文档名:基于MaxViT和改进几何特征点法的车载单目视觉测速方法研究
摘要:车载视觉测速技术作为自动驾驶车辆组合测速技术的重要组成,具有硬件成本低、算法拓展性强、低速下测量准确等特点,应用前景广阔.为进一步提高视觉测速算法在各类工况下的精度和鲁棒性,将几何特征点法在特征点充足时测速精度高和深度学习方法在多场景下测速稳定的优势进行结合,提出一种基于MaxViT和改进几何特征点法的车载单目视觉测速算法.该算法构建基于双输入MaxViT网络和改进几何特征点法的双通道,并行处理车载前视相机获取的连续3帧输入图像序列,滚动估计车辆当前速度,其中双输入MaxViT网络差异化提取图像不同区域的光流特征,估计当前速度所在的置信度为90%的速度区间,改进特征点法基于特征点运动计算当前速度估计值.当速度估计值落在双输入MaxViT网络估计的速度区间时,以该估计值作为实时车速测量值,否则以速度区间中值作为实时车速测量值.当算法迭代运行多帧后,将速度区间中值作为本帧速度输出以减小累积误差.使用6个车速小于40km/h且包括加减速等工况与直弯道等场景的自建数据集进行实验验证,以理论测速精度0.1m/s的GPS速度信号为参考速度,本文方法平均相对测速误差少于1.37%,最大相对测速误差少于6.13%.实验结果表明,提出的新方法有效提高了车载视觉测速精度与鲁棒性,可为多元车载视觉测速方法融合提供理论支撑.
Abstract:Vehicle-mountedvisualspeedmeasurementtechnology,asanimportantcomponentofcompositespeedmeasurementtechnologyforautonomousvehicles,ischaracterizedbylowhardwarecosts,strongalgorithmscalability,andaccuratemeasurementsatlowspeeds,offeringbroadapplicationprospects.Tofurtherimprovetheaccuracyandrobustnessofvisualspeedmeasurementalgorithmsundervariousworkingconditions,thisarticlecombinedtheadvantagesofgeometricfeaturepointmethodsforhighaccuracywhensufficientfeaturepointswereavailableanddeeplearningmethodsforstablespeedmeasurementinmultiplescenarios.Avehicle-mountedmonocularvisualspeedmeasurementalgorithmwasproposebasedonMaxViTandanimprovedgeometricfeaturepointmethod.Thisalgorithmconstructedadual-channel,parallelprocessingsystembasedonadual-inputMaxViTnetworkandanimprovedgeometricfeaturepointmethod.Thiscouldcontinuouslyprocessasequenceofthreeinputimagesobtainedfromtheforward-facingcameraofthevehicle.Itestimatedthecurrentvehiclespeedinarollingmanner.Thedual-inputMaxViTnetworkdifferentiallyextractedopticalflowfeaturesfromdifferentregionsoftheimages,andestimatedthevelocityintervalwithaconfidencelevelof90%wherethecurrentvelocitylied.Theimprovedfeaturepointmethodcouldcalculatethecurrentvelocityestimatebasedonthemotionoffeaturepoints.Ifthevelocityestimatefallswithinthevelocityintervalestimatedbythedual-inputMaxViTnetwork,itwasusedasthereal-timevehiclevelocitymeasurementvalue.Otherwise,themidpointofthevelocityintervalwasusedasthereal-timevehiclevelocitymeasurementvalue.Afterrunningthealgorithmformultipleframes,themidpointofthevelocityintervalwasusedasthevelocityoutputforthecurrentframetoreducecumulativeerrors.Experimentalverificationwasconductedusingaself-builtdatasetcontainingsixdifferentvehiclevelocity,includingvelocitybelow40kilometersperhourandvariousdrivingscenariossuchasacceleration,deceleration,andstraightandcurvedroads.ThetheoreticalvelocityaccuracyofGPSvelocitysignalswasusedasareferencevelocitywithanaccuracyof0.1m/s.Theproposedmethodcanachieveanaveragerelativevelocitymeasurementerroroflessthan1.37%andamaximumrelativevelocitymeasurementerroroflessthan6.13%.Theexperimentalresultsdemonstratethattheproposedapproacheffectivelyenhancestheaccuracyandrobustnessofvehicle-mountedvisualvelocitymeasurement,providingtheoreticalsupportfortheintegrationofdiversevehicle-mountedvisualvelocitymeasurementmethods.
作者:韩锟 田文涛 李蔚 樊运新 张浩波 Author:HANKun TIANWentao LIWei FANYunxin ZHANGHaobo
作者单位:中南大学交通运输工程学院,湖南长沙410075重载快捷大功率电力机车全国重点实验室,湖南株洲412000
刊名:铁道科学与工程学报 ISTICPKU
Journal:JournalofRailwayScienceandEngineering
年,卷(期):2024, 21(5)
分类号:U495
关键词:车载车辆测速 视觉测速 双输入MaxViT网络 特征点法 验证输出
Keywords:vehiclevelocitymeasurement visualvelocitymeasurement dual-input-MaxViTnetwork characteristicpointmethod verificationoutput
机标分类号:U491TP273.4TP391.41
在线出版日期:2024年7月8日
基金项目:重载快捷大功率电力机车全国重点实验室开放基金资助项目基于MaxViT和改进几何特征点法的车载单目视觉测速方法研究[
期刊论文] 铁道科学与工程学报--2024, 21(5)韩锟 田文涛 李蔚 樊运新 张浩波车载视觉测速技术作为自动驾驶车辆组合测速技术的重要组成,具有硬件成本低、算法拓展性强、低速下测量准确等特点,应用前景广阔.为进一步提高视觉测速算法在各类工况下的精度和鲁棒性,将几何特征点法在特征点充足时测速精...参考文献和引证文献
参考文献
引证文献
本文读者也读过
相似文献
相关博文
基于MaxViT和改进几何特征点法的车载单目视觉测速方法研究 Vehicle-mounted monocular visual velocity measurement method based on MaxViT and improved geometric feature point method
基于MaxViT和改进几何特征点法的车载单目视觉测速方法研究.pdf
- 文件大小:
- 23.94 MB
- 下载次数:
- 60
-
高速下载
|
|