VISION-BASED WARNING SYSTEM FOR FALL DETECTION

Recently, the number of elderly people has been increasing, indicating that the number of falls per year increases. The major consequences of these falls that may lead to death have motivated us to implement a system that keeps monitoring the elderly in the house all the time without any human intervention. Having a fall detection system based on wearable or non-wearable sensors has countless drawbacks such as forgetting to wear them and replacing/recharging their batteries thus our system is a passive system that does not require using any sensors, neither wearable nor non-wearable sensors in order to overcome these drawbacks. The respective person will only have to register through our web application to receive a notification through his email and mobile number in case an emergency occurs. The system detects the falls using an action recognition based approach by detecting the person in the room using the CenterNet model, extracting the coordinates of the bounding box that encloses the person, and keeps monitoring that person. By analyzing the coordinates of the bounding box in case of detecting a fall for 4 consecutive frames, the respective person will be notified. The respective person also receives a notification in case of not detecting the person in the room for 4 consecutive frames. Our system was tested using a custom dataset of real-life falls from different angles and showed impressive results compared to existing systems in the literature.