Fast and High-Precision Human Fall Detection Using Improved YOLOv8 Model

Abstract

Human falls refer to a sudden, accidental, unintentional, involuntary, and traumatizing action of a person losing stability and ending up lying down (often on the ground). These incidents of collapsing, tripping, crashing, tumbling, falling, etc. cause major harm and injuries to the elderly in our society. According to WHO, it is the second-leading cause of death worldwide. Therefore, having a fast and high-precision fall detection system operating in real time is crucial. It will ensure quick assistance and early treatment are provided to the impacted elderly individuals. Recently, Deep Learning based systems have shown promising results for detecting changes in postures and promptly responding to them in real-time. The You Only Look Once (YOLO) models have been widely used previously for designing and implementing human fall detection systems. This research presents an improved YOLOv8 model for fast and highly accurate human fall detection. The improvements include reducing the number of layers in the backbone of YOLOv8 and incorporating the attention mechanism in the head of the network. For training and evaluation, the CAUCAFall dataset is used. The original YOLOv8 and improved YOLOv8 are evaluated on the same dataset, achieving mAP of 0.995. Also, the training time of the improved YOLOv8 is 0.457 hours, faster than other YOLO models. Therefore, the presented model is a practical approach for human fall detection and can serve in public places for safeguarding the elderly of our society against potential injuries.

Keywords

• Human fall detection
• deep learning
• YOLOv8
• attention mechanism
• lightweight network
• parameter reduction