A deep-learning-based fall-detection system to support aging-in-place
Abstract
Description
A thesis Submitted in Partial Fulfillment of the Requirements for the Degree of MASTER OF SCIENCE in COMPUTER SCIENCE from Texas A&M University-Corpus Christi in Corpus Christi, Texas.
Emergency departments treat around 2.5 million older people for fall injuries each year. Serious head and broken bones injuries occur in 20% of falls. Fall injuries, adjusted for inflation, has direct medical costs of $34 billion a year. Taking into account that people 65 and older are expected to comprise 21.7% of the U.S population in 2040, compared to 14.4% in year 2013, the numbers presented in the statistics will dramatically increase as well. Preserving the elderlys' right of aging in a home of their own choice is mandatory in today's world, as more elderly people are willing to live independently. But, with the statistics showing that falling is a major health problem that has a huge non-desirable impact on elderly lives, fall detection systems become a necessity. Different approaches have been used to design fall detection systems. One approach depends on wearable sensors that measure different physical parameters of a human body or the environment around it, such as the body acceleration or its pressure on the floor. A second approach depends on sensors employed in the environment. These sensors mainly include wide-angle cameras, depth cameras, and microphones. Different approaches used different classifiers for training the system to detect falls. Despite these efforts to detect falls, it is possible that other naturally occurring falls trigger false alarms. Thus, the current implementations of fall detection systems need to be improved. Most recently, computer vision based approaches using depth cameras are the mostly used for such improvement. Using deep neural networks to learn features from video frames have a potential to improve the fall-detection accuracy and reduce triggering false alarms. In this study, a more robust and deep fall-detection system was designed. This approach extends deep convolutional neural networks in time. This extension allows capturing the spatial and temporal information presented through successive video frames. The result of the new approach can be used to implement a reliable surveillance system in a real-world environment.
Computing Sciences
College of Science and Engineering
Emergency departments treat around 2.5 million older people for fall injuries each year. Serious head and broken bones injuries occur in 20% of falls. Fall injuries, adjusted for inflation, has direct medical costs of $34 billion a year. Taking into account that people 65 and older are expected to comprise 21.7% of the U.S population in 2040, compared to 14.4% in year 2013, the numbers presented in the statistics will dramatically increase as well. Preserving the elderlys' right of aging in a home of their own choice is mandatory in today's world, as more elderly people are willing to live independently. But, with the statistics showing that falling is a major health problem that has a huge non-desirable impact on elderly lives, fall detection systems become a necessity. Different approaches have been used to design fall detection systems. One approach depends on wearable sensors that measure different physical parameters of a human body or the environment around it, such as the body acceleration or its pressure on the floor. A second approach depends on sensors employed in the environment. These sensors mainly include wide-angle cameras, depth cameras, and microphones. Different approaches used different classifiers for training the system to detect falls. Despite these efforts to detect falls, it is possible that other naturally occurring falls trigger false alarms. Thus, the current implementations of fall detection systems need to be improved. Most recently, computer vision based approaches using depth cameras are the mostly used for such improvement. Using deep neural networks to learn features from video frames have a potential to improve the fall-detection accuracy and reduce triggering false alarms. In this study, a more robust and deep fall-detection system was designed. This approach extends deep convolutional neural networks in time. This extension allows capturing the spatial and temporal information presented through successive video frames. The result of the new approach can be used to implement a reliable surveillance system in a real-world environment.
Computing Sciences
College of Science and Engineering