Spatiotemporal Analysis for Accurate Real-Time Fall Identification in Elderly Care

Falls result in the most severe injuries in older adults. Data shows that more than one in five older adults experience at least one fall each year, and nearly one in three older adults fall annually. Almost half of older adults 42% over 70 years of age fall at least once each year. Being able to recognize falls quickly is very important. A quick response can help prevent serious complications from falls. This journal presents a real-time visual fall identification system based on YOLOv8, with the Yolo Model scoring system and recovery metrics to accurately identify falls without using any wearable sensors. This proposed framework utilizes several different analysis methods for identifying the different stages of falling, standing, sitting, lying down, and recovery. It does this by analyzing multiple spatio-temporal indicators such as height loss, bounding box aspect, displacement of body mass, and time of persistence between points in determining whether someone has fallen or appears to have fallen. The scoring system helps to provide a more accurate means of identifying a fall based on the spatiotemporal factors reduced rate of false alarms and improved accuracy of identifying elderly individuals before or after they fall, by using many different spatiotemporal features as opposed to using only a single one or threshold method. The developed system can operate in real-time using a webcam and provides visual alerts of detected falls with a method for providing confidence ranging from 0% to 100% to the viewer. The system developed in this work is capable of being used in elder care type facilities and for the wider implementation of smart surveillance systems that monitor areas where older adults frequently spend time. However, the reliability of detecting falls and post-fall recovery could be negatively affected due to severe occlusions, poor lighting conditions, or due to too many individuals in the same location. Despite these limitations, the framework offers improved interpretability and robustness compared to existing vision-based fall detection systems.