mean shift算法matlab源代码及演示图片

clc;clear all;close all; 清除所有 y_Center = 115; x_Center = 210; % center;285445 以上两句为中心点的坐标 Hy = 10; Hx = 10; 定义核函数的带宽，一般为搜索窗口的一半 NImage = 11; % number of frame; 处理图像的总帧数 bin = 32; % number of bin; 灰度级分组， 32组 path = 'hall_bitmap_cif\0'; 图像的路径 weighed_each_bins = zeros(bin,NImage); 定义一个0矩阵weighed_each_bins 每列为相应图像的一个直方图 % % show the target region of first frame; iImage0 = 25; ImageFileName =[path, num2str(iImage0),'.bmp']; 以上两句将iImage0作为图像名，并将图像赋给ImageFileName Image0 = imread(ImageFileName); 读取图像ImageFileName ，作为第一帧图像 Image0 = rgb2gray(Image0); 将图像转化为灰度图像 weighed_each_bins(:,1) = kernelweigh( y_Center, x_Center, Hy, Hx, Image0); 对第一帧图像建立直方图 % define the target region; Image0((x_Center - Hx):(x_Center + Hx), (y_Center - Hy)) = zeros(length((x_Center - Hx):(x_Center + Hx)),1); Image0((x_Center - Hx):(x_Center + Hx), (y_Center + Hy)) = zeros(length((x_Center - Hx):(x_Center + Hx)),1); Image0((x_Center - Hx), (y_Center - Hy):(y_Center + Hy)) = zeros(1,length((y_Center - Hy):(y_Center + Hy))); Image0((x_Center + Hx), (y_Center - Hy):(y_Center + Hy)) = zeros(1,length((y_Center - Hy):(y_Center + Hy))); figure; 以上几句确定第一帧图像的目标区域 imshow(Image0); hold on; plot(y_Center, x_Center, 'o'); 目标区域中心点处画上一个圈 title('Frame1') 确定标题Frame1 % for iImage = iImage0+1 : iImage0 + NImage % ImageFileName =[path, num2str(iImage),'.bmp']; Image = imread(ImageFileName); Image = rgb2gray(Image); 取当前图像存入Image index = iImage-iImage0; % Step 1: Initialize the location of the target in the current frame index表示当前是第几帧图像 % with y0, compute pdf % 在当前窗口得到目标的初始位置 weighed_each_bins(:,index) = kernelweigh( y_Center, x_Center, Hy, Hx, Image); 对当前图像建立直方图 %bar(weighed_each_bins(:,iImage)) % Evalute the Bhattacharyya coefficient Rou0 = 0; for iBins = 1:bin Rou0 = Rou0 + sqrt(weighed_each_bins( iBins, 1) * weighed_each_bins( iBins, index)); end 此for循环用于计算相似性函数，将每组灰度级下的相似性函数值计算出来后相加 % Calculate the weight of each pixel position in the candidate target x_Center_tmp = 0; y_Center_tmp = 0; iteration = 0; 定义了三个变量并初始化为0 % iterate for the optimal solution while (sqrt((x_Center_tmp - x_Center)^2 +(y_Center_tmp - y_Center)^2) > 1)&&(iteration < 10) iteration为最大迭代次数 % Step 2: Derive the weights wi and find the next location of the target candidate; wi = 0; % x_Center1_Den: denominator of formula to compute new center y1; % x_Center1_Nom: numerator of formula to compute new center y1; % <Renference> x_Center1_Nom = 0; x_Center1_Den = 0; y_Center1_Nom = 0; y_Center1_Den = 0; % for ix = (x_Center - Hx):(x_Center + Hx) for iy = (y_Center - Hy):(y_Center + Hy) wi = 0; for iBins = 1:bin % the interval segment of every "bin"; BoundBins = round((255 - 0 + 1)/bin); % the "bin" size; 计算每组的灰度级个数 % the upper bound and the lower bound of the current bin valueBinsLow = BoundBins * (iBins - 1); 每组灰度级的最小值 valueBinsHigh = BoundBins * (iBins) - 1; 每组灰度级的最大值 % compute wi; if (Image(ix,iy) >= valueBinsLow) && (Image(ix,iy) <= valueBinsHigh) % whether the pixel belongs to this "bin"; 如果像素属于该组 if weighed_each_bins(iBins, 1)==0 && weighed_each_bins(iBins, index)==0 如果第一帧图像且当前图像在该灰度组的像素为0 wi = wi + 1; 权重值就加1 else if weighed_each_bins(iBins, index) == 0 wi = wi +1; %weighed_each_bins(iBins, 1) else wi = wi + sqrt( weighed_each_bins(iBins, 1)/weighed_each_bins( iBins, index)); end end break; end 整个if嵌套就是在计算该组的权重值 end x_Center1_Nom = x_Center1_Nom + (ix - x_Center)*wi; x_Center1_Den = x_Center1_Den + wi; y_Center1_Nom = y_Center1_Nom + (iy - y_Center)*wi; y_Center1_Den = y_Center1_Den + wi; end % end for iy = (y_Center - Hy):(y_Center + Hy) end % end for ix = (x_Center - Hx):(x_Center + Hx) %Step3: modify the new center candidate y1 (x_Center1, y_Center1); x_Center1 = x_Center1_Nom/x_Center1_Den + x_Center; y_Center1 = y_Center1_Nom/y_Center1_Den + y_Center; 根据算出的mean shift向量计算新的中心位置 % % Step 4: Compute the newer target model at location (x_Center1, y_Center1) and the Bhattacharyya coefficient % weighed_each_bins(:,index) = kernelweigh( round(y_Center1), round(x_Center1), Hy, Hx, Image);中心移动了，窗口也跟着滑动，在新窗口下建立直方图 % Evalute the Bhattacharyya coefficient Rou1 = 0; for iBins = 1:bin Rou1 = Rou1 + sqrt(weighed_each_bins(iBins, 1) * weighed_each_bins( iBins, index)); end 计算新的相似性函数Rou1 %Step5: if (Rou1<Rou0) x_Center1 = (x_Center1 + x_Center)/2; y_Center1 = (y_Center1 + y_Center)/2; 如果相似性降低，那么窗口中心位置在原来中心和当前中心的中点处 weighed_each_bins(:, index) = kernelweigh( round(y_Center1), round(x_Center1), Hy, Hx, Image);在对当前候选区建直方图模型 Rou1 = 0; for iBins = 1:bin Rou1 = Rou1 + sqrt(weighed_each_bins(iBins, 1) * weighed_each_bins( iBins, index)); end end; 再计算当前的相似性函数值 %上次的中心位置 x_Center_tmp = x_Center; y_Center_tmp = y_Cen