High-Performance Lightweight Fall Detection with an Improved YOLOv5s Algorithm

Abstract

The aging population has drastically increased in the past two decades, stimulating the development of devices for healthcare and medical purposes. As one of the leading potential risks, the injuries caused by accidental falls at home are hazardous to the health (and even lifespan) of elderly people. In this paper, an improved YOLOv5s algorithm is proposed, aiming to improve the efficiency and accuracy of lightweight fall detection via the following modifications that elevate its accuracy and speed: first, a k-means++ clustering algorithm was applied to increase the accuracy of the anchor boxes; the backbone network was replaced with a lightweight ShuffleNetV2 network to embed simplified devices with limited computing ability; an SE attention mechanism module was added to the last layer of the backbone to improve the feature extraction capability; the GIOU loss function was replaced by a SIOU loss function to increase the accuracy of detection and the training speed. The results of testing show that the mAP of the improved algorithm was improved by 3.5%, the model size was reduced by 75%, and the time consumed for computation was reduced by 79.4% compared with the conventional YOLOv5s. The algorithm proposed in this paper has higher detection accuracy and detection speed. It is suitable for deployment in embedded devices with limited performance and with lower cost.

Integrating attention mechanism and multi-scale feature extraction for fall detection

“… [ [21] , [22] , [23] ] Enhanced YOLO with multi-scale features. [ [24] , [25] , [26] ] [ 27 , 28 ] Enhanced YOLO with attention modules. …”

“…Zhao et al [ 27 ] improved the feature extraction ability for fall target detection by enhancing both the coordinate attention and shuffle attention mechanisms. Wang et al [ 28 ] proposed adding squeeze-and-excitation networks to the last layer of the backbone network to further enhance feature extraction abilities. Given the features of fall movements, it is crucial to pay attention to both spatial and positional information.…”