Human fall detection based on posture estimation and infrared thermography

Fall detection holds significant practical value in the fields of medicine and security. In this study, we propose a pose-based human fall detection system that incorporates an infrared thermal imaging camera to overcome challenges faced during nocturnal and poorly illuminated scenarios. To facilitate fall detection, we have integrated a classification function into the YOLOv7POSE network. Furthermore, the introduction of SimAM to the backbone network substantially enhances its capability to discern human features. In terms of image processing, we optimize the ConvNeXt structure and integrated it into the backbone network to improve the accuracy of key point detection and fall classification. Our enhancements also encompass the redesign of the spatial pyramid pool structure within the head layer, resulting in improved model accuracy. We also replace the regression function of the target detection framework with SIOU, and correspondingly adjusted the calculation formula to enhance the accuracy of coordinate regression. Experiments show that the mAP of the improved model reaches 0.964, which is 0.045 higher than that of the original YOLOv7POSE model, the accuracy of fall pose recognition and the confidence of detection are significantly increased.