干货！升级Opencv的多目标跟踪(多线程，可根据需要实时删减和新增跟踪目标)

Keywords： Multitracker 多线程跟踪 跟踪目标删除 跟踪目标新增 opencv跟踪算法 多线程 动态删减新增跟踪目标 目标跟踪 opencv 多目标跟踪 c++ opencv 多目标跟踪 动态删减目标 动态新增目标 新增目标 删减目标 多个目标跟踪 多个目标跟踪器 目标跟踪器

程序的功能和框架 本源码分三个功能： 1，基于固定场景的多帧中值滤波 2，基于高斯背景模型的前景提取(可以换成其他方法如：深度学习或机器学习的检测结果ROI) 3，根据高斯背景模型提取的ROI框进行目标跟踪。本文选用的是CSTR跟踪器，可根据需要换成KCF等其他跟踪器。 相当于重写了Multitracker 的新增跟踪目标和删减部分。 性能优于opencv原版的Multitracker ，因为Multitracker 跟踪多个目标是串行

要说的话 源码未整理，只保留功能，没有封装。多线程仅限于多跟踪器的多线程，其他部分没加多线程，有兴趣的可以优化。建议用Release模式跑，不然卡死。

opencv源码里的多目标跟踪器并没有删减目标跟踪的功能，本源码加上这以功能。 并可根据需要新增和删减跟踪目标

依赖环境 opencv342 X64 Release版本 VS2019 opencv3 vs各版本下载地址： 链接: https://pan.baidu.com/s/1f5oAFqs-u15vkD5LNTcxtw 提取码: 2qj9 上面的百度网盘应该没有我调试用的opencv版本 VS2019_release_X64_opencv342，可以加QQ群去群文件找:539308722

话不多说，具体请参考下面的源码

随便抄，原创不易，点个赞或者关注一下 另：两个可以水中文核心的代码。懒得写论文了，有需要的也可以进群联系 给我挂个共同一作或者二作就成 源码如果有错的地方，那就是csdn编辑器编辑起来很别扭，复制粘贴时乱了一点。也可以加群找我要代码。

#include <opencv2/opencv.hpp> #include <iostream> #include <math.h> #include <thread> #include <opencv2/tracking.hpp> #include <opencv2/core/ocl.hpp> using namespace cv; using namespace std; #define COLS 960 #define ROWS 540 #define FN 10//时序中值的图像数目 Mat matRotateClockWise180(Mat src)//顺时针180 { flip(src, src, 0); flip(src, src, 1); return src; } void BaseLine_Generator_1(Mat* BaseLine,vector<Mat>* vMat) { for (int i = 0; i < ROWS; ++i) { for (int j = 0; j < COLS; ++j) { uchar Mid_value_[FN]; for (int k = 0; k < FN; k++) { Mid_value_[k] = (*vMat)[k].at<Vec3b>(i, j)[0]; } sort(Mid_value_, Mid_value_ + FN); BaseLine->at<Vec3b>(i, j)[0] = Mid_value_[FN / 2]; for (int k = 0; k < FN; k++) { Mid_value_[k] = (*vMat)[k].at<Vec3b>(i, j)[1]; } sort(Mid_value_, Mid_value_ + FN); BaseLine->at<Vec3b>(i, j)[1] = Mid_value_[FN / 2]; for (int k = 0; k < FN; k++) { Mid_value_[k] = (*vMat)[k].at<Vec3b>(i, j)[2]; } sort(Mid_value_, Mid_value_ + FN); BaseLine->at<Vec3b>(i, j)[2] = Mid_value_[FN / 2]; } } } float RectOverlap(cv::Rect r, cv::Rect rOther) { int x0 = std::max(r.x, rOther.x); int x1 = std::min(r.x + r.width, rOther.x + rOther.width); int y0 = std::max(r.y, rOther.y); int y1 = std::min(r.y + r.height, rOther.y + rOther.height); if (x0 >= x1 || y0 >= y1) return 0.f; float areaInt = (x1 - x0) * (y1 - y0); float MINArea= MIN((float)r.width * r.height, (float)rOther.width * rOther.height); return areaInt / MINArea; } //删除跟踪器 void CSRT_tracker(vector<Ptr<Tracker>>* tracker, Mat* frame, vector<Rect2d>* bbox, vector<bool>* ok_CSRT, int i) { (*ok_CSRT)[i] = (*tracker)[i]->update(*frame, (*bbox)[i]); } void CSRT_tracker_erase_(vector<Ptr<Tracker>>* tracker, vector<Rect2d>* bbox, vector<bool>* ok_CSRT, vector<clock_t>* Start_time_record, vector<double>* fish_distance, vector<Rect2d>* ptr_bboxCSRT,int i) { (*tracker).erase((*tracker).begin() + i); (*bbox).erase((*bbox).begin() + i); (*ok_CSRT).erase((*ok_CSRT).begin() + i); (*Start_time_record).erase((*Start_time_record).begin() + i); (*fish_distance).erase((*fish_distance).begin() + i); (*ptr_bboxCSRT).erase((*ptr_bboxCSRT).begin() + i); } int main(int argc, char** argv) { Ptr<BackgroundSubtractorMOG2> bgsubtractor = createBackgroundSubtractorMOG2(); bgsubtractor->setVarThreshold(20); VideoCapture video("C:/testv/5.mp4"); if (!video.isOpened()) { cout << "Could not read video file" << endl; return 1; } Mat frame; Mat src; vector<Mat> Filtering_Mat; //*******************8 vector<thread> TrackerThread_; vector<Ptr<Tracker>> trackerCSRT; vector<Rect2d> bboxCSRT; vector<double> fish_distance; vector<Rect2d> ptr_bboxCSRT; Rect2d bbox; bool sign = false; vector<bool> OK_; vector<clock_t> Start_time_record; bool OK_0 = false; long int fishnum = 0; /// for (int i = 0; i < FN; i++) { Mat src1; video.read(frame); resize(frame, src1, Size(960, 540)); src1 = matRotateClockWise180(src1); imshow("check", src1); waitKey(100); Filtering_Mat.push_back(src1); } while (video.read(frame)) { resize(frame, src, Size(960, 540)); src = matRotateClockWise180(src); imshow("src", src); Mat filter_src1 = src.clone(); Filtering_Mat.erase(Filtering_Mat.begin()); Filtering_Mat.push_back(filter_src1); Mat filter_src = src.clone(); BaseLine_Generator_1(&filter_src, &Filtering_Mat); imshow("Final result", filter_src); Mat mask; Mat Gfilter_src; cvtColor(filter_src, Gfilter_src, CV_RGB2GRAY); bgsubtractor->apply(filter_src, mask, 0.01); imshow("GMask", mask); std::vector<std::vector<cv::Point> > cnts; Mat kernel = getStructuringElement(MORPH_CROSS, Size(13, 13), Point(-1, -1)); cv::dilate(mask, mask, kernel); Mat kernel1 = getStructuringElement(MORPH_CROSS, Size(12,12), Point(-1, -1)); cv::erode(mask, mask, kernel1); cv::threshold(mask, mask, 128, 255, THRESH_BINARY); imshow("GMask1", mask); findContours(mask, cnts, RETR_EXTERNAL, CHAIN_APPROX_SIMPLE); float Area; Rect rect; vector<Point> m; for (int i = cnts.size() - 1; i >= 0; i--) { vector<Point> c = cnts[i]; Area = contourArea(c); if (Area < 1600)//目标的大小 { continue; } else { m = c; } rect = boundingRect(m); //高斯背景求的前景目标比例限制，可根据需要调整 if(double(MAX(rect.width, rect.height)) / double(MIN(rect.width, rect.height)) <1.5) continue; if (double(MAX(rect.width, rect.height)) / double(MIN(rect.width, rect.height)) > 3) continue; if((rect.width + rect.height)>400)continue; // Rect rect_track = rect; double cx = double(rect.x + double(rect.width)/2); double cy = double(rect.y + double(rect.height) / 2); bool Needcontinue = false; for (int kk =0;kk<bboxCSRT.size();kk++) { //与已有的目标中心点曼哈顿距离小于100 if ((abs((bboxCSRT[kk].x + bboxCSRT[kk].width / 2) - cx) + abs((bboxCSRT[kk].y + bboxCSRT[kk].height / 2) - cy)) < 100) { Needcontinue = true; continue; } //于已有框体交叉面积大于20% float verlap_ = RectOverlap(rect, bboxCSRT[kk]); if (verlap_ >0.2) { Needcontinue = true; continue; } } if (Needcontinue) { continue; } Ptr<Tracker > tracker_0; tracker_0 = TrackerCSRT::create(); bbox = rect; rectangle(filter_src, bbox, Scalar(255, 0, 0), 2, 1); bboxCSRT.push_back(bbox); tracker_0->init(src, bbox); trackerCSRT.push_back(tracker_0); OK_.push_back(OK_0); clock_t start_time = clock(); double zero_ = 0; fish_distance.push_back(zero_); ptr_bboxCSRT.push_back(bbox); Start_time_record.push_back(start_time); } imshow("CSRT Tracking src", filter_src); int k1 = waitKey(1); Mat tracksrc; bilateralFilter(src, tracksrc, 5, 2, 2); if(OK_.size() > 0){ double timer = (double)getTickCount(); for (int i = 1; i < trackerCSRT.size(); i++) { TrackerThread_.push_back(thread(CSRT_tracker, &trackerCSRT, &tracksrc, &bboxCSRT, &OK_, i)); } OK_[0] = trackerCSRT[0]->update(tracksrc, bboxCSRT[0]); for (auto& ite : TrackerThread_) { ite.join(); } TrackerThread_.clear(); float fps = getTickFrequency() / ((double)getTickCount() - timer); Mat frameCSRT = tracksrc.clone(); for (int i = 0; i < trackerCSRT.size(); i++) { if (OK_[i]) { rectangle(frameCSRT, bboxCSRT[i], Scalar(0, 255, 255), 2, 1); } else { putText(frameCSRT, "Tracking failure detected" + to_string(i), Point(100, 80 + i * 10), FONT_HERSHEY_SIMPLEX, 0.75, Scalar(0, 0, 255), 2); } double x_b; double y_b;//框体的右下点坐标 x_b = bboxCSRT[i].x + bboxCSRT[i].width; y_b = bboxCSRT[i].y + bboxCSRT[i].height; string str; clock_t time_out = clock() - Start_time_record[i]; double cx0 = double(bboxCSRT[i].x + double(bboxCSRT[i].width) / 2); double cy0 = double(bboxCSRT[i].y + double(bboxCSRT[i].height) / 2); double cx1 = double(ptr_bboxCSRT[i].x + double(ptr_bboxCSRT[i].width) / 2); double cy1 = double(ptr_bboxCSRT[i].y + double(ptr_bboxCSRT[i].height) / 2); fish_distance[i]= fish_distance[i] + abs(cx0 - cx1) + abs(cy0 - cy1); ptr_bboxCSRT[i].x = bboxCSRT[i].x; ptr_bboxCSRT[i].y= bboxCSRT[i].y; ptr_bboxCSRT[i].width = bboxCSRT[i].width; ptr_bboxCSRT[i].height = bboxCSRT[i].height; bool dis_time = false; dis_time = ((time_out / CLOCKS_PER_SEC) > 30) && (fish_distance[i] < 300);//出现在场景中的时间超过30秒且目标的距离小于300 //目标越界，且走过de的距离大于400，删除目标，计数+1 if ((bboxCSRT[i].x < 0 || bboxCSRT[i].y < 0 || x_b >960 || y_b>540) && (fish_distance[i] > 400)) { CSRT_tracker_erase_(&trackerCSRT, &bboxCSRT, &OK_, &Start_time_record,&fish_distance,&ptr_bboxCSRT,i); str = "Find Fish, delete Rect!"; fishnum++; } else if(!OK_[i] || dis_time|| ((time_out / CLOCKS_PER_SEC) > 80)){ CSRT_tracker_erase_(&trackerCSRT, &bboxCSRT, &OK_, &Start_time_record, &fish_distance, &ptr_bboxCSRT, i); str = "Not Fish, delete Rect!"; } //跟踪目标越界，在场景中走过的距离不超过300个像素 曼哈顿距离 else if((bboxCSRT[i].x < 0 || bboxCSRT[i].y < 0 || x_b >960 || y_b>540) && (fish_distance[i] < 300)){ CSRT_tracker_erase_(&trackerCSRT, &bboxCSRT, &OK_, &Start_time_record, &fish_distance, &ptr_bboxCSRT, i); str = "Not Fish, delete Rect!"; } putText(frameCSRT, "CSRT Tracker Number:" + to_string(OK_.size()), Point(100, 20), FONT_HERSHEY_SIMPLEX, 0.75, Scalar(50, 170, 50), 2); putText(frameCSRT, "FPS : " + to_string((int(fps))) + str, Point(100, 50), FONT_HERSHEY_SIMPLEX, 0.75, Scalar(50, 170, 50), 2); putText(frameCSRT, "Found fish: " + to_string(fishnum) , Point(100, 80), FONT_HERSHEY_SIMPLEX, 0.75, Scalar(50, 170, 50), 2); imshow("CSRT Tracking", frameCSRT); waitKey(1); } } } return 0; }

结果图示意:貌似上面这个代码是没有鱼轨迹的，忘记了. 获取有轨迹的版本：https://download.csdn.net/download/weixin_37918890/12593594