预览加载中，请您耐心等待几秒...

一种多运动目标检测方法在FPGA上的实现.docx

预览
1/2
2/2

在线预览结束，喜欢就下载吧，查找使用更方便

下载提示 文本预览

如果您无法下载资料，请参考说明：

1、部分资料下载需要金币，请确保您的账户上有足够的金币

2、已购买过的文档，再次下载不重复扣费

3、资料包下载后请先用软件解压，在使用对应软件打开

一种多运动目标检测方法在FPGA上的实现 Abstract Multipleobjectdetectionisacriticalareaincomputervisionandimageprocessing.Asdeeplearningalgorithmscontinuetoadvance,objectdetectionaccuracyhasgreatlyimproved.However,thevastmajorityofthesealgorithmsrequiremassivecomputingresourcesandhighinferencelatency.ThispaperpresentsanovelmethodformultipleobjectdetectiononFPGA,whichenableshigh-speed,low-latencyobjectdetection.TheimplementationoftheproposedmethodwasachievedontheZynq-7000SoCdevelopmentboard,andtheexperimentalresultsshowedthattheproposedmethodsignificantlyimprovesthedetectionefficiencyandreduceslatencycomparedwithtraditionalmethods. Introduction Objectdetectionisacrucialtaskincomputervisionandimageprocessing.Recentadvancementsindeeplearninghavesignificantlyimprovedtheaccuracyofobjectdetectionalgorithms.However,mostoftheseapproachesrequiremassiveamountofcomputingresourcesandhighinferencelatency.Toaddresstheseissues,hardwareaccelerationtechniqueshavebeenappliedinobjectdetection. FPGAisapopularhardwareacceleratorthatsupportsparallelprocessingandlow-latencycomputation.Inthispaper,weproposeamultipleobjectdetectionmethodimplementedonFPGA.TheproposedmethodusesamodifiedYOLOv4-tinynetworkforobjectdetectionandoptimizationtechniquestoreducecomputationcomplexity.Thegoalistoimprovedetectionaccuracy,reducecomputationtime,andlowerlatency. Methodology Theproposedmethodconsistsofmultiplestages:inputdatapreprocessing,modifiedYOLOv4-tinynetwork,non-maximumsuppression,andpost-processing. InputDataPreprocessing Theinputdataispreprocessedtoreducetheamountofcomputationrequired.First,animageisresizedto416x416andconvertedtoYOLOformat.Next,theimageisdividedinto13x13grids,andeachgridpredictsfiveboundingboxes. YOLOv4-tinyNetwork Forobjectdetection,weuseamodifiedversionoftheYOLOv4-tinynetwork,whichhasfewerconvolutionlayersandasmallernumberoffiltersthantheoriginalnetwork.Themodifiednetworkconsistsofeightconvolutionallayers,followedbytwofullyconnectedlayersandasoftmaxactivationlayer.Thefinallayerofthenetworkpredictso