Real-Time Human Fall Recognition based on Deep Learning Methods and Single Depth Image with Privacy Requirements

With the decline in physical function of the elderly, the risk of falling down increases and threatening their health. In the traditional fall recognition methods, multi-frame RGB image sequences are obtained to recognize falls, which may lead to problems of a high computational cost, real-time recognition, privacy leakages. In this paper, a novel fall recognition algorithm based on a single frame human binary image is proposed. Especially, the method is suitable for scenes with a basically unchanged background and privacy protection requirements, such as bathroom and toilet. The experimental results show that the precision and recall of the model achieves 97.3% and 99.4% on the public URFD dataset, which meets the practical application requirements. Furthermore, a posture recognition algorithm based on yolov5s and depth images is designed to distinguish between falling and lying down, which can be applied to bedrooms, living rooms and other places. On the basis of these algorithms, a depth image fall recognition software based on the flask framework is developed and deployed to the Alibaba cloud platform. Finally, the proposed fall recognition approach is validated by experiments and realized in real time.