Optimized YOLOv Algorithm for Fall Detection in Elderly Care Scenarios

Abstract

With the increasing trend of global population aging, falls among the elderly have become a major public health challenge. Traditional fall detection methods have problems like high false alarm rates, susceptibility to lighting, occlusion restrictions, and easy invasion of privacy. To optimize the detection accuracy and robustness, an optimized real-time object detection algorithm (You Only Look Once version, YOLOv) for elderly care scenarios was proposed. The research develops an environment adaptive fall detection model based on optimized YOLOv algorithm, which intelligently selects CSCD-YOLOv5 or MDF-YOLOv8 algorithm through "environment judgment". The CSCD-YOLOv5 algorithm utilizes the multi-scale pooling of Spatial Pyramid Pooling with Expanded Layer Aggregation Network to enhance feature expression, and combines context guidance to improve detection accuracy in complex environments. The MDF-YOLOv8 algorithm enhances the ability to recover details in low light conditions through low light enhancement, lightweight design, and attention mechanism. The the area under the recall-precision curve of the research method was close to a rectangle, with an area value of about 97.6%. The missed detection rate was 7.2% in the composite posture, and the false alarm rate was 3.9% in the supporting behavior. When the light intensity was 400Lux, the average accuracy reached the maximum of 97.4%. The proposed method has good accuracy, robustness, generalization, and stability, effectively solving the insufficient accuracy and poor robustness of traditional methods, and enhancing the reliability of fall detection in elderly care scenarios.