Human Fall Detection Algorithm Based on YoloX-s and Lightweight OpenPose

Abstract

The existing research shows that falls account for a significant proportion of safety accidents. At the same time, as many countries enter an aging society, falls have increasingly become a non-negligible safety issue affecting the lives and health of the elderly. To address the current problems of human fall detection, we propose to extract a human skeleton model based on YoloX-s in combination with Lightweight OpenPose. This model can identify human fall by the difference values of angle change’s rate between the key points of the neck and knees. The results demonstrate that the accuracy rate for fall detection is 97.92% and that for normal behavior detection is 96.46%. The computing speed of the method satisfies the need for real-time processing with satisfactory robustness.