A High-Precision Fall Detection Model Based on Dynamic Convolution in Complex Scenes

Abstract

Falls can cause significant harm, and even death, to elderly individuals. Therefore, it is crucial to have a highly accurate fall detection model that can promptly detect and respond to changes in posture. The YOLOv8 model may not effectively address the challenges posed by deformation, different scale targets, and occlusion in complex scenes during human falls. This paper presented ESD-YOLO, a new high-precision fall detection model based on dynamic convolution that improves upon the YOLOv8 model. The C2f module in the backbone network was replaced with the C2Dv3 module to enhance the network’s ability to capture complex details and deformations. The Neck section used the DyHead block to unify multiple attentional operations, enhancing the detection accuracy of targets at different scales and improving performance in cases of occlusion. Additionally, the algorithm proposed in this paper utilized the loss function EASlideloss to increase the model’s focus on hard samples and solve the problem of sample imbalance. The experimental results demonstrated a 1.9% increase in precision, a 4.1% increase in recall, a 4.3% increase in mAP0.5, and a 2.8% increase in mAP0.5:0.95 compared to YOLOv8. Specifically, it has significantly improved the precision of human fall detection in complex scenes.