基于改进YOLOv8的复杂场景跌倒检测算法  

Fall Detection Algorithm in Complex Scenes Based on Improved YOLOv8

作　　者：曹萌迪 杨梦凡 王留毅 王东洋 CAO Mengdi;YANG Mengfan;WANG Liuyi;WANG Dongyang(North China University of Water Resources and Electric Power,Zhengzhou 450046,China;Henan Vocational College of Water Conservancy and Environment,Zhengzhou 450008,China)

机构地区：[1]华北水利水电大学,河南郑州450046 [2]河南水利与环境职业学院,河南郑州450008

出　　处：《现代信息科技》2025年第5期66-71,共6页Modern Information Technology

摘　　要：针对由于光照变化、人员密集以及人体形态被遮挡等因素导致跌倒检测精度低、实时性差的问题,文章提出了一种基于改进YOLOv8的复杂场景跌倒检测算法:SLG-YOLO。设计了SC模块并嵌入C2f模块,利用边缘检测算法来提取跌倒行为的边缘特征,在SPPF模块引入Large Separable Kernel Attention(LSKA)注意力机制,利用大型可分离卷积捕捉图像广泛的上下文信息,在增强对于重要特征关注度的同时不会增加计算复杂度;同时,采用Gather and Distribute(GD)机制,结合3D卷积的多尺度特征序列融合方法,取代原有的Neck部分,提升了模型对于复杂场景和不同尺度目标的检测精度。实验结果表明,在参数量和计算量为6.2 M和10.7 FLOPs的基础上,SLG-YOLO跌倒检测算法与YOLOv8相比,准确率提升了2.2%,召回率提升了3.0%,mAP@0.5提升了3.6%,mAP@0.5:0.95提升了3.1%。Aiming at the problems of low fall detection accuracy and poor real-time performance due to factors such as lighting changes,dense population,and occluded human form,this paper proposes a fall detection algorithm in complex scenes based on YOLOv8 of SLG-YOLO.The SC module is designed and embedded in the C2f module to extract edge features of fall behavior using edge detection algorithms,and the Large Separable Kernel Attention(LSKA)is introduced in the SPPF module,which utilizes large separable convolution to capture a wide range of contextual information of images,enhancing the attention to important features without increasing the computational complexity.At the same time,it adopts the Gather and Distribute(GD)mechanism,and combines with the Multi-scale Sequence Feature Fusion method(SSFF)of 3D convolution to replace the original Neck part,which improves the detection accuracy of the model for complex scenes and targets of different scales.The experimental results show that the SLG-YOLO fall detection algorithm improves the precision by 2.2%,recall by 3.0%,mAP@0.5 by 3.6%,and mAP@0.5:0.95 by 3.1%based on the number of parameters and calculated quantities of 6.2 M and 10.7 FLOPs.

关 键 词：跌倒检测 边缘检测 注意力机制 YOLOv8 

分 类 号：TP311.5[自动化与计算机技术—计算机软件与理论]