Robust elderly fall detection using CARE-light–enhanced YOLOv8 with triplet attention and BMCNet

As the global population ages, the demand for fall detection systems for the elderly in both home and institutional settings is continuously rising. However, current visual solutions often face issues such as missed detections and false alarms in real-world environments, especially under low light, complex backgrounds, and varying fall postures. To address these challenges, this paper presents an innovative fall detection system for the elderly, designed for edge devices, with a focus on achieving efficient and robust fall detection. The core innovations include: 1) the CARE-Light (Constraint Adaptive Re-illumination and Enhancement for Low-light) mechanism, which significantly enhances the detectability of human features under low light by integrating adaptive brightness-dynamic gamma, local contrast enhancement, denoising, and online enhancement strength constraints; 2) the integration of the TripletAttention module into the YOLOv8 backbone network, which focuses on foreground attention through three-dimensional attention and significantly improves the model's ability to distinguish human targets, solving the problem of interference from complex backgrounds; and 3) the introduction of a boundary-aware multi-scale context network (BMCNet) at the YOLOv8 detection head, which comprehensively understands the varying fall postures and target scales, refining the human edge and posture structure to address inaccurate localization under varying fall postures. Comparative and ablation experiments on multi-scene datasets demonstrate that, while maintaining real-time performance, this method achieves an mAP@0.5 of 0.860, a Precision of 0.983, and a Recall of 0.822, delivering excellent detection performance. This method significantly improves fall detection accuracy in low light, complex backgrounds, and variable postures, providing an efficient and robust solution for elderly fall monitoring.