Fall detection system based on real-time pose estimation and SVM

With the rapid growth of the elderly population, fall detection has become a key issue in the medical and health field. Accurately detecting fall behavior in surveillance video and timely feedback can effectively reduce the injury and even death of the elderly due to falls. For the complex scenes in surveillance video and the interference of multiple similar human behaviors, this paper proposes a method based on pose estimation and the auxiliary detection method based on yoloV5. First, extract video frames from different falling video sequences to form a data set; then, input the training sample set into the improved network for training until the network converges; finally, test the category of the target in the video according to the optimized network model and locate the target. Experimental results show that the improved algorithm can effectively detect falls or Activities of Daily Living (ADL) events in each frame of the image and give real-time feedback. The detection of falling behavior in the video further verifies the feasibility and efficiency of the recognition method based on our deep learning methods.