Date of Graduation
Master of Natural and Applied Science in Computer Science
Falling is one of the leading causes of death from unintentional injuries in older adults. They are more common in people over the age of 65. Wearable sensor-based solutions are commercially available, but they possess limitations like recharging the sensors, and wearing them can be intrusive to the user. Consequently, vision-based fall detection approaches offer a feasible alternative due to the ever-increasing presence of cameras in smart homes. This thesis presents a novel two-stage human fall detection system for smart homes. The proposed approach uses humans as a sensor. It is a vision-based two-stage process where Stage-1 is dedicated to detecting fall-like events at the edge of the network, and Stage-2 is hosted in the cloud to confirm the fall. I propose a template matching technique in the first stage and a model based on LiteFlowNet and LRCN in the second stage. The proposed deployment reduces the workload on cloud servers while ensuring service availability at the edge level if the cloud service is inaccessible. I evaluated this approach using publicly available datasets and real-time videos. I have also compared the model performance with existing state-of-the-art vision-based fall detection systems that used the same publicly available dataset. Results accumulated from the experiments show the efficacy of the proposed approach for smart home deployment.
CNN, contour detection, edge detection, foreground detection, LiteFlowNet, LRCN, template matching
Other Computer Sciences
© Snigdha Chaudhari
Chaudhari, Snigdha, "Vision-Based Human Fall Detection in Smart Homes" (2022). MSU Graduate Theses. 3767.
Available for download on Monday, August 05, 2024