Vision-Based Elderly Fall Detection Algorithm for Mobile Robot

Abstract

Fall ranks first among the elderly aged 65 and above. In order to detect falls of the elderly living alone, we use a vision-based detection method to complete fall detection for the elderly on a low-cost small mobile robot. We propose a deep learning fall detection framework for mobile robot. In this framework, Raspberry Pi 4 Model B is selected as hardware platform for mobile robot, and lightweight NanoDet-Lite is used for fall detection. The mAP of our method is 0.912 and model size is only 2.17MB. Our method works more than 3 times faster than YOLOv3-tiny on Raspberry Pi without any hardware accelerator. In ncnn framework, NanoDet-Lite works at 22.03 FPS on Raspberry Pi and mAP reaches 0.902. The results show that our method not only can be applied to the low-cost mobile robot, but also has a good detection performance.