Fast-YOLO-FD: An improved fall detection model based on YOLOv10

Falls result in serious injuries such as fractures and head trauma, placing a substantial financial and operational burden on healthcare systems due to the costs of both treatment and prevention. For deploying fall detection algorithms on edge computing devices with limited processing capabilities, this paper proposed a novel lightweight algorithm called Fast-YOLO-FD based on YOLOv10n. First, by deconstructing the block proposed in RepViT block, C2f-RTB module was developed. Second, a frequency-aware feature fusion method was introduced for the first time in the field of fall detection, resulting in a novel and first-proposed FFF-FPN. Third, in the field of fall detection, the TADDH is innovatively integrated into the YOLOv10 architecture. By utilizing shared convolutions, it significantly reduces the number of parameters, resulting in a more lightweight model suitable for resource-constrained devices. In order to address scale inconsistency across detection heads, scale layers were incorporated for feature adaptation. Finally, to improve the accuracy achieved by these lightweight optimizations, this paper proposes a novel FDIoU loss function and applies it to bounding-box regression. The FDIoU is optimized for standard wall-mounted side-view surveillance, the predominant setup in elderly care. The geometric penalties, with a conservative weight of λ_gnd=0.01, serve as a form of soft regularization. Experimental results show that Fast-YOLO-FD reduces the parameters and FLOPs to 1.553M and 5.9G, while achieving an mAP@0.5:0.95 of 75%. On edge devices, Fast-YOLO-FD shows superior speed performance compared to other lightweight models.