Study of fall detection in infrared scenes based on improved YOLOv8-POSE

Accidental fall is one of the main reasons that threaten the health of the elderly and even cause death. The hazards caused by falls can be minimized by timely detection and timely rescue through monitoring equipment. Significant progress has been made in fall detection algorithms based on image processing. Still, fall detection algorithms under natural light do not apply to dark scenes and scenes that need to protect people's privacy. To address the identified limitations, a fall detection algorithm based on infrared surveillance devices is proposed. An infrared human posture dataset is developed, and the YOLOv8-PCS key point detection model is introduced. Additionally, the CBAM and Slim-neck lightweight architecture are integrated to maintain accuracy while improving the model's efficiency. Compared with the YOLOv8-POSE, mAP@50 improves by 2.5%, and the detection speed FPS reaches 27.8. And a multi-feature based fall detection decision strategy is proposed, and the experimental results show that the fall strategy has high accuracy and response speed, the accuracy rate reaches 91.6%, and the average detection time consumes 36ms, which meets the basic requirements of fall detection.