require the identity of the fallen person to provide personalized medical help. It is comprised of a two-dimensional (2D) convolutional neural network (CNN) that extracts features from the frames and feeds them to an LSTM neural network. To address privacy concerns, edge computing is essential for models working with RGB videos to avoid video transmission. In designing our third method, we paid specific attention to this aspect by keeping the computational complexity of the model so low that it can be implemented on cameras built-in boards without losing its real-time performance. Spatial features extracted from the frames by a lightweight, yet powerful 2D CNN are sequentially processed by a transformer. Evaluations prove the competing performance of our models with regards to the state-of-the-art methods. We also present an embedded system prototype we have built to verify the capability of our RGB models to perform in real time on limited computational resource. Finally, we introduce the UBC-DML Fall Dataset, a large comprehensive video dataset we have captured to address the shortcomings in the current public datasets for fall detection.

Text

2025-04-30

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/84555

Electrical and Computer Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/