基于自学习区域选择与边缘聚焦的单目3D检测

打开文本图片集

关键词：单目3D检测；自学习区域选择；边缘融合；数据增强；注意力机制中图分类号：TP391.4 文献标志码：A 文章编号：1001-3695（2025）08-041-2552-09doi：10.19734/j.issn.1001-3695.2024.09.0439

Monocular 3D detection based on self-learning region selection and edge focus

Wang Xinwei1， Zhang Youbing ， Zhou Kui1 （1.Shring-XJtinbeie yanHubei 44200o，China；2. Jingchu Universityof Technology，JingmenHubei 4480oo，China）

Abstract：This study proposedamonocular3Ddetection algorithmbasedonself-learning regionselectionandedge focusing. UnliketraditionalmethodsthatusedtheentireRolfor3Datributelearning，theproposedalgorithmleveragedadata-driven modeltoself-learnandselectvaluableregions，mitigatingthenegativeinfluenceofirelevantareas.Aditionally，thealgo rithmenhancededgeregions bymodelingthedistributioncharacteristicsof truncatedtargets，improving thefocusonhighfrequencyregions.Furthermore，itintroducedadataaugmentationmethodincorporatingspatialconsistencyconstraints，xtendingthe3Dsamplesetbyaddingspatialconstraints tothecut-and-paste method，thusensuring consistencywith imaging principles.Experimentalresultsonthe KITIdatasetdemonstratethattheproposed methodoutperforms baseline models，witha （204 29.8% improvement in accuracy for truncated targets compared to the baseline model.

Key words：monocular3Ddetection；self-learning region selection；edge fusion；dataaugmentation；atention mechanism

0引言

随着自动驾驶和机器人系统的迅速发展，环境感知成为智能系统的关键环节。（剩余19750字）