基于轻量化Y0L0v8的行人跌倒检测算法的研究与实现

打开文本图片集

中图分类号：TP391.4 文献标识码：A 文章编号：2096-4706（2025）22-0040-05

Research and Implementation of Pedestrian Fall Detection Algorithm Based on Lightweight YOLOv8

KONG Qin， WU Haoyang （NanjingVocational CollegeofInformation Technology，Nanjing21oo23，China）

Abstract：With the acceleration of the global aging process，thefallof the elderlyhas become a major public health problem.Traditional monitoring methods face multiple challenges such as insuffcient accuracy，poorreal-timeperformance， andhighdeploymentcosts incomplex scenarios.Aiming atthe problemthatthe detectionperformanceof theexisting visual monitoring system decreases incomplex environments such as low illumination and oclusion，apedestrian falldetection algorithmbasedonlightweightYOLOv8isproposed.Basedontheoriginalnetwork architecture，byreconstructingalletwork modules，theGSConv lightweightconvolution module isused toreplacethestandardconvolution，and thechannelsplitingfeaturecompensation mechanism isusedtooptimizetheoriginal C2ffeature extraction module.Theexperimentalresultsshow that the improved model reduces the computational complexityby 68% while retaining 97.3% ofthe feature representation eficiency.Theoriginal SPPFmoduleconstructsahybridSPPF_Hybrid pyramid poling structurebycombiningDynamic Spatial PyramidPooling（DsPP） andattentionmechanism.The mAP（∅0.5 indexofsinglesourcereaches O.992，which greatlyreduces the model volumeand computational complexity，and significantly improves the single frame inference speed.

Keywords： Deep Learning; Object Detection; Lightweight Neural Network; Dynamic Spatial Pyramid Poling; edge computing

0 引言

随着全球老龄化进程的加速推进，老年人在未经监管的情况下容易发生跌倒，造成老年人伤亡，已成为当今社会面临的重大公共卫生挑战。（剩余6394字）