融合全局子采样注意力与定向随机失活的电梯场景目标检测模型

打开文本图片集

本文引用格式：，，，等.融合全局子采样注意力与定向随机失活的电梯场景目标检测模型[J.自动化与信息工程 2025，46（6）：24-31. WU Bin，YU Bin， CAO Jianbo，et al.An elevator scene object detection model integrating global subsampling attentionand targeted dropout[J].Automation& Information Engineering，2O25，46（6）：24-31.

关键词：目标检测；电梯场景；注意力机制；全局子采样注意力；定向随机失活中图分类号：TP391.4 文献标志码：A 文章编号：1674-2605（2025）06-0004-08DOI：10.12475/aie.20250604 开放获取

An Elevator Scene Object Detection Model Integrating Global Subsampling Attention and Targeted Dropout

WUBinlYUBin²CAO Jianbo²ZHUYixial FANG Jiansheng² CHEN Zaili³

（1.Guangzhou Guangri Stock Co.，Ltd.，Guangzhou 51oo45，China 2.Guangzhou Guangri Stock Co.，Ltd.，Guangri Research and Development Institute，Guangzhou 51O045，China 3.Institute ofIntelligent Manufacturing， Guangdong Academy of Sciences， Guangzhou 51Oo70， China）

Abstract：Addressing thechallenges inelevatorscenarios—such asuneven illumination inside thecabin，high-density oclusions，andstringentreal-timedetectionrequirements—whichmakeitdificultforobjectdetectionmodelstobalance computational eficiencywithgeneralizationcapability，GSAODmodelframework integratingGlobalSubsamplingAtention（GSA） andDirectedRandomDropoutisproposed.TheGSAmoduleisembeddedintotheobjectdetectionframework，wheremulti-scale dynamicsub-windosamplingisemploedtoapturekeyatentionregions，preservingspatialawarenesandlong-rangedependencies whilereducingthecomputationalcomplexityoftraditionalself-atentiontolinearlevel.Duringmodel training，adirectedrandom dropoutmechanismisintroducedandspecificallyapliedtoatentionlayersandthefeaturepyramidnetworktomitigateoverfiing risks.This frameworkisimplementedontheclasicYOLOmodelandvalidatedthroughexperimentsontheCOCOdatasetandaselfbuiltelevatorscenedatasetExperimentalresultsshowthattheproposedmodelmaintainsconsistentdetectionperformanceingeneral scenarios; in elevator scenarios，the YOLOl_GD model improves mAP（∅50 andmAP by2.6and 2.2 respectively compared to thebaselineYOLOlodel，ectielyancingobustsiomplexidustrlnviomentsschsevatosilerigal time detection speed.

Keywords： object detection; elevator scene;atention mechanism; global subsampling atention; targeted dropout

0 引言

电梯作为高频使用的特种设备，其运行安全与智能化管理可通过目标检测技术来实现[1]，如实时检测、统计轿厢内乘客数量以优化电梯调度；识别电动车或轮椅等特殊目标以触发预警等。（剩余9730字）