outputFolder=fullfile（tempdir，“潘达塞特”）;lidarURL = ['https://ssd.mathworks.com/金宝appsupportfiles/lidar/data/'．..“Pandaset_LidarData.tar.gz”];helperDownloadPandasetData (outputFolder lidarURL);

pretrainedNetURL = ['https://ssd.mathworks.com/金宝appsupportfiles/lidar/data/'．..“trainedPointPillarsPandasetNet.zip”]；doTraining=false；如果~doTraining helperDownloadPretrainedPointPillarsNet（outputFolder，pretrainedNetURL）；结束

路径= fullfile (outputFolder,激光雷达的）;lidardata = filedatastore（路径，“ReadFcn”@ (x) pcread (x));

gtPath = fullfile (outputFolder,“长方体”,“PandaSetLidarGroundTruth.mat”）;data =负载(gtPath“lidarGtLabels”)；Labels=timetable2table（data.lidarGtLabels）；boxLabels=Labels（：，2:3）；

图ptld = read(lidarData);ax = pcshow (ptCld.Location);集(ax,“XLim”50 [-50],“YLim”, 40 [-40]);变焦(ax, 2.5);轴从；

重置（Lidardata）;

xMin=0.0；% x轴上最小值。yMin = -39.68;% y轴最小值。zMin = -5.0;%沿Z轴的最小值。Xmax = 69.12;% x轴上的最大值。yMax=39.68；% y轴上的最大值。zMax = 5.0;沿z轴的％最大值。xStep = 0.16;沿x轴的分辨率。yStep = 0.16;沿y轴的分辨率。dsFactor=2.0；百分比下采样因子。%计算伪图像的尺寸。Xn = round(((xMax - xMin) / xStep));Yn = round(((yMax - yMin) / yStep));%定义柱提取参数。gridparams = {{xmin，ymin，zmin}，{xmax，ymax，zmax}，{xstep，ystep，dsfactor}，{xn，yn}};

[CROMPEDPOINTCLOUDOBJ，PROCESSEDLABELS] = CROMFRONTVIEWFROMLIDADDATA（．..lidarData、boxLabels gridParams);

pc=croppedPointCloudObj{1,1}；gtLabelsCar=processedLabels.Car{1}；gtLabelsTruck=processedLabels.Truck{1}；HelperDisplay3BoxesOverlaidPointCloud（pc.Location，gtLabelsCar，．..“绿色”gtLabelsTruck,“红色”,“裁剪点云”）;

重置（Lidardata）;

rng (1);shuffledIndices = randperm(大小(processedLabels, 1));idx = floor(0.7 * length(shuffledIndices));trainData = croppedPointCloudObj (shuffledIndices (1: idx):);testData = croppedPointCloudObj (shuffledIndices (idx + 1:结束):);trainLabels = processedLabels (shuffledIndices (1: idx):);testLabels = processedLabels (shuffledIndices (idx + 1:结束):);

writefile = true;dataLocation = fullfile (outputFolder,“InputData”）;[trainData, trainLabels] = saveptCldToPCD (trainData trainLabels,．..Datalocation，Writefiles）;

lds=文件数据存储（数据位置，“ReadFcn”@ (x) pcread (x));

bds = boxLabelDatastore (trainLabels);

cd =结合(摩门教,bds);

augData =阅读(cds);augptCld = augData {1};augLabels = augData {1,2};augClass = augData {1,3};labelsCar = augLabels (augClass = ='车':);labelsTruck = augLabels (augClass = ='卡车':);helperDisplay3DBoxesOverlaidPointCloud (augptCld。位置、labelsCar“绿色”,．..标签车，“红色”,“数据扩充前”）;

重置（CD）；

gtData = generateGTDataForAugmentation (trainData trainLabels);

samplesToAdd =结构('车'10，'卡车'10);cdsAugmented = transform(cds，@(x) groundTruthDataAugmenation(x,gtData,samplesToAdd));

cdsAugmented = transform(cdsAugmented，@(x) augmentData(x));

奥杜塔=读（Cdsaugmented）;augptCld = augData {1};augLabels = augData {1,2};augClass = augData {1,3};labelsCar = augLabels (augClass = ='车':);labelsTruck = augLabels (augClass = ='卡车':);helperDisplay3DBoxesOverlaidPointCloud (augptCld (: 1:3), labelsCar,“绿色”,．..标签车，“红色”,数据增加的）;

重置(cdsAugmented);

定义突出柱子的数量。P = 12000;定义每根柱子的点数。N=100；cdsTransformed=transform（cdsAugmented，@（x）createPillars（x，gridParams，P，N））；

锚具箱=计算锚点支柱（列车标签）；numAnchors=尺寸（锚箱，2）；classNames=trainLabels.Properties.VariableNames；numClasses=numel（类名）；

lgraph=点柱面网络（numAnchors、gridParams、P、N、NumClass）；

numEpochs = 60;miniBatchSize = 2;learningRate = 0.0002;learnRateDropPeriod = 15;learnRateDropFactor = 0.8;gradientDecayFactor = 0.9;squaredGradientDecayFactor = 0.999;trailingAvg = [];trailingAvgSq = [];

执行环境=“汽车”；如果canUseParallelPool dispatchInBackground = true;其他的Disparctinbackground = false;结束mbq=小型批处理队列(．..cdsTransformed 3．..“MiniBatchSize”miniBatchSize,．..“ouppantenvironment”，execultionenvironment，．..“MiniBatchFcn”@(特性指标,盒子,标签)．..helperCreateBatchData(特性、指标框、标签类名),．..“MiniBatchFormat”，[“SSCB”,“SSCB”,""]，．..“DispatchInBackground”,真正的);

如果溺爱将层图转换为dlnetwork。net=dlnetwork（lgraph）；%初始化情节。无花果=图;lossPlotter = helperConfigureTrainingProgressPlotter(图);迭代= 0;%自定义训练循环。对于epoch = 1：numepochs%重置数据存储。重置(兆贝可);而(hasdata(mbq)) iteration = iteration + 1;％读取数据批次。[Pillarfeatures，Pillarindices，Boxlabels] =下一个（MBQ）;%使用dlfeval评估模型梯度和损失％和MapeStradients函数。[梯度、损失、状态]=dlfeval（@modelGradients，net，．..Pillarfeatures，Pillarindices，Booklabels，Gridparams，．..anchorBoxes、numClasses executionEnvironment);%如果是NaN值，则不更新网络可学习参数%以梯度或损失值表示。如果helperCheckForNaN(梯度、损失)继续；结束％更新DLNetwork的状态参数。网。=状态;％使用adam更新网络学习参数%的优化器。[net.Learnables，trailingAvg，trailingAvgSq]=．..adamupdate(净。可学的、渐变trailingAvg,．..跟踪AVGSQ，迭代，学习率，．..梯度Dayfactor，SquaredgradientDecayfactor）;%使用新点更新训练图。addpoints (lossPlotter、迭代、双(收集(extractdata(损失))));标题(“训练时代”+纪元+“的”+ numEpochs);drawnow;结束%每次learnRateDropPeriod后更新学习率。如果mod(epoch,learnRateDropPeriod) == 0 learningRate = learningRate * learnRateDropFactor;结束结束其他的preTrainedMATFile=fullfile（outputFolder，“trainedPointPillarsPandasetNet.mat”）;pretrainedNetwork =负载(preTrainedMATFile,“净”）;网= pretrainedNetwork.net;结束

45 ptCloud = testData {1};gtLabels = testLabels(45岁);%generatePointPillarDetections功能检测%包围框，并为给定的点云得分。confidenceThreshold = 0.5;overlapThreshold = 0.1;(盒、分数、标签)= generatePointPillarDetections(净、ptCloud anchorBoxes,．..Gridparams，ClassNames，ConcidenceThreshold，Overlapthreshold，P，N，．..执行环境）；boxlabelsCar=盒子（标签'=='车':);boxlabelsTruck(标签' = = =框'卡车':);%在点云上显示预测。Helperdisplay3dboxesoverlaidPointCloud（PTCloud.Location，Boxlabelscar，“绿色”,．..Boxlabelstruck，“红色”,“预测边界框”）;

numInputs=numel（测试数据）；%从长方体标签生成旋转的矩形。bds = boxLabelDatastore (testLabels);groundTruthData = transform(bds，@(x) createRotRect(x));%设置阈值。nmsPositiveIoUThreshold = 0.5;confidenceThreshold = 0.25;overlapThreshold = 0.1;％SET NUMSAMPLESTOTEST到NUMInputs以评估整个模型％测试数据集。numSamplesToTest = 50;detectionResults =表(“大小”numSamplesToTest [3],．..'variabletypes'，{“细胞”,“细胞”,“细胞”}，．..“VariableNames”，{“盒子”,“分数”,“标签”})；对于num=1:numSamplesToTest ptCloud=testData{num，1}；[box，score，labels]=generatePointPillarDetections（net，ptCloud，anchorBoxes，．..Gridparams，ClassNames，ConfidenceThreshold，Orectapthreshold，．..P, N, executionEnvironment);%将检测到的框转换为旋转矩形格式。如果detectionResults ~ isempty(盒)。盒子{num} =盒(:,(1、2、4、5、7]);其他的detectionResults。盒子{num} =框;结束detectionResults.Scores{num}=分数；detectionResults.Labels{num}=标签；结束度量=评估的检测AOS（检测结果、地面数据、，．..nmsPositiveIoUThreshold);disp(指标(:,1:2))

函数(渐变、损失、状态)= modelGradients (pillarFeatures净,．..pillarIndices、boxLabels gridParams anchorBoxes,．..numClasses,executionEnvironment) numAnchors = size(锚盒，2);%从网络中提取预测。ypredictions = cell（size（net.outputnames））;[ypredictions {：}，州] =前进（净，柱，支柱）;%从地面真实数据生成预测目标。YTargets = generatePointPillarTargets (YPredictions boxLabels pillarIndices,．..Gridparams，锚盒，numcrasses）;ytargets = Cellfun（@ Dlarray，Ytargets，'统一输出'、假);如果（execultenvironment ==.“汽车”&& canusegpu）||executionenvironment ==.“GPU”YTargets = cellfun(@ gpuArray,YTargets，'统一输出'、假);结束[盎格鲁人,occLoss locLoss、szLoss hdLoss, clfLoss] =．..computePointPillarLoss（预测、目标、网格参数、，．..numClasses numAnchors);%计算总损失。loss=angLoss+occLoss+locLoss+szLoss+hdLoss+clfLoss；％在丢失方面计算了学习栏的渐变。梯度= dlgradient(损失、net.Learnables);结束函数[pillarFeatures、PillarDice、labels]=helperCreateBatchData(．..特性、指标groundTruthBoxes groundTruthClasses,类名)%沿批次尺寸组合的返回柱特征和索引沿标签中的批量维度连接的％和边界框。沿批处理维度的％连接功能和指标。pillarFeatures =猫(4特性{:1});pillarIndices =猫(4,指标{:1});%从类名获取类id。一会= repmat({分类(类名)},大小(groundTruthClasses));[~, classIndices] = cellfun (@ (a, b) ismember (a、b), groundTruthClasses,．..一会,'统一输出'、假);％附加类索引并创建单个响应数组。= cellfun(@(bbox,classid) [bbox,classid]，．..土拨鼠箱，分类，'统一输出'，false）；len=max（cellfun（@（x）大小（x，1），组合响应））；paddedboxes=cellfun（@（v）焊盘阵列（v，[len大小（v，1），0]，0，“职位”)，．..组合响应，'统一输出'、假);标签=猫(4,paddedBBoxes {: 1});结束函数helperDownloadPandasetData (outputFolder lidarURL)%将数据集从给定URL下载到输出文件夹。lidarDataTarFile = fullfile (outputFolder,“Pandaset_LidarData.tar.gz”）;如果〜存在（LidardataTarfile，'文件'mkdir (outputFolder);disp (“正在下载PandaSet激光雷达驱动数据(5.2 GB)……”）;Websave（LidardataTarfile，Lidarurl）;Untar（LidardataTarfile，Outputfolder）;结束%解压文件。如果(~存在(fullfile (outputFolder激光雷达的)，“dir”））．..&&（〜存在（fullfile（outputfile，“长方体”)，“dir”)解压(lidarDataTarFile outputFolder);结束结束函数helperDownloadPretrainedPointPillarsNet (outputFolder pretrainedNetURL)%下载预先训练过的PointPillars网络。preTrainedMATFile=fullfile（outputFolder，“trainedPointPillarsPandasetNet.mat”)；preTrainedZipFile=fullfile（outputFolder，“trainedPointPillarsPandasetNet.zip”）;如果~exist（preTrainedMATFile，'文件')如果~exist（预训练Zipfile，'文件') disp ('下载掠夺探测器（8.4 MB）......'）;websave (preTrainedZipFile pretrainedNetURL);结束解压（preTrainedZipFile，outputFolder）；结束结束函数lossPlotter = helperConfigureTrainingProgressPlotter (f)%配置各种损耗的训练进度表。图(f);clf ylabel (“全损”)；xlabel(“迭代”）;lossPlotter = animatedline;结束函数retValue = helperCheckForNaN(梯度、损失)%已知最后一个卷积头'occupancy|conv2d'包含nan%的梯度。这个函数检查梯度值是否包含％纳尼。如果NANS将其他卷积头值添加到条件下%出现在梯度中。gradValue = gradients.Value(梯度。层= =“入住率| conv2d”）＆．..(渐变。参数= =“偏见”))；如果（总和（isnan（提取数据（损失）），“所有”) > 0) ||．..(和(isnan (extractdata (gradValue {1 1})),“所有”)>0）retValue=true；其他的etvalue = false;结束结束函数Helperdisplay3dboxesoverlaidPointPlod（PTCLD，Labelscar，Carcolor，．..LabelStruck，TruckColor，TitleFutigure）%显示具有不同颜色边界框的点云，以显示不同颜色的点云％班级。数字;AX = PCSHOW（PTCLD）;showshape（'长方体'labelsCar,'父母'ax,'不透明度',0.1,．..“颜色”，arcolor，“线宽”，0.5）；保持在；showshape（'长方体'，labelstruck，'父母'ax,'不透明度',0.1,．..“颜色”，卡车色，“线宽”, 0.5);标题(titleForFigure);变焦(ax, 1.5);结束