评估和可视化车道边界检测对地面真理

%与地面实况数据加载垫文件。加载=负载(“caltech_cordova1_laneAndVehicleGroundTruth.mat”);

%创建VideoReader对象读取视频的帧。videoName = loaded.videoName;fileReader = VideoReader (videoName);%地面实况数据被组织在一个时间表。gtdata = loaded.groundTruthData;%显示地面真理的前几行数据。头(gtdata)

%的第一帧视频。帧= readFrame (fileReader);%在第一帧提取所有车道点。lanePoints = gtdata.LaneBoundaries {1};%提取车辆在第一帧边界框。vehicleBBox = gtdata.Vehicles {1};%重叠和车辆边界框的右车道点。= insertMarker帧(帧,lanePoints {2},“X”);= insertObjectAnnotation帧(帧,“矩形”vehicleBBox,“汽车”);%显示地面实况第一帧数据。图imshow(框架)

%建立monoSensorHelper处理视频。monoCameraSensor = loaded.sensor;monoSensorHelper = helperMonoSensor (monoCameraSensor);%创建新的时间表,同时测量向量。测量=时间表(gtdata.Time);%设置时间表列控股车道边界和车辆数据。numFrames =地板(fileReader.FrameRate * fileReader.Duration);测量。LaneBoundaries =细胞(numFrames, 2);测量。VehicleDetections =细胞(numFrames, 1);gtdata。LanesInVehicleCoord =细胞(numFrames, 2);%回放视频t = 0时,创建一个帧索引来控制电流%框架。fileReader。CurrentTime = 0;frameIndex = 0;%循环videoFile直到没有新的帧。而hasFrame (fileReader) frameIndex = frameIndex + 1;帧= readFrame (fileReader);%使用processFrame方法计算检测。%这个方法可以取代一个定制的车道检测方法。检测= processFrame (monoSensorHelper,框架);%存储估计车道边界和车辆检测。测量。(detections.leftEgoBoundary LaneBoundaries {frameIndex} =…detections.rightEgoBoundary];测量。VehicleDetections {frameIndex} = detections.vehicleBoxes;%为了便于比较,将地面真理莱恩指出%车辆坐标系。gtPointsThisFrame = gtdata.LaneBoundaries {frameIndex};vehiclePoints =细胞(1,元素个数(gtPointsThisFrame));为2 = 1:元素个数(gtPointsThisFrame) vehiclePoints {2} = imageToVehicle (monoCameraSensor, gtPointsThisFrame {2});结束%地面实况点表示存储在车辆坐标。gtdata。LanesInVehicleCoord {frameIndex} = vehiclePoints;结束

%回放视频t = 0。fileReader。CurrentTime = 0;%的第一帧视频。帧= readFrame (fileReader);birdsEyeImage = transformImage (monoSensorHelper。BirdsEyeConfig,框架);%提取右车道点第一帧的鸟瞰图。firstFrameVehiclePoints = gtdata.LanesInVehicleCoord {1};pointsInBEV = vehicleToImage (monoSensorHelper。BirdsEyeConfig, firstFrameVehiclePoints {2});%安装在框架上点。birdsEyeImage = insertMarker (birdsEyeImage pointsInBEV,“X”,“大小”6);%在鸟瞰图显示转变点。图imshow (birdsEyeImage)

阈值= 0.25;%在米[numMatches, numMisses numFalsePositives、作业]=…evaluateLaneBoundaries (measurements.LaneBoundaries…gtdata.LanesInVehicleCoord,…阈值);disp ([的匹配:num2str (numMatches)]);disp ([的失误:num2str (numMisses)]);disp ([的假阳性的数量:num2str (numFalsePositives)]);

averageDistance = helperComputeLaneStatistics (measurements.LaneBoundaries,…gtdata.LanesInVehicleCoord,…作业,@mean);%情节之间的平均距离估计和地面真理。图阻止(gtdata。时间,averageDistance)标题(估计和地面之间的平均距离真理”网格)在ylabel (“米距离”)传说(的左边界,“右边界”)

%计算估计使用monoCamera点位置。[estVehiclePoints, estImagePoints] = helperGetCorrespondingPoints (monoCameraSensor,…measurements.LaneBoundaries,…gtdata.LanesInVehicleCoord,…作业);%估计通道添加到测量的时间表。测量。EstimatedLanes = estImagePoints;测量。LanesInVehicleCoord = estVehiclePoints;%创建一个新的时间表与可视化所需的所有变量。名称= {“LanePoints”;“DetectedLanePoints”};类型= labelType ({“行”;“行”});labelDefs =表(名称,类型,“VariableNames”,{“名字”,“类型”});visualizeInFrame =时间表(gtdata.Time,…gtdata.LaneBoundaries,…measurements.EstimatedLanes,…“VariableNames”、名称);%创建groundTruth对象。数据源= groundTruthDataSource (videoName);dataToVisualize = groundTruth(数据源、labelDefs visualizeInFrame);%保存所有的结果在distanceData前一节。垫在%的临时文件夹。dataToLoad = [tempdir“distanceData.mat”];保存(dataToLoad,“monoSensorHelper”,“videoName”,“测量”,“gtdata”,“averageDistance”);

函数统计= helperComputeLaneStatistics (estModels gtPoints,作业,fcnHandle) numFrames =长度(estModels);%让左派和右派估计南默认代表缺乏%的数据。统计=南*的(numFrames 2);为frameInd = 1: numFrames%左右估计南默认情况下。stat (frameInd:) =南*的(2,1);为idx = 1:长度(estModels {frameInd})%忽视假阳性作业。如果作业{frameInd} (idx) = = 0继续;结束% k边界estModelInFrame匹配k%元素索引gtPointsInFrame作业。这种模式下= estModels {frameInd} (idx);thisGT = gtPoints {frameInd}{作业{frameInd} (idx)};thisGTModel = driving.internal.piecewiseLinearBoundary (thisGT);如果意味着(thisGTModel.Points (:, 2)) > 0%左车道xPoints = thisGTModel.Points (: 1);yDist = 0(大小(xPoints));为指数= 1:元素个数(xPoints) gtYPoints = thisGTModel.computeBoundaryModel (xPoints(指数));testYPoints = thisModel.computeBoundaryModel (xPoints(指数));yDist(指数)= abs (testYPoints-gtYPoints);结束stat (frameInd, 1) = fcnHandle (yDist);其他的%右车道xPoints = thisGTModel.Points (: 1);yDist = 0(大小(xPoints));为指数= 1:元素个数(xPoints) gtYPoints = thisGTModel.computeBoundaryModel (xPoints(指数));testYPoints = thisModel.computeBoundaryModel (xPoints(指数));yDist(指数)= abs (testYPoints-gtYPoints);结束stat (frameInd 2) = fcnHandle (yDist);结束结束结束结束

函数[vehiclePoints, imagePoints] = helperGetCorrespondingPoints (monoCameraSensor、estModels gtPoints,作业)numFrames =长度(estModels);imagePoints =细胞(numFrames, 1);vehiclePoints =细胞(numFrames, 1);为frameInd = 1: numFrames如果isempty(作业{frameInd}) imagePointsInFrame = [];vehiclePointsInFrame = [];其他的estModelInFrame = estModels {frameInd};gtPointsInFrame = gtPoints {frameInd};imagePointsInFrame =细胞(长度(estModelInFrame), 1);vehiclePointsInFrame =细胞(长度(estModelInFrame), 1);为idx = 1:长度(estModelInFrame)%忽视假阳性作业。如果作业{frameInd} (idx) = = 0 imagePointsInFrame {idx} = (NaN南);继续;结束% k边界estModelInFrame匹配k%元素索引gtPointsInFrame作业。这种模式下= estModelInFrame (idx);thisGT = gtPointsInFrame{作业{frameInd} (idx)};xPoints = thisGT (: 1);yPoints = thisModel.computeBoundaryModel (xPoints);vehiclePointsInFrame {idx} = [xPoints, yPoints];imagePointsInFrame {idx} = vehicleToImage (monoCameraSensor [xPoints yPoints]);结束结束vehiclePoints {frameInd} = vehiclePointsInFrame;imagePoints {frameInd} = imagePointsInFrame;%让imagePoints[],而不是{}遵守groundTruth对象。如果isempty (imagePoints {frameInd}) imagePoints {frameInd} = [];结束如果isempty (vehiclePoints {frameInd}) vehiclePoints {frameInd} = [];结束结束结束