downloadDirectory = tempdir;datapath公司= downloadSydneyUrbanObjects (downloadDirectory);

foldsTrain=1:3；foldsVal=4；dsTrain=loadSydneyUrbanObjectsData（数据路径，foldsTrain）；dsVal=loadSydneyUrbanObjectsData（数据路径，foldsVal）；

data =阅读(dsTrain);ptCloud数据= {1};标签={1,2}数据;图pcshow（ptcloud.location，[0 1 0]，“MarkerSize”，40，“VerticalAxisDir”，“向下”)包含(“X”) ylabel (“Y”）Zlabel（“Z”标题(标签)

dsLabelCounts = transform(dsTrain，@(data){data{2} data{1}.Count}); / /显示数据labelCounts = readall (dsLabelCounts);标签= vertcat (labelCounts {: 1});数量= vertcat (labelCounts {: 2});

图直方图(标签)

rng（0）[G，类]=FindGroup（标签）；numObservations=splitapply（@numel，labels，G）；desiredNumObservationsPerClass=最大值（numObservations）；files=splitapply（@（x）{randReplicateFiles（x，desiredNumObservationsPerClass）}，dsTrain.files，G）；files=vertcat（files{:}）；dsTrain.Files=文件；

dsTrain =变换(dsTrain @augmentPointCloud);

data =预览(dsTrain);ptCloud数据= {1};标签={1,2}数据;图pcshow（ptcloud.location，[0 1 0]，“MarkerSize”，40，“VerticalAxisDir”，“向下”)包含(“X”) ylabel (“Y”）Zlabel（“Z”标题(标签)

minPointCount=splitapply（@min，counts，G）；maxPointCount=splitapply（@max，counts，G）；meanPointCount=splitapply（@x）轮（平均数（x）），计数，G）；stats=table（类，numObservations，minPointCount，maxPointCount，meanPointCount）

numPoints = 1024;dsTrain =变换(dsTrain @(数据)selectPoints(数据、numPoints));dsVal =变换(dsVal @(数据)selectPoints(数据、numPoints));

dsTrain =变换(dsTrain @preprocessPointCloud);dsVal =变换(dsVal @preprocessPointCloud);

data =预览(dsTrain);图pcshow(data{1,1}，[0 1 0]，“MarkerSize”，40，“VerticalAxisDir”，“向下”）;包含(“X”) ylabel (“Y”）Zlabel（“Z”{1,2})标题(数据)

inputChannelSize = 3;hiddenChannelSize1 = [64128];hiddenChannelSize2 = 256;(参数。在putTransform, state.InputTransform] = initializeTransform(inputChannelSize,hiddenChannelSize1,hiddenChannelSize2);

inputChannelSize = 3;hiddenchannelsize = [64 64];[parameters.sharedmlp1，state.sharedmlp1] = initializeSharedmlp（InputChannelsize，HiddenChannelsize）;

inputChannelSize = 64;hiddenChannelSize1 = [64128];hiddenChannelSize2 = 256;(参数。FeatureTransform,状态。FeatureTransform] = initializeTransform (inputChannelSize, hiddenChannelSize hiddenChannelSize2);

inputChannelSize=64；hiddenChannelSize=64；[parameters.SharedMLP2，state.SharedMLP2]=初始化SharedMLP（inputChannelSize，hiddenChannelSize）；

inputChannelSize = 64;hiddenChannelSize = [512256];numClasses =元素个数(类);(参数。ClassificationMLP,状态。ClassificationMLP] = initializeClassificationMLP (inputChannelSize, hiddenChannelSize numClasses);

numEpochs = 10;learnRate = 0.002;miniBatchSize = 128;l2Regularization = 0.01;learnRateDropPeriod = 15;learnRateDropFactor = 0.5;

gradientDecayFactor = 0.9;squaredGradientDecayFactor = 0.999;

avgGradients = [];avgSquaredGradients = [];

doTraining = false;如果doTraining%创建一个minibatchqueue来批处理来自培训和验证的数据%数据存储。使用batchData函数，列在%例如，批处理点云数据并对标签进行热编码%的数据。numoutsfromdsread=2；mbqTrain=minibatchqueue（dsTrain，numoutsfromdsread，．..“迷你atchsize”，小批量，．..“MiniBatchFcn”@batchData,．..“MiniBatchFormat”，[“SCSB”“公元前”])；mbqVal=minibatchqueue（dsVal，numoutsfromdsread，．..“迷你atchsize”，小批量，．..“MiniBatchFcn”@batchData,．..“MiniBatchFormat”，[“SCSB”“公元前”]）;%使用configureTrainingProgressPlot函数，列在文件末尾%示例，初始化训练进度图以显示训练%损失，培训准确性和验证准确性。[lossPlotter, trainAccPlotter,valAccPlotter] = initializeTrainingProgressPlot;numClasses =元素个数(类);迭代= 0;开始=抽搐;为时代= 1:numEpochs%每个epoch Shuffle data。洗牌(mbqTrain);%遍历数据集。尽管hasdata(mbqTrain) iteration = iteration + 1;%读取下一批训练数据。[XTrain, YTrain] = next(mbqTrain);使用dlfeval和% modelGradients函数。[gradient, loss, state, acc] = dlfeval(@modelGradients,XTrain,YTrain,parameters,state);% L2正规化。梯度= dlupdate（@（g，p）g + l2regularization * p，梯度，参数）;%使用Adam优化器更新网络参数。[parameter, avgGradients, avgSquaredGradients] = adamupdate(parameter, gradients, avgSquaredGradients)．..avgGradients avgSquaredGradients,迭代,．..学习，梯度dayfactor，squaredgradientdecayfactor）;%更新培训进度。d =持续时间（0,0，toc（start），“格式”，“hh: mm: ss”）;标题(lossPlotter。父母,”时代:“+时代+”,过去:“+字符串（d））addpoints（lockplotter，reateration，double（收集（提取数据（丢失））））等分点（训练料accplotter，迭代，acc）;drawn结束%根据验证数据评估模型。Cmat =稀疏（Numcrasses，Numcrasses）;尽管hasdata (mbqVal)读取下一批验证数据。[XVal, YVal] = next(mbqVal);%计算标签预测。isTraining = false;YPred = pointnetClassifier (XVal、参数状态,isTraining);%选择得分最高的预测作为类标签% XTest。[~, YValLabel] = max (YVal [], 1);[~, YPredLabel] = max (YPred [], 1);收集混淆指标。cmat = aggreateConfusionMetric (cmat YValLabel YPredLabel);结束%以平均分类精度更新训练进度图。acc=总和（诊断（cmat））。/总和（cmat，“所有”）;addpoints (valAccPlotter迭代,acc);%更新学习率。如果mod（epoch，sechnratedropperiod）== 0 snownrate = lavingrate * shearchratedropfactor;结束%重置训练和验证数据队列。重置（MBQTrain）;重置（MBQVAL）;结束别的％下载返回售价模型参数，模型状态和验证%的结果。pretrainedURL =“https://ssd.mathworks.com/金宝appsupportfiles/vision/data/pointnetSydneyUrbanObjects.zip”；pretrainedResults = downloadPretrainedPointNet (pretrainedURL);参数= pretrainedResults.parameters;状态= pretrainedResults.state;cmat = pretrainedResults.cmat;%如果可能的话，将模型参数移动到GPU并转换为dlarray。参数= prepareForPrediction(参数,@ (x) dlarray (toDevice (x, canUseGPU)));%如果可能，将模型状态移到GPU。状态= prepareForPrediction(状态,@ (x) toDevice (x, canUseGPU));结束

Figure chart = confusionchart(cma,classes);

acc=总和（诊断（cmat））。/总和（cmat，“所有”）

ptCloud = pcread (“car.pcd”）;X = preprocessPointCloud (ptCloud);dlX = dlarray (X {1},“SCSB”）;

YPred = pointnetClassifier (dlX、参数、状态、错误);[~, classIdx] = max (YPred [], 1);

图pcshow（ptcloud.location，[0 1 0]，“MarkerSize”，40，“VerticalAxisDir”，“向下”标题(类(classIdx))

功能[gradient, loss, state, acc] = modelGradients(X,Y,parameters,state)%执行模型函数。isTraining = true;[YPred、州dlT] = pointnetClassifier (X,参数,状态,isTraining);％添加正则化术语以确保功能变换矩阵是%大致正交的。K =大小(dlT, 1);B = size(dlT, 4);我= repelem(眼(K), 1, 1, 1, B);dlI = dlarray(我“SSCB”）;treg = mse(dlI,pagemtimes(dlT,permute(dlT，[2 1 3 4])));系数= 0.001;％计算丢失。loss = crossentropy(YPred,Y) + factor*treg;％在丢失上计算参数梯度。gradient = dlgradient(loss, parameters);％计算训练精度度量。[~,欧美]= max (Y, [], 1);[~, YPred] = max (YPred [], 1);acc = gather(extractdata(sum(YTest == YPred)./numel(YTest)));结束

功能(海底、州dlT) = pointnetClassifier (dlX、参数、状态、isTraining)%调用PointNet编码器。[dlY，state，dlT]=pointnetEncoder（dlX，参数，状态，正在训练）；%调用分类器。p = parameters.ClassificationMLP.Perceptron;s = state.ClassificationMLP.Perceptron;为k = 1:元素个数(p)[海底,s (k)] =感知器(海底,p (k), s (k), isTraining);%如果进行训练，则应用倒退的概率为0.3。如果约束概率= 0.3;dropoutScaleFactor = 1 -概率;dropoutMask = (rand(size(dlY)，“喜欢”， d) > probability) / dropoutScaleFactor;海底=海底。* dropoutMask;结束结束state.ClassificationMLP.Perceptron = s;%应用最终的完全连接和softmax操作。权重=parameters.ClassificationMLP.FC.weights；bias=parameters.ClassificationMLP.FC.bias；dlY=完全连接（dlY，权重，bias）；dlY=softmax（dlY）；结束

功能(海底,状态,T) = pointnetEncoder (dlX、参数、状态、isTraining)％输入变换。[dlY，state.InputTransform]=数据转换（dlX，parameters.InputTransform，state.InputTransform，isTraining）；%共享延时。(海底,state.SharedMLP1。感知器]= sharedMLP(海底,parameters.SharedMLP1.Perceptron, state.SharedMLP1.Perceptron isTraining);%功能变换。(海底,状态。FeatureTransform T] = dataTransform(海底,parameters.FeatureTransform, state.FeatureTransform isTraining);%共享延时。(海底,state.SharedMLP2。感知器]= sharedMLP(海底,parameters.SharedMLP2.Perceptron, state.SharedMLP2.Perceptron isTraining);%最大操作。海底= max(海底,[],1);结束

功能[dlY,state] = sharedMLP(dlX,parameters,state, is拴住)为k = 1:numel(parameters) [dlY, state(k)] =感知机(dlY,parameters(k)，state(k)， istrain);结束结束

功能(海底、州)=感知器(dlX、参数、状态、isTraining)%卷积。W = parameters.Conv.Weights;B = parameters.Conv.Bias;海底= dlconv (dlX, W, B);%批正常化。在培训时更新批标准化状态。抵消= parameters.BatchNorm.Offset;规模= parameters.BatchNorm.Scale;trainedMean = state.BatchNorm.TrainedMean;trainedVariance = state.BatchNorm.TrainedVariance;如果isTraining[Dy，trainedMean，trainedVariance]=批次标准（Dy，偏移量，比例，trainedMean，trainedVariance）；%更新状态。state.BatchNorm.TrainedMean = trainedMean;state.BatchNorm.TrainedVariance = trainedVariance;别的海底= batchnorm(海底、抵消、规模、trainedMean trainedVariance);结束% ReLU。海底= relu(海底);结束

功能(海底,状态,T) = dataTransform (dlX、参数、状态、isTraining)%共享延时。(海底,state.Block1。感知器]= sharedMLP (dlX、parameters.Block1.Perceptron state.Block1.Perceptron, isTraining);%最大操作。海底= max(海底,[],1);%共享延时。(海底,state.Block2。感知器]= sharedMLP(海底,parameters.Block2.Perceptron, state.Block2.Perceptron isTraining);%变换网(T-Net)。将上次完全连接的操作(如W*X)应用到%预测变换矩阵T。海底=挤压(海底);%N×BT =参数。变换* stripdims(海底);% K ^ 2-by-B%将T重塑为一个方阵。K =√大小(T, 1));T =重塑(T, K, K, 1, []);% k + k + bT = T +眼(K);%使用批处理矩阵乘法输入dlX。[C，B]=尺寸（dlX，[34]）；%[M 1 K B]dlX =重塑(dlX, [], C, 1 B);(b) n ' m / bY = pagemtimes (dlX T);海底= dlarray (Y,“SCSB”）;结束

功能[params,state] = initializeTransform(inputChannelSize,block1,block2)Block1] = initializeSharedMLP (inputChannelSize Block1);[params.Block2状态。Block2] = initializeSharedMLP (block1(结束),Block2);%变换矩阵的参数。参数个数。变换= dlarray (0 (inputChannelSize ^ 2, block2(结束)));结束

功能[params，state] = initializesharedmlp（inputchcannelsize，hiddenhennelsize）权重= initializewightshe（[1 1 inputchcannelsize hiddenhannelsize（1）]）;BIAS = Zeros（HiddenChannelsize（1），1，“单身”）;p.Conv.Weights = dlarray(重量);p.Conv.Bias = dlarray(偏差);p.BatchNorm.Offset = dlarray (0 (hiddenChannelSize (1), 1,“单身”)); p、 BatchNorm.Scale=dlarray（个数（hiddenChannelSize（1），1，“单身”））;s.BatchNorm.TrainedMean = 0 (hiddenChannelSize (1), 1,“单身”）;s.BatchNorm.TrainedVariance = 1 (hiddenChannelSize (1), 1,“单身”）;params.Perceptron (1) = p;state.Perceptron (1) = s;为k = 2:numel(hiddenChannelSize) weights = initializeWeightsHe([1 1 hiddenChannelSize(k-1))]);偏见= 0 (hiddenChannelSize (k), 1,“单身”）;p.Conv.Weights = dlarray(重量);p.Conv.Bias = dlarray(偏差);p.BatchNorm.Offset = dlarray (0 (hiddenChannelSize (k), 1,“单身”))；p.BatchNorm.Scale=dlarray（个数（hiddenChannelSize，k），1，“单身”））;s.BatchNorm.TrainedMean = 0 (hiddenChannelSize (k), 1,“单身”)；s.BatchNorm.TrainedVariance=个（hiddenChannelSize（k），1，“单身”）;params.Perceptron (k) = p;state.Perceptron (k) = s;结束结束

功能[params,state] = initializeSharedMLP(inputChannelSize,hiddenChannelSize,numClasses);weights = initializeWeightsGaussian([numClasses hiddenChannelSize(end)]);偏见= 0 (numClasses 1“单身”）;params.fc.weights = dlarray（重量）;params.fc.bias = dlarray（偏见）;结束

功能x = initializeWeightsHe(sz) fanIn = prod(sz(1:3));stddev =√2 /扇入);X = stddev .* randn(sz); / /指定一个数组结束

功能x = initializeWeightsGaussian(sz)“单身”) * 0.01;结束

功能数据=预处理点云（数据）如果〜Iscell（数据）数据= {data};结束numObservations =大小(数据,1);为i = 1：numobservations百分数在0到1之间。xlim ={我1}.XLimits数据;ylim ={我1}.YLimits数据;zlim ={我1}.ZLimits数据;xyzMin = [xlim(1) ylim(1) zlim(1)];drawtext (xyzDiff = [diff(xlim) diff(ylim) diff(zlim)]];{我1}=(数据{1}我。位置- xyzMin) ./ xyzDiff;结束结束

功能numPoints data = selectPoints(数据)%通过向下采样或复制选择所需的点数%点云数据。numObservations =大小(数据,1);为i = 1:numObservations ptCloud = data{i,1};如果ptCloud。Count > numPoints percentage = numPoints/ptCloud.Count;数据{1}= pcdownsample (ptCloud,“随机”、比例);别的replicationFactor=ceil（numPoints/ptCloud.Count）；ind=repmat（1:ptCloud.Count，1，replicationFactor）；data{i，1}=select（ptCloud，ind（1:numPoints））；结束结束结束

功能data=augmentPointCloud（data）numObservations=size（data，1）；为i = 1:numObservations ptCloud = data{i,1};%旋转点云的“上轴”，也就是这个数据集的Z轴。tform = randomAffine3d (．..“XReflection”,真的,．..“yreflection”,真的,．..“旋转”, @randomRotationAboutZ);ptCloud = pctransform (ptCloud tform);％随机丢弃了分数的30％。如果rand > 0.5 ptCloud = pcdownsample(ptCloud，“随机”, 0.3);结束如果兰德> 0.5%抖动的点位置与高斯噪声的平均值为0和通过创建一个随机位移场，得到0.02的标准偏差。D = 0.02 * randn(size(ptCloud.Location));ptCloud = pctransform (ptCloud D);结束数据{1}= ptCloud;结束结束功能[旋转轴，theta] = randomRotationAboutZ()旋转轴= [0 0 1];θ= 2 *π*兰特;结束

功能cmat = aggreateConfusionMetric(cmat,YTest,YPred) YTest = gather(extractdata(YTest));YPred =收集(extractdata (YPred));[m, n] =大小(cmat);cmat = cmat + full(稀疏(YTest,YPred,1,m,n));结束

功能[绘图仪,trainAccPlotter valAccPlotter] = initializeTrainingProgressPlot ()%绘制损失、培训准确度和验证准确度。图形%损失情节Subplot (2,1,1) = animatedline;包含(“迭代”) ylabel (“损失”）％准确性图subplot(2,1,2) trainAccPlotter = animatedline(“颜色”，“b”）;valAccPlotter = animatedline (“颜色”，“g”）;传奇(“训练的准确性”，“验证准确性”，“位置”，“西北”）;包含(“迭代”) ylabel (“准确性”）结束

功能n = nummel (files); / /文件的数量印第安纳州=兰迪(n, numDesired, 1);文件=文件(印第安纳州);结束

功能datapath公司= downloadSydneyUrbanObjects (dataLoc)如果nargin == 0 dataLoc = pwd;结束dataLoc=字符串（dataLoc）；url=“http://www.acfr.usyd.edu.au/papers/data/”；名称=“sydney-urban-objects-dataset.tar.gz”；datapath公司= fullfile (dataLoc,“sydney-urban-objects-dataset”）;如果~存在(datapath公司,“dir”）disp（“正在下载悉尼城市物体数据集……”）;解压(url +名字,dataLoc);结束结束

功能ds = loadSydneyUrbanObjectsData (datapath公司、折叠)如果nargin == 0.返回；结束如果Nargin < 2 fold = 1:4;结束datapath公司=字符串(datapath公司);路径= fullfile (datapath公司,'物体', filesep);%向数据存储中添加折叠。foldNames {1} = importdata (fullfile (datapath公司,“折叠”，“fold0.txt”））;foldNames {2} = importdata (fullfile (datapath公司,“折叠”，“fold1.txt”））;foldNames {3} = importdata (fullfile (datapath公司,“折叠”，“fold2.txt”))；foldNames{4}=importdata（完整文件（数据路径，“折叠”，“fold3.txt”））;名称= foldNames(折叠);名称= vertcat(名字{:});fullFilenames = append(路径、名称);ds = fileDatastore (fullFilenames,“ReadFcn”，@ExtrainingData，“FileExtensions”，“。斌”）;结束

功能[X,Y] = batchData(ptCloud,labels) X = cat(4,ptCloud{:});标签=猫(1、标签{:});Y = onehotencode(标签,2);结束

功能dataOut=extractTrainingData（fname）[pointData，intensity]=readbin（fname）；[~，name]=fileparts（fname）；name=string（name）；name=extractBefore（name，“。”)；名称=替换（名称，“_”，＇ ＇）;labelNames = [“四轮驱动”，“建筑”，“公共汽车”，“汽车”，“行人”，“支柱”，．..“极”，“红绿灯”，“交通标志”，“树”，“卡车”，“树干”，“哑巴”，“范”];标签=分类(名称、labelNames);dataOut = {pointCloud (pointData'强度'，强度），标签}；结束

功能(pointData、强度)= readbin(帧)% readbin从悉尼城市目标二进制中读取点和强度数据%的文件。% name = ['t'，'intensity'，'id'，…]% ' x ', ' y ', ' z ',……%“方位角”、“距离”、“pid”]％% formats = ['int64'， 'uint8'， 'uint8'，…% ` float32 `， ` float32 `， ` float32 `，…% ` float32 `， ` float32 `， ` int32 `]fid = fopen(帧,“r”）;c = oncleanup（@（）fclose（fid））;FSEEK（FID，10，-1）;从开始移动到第一个X点位置10个字节x = fread（fid，inf，“单身”, 30);fseek (fid, 14日1);Y =从文件中读(fid,正无穷,“单身”, 30);fseek(支撑材,18岁,1);Z =从文件中读(fid,正无穷,“单身”, 30);fseek(支撑材8 1);强度=从文件中读(fid,正无穷,“uint8”33);pointData = [X, Y, Z];结束

功能data = downloadPretrainedPointNet (pretrainedURL)%下载和加载一个预先训练的pointnet模型。如果~ ('pointnetSydneyUrbanObjects.mat'，“文件”）如果~ (“pointnetSydneyUrbanObjects.zip”，“文件”）disp（“下载预训练探测器(5mb)……”)；韦伯斯韦(“pointnetSydneyUrbanObjects.zip”，pretrowsurl）;结束解压(“pointnetSydneyUrbanObjects.zip”）;结束data =负载(“pointnetSydneyUrbanObjects.mat”）;结束

功能p = prepareForPrediction (p, fcn)为i = 1:元素个数(p) p (i) = structfun (@ (x)调用(fcn x)、p (i),“UniformOutput”，0）;结束功能data =调用(fcn、数据)如果isstruct（数据）数据=准备预测（数据，fcn）；别的data = fcn(数据);结束结束结束％将数据移动到GPU。功能x = toDevice (x, useGPU)如果useGPU x = gpuArray(x);结束结束