资源描述：
《Fusing LIDAR and Images for Pedestrian Detection using Convolutional Neural Networks》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。
1、2016IEEEInternationalConferenceonRoboticsandAutomation(ICRA)Stockholm,Sweden,May16-21,2016FusingLIDARandImagesforPedestrianDetectionusingConvolutionalNeuralNetworks*JoelSchlosser,ChristopherK.Chow,andZsoltKira1Abstract—Inthispaper,weexplorevariousaspect
2、soffusingthisrepresentationissuccessfulevenonup-sampledLIDARLIDARandcolorimageryforpedestriandetectioninthedata.contextofconvolutionalneuralnetworks(CNNs),whichhaveWethendesignandtrainseveralCNNsbasedon[11]recentlybecomestate-of-artformanyvisionproblems
3、.WethatoperateonbothRGBandHHAchannels,distinguishedincorporateLIDARbyup-samplingthepointcloudtoadensedepthmapandthenextractingthreefeaturesrepresentingbythepointatwhichtheRGBandHHAdatachannelsdifferentaspectsofthe3Dscene.Wethenusethosefeaturesasarefused
4、,andanalyzetheresultstoempiricallydetermineextraimagechannels.Specifically,weleveragerecentworkonthemodelthatachievesthebestresults.WeperformthisHHA[9](horizontaldisparity,heightaboveground,andangle)evaluationonatraining/validationsplitoftheKITTIurbanrep
5、resentations,adaptingthecodetoworkonup-sampledpedestriandetection,whichhasavailableLIDARdata.UsingLIDARratherthanMicrosoftKinectdepthmaps.Weshow,forthefirsttime,thatsucharepresentationisapplicablethisdata,wecontributeseveralfindings,namelythat:1)toup-samp
6、ledLIDARdata,despiteitssparsity.SinceCNNsusingHHAfeaturesandRGBimagesperformsbetterthanlearnadeephierarchyoffeaturerepresentations,wethenRGB-only,evenwithoutanyfine-tuningusinglargeRGBexplorethequestion:Atwhatlevelofrepresentationshouldwewebdata,2)fusing
7、RGBandHHAachievesthestrongestfusethisadditionalinformationwiththeoriginalRGBimageresultsifdonelate,butunderaparameterorcomputationalchannels?WeusetheKITTIpedestriandetectiondatasetforourexploration.Wefirstreplicatethefindingthatregion-CNNsbudgetisbestdone
8、attheearlytomiddlelayersofthe(R-CNNs)[8]canoutperformtheoriginalproposalmechanismhierarchicalrepresentation(representingmid-levelfeaturesusingonlyRGBimages,butonlyiffine-tuningisemployed.ratherthanloworhigh-levelones)withonlysmall
