Fall Detection in Stairwells Based on Depth Images and Deep Learning

In response to the increasing proportion of the elderly population, the problem of long-term care at home has arisen. To prevent injuries caused by falls in dangerous areas of the house when caregivers are not paying attention, this paper introduces a video-based system for recognizing hazards on staircases. The proposed system employs depth imaging technology, enabling nighttime functionality. Utilizing human body pose detection aims to identify potential falls. For this research, two experimenters collected data from two distinct datasets of depth images. The system was trained using YOLO and supplemented with MeanShift algorithm for people's body tracking during image detection. Different deep learning models were utilized for predicting human postures, resulting in superior accuracy. The system demonstrated optimal accuracy rates of 90.4% and 77.5% across different experimental scenarios while maintaining a consistent real-time recognition speed of 30 FPS. This paper's proposed application enriches conventional RGB imaging with depth imaging, mitigating issues such as low recognition rates and image distortion caused by indoor lighting conditions. Furthermore, it enables caregivers to receive instant alerts, even when away from home.