Elderly fall behavior recognition algorithm based on improved YOLOv8-Pose

With the global population aging, monitoring the health and safety of the elderly has become increasingly important, especially since falls may lead to serious consequences. This paper improves the YOLOv8-Pose model with the help of the study of posture estimation methods to improve the accuracy of identifying falls of the elderly. Existing models have limited ability to fuse multi-scale features and insufficient detection of subtle behavioral differences. To address these issues, this paper adopts a series of technical improvements, including the introduction of the C2f_DBB module and the use of the PIoUv2 loss function, to enhance the model's sensitivity to the characteristics of fall behavior and the ability to locate key points. The C2f_DBB module introduces the Diverse Branch Block on the basis of the C2f module to improve the spatial perception of feature extraction. Its multi-scale fusion strategy effectively captures contextual information and enhances the robustness and accuracy of the model in identifying falls. The PIoUv2 loss function improves the traditional IoU through a dynamic penalty mechanism to improve the regression performance of body posture, especially in complex environments. Experimental results show that the improved YOLOv8-Pose model has obvious advantages in multiple evaluation indicators, especially in accuracy, recall and mAP50, and its performance is better than other existing models, which proves the superiority of the proposed method and its application potential.