Fall detection algorithm based on improved YOLOv8

Falls are an important factor threatening the physical health of elderly people. Aiming at the shortcomings of low accuracy and poor robustness of existing models for fall detection, an improved YOLOv8 fall detection algorithm is proposed. Firstly, in the multi-scale fusion stage at the Neck, the CBS module used for downsampling is replaced with a dynamic deformable convolution of size 3×3 with a stride of 2, replacing the original regular convolution model. Meanwhile, we have also added a shuffle attention mechanism (SAttention) to enhance spatial attention and extract more accurate detection information. In this paper, the loss function is optimized using Focal Loss, and the detection performance of the model is improved by optimizing the loss function CIoU. Experimental results have shown that the improved model has a fall detection detection rate of 91.2%. At the same time, the improved algorithm in this paper achieved an improvement in detection rate in three states: falling, standing, and squatting, with increases of 3.9%, 3.4%, and 3.7%, respectively.