简介
- MIL: TrackerMIL 以在线方式训练分类器将对象与背景分离;多实例学习避免鲁棒跟踪的漂移问题.
- OLB: TrackerBoosting 基于AdaBoost算法的在线实时对象跟踪.分类器在更新步骤中使用周围背景作为反例以避免漂移问题.
- MedianFlow: TrackerMedianFlow 跟踪器适用于非常平滑和可预测的运动,物体在整个序列中可见.
- TLD: TrackerTLD 将长期跟踪任务分解为跟踪,学习和检测.跟踪器在帧之间跟踪对象.探测器本地化所观察到的所有外观,并在必要时纠正跟踪器.学习估计检测器的错误并进行更新以避免再出现这些错误.追踪器能够处理快速运动,部分遮挡,物体缺失等情况.
- KCF: TrackerKCF 使用目标周围区域的循环矩阵采集正负样本,利用脊回归训练目标检测器,并成功的利用循环矩阵在傅里叶空间可对角化的性质将矩阵的运算转化为向量的Hadamad积,即元素的点乘,大大降低了运算量,提高了运算速度,使算法满足实时性要求.
部分相关API:
TrackerMIL
static Ptr<TrackerMIL> create(const TrackerMIL::Params ¶meters);
CV_WRAP static Ptr<TrackerMIL> create();
struct CV_EXPORTS Params
{
PARAMS();
float samplerInitInRadius;
int samplerInitMaxNegNum;
float samplerSearchWinSize;
float samplerTrackInRadius;
int samplerTrackMaxPosNum;
int samplerTrackMaxNegNum;
int featureSetNumFeatures;
void read(const FileNode&fn);
void write(FileStorage&fs)const;
};
TrackerBoosting
static Ptr<TrackerBoosting> create(const TrackerBoosting::Params ¶meters);
CV_WRAP static Ptr<TrackerBoosting> create();
struct CV_EXPORTS Params
{
PARAMS();
int numClassifiers;
float samplerOverlap;
float samplerSearchFactor;
int iterationInit;
int featureSetNumFeatures;
void read(const FileNode&fn);
void write(FileStorage&fs)const;
};
示例
首先获取视频的第一帧,通过点击左键框选选择要跟踪的目标,点击右键确认并使用MIL开始跟踪.(从实际情况看来,算法对过程中有遮挡的情况跟踪能力较差.)
(环境:Ubuntu16.04+QT5.8+opencv3.3.1)
#include <opencv2/opencv.hpp>
#include <opencv2/video.hpp>
#include <opencv2/tracking.hpp>
#include <opencv2/tracking/tracker.hpp>
using namespace cv;
void draw_rectangle(int event, int x, int y, int flags, void*);
Mat firstFrame;
Point previousPoint, currentPoint;
Rect2d bbox;
int main(int argc, char *argv[])
{
VideoCapture capture;
Mat frame;
frame = capture.open("/home/w/mycode/QT/img/runners.avi");
if(!capture.isOpened())
{
printf("can not open ...\n");
return -1;
}
capture.read(firstFrame);
if(!firstFrame.empty())
{
namedWindow("output", WINDOW_AUTOSIZE);
imshow("output", firstFrame);
setMouseCallback("output", draw_rectangle, 0);
waitKey();
}
Ptr<TrackerMIL> tracker= TrackerMIL::create();
capture.read(frame);
tracker->init(frame,bbox);
namedWindow("output", WINDOW_AUTOSIZE);
while (capture.read(frame))
{
tracker->update(frame,bbox);
rectangle(frame,bbox, Scalar(255, 0, 0), 2, 1);
imshow("output", frame);
if(waitKey(20)=='q')
return 0;
}
capture.release();
destroyWindow("output");
return 0;
}
void draw_rectangle(int event, int x, int y, int flags, void*)
{
if (event == EVENT_LBUTTONDOWN)
{
previousPoint = Point(x, y);
}
else if (event == EVENT_MOUSEMOVE && (flags&EVENT_FLAG_LBUTTON))
{
Mat tmp;
firstFrame.copyTo(tmp);
currentPoint = Point(x, y);
rectangle(tmp, previousPoint, currentPoint, Scalar(0, 255, 0, 0), 1, 8, 0);
imshow("output", tmp);
}
else if (event == EVENT_LBUTTONUP)
{
bbox.x = previousPoint.x;
bbox.y = previousPoint.y;
bbox.width = abs(previousPoint.x-currentPoint.x);
bbox.height = abs(previousPoint.y-currentPoint.y);
}
else if (event == EVENT_RBUTTONUP)
{
destroyWindow("output");
}
}
![这里写图片描述](https://img-blog.csdn.net/20180613145744309?watermark/2/text/aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FrYWRpYW8=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70)
框选目标:
跟踪效果:
![这里写图片描述](https://img-blog.csdn.net/20180613145446713?watermark/2/text/aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FrYWRpYW8=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70)
实验对比发现:KCF速度最快,MedianFlow的速度也较快,对于无遮挡情况跟踪效果较好;TLD对部分遮挡处理的效果最好,处理时间相对较慢.
部分遮挡处理效果
MIL对部分遮挡的处理效果:
KCF对部分遮挡的处理效果:
TLD对部分遮挡的处理效果:
![这里写图片描述](https://img-blog.csdn.net/20180613160419629?watermark/2/text/aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FrYWRpYW8=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70)
opencv::Tracker Algorithms
本文内容由网友自发贡献,版权归原作者所有,本站不承担相应法律责任。如您发现有涉嫌抄袭侵权的内容,请联系:hwhale#tublm.com(使用前将#替换为@)