opencv的meanshift使用问题

[hnwlxyzhl]

小弟正在学习opencv2，我在看《OpenCV 2 Computer Vision Application Programming Cookbook》中的meanshift算法时，尝试了下其中的示例代码，这个代码的目的是跟踪下图中红框的猩猩的脸庞，
其示例代码运行结果如下：
，
可以看到示例代码的效果很不错。
但是当我稍微修改代码中（如代码红色部分），meanshift算法的红框的初始搜索位置后，就无法实现跟踪了，
请问这是为什么呢？泛化能力如此差，meanshift算法还有用啊？


源代码如下：#include <iostream>
#include <vector>
using namespace std;

#include <opencv2\core\core.hpp>
#include <opencv2\highgui\highgui.hpp>
#include <opencv2\imgproc\imgproc.hpp>
#include <opencv2\video\tracking.hpp>

#include "objectFinder.h"
#include "colorhistogram.h"

int main()
{
// Read reference image
cv::Mat image= cv::imread("../baboon1.jpg");
if (!image.data)
return 0;

// Define ROI
cv::Mat imageROI= image(cv::Rect(110,260,35,40));
cv::rectangle(image, cv::Rect(110,260,35,40),cv::Scalar(0,0,255));

// Display image
cv::namedWindow("Image");
cv::imshow("Image",image);

// Get the Hue histogram
int minSat=65;
ColorHistogram hc;
cv::MatND colorhist= hc.getHueHistogram(imageROI);

ObjectFinder finder;
finder.setHistogram(colorhist);
finder.setThreshold(0.2f);

// Convert to HSV space
cv::Mat hsv;
cv::cvtColor(image, hsv, CV_BGR2HSV);

// Split the image
vector<cv::Mat> v;
cv::split(hsv,v);

// Eliminate pixels with low saturation
cv::threshold(v[1],v[1],minSat,255,cv::THRESH_BINARY);
cv::namedWindow("Saturation");
cv::imshow("Saturation",v[1]);

// Get back-projection of hue histogram
int ch[1]={0};
cv::Mat result= finder.find(hsv,0.0f,180.0f,ch,1);

cv::namedWindow("Result Hue");
cv::imshow("Result Hue",result);

cv::bitwise_and(result,v[1],result);
cv::namedWindow("Result Hue and");
cv::imshow("Result Hue and",result);

// Second image
image= cv::imread("../baboon3.jpg");

// Display image
cv::namedWindow("Image 2");
cv::imshow("Image 2",image);

// Convert to HSV space
cv::cvtColor(image, hsv, CV_BGR2HSV);

// Split the image
cv::split(hsv,v);

// Eliminate pixels with low saturation
cv::threshold(v[1],v[1],minSat,255,cv::THRESH_BINARY);
cv::namedWindow("Saturation");
cv::imshow("Saturation",v[1]);

// Get back-projection of hue histogram
result= finder.find(hsv,0.0f,180.0f,ch,1);

cv::namedWindow("Result Hue");
cv::imshow("Result Hue",result);

// Eliminate low stauration pixels
cv::bitwise_and(result,v[1],result);
cv::namedWindow("Result Hue and");
cv::imshow("Result Hue and",result);

// Get back-projection of hue histogram
finder.setThreshold(-1.0f);
result= finder.find(hsv,0.0f,180.0f,ch,1);
cv::bitwise_and(result,v[1],result);
cv::namedWindow("Result Hue and raw");
cv::imshow("Result Hue and raw",result);

cv::Rect rect(110,260,35,40);
//cv::Rect rect(120,260,35,40);如果我把此处修改为120,260....就会出现无法准确跟踪区域的情况

cv::rectangle(image, rect, cv::Scalar(0,0,255));

cv::TermCriteria criteria(cv::TermCriteria::MAX_ITER,10,0.01);
cout << "meanshift= " << cv::meanShift(result,rect,criteria) << endl;

cv::rectangle(image, rect, cv::Scalar(0,255,0));

// Display image
cv::namedWindow("Image 2 result");
cv::imshow("Image 2 result",image);

cv::waitKey();
return 0;
}

[yo_yo77]

实际上，超出了迭代的条件。

[程序狗觅食中]

首先你要知道meanshift的原理。这里是根据颜色直方图跟踪的，猴子屁股是红的，跟图片其他部分差异很大，所以效果好，你移动了初始位置之后就变成了猴子的身体的一部分，与身体其他部分是相似的，所以效果不好