SAMF tracker代码阅读

**SAMF tracker代码阅读笔记（matlab）**如有错误，请不吝赐教

function [positions, rect_results, time] = tracker(video_path, img_files, pos, target_sz, ...
	padding, kernel, lambda, output_sigma_factor, interp_factor, cell_size, ...
	features, show_visualization)

简单的定义function

1.addpath('./utility');
2.temp = load('w2crs'); 
3.w2c = temp.w2crs;

1.添加脚本文件utility到工作路径
2.temp是一个返回值，load是将w2crs变量加载到工作区
3.w2crs,一个数据集 需要matlab分帧的话可以点击
链接: [link]https://download.csdn.net/download/CloudVans/20354273?spm=1001.2014.3001.5503

4.resize_image = (sqrt(prod(target_sz)) >= 100);

(sqrt(prod(target_sz))是追踪框尺寸
resize_image是图像缩放函数，后半部分是对角线尺寸≥是阈值

5.   if resize_image
6.		   pos = floor(pos / 2);                     
7.		   target_sz = floor(target_sz / 2);
8.   end
9.         target_sz_back = target_sz;
10.        window_sz = floor(target_sz * (1 + padding));
11.	       current_size =1;                            

target_sz是目标初始大小
window_sz是窗口的尺寸
pos是目标初始位置
padding:对目标附近附加跟踪区域
10.窗口大小，考虑填充
11.设置初始尺寸为1，但是后面没用到这个赋值

12.output_sigma = sqrt(prod(target_sz)) * output_sigma_factor / cell_size; 

用空间带宽，像素大小，跟踪框尺寸计算高斯标签的带宽

13.yf = fft2(gaussian_shaped_labels(output_sigma, floor(window_sz / cell_size)));

fft2是傅里叶变换，用于返回矩阵

14.cos_window = hann(size(yf,1)) * hann(size(yf,2))';	

生成汉宁窗cos_window，尺寸与yf相同，即floor(window_sz / cell_size)，检测范围window_sz中cell的个数。

15.search_size = [1  0.985 0.99 0.995 1.005 1.01 1.015];

没读懂

16.  if  show_visualization 
17.		 update_visualization = show_video(img_files, video_path, resize_image);
18.  end

创建视频界面，对上一个可视化进行调整
show_video是将跟踪结果可视化，通过后面的几个参数生成新视频界面

time = 0; 

用于计时

19.positions = zeros(numel(img_files), 2);
20.rect_results = zeros(numel(img_files), 4);
21.response = zeros(size(cos_window,1),size(cos_window,2),size(search_size,2);
22.szid = 1;

zeros函数用于返回全零矩阵
19.计算每帧目标的中心位置
20.结果矩阵
21.得到了一个回应的行矩阵
22.这个赋值没懂

23.for frame = 1:numel(img_files)
24.im = imread([video_path img_files{frame}]);

23.简单的for循环从第一帧到最后一帧
24.读取图片为im

25.if resize_image
26.			im = imresize(im, 0.5);  
27.end

将图片变成0.5倍返还给im，a= imresize(b, c);将b的大小变成c给a

28.tic() 

计时开始

29. if frame > 1
30.    for i=1:size(search_size,2)                                      
31.                tmp_sz = floor((target_sz * (1 + padding))*search_size(i));
32.               param0 = [pos(2), pos(1), tmp_sz(2)/window_sz(2), 0,...
33.                        tmp_sz(1)/window_sz(2)/(window_sz(1)/window_sz(2)),0];
34.               param0 = affparam2mat(param0); 
35.                patch = uint8(warpimg(double(im), param0, window_sz));
36.                zf = fft2(get_features(patch, features, cell_size, cos_window,w2c));
             
37.                switch kernel.type
38.                case 'gaussian'
39.                   kzf = gaussian_correlation(zf, model_xf, kernel.sigma);
40.                case 'polynomial'
41.                    kzf = polynomial_correlation(zf, model_xf, kernel.poly_a, kernel.poly_b);
42.                case 'linear'
43.                    kzf = linear_correlation(zf, model_xf);
44.     end
45.                response(:,:,i) = real(ifft2(model_alphaf .* kzf));  
46. end

从第二帧开始进入循环，计算分类器在所有班次的响应，回应快速检测方程

47.[vert_delta,tmp, horiz_delta] = find(response == max(response(:)), 1);
47. szid = floor((tmp-1)/(size(cos_window,2)))+1;
49.horiz_delta = tmp - ((szid -1)* size(cos_window,2));

没读懂，猜测是目标在最大响应值，峰值出现在左上角而不是中间的处理方法

48.  if vert_delta > size(zf,1) / 2  
49.				vert_delta = vert_delta - size(zf,1);
50.  end
51.  if horiz_delta > size(zf,2) / 2   
52.				horiz_delta = horiz_delta - size(zf,2);
53.  end
54.  tmp_sz = floor((target_sz * (1 + padding))*search_size(szid));
55.             current_size = tmp_sz(2)/window_sz(2);
56.			    pos = pos + current_size*cell_size * [vert_delta - 1, horiz_delta - 1];
57.	 end

垂直与水平的矫正

58.        target_sz = target_sz * search_size(szid);
59.        tmp_sz = floor((target_sz * (1 + padding)));
60.        param0 = [pos(2), pos(1), tmp_sz(2)/window_sz(2), 0,...
61.        tmp_sz(1)/window_sz(2)/(window_sz(1)/window_sz(2)),0];
62.        param0 = affparam2mat(param0); 
63.        patch = uint8(warpimg(double(im), param0, window_sz));
64.		   xf = fft2(get_features(patch, features, cell_size, cos_window,w2c));

在新估计的目标位置获得用于训练的子窗口

65.switch kernel.type
66.		case 'gaussian'
67.			kf = gaussian_correlation(xf, xf, kernel.sigma);
68.		case 'polynomial'
69.			kf = polynomial_correlation(xf, xf, kernel.poly_a, kernel.poly_b);
70.		case 'linear'
71.			kf = linear_correlation(xf, xf);
72.		end
73.		alphaf = yf ./ (kf + lambda);   

核岭回归，计算 alphas（在傅立叶域中）

74.     if frame == 1  
75.			model_alphaf = alphaf;
76.			model_xf = xf;
77.		else
78.			model_alphaf = (1 - interp_factor) * model_alphaf + interp_factor * alphaf;
79.			model_xf = (1 - interp_factor) * model_xf + interp_factor * xf;
80.		end

第一帧，用单个图像训练，后续帧用后续插值模型

81.positions(frame,:) = pos;
82.time = time + toc();
83.box = [pos([2,1]) - target_sz([2,1])/2, target_sz([2,1])];
84.rect_results(frame,:)=box;

保存位置和时间

85.  if show_visualization  
86.     stop = update_visualization(frame, box);
87.			 if stop, break, end
88.			       drawnow             
89.		     end
90.	 end

令结果可视化并跟新数据，处理回调>绘图

91.  if resize_image
92.		   positions = positions * 2;
93.        rect_results = rect_results*2;
94.	 end

最后再次调整图像大小