函数obj = setupSystemObjects()初始化视频I/O创建用于从文件中读取视频的对象，绘制跟踪对象%对象，并播放视频。创建一个视频阅读器。obj。阅读器=视频阅读器(“atrium.mp4”）;创建两个视频播放器，一个用于显示视频，%和一个显示前景蒙版。obj。maskPlayer =视觉。放像机(“位置”， [740, 400, 700, 400]);obj。videoPlayer =视觉。放像机(“位置”， [20, 400, 700, 400]);创建用于前景检测和斑点分析的系统对象前景检测器用于分割移动的物体%背景。它输出一个二进制掩码，其中像素值1的%对应前台，0的值对应前台%到后台。Obj.detector =视觉。ForegroundDetector (“NumGaussians”3,.．.“NumTrainingFrames”现年40岁的“MinimumBackgroundRatio”, 0.7);前景像素的连接组很可能对应于移动%的对象。对象blob analysis System用于查找这样的组%(称为“blobs”或“连接组件”)，并计算它们的%特征，如面积、质心和包围框。obj。blobAnalyser =视觉。BlobAnalysis (“BoundingBoxOutputPort”,真的,.．.“AreaOutputPort”,真的,“CentroidOutputPort”,真的,.．.“MinimumBlobArea”, 400);结束

函数tracks = initializeTracks()创建一个空的音轨数组Tracks = struct(.．.“id”{},.．.“bbox”{},.．.“kalmanFilter”{},.．.“年龄”{},.．.“totalVisibleCount”{},.．.“consecutiveInvisibleCount”, {});结束

函数[质心，框，掩码]= detectObjects(帧)%检测前景。掩码= obj.detect .step(帧);应用形态学操作去除噪声和填充洞。。掩码= imopen(掩码，strel(“矩形”[3 3]));掩码= imclose(掩码，strel(“矩形”， [15,15]));掩码=填充(掩码，“黑洞”）;执行blob分析以查找连接的组件。[~， centroids, bboxes] = obj.blobAnalyser.step(mask);结束

函数predictNewLocationsOfTracks ()为I = 1:长度(轨道)bbox =轨道(I).bbox;预测轨道的当前位置。predictedCentroid = predict(tracks(i).kalmanFilter);移动边框，使其中心位于%预测的位置。predictedCentroid = int32(predictedCentroid) - bbox(3:4) / 2;跟踪(i)。bbox= [predictedCentroid, bbox(3:4)];结束结束

函数[assignments, unassigndtracks, unassignddetections] =.．.detectionToTrackAssignment() nTracks =长度(轨道);nDetections = size(质心，1);计算分配每个检测到每个轨道的成本。。成本=零(nTracks, nDetections);为i = 1:nTracks成本(i，:) =距离(轨道(i)。kalmanFilter、质心);结束解决分配问题。costOfNonAssignment = 20;[assignments, unassigndtracks, unassignddetections] =.．.assignDetectionsToTracks(成本、costOfNonAssignment);结束

函数updateAssignedTracks() numAssignedTracks = size(assignments, 1);为trackIdx =赋值(i, 1);detectionIdx =赋值(i, 2);质心=质心(detectionIdx，:);bbox = bboxes(detectionIdx，:);修正物体位置的估计%使用新的检测。正确的(跟踪(trackIdx)。kalmanFilter、质心);将预测的包围框替换为检测到的%包围框。跟踪(trackIdx)。Bbox = Bbox;%更新轨道的年龄。跟踪(trackIdx)。年龄= tracks(trackIdx).age + 1;%更新可见性。跟踪(trackIdx)。totalVisibleCount=.．.跟踪(trackIdx)。totalVisibleCount+ 1; tracks(trackIdx).consecutiveInvisibleCount = 0;结束结束

函数updateUnassignedTracks ()为i = 1:长度(unassigndtracks) ind = unassigndtracks (i);跟踪(印第安纳州)。年龄= tracks(ind).age + 1; tracks(ind).consecutiveInvisibleCount =.．.跟踪(印第安纳州)。consecutiveInvisibleCount + 1;结束结束

函数deleteLostTracks ()如果isempty(跟踪)返回；结束invisibleForTooLong = 20;ageThreshold = 8;计算轨道可见年龄的百分比。%年龄= [tracks(:).age];totalVisibleCounts = [tracks(:).totalVisibleCount];可见度= totalVisibleCounts ./ ages;查找“丢失”曲目的索引。lostInds =(年龄<年龄阈值&能见度< 0.6)|.．.(跟踪(:)。consecutiveInvisibleCount] >= invisibleForTooLong;删除丢失的轨道。tracks = tracks(~lostInds);结束

函数createNewTracks() centroids = centroids(unassignddetections，:);bboxes = bboxes(unassignddetections，:);为I = 1:大小(质心，1)质心=质心(I，:);Bbox = bboxes(i，:);创建一个卡尔曼过滤器对象。kalmanFilter = configureKalmanFilter(“ConstantVelocity”，.．.质心，[200,50]，[100,25]，100);创建一个新轨道。newTrack = struct(.．.“id”nextId,.．.“bbox”bbox,.．.“kalmanFilter”kalmanFilter,.．.“年龄”, 1.．.“totalVisibleCount”, 1.．.“consecutiveInvisibleCount”, 0);将其添加到曲目数组中。。tracks(end + 1) = newTrack;自增下一个id。nextId = nextId + 1;结束结束

函数displayTrackingResults ()将帧和蒙版转换为uint8 RGB。Frame = im2uint8(Frame);掩码= uint8(repmat(掩码，[1,1,3])).* 255;minVisibleCount = 8;如果~ isempty(跟踪)噪声检测往往会导致短时间的轨迹。。只显示已可见超过。%最小帧数。reliableTrackInds =.．.(跟踪(:)。totalVisibleCount] > minVisibleCount;reliableTracks = tracks(reliableTrackInds);显示对象。如果未检测到对象%，显示其预测的边界框。如果~ isempty (reliableTracks)获取边界框。bboxes = cat(1, reliableTracks.bbox);获取id。ids = int32([reliableTracks(:).id]);为对象创建标签我们展示的是预测值而不是实际值。%的位置。标签= cellstr(int2str(ids'));predictedTrackInds =.．.[reliableTracks(:)。consecutiveInvisibleCount] > 0;isexpected = cell(size(labels));ispredict (predictedTrackInds) = {“预测”};标签= strcat(标签，isexpected);在框架上绘制物体。帧= insertObjectAnnotation(帧，“矩形”，.．.bboxes、标签);在蒙版上绘制对象。掩码= insertObjectAnnotation“矩形”，.．.bboxes、标签);结束结束显示掩码和帧。obj.maskPlayer.step(面具);obj.videoPlayer.step(框架);结束

结束