CAGn: High-Order Coordinated Attention Module for Improving Fall Detection Models

In order to quickly and accurately detect the occurrence of accidental falls and save the lives of more elderly people living alone, this paper proposes a new convolutional network module Higher-order Coordinated Attention module (CAGn) based on the latest research, and uses it and the lightweight convolutional module (GSConv) to improve the YOLOv5s model to construct a new fall detection model YOLOv5s-CAGn-GSConv. In addition, a new fall detection dataset Fall-Dataset is produced, and a fall event triggering mechanism based on queue window is proposed to be applied to the fall detection system. Experimental results show that the CAGn module can capture cross-channel information and high-order spatial information at the same time, thereby effectively improving the accuracy of the detection model, and the introduction of GSConv module further improves the accuracy and greatly reduces the number of parameters of the detection model. The final YOLOv5s-CAGn-GSConv model only increases the number of parameters by 4.2% but improves the accuracy by 2.1% and the average accuracy by 0.9% (mAP@0.5) and 2.6% (mAP@0.5:0.95).