Embedded Fall Detection System using Deep Learning

The balance control system, crucial for maintaining stability and preventing falls, has garnered significant attention due to its decline in the elderly and individuals with various pathologies. As the global aging population continues to grow and life expectancy increases, preserving mobility has become increasingly vital. This thesis addresses the pressing need for an accurate, portable, and affordable device for fall detection. The World Health Organization reports alarming statistics, with a substantial percentage of individuals over 65 experiencing falls annually, often leading to severe consequences. Traditional balance assessment methods rely on costly and immobile dynamometric force platforms, limiting their practicality for widespread use. Fall detection systems can help to mitigate this risk by alerting caregivers when a fall has occurred. However, many existing fall detection systems are expensive, bulky, and/or require specialized installation. Recent years have witnessed a remarkable shift towards body-wearable sensor technology, offering an innovative approach to assessing an individual's motion and activity. These sensors, including accelerometers and gyroscopes, enable cost-effective, portable, and versatile solutions for measuring three-dimensional movements. This development aligns with the increasing importance of fall detection as falls become a growing public health concern. In this thesis, we propose a fall detection system that uses a neural network to classify data from a tri-axial accelerometer tri-axial gyroscope, and a pressure sensor. Our system is designed to be low-cost, portable, and easy to install. The aim of this study is to develop a high-accuracy model that is small enough to run on an embedded controller and also to improve the model to work with different datasets. Lastly, assess the benefit of adding a pressure signal to the dataset. We evaluated our system on two datasets the SisFall dataset, and a self-collected dataset. A combined dataset that includes data from SisFall and self-collected datasets was also assessed. Moreover, the self-collected dataset contains in addition to a pressure signal a new fall type which was not introduced in the SisFall dataset which is Syncope. Our system achieved a testing accuracy of 99.38\% on the combined dataset, demonstrating its potential for use in a real-world setting. moreover, adding the pressure signal to the training data led to improving the accuracy slightly and lowered the false positive and false negative when compared to the results of training the model without the inclusion of the pressure signal. Our system has the potential to improve the safety and independence of individuals of all ages, including elderly individuals and workers, by ensuring immediate assistance is provided when a fall incident happens. This rapid response not only reduces the time between the incident and help arriving but also significantly diminishes the potential consequences of the fall.