基于目标检测和光流法的动态SLAM算法

打开文本图片集

关键词：目标检测；视觉SLAM；光流法；动态环境；八叉树地图；ORB特征点中图分类号：TN911.73-34；TP242.6 文献标识码：A 文章编号：1004-373X（2025）17-0048-07

Dynamic SLAM algorithm based onobject detection andoptical flow method

YUXiayi， ZHENGEnhui， SUNHaidi （SchoolofMechanicalandElectricalEngineering，ChinaJiliangUniversity，Hangzhou31Oo18，China）

Abstract：Inviewofthefactthat dynamicobjectsaffect theaccuracyofcamerapose estimationinsimultaneous localization andmapping （SLAM） system，a dynamic SLAM system basedon object detectionandoptical flow method is proposed.Thesystemaims toadddynamicobjectdetectionthreadsandmapconstructionthreads totheORB-SLAM2（oriented FASTandrotated BRIEF-simultaneouslocalizationandmapping2）framework.Firstly，thesceneopticalflowinformationis obtained bythePWC-Netopticalflownetwork，andthepotential dynamicarea intheenvironmentisobtainedbytheYOLOv5s objectdetectionnetwork，andthescenedynamicareaisobtainedbythecombinationofthetwo.Secondly，topreventtheissed detectionintheobjectdetectionnetwork，theopticalflowinformationisusedforspeedconstraint，andthedynamicfeature points areeliminatedtwice.Finally，thecameraposeestimationandoctreemapconstructionafterremoving dynamicobjectsare completed.TheEVOisusedtoevaluatethesystemontheTUMRGB-Dpublicdataset.Theexperimentaldatashowthecamera poseestimationaccuracyof theproposedsystemin highdynamicenvironmentisimproved incomparison withtheORB-SLAM2， Dynamic-VINSandDS-SLAMsystems，which indicates thattheimprovement canimprovethe performance of SLAMsystem effectivelyindynamic environment.

Keywords：object detection; visual SLAM;optical flow method;dynamic environment;octree map; ORB feature point

0 引言

同时定位与地图构建（SimultaneousLocalizationandMapping，SLAM）是指在未知的环境中，移动机器人在运动时通过观察环境来确定自己的运动路径，并同时创建出环境地图。（剩余8247字）