激光雷达目标检测使用Complex-YOLO v4网络

outputFolder = fullfile (tempdir,“Pandaset”);lidarURL = [“https://ssd.mathworks.com/万博1manbetxsupportfiles/lidar/data/”…“Pandaset_LidarData.tar.gz”];helperDownloadPandasetData (outputFolder lidarURL);

modelName =“tiny-complex-yolov4-pandaset”;mdl = downloadPretrainedComplexYOLOv4 (modelName);网= mdl.net;

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”);标签= timetable2table (data.lidarGtLabels);boxLabels =标签(:,2:结束);

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

xMin = -25.0;xMax = 25.0;yMin = 0.0;yMax = 50.0;zMin = -7.0;zMax = 15.0;

bevHeight = 608;bevWidth = 608;

gridW = (yMax - yMin) / bevWidth;gridH = (xMax - xMin) / bevHeight;

yMin gridParams = {{xMin, xMax, yMax, zMin, zMax}, {bevWidth, bevHeight}, {gridW, gridH}};

writefile = true;如果writefile transformPCtoBEV (lidarData boxLabels、gridParams outputFolder);结束

dataPath公司= fullfile (outputFolder,“BEVImages”);imd = imageDatastore (dataPath公司);

labelPath = fullfile (outputFolder,“长方体”,“BEVGroundTruthLabels.mat”);负载(labelPath“processedLabels”);建筑物= boxLabelDatastore (processedLabels);

(imd,建筑物)= removeEmptyData (imd,建筑物);

rng (0);shuffledIndices = randperm(大小(imds.Files, 1));地板idx =(0.6 *长度(shuffledIndices));

imdsTrain =子集(imd, shuffledIndices (1: idx));imdsTest =子集(imd, shuffledIndices (idx + 1:结束);

bldsTrain =子集(建筑物,shuffledIndices (1: idx));bldsTest =子集(建筑物,shuffledIndices (idx + 1:结束);

trainData =结合(imdsTrain bldsTrain);testData =结合(imdsTest bldsTest);

validateInputDataComplexYOLOv4 (trainData);validateInputDataComplexYOLOv4 (testData);

networkInputSize = (608 608 3);isRotRect = true;preprocessedTrainingData =变换(trainData @(数据)preprocessData(数据、networkInputSize isRotRect));

data =阅读(preprocessedTrainingData);

我={1 1}数据;bbox ={1,2}数据;标签={1,3}数据;helperDisplayBoxes (bbox,标签);

重置(preprocessedTrainingData);

rng (0) isRotRect = false;trainingDataForEstimation =变换(trainData @(数据)preprocessData(数据、networkInputSize isRotRect));numAnchors = 6;[anchorBoxes, meanIoU] = estimateAnchorBoxes (trainingDataForEstimation numAnchors)

%培训指定要使用的类名。一会= {“汽车”“卡车”“Pedestrain”};[净,networkOutputs anchorBoxes] = configureComplexYOLOv4(网,一会,anchorBoxes modelName);

maxEpochs = 90;miniBatchSize = 8;learningRate = 0.001;warmupPeriod = 1000;l2Regularization = 0.001;penaltyThreshold = 0.5;速度= [];

如果canUseParallelPool dispatchInBackground = true;其他的dispatchInBackground = false;结束mbqTrain = minibatchqueue (preprocessedTrainingData 2…“MiniBatchSize”miniBatchSize,…“MiniBatchFcn”@(图片,盒子,标签)createBatchData(图片,盒子,标签,类名),…“MiniBatchFormat”,(“SSCB”,”“),…“DispatchInBackground”dispatchInBackground,…“OutputCast”,(”“,“替身”]);

doTraining = false;如果doTraining迭代= 0;%为学习速率创建次要情节和mini-batch损失。无花果=图;[lossPlotter, learningRatePlotter] = configureTrainingProgressPlotter(图);%的定制培训循环。为时代= 1:maxEpochs重置(mbqTrain);洗牌(mbqTrain);而(hasdata (mbqTrain)) =迭代+ 1;[XTrain, YTrain] =下一个(mbqTrain);%计算模型使用dlfeval和梯度和损失% modelGradients函数。(渐变、州lossInfo) = dlfeval (@modelGradients,净、XTrain YTrain, anchorBoxes, penaltyThreshold, networkOutputs);%应用L2正规化。梯度= dlupdate (@ (g, w) g + l2Regularization * w,渐变net.Learnables);%确定当前的学习速率值。currentLR = piecewiseLearningRateWithWarmup(迭代,时代、learningRate warmupPeriod, maxEpochs);%更新使用个网络可学的参数优化。(净、速度)= sgdmupdate(净、渐变速度,currentLR);%更新dlnetwork的状态参数。网。=状态;%显示进展。如果国防部(迭代,10)= = 1 displayLossInfo(时代、迭代、currentLR lossInfo);结束%绘制新的点更新培训。updatePlots (lossPlotter learningRatePlotter、迭代、currentLR lossInfo.totalLoss);结束结束其他的网= mdl.net;anchorBoxes = mdl.anchorBoxes;结束

%创建一个表来保存边界框,返回的分数,和标签%的探测器。结果=表(“大小”[0 3],…“VariableTypes”,{“细胞”,“细胞”,“细胞”},…“VariableNames”,{“盒子”,“分数”,“标签”});%在图像上运行探测器测试集和收集结果。重置(testData)而hasdata (testData)%读取数据存储和图像。data =阅读(testData);形象={1 1}数据;%运行探测器。executionEnvironment =“汽车”;(bboxes、分数、标签)= detectComplexYOLOv4(净、图像、anchorBoxes一会,executionEnvironment);%收集结果。台=表({bboxes},{},{}标签,“VariableNames”,{“盒子”,“分数”,“标签”});结果=[结果;(资源);结束%评估对象探测器使用的平均精度指标。testData指标= evaluateDetectionAOS(结果)

%读取数据存储。重置(testData)数据=阅读(testData);%的形象。我={1 1}数据;%运行探测器。executionEnvironment =“汽车”;(bboxes、分数、标签)= detectComplexYOLOv4 (anchorBoxes,净,我一会,executionEnvironment);%显示输出。图helperDisplayBoxes(我bboxes标签);

lidarTestData =子集(lidarData shuffledIndices (idx + 1:结束);ptCld =阅读(lidarTestData);[ptCldOut, bboxCuboid] = transferbboxToPointCloud (bboxes、gridParams ptCld);helperDisplayBoxes (ptCldOut、bboxCuboid、标签);

函数(梯度、状态信息)= modelGradients(净,XTrain YTrain,锚、penaltyThreshold networkOutputs) inputImageSize =大小(XTrain 1:2);%收集地面真理CPU的后处理。YTrain =收集(extractdata (YTrain));%从网络中提取的预测。[YPredCell、州]= complexYOLOv4Forward(网络,XTrain networkOutputs,锚);%收集CPU的激活后加工和提取dlarray数据。gatheredPredictions = cellfun(@收集、YPredCell (: 1:8)“UniformOutput”、假);gatheredPredictions = cellfun (@ extractdata gatheredPredictions,“UniformOutput”、假);%将预测从网格单元坐标框坐标。tiledAnchors = generateTiledAnchorsComplexYolov4 (gatheredPredictions(:, 2:5),锚);gatheredPredictions (:, 2:5) = applyAnchorBoxOffsetsComplexYolov4 (tiledAnchors, gatheredPredictions (:, 2:5) inputImageSize);%从地面实况数据生成目标预测。[boxTarget, objectnessTarget classTarget、objectMaskTarget boxErrorScale] = generateComplexYOLOv4Targets (gatheredPredictions、YTrain inputImageSize,锚,penaltyThreshold);%计算损失。boxLoss = bboxOffsetLoss (YPredCell (: [2 3 9 10 6 7]), boxTarget, objectMaskTarget, boxErrorScale);objLoss = objectnessLoss (YPredCell (: 1), objectnessTarget, objectMaskTarget);clsLoss = classConfidenceLoss (YPredCell (:, 8), classTarget, objectMaskTarget);totalLoss = boxLoss + objLoss + clsLoss;信息。boxLoss = boxLoss;信息。objLoss = objLoss;信息。clsLoss = clsLoss; info.totalLoss = totalLoss;%的梯度计算可学的有关损失。梯度= dlgradient (totalLoss net.Learnables);结束

函数boxLoss = bboxOffsetLoss (boxPredCell boxDeltaTarget、boxMaskTarget boxErrorScaleTarget) lossX =总和(cellfun (@ (a, b, c, d) mse (a。* c。* d b。* c。* d), boxPredCell (: 1), boxDeltaTarget (: 1), boxMaskTarget (: 1), boxErrorScaleTarget));有损=总和(cellfun (@ (a, b, c, d) mse (a。* c。* d b。* c。* d), boxPredCell (:, 2), boxDeltaTarget (:, 2), boxMaskTarget (: 1), boxErrorScaleTarget));lossW =总和(cellfun (@ (a, b, c, d) mse (a。* c。* d b。* c。* d), boxPredCell (:, 3), boxDeltaTarget (:, 3), boxMaskTarget (: 1), boxErrorScaleTarget));lossH =总和(cellfun (@ (a, b, c, d) mse (a。* c。* d b。* c。* d), boxPredCell (:, 4), boxDeltaTarget (:, 4), boxMaskTarget (: 1), boxErrorScaleTarget));lossYaw1 =总和(cellfun (@ (a, b, c, d) mse (a。* c。* d b。* c。* d), boxPredCell (:, 5), boxDeltaTarget (:, 5), boxMaskTarget (: 1), boxErrorScaleTarget));lossYaw2 =总和(cellfun (@ (a, b, c, d) mse (a。* c。* d b。* c。* d), boxPredCell (:, 6), boxDeltaTarget (:, 6), boxMaskTarget (: 1), boxErrorScaleTarget));boxLoss = lossX +损耗+ lossW + lossH + lossYaw1 + lossYaw2;结束

函数clsLoss = classConfidenceLoss (classPredCell classTarget boxMaskTarget) clsLoss =总和(cellfun (@ (a, b, c) crossentropy (c。*, * c, b。“TargetCategories”,“独立”)、classPredCell classTarget boxMaskTarget (:, 3)));结束

函数objLoss = objectnessLoss (objectnessPredCell objectnessDeltaTarget boxMaskTarget) objLoss =总和(cellfun (@ (a, b, c) crossentropy (c。*, * c, b。“TargetCategories”,“独立”)、objectnessPredCell objectnessDeltaTarget boxMaskTarget (:, 2)));结束

函数data = preprocessData(数据、targetSize isRotRect)%改变图像的大小和规模在0和1之间的像素。同时规模%相应的边界框。为2 = 1:尺寸数据(数据,1)I = {ii, 1};imgSize =大小(I);%将一个单通道的输入图像转换成三个渠道。如果元素个数(imgSize) < 3 I = repmat(我,1,1,3);结束bboxes = {ii, 2}数据;我= im2single (imresize(我targetSize (1:2)));规模= targetSize (1:2)。/ imgSize (1:2);bboxes = bboxresize (bboxes、规模);如果~ isRotRect bboxes = bboxes (:, 1:4);结束数据(ii, 1:2) ={我bboxes};结束结束函数[XTrain, YTrain] = createBatchData(数据、groundTruthBoxes groundTruthClasses一会)%返回沿着XTrain批尺寸和图片相结合%规范化边框连接YTrain classIDs。%连接图像沿批维度。XTrain =猫(4、数据{:1});%获得类id和类名称。一会= repmat({分类(类名)},大小(groundTruthClasses));[~,classIndices] = cellfun (@ (a, b) ismember (a、b), groundTruthClasses,一会,“UniformOutput”、假);%附加标签索引和训练图像尺寸按比例缩小的边界框%和创建一个单一单元阵列的响应。combinedResponses = cellfun (@ (bbox, classid) [bbox, classid], groundTruthBoxes, classIndices,“UniformOutput”、假);len = max (cellfun (@ (x)大小(x, 1), combinedResponses));paddedBBoxes = cellfun (@ (v) padarray (v, [len-size (v, 1), 0], 0,“职位”)、combinedResponses“UniformOutput”、假);YTrain =猫(4,paddedBBoxes {: 1});结束

函数currentLR = piecewiseLearningRateWithWarmup(迭代,时代、learningRate warmupPeriod, numEpochs)% piecewiseLearningRateWithWarmup函数计算电流%学习速率基于迭代数。持续的warmUpEpoch;如果迭代< = warmupPeriod%的学习速率增加预热阶段的迭代次数。currentLR = learningRate *((迭代/ warmupPeriod) ^ 4);warmUpEpoch =时代;elseif迭代> = < warmUpEpoch warmupPeriod & &时代+地板(0.6 * (numEpochs-warmUpEpoch))%预热阶段后,保持学习速率常数如果剩余数量的时代还不到60%。currentLR = learningRate;elseif时代> = warmUpEpoch +地板(0.6 * (numEpochs-warmUpEpoch)) & & <时代warmUpEpoch +地板(0.9 * (numEpochs-warmUpEpoch))%如果剩余数量的时代但不超过60%% 90%以上,学习速率乘以0.1。currentLR = learningRate * 0.1;其他的%如果仍然有超过90%的时代,用学习% 0.01。currentLR = learningRate * 0.01;结束结束

函数[lossPlotter, learningRatePlotter] = configureTrainingProgressPlotter (f)%创建一个次要情节显示丧失和学习速率。图(f);clf次要情节(2,1,1);ylabel (学习速率的);包含(“迭代”);learningRatePlotter = animatedline;次要情节(2,1,2);ylabel (“全损”);包含(“迭代”);lossPlotter = animatedline;结束函数displayLossInfo(时代、迭代,currentLR lossInfo)%显示损失信息为每个迭代。disp (”时代:“+时代+”|迭代:“+迭代+“|学习速率:“+ currentLR +…”|全损:“+双(收集(extractdata (lossInfo.totalLoss))) +…“|盒子损失:“+双(收集(extractdata (lossInfo.boxLoss))) +…“|对象损失:“+双(收集(extractdata (lossInfo.objLoss))) +…“|类损失:“+双(收集(extractdata (lossInfo.clsLoss))));结束函数updatePlots (lossPlotter learningRatePlotter、迭代、currentLR totalLoss)%更新丧失和学习速率的情节。addpoints (lossPlotter、迭代、双(extractdata(收集(totalLoss))));addpoints (learningRatePlotter迭代,currentLR);drawnow结束函数helperDisplayBoxes (obj bboxes、标签)%显示图像和点云的盒子。图如果~ isa (obj,“pointCloud”=)imshow (obj)形状“矩形”;其他的pcshow (obj.Location);形状=“长方体”;结束showShape(形状,bboxes(标签= =“汽车”:)…“颜色”,“绿色”,“线宽”,0.5);在;showShape(形状,bboxes(标签= =“卡车”:)…“颜色”,“红色”,“线宽”,0.5);showShape(形状,bboxes(标签= =“Pedestrain”:)…“颜色”,“黄色”,“线宽”,0.5);持有从;结束函数helperDownloadPandasetData (outputFolder lidarURL)%从给定的URL下载数据集的输出文件夹。lidarDataTarFile = fullfile (outputFolder,“Pandaset_LidarData.tar.gz”);如果~存在(lidarDataTarFile“文件”mkdir (outputFolder);disp (“下载PandaSet激光雷达驾驶数据(5.2 GB)……”);websave (lidarDataTarFile lidarURL);解压(lidarDataTarFile outputFolder);结束%提取文件。如果(~存在(fullfile (outputFolder激光雷达的),“dir”))…& & (~ (fullfile (outputFolder,存在“长方体”),“dir”)解压(lidarDataTarFile outputFolder);结束结束