KCF代码阅读笔记（matlab版）
1 主函数run_tracker.m
提取HOG特征用的函数和CPP文件作者已经给出并编译
什么也不输入时，默认为使用HOG特征和高斯核函数（还可以选择使用灰度特征和线性核）
要把base_path改成自己的benchamark文件夹所在路径（第43行），比如：
base_path = ‘F:\track\Benchmark\’;
主函数中给出的相关参数：
padding 目标搜索范围系数，决定了每帧进行检测的范围
Lambada 正则化参数λ
output_sigma_factor 空间带宽：与目标大小成比例
interp_factor 时间更新系数
sigma 高斯核函数的带宽/方差
poly_a/poly_b
cell_size 求HOG特征时cell的尺寸
% High-Speed Tracking with Kernelized Correlation Filters
%
% Joao F. Henriques, 2014
% http://www.isr.uc.pt/~henriques/
%
%------------------------------------------------------------------------------------------------------------
% 内核化/双相关滤波器的主函数(KCF/DCF)。这个函数负责设置参数，加载视频信息和计算精度。对于实际的跟踪代码进行检查跟踪器函数。
%
% Main interface for Kernelized/Dual Correlation Filters (KCF/DCF).
% This function takes care of setting up parameters, loading video
% information and computing precisions. For the actual tracking code,
% check out the TRACKER function.
%
%--------------------------------------------------------------------------------------------------------------
% 没有任何参数，会要求您选择视频，跟踪使用的高斯KCF上的HOG，并在交互中显示结果的人物。
% 按下“Esc”可以提前停止跟踪器。你可以浏览视频使用滚动条在底部。
%
% RUN_TRACKER
% Without any parameters, will ask you to choose a video, track using
% the Gaussian KCF on HOG, and show the results in an interactive
% figure. Press 'Esc' to stop the tracker early. You can navigate the
% video using the scrollbar at the bottom.
%
%--------------------------------------------------------------------------------------------------------------
% 允许您选择一个视频的名称。all将运行所有视频，并显示平均统计数据。“选择”将交互式地选择一个。
%
% RUN_TRACKER VIDEO
% Allows you to select a VIDEO by its name. 'all' will run all videos
% and show average statistics. 'choose' will select one interactively.
%
%---------------------------------------------------------------------------------------------------------------
% 选择一个内核。'高斯'/'多项式'运行KCF， '线性'为DCF。
%
% RUN_TRACKER VIDEO KERNEL
% Choose a KERNEL. 'gaussian'/'polynomial' to run KCF, 'linear' for DCF.
%
%---------------------------------------------------------------------------------------------------------------
% 选择一个特征类型，要么是“hog”，要么是“gray”(原始像素)。
%
% RUN_TRACKER VIDEO KERNEL FEATURE
% Choose a FEATURE type, either 'hog' or 'gray' (raw pixels).
%
%---------------------------------------------------------------------------------------------------------------
% 决定是否显示可滚动图形和精度图。
%
% RUN_TRACKER(VIDEO, KERNEL, FEATURE, SHOW_VISUALIZATION, SHOW_PLOTS)
% Decide whether to show the scrollable figure, and the precision plot.
%
%---------------------------------------------------------------------------------------------------------------
% 有用的组合:
%
% Useful combinations:
% >> run_tracker choose gaussian hog %Kernelized Correlation Filter (KCF)
% >> run_tracker choose linear hog %Dual Correlation Filter (DCF)
% >> run_tracker choose gaussian gray %Single-channel KCF (ECCV'12 paper)
% >> run_tracker choose linear gray %MOSSE filter (single channel)
%
function [precision, fps] = run_tracker(video, kernel_type, feature_type, show_visualization, show_plots)
%path to the videos (you'll be able to choose one with the GUI).
base_path = './data/Benchmark/';
%default settings %默认设置
if nargin < 1, video = 'choose'; end%如果输入变量的个数小于1个，那么给video赋默认值，只执行这一条语句
if nargin < 2, kernel_type = 'gaussian'; end
if nargin < 3, feature_type = 'hog'; end
if nargin < 4, show_visualization = ~strcmp(video, 'all'); end
if nargin < 5, show_plots = ~strcmp(video, 'all'); end
%根据论文定义的参数，在这里我们基于已经选择的核函数与特征类型 写入参数
%parameters according to the paper. at this point we can override
%parameters based on the chosen kernel or feature type
kernel.type = kernel_type;
features.gray = false;
features.hog = false;
padding = 1.5; %extra area surrounding the target 目标搜索范围系数，决定了每帧进行检测的范围
lambda = 1e-4; %regularization 正则化参数λ
output_sigma_factor = 0.1; %spatial bandwidth (proportional to target) 空间带宽：与目标大小成比例
switch feature_type
case 'gray',
interp_factor = 0.075; %linear interpolation factor for adaptation 时间更新系数
kernel.sigma = 0.2; %gaussian kernel bandwidth 高斯核函数的带宽/方差
kernel.poly_a = 1; %polynomial kernel additive term
kernel.poly_b = 7; %polynomial kernel exponent
features.gray = true;
cell_size = 1; 求HOG特征时cell的尺寸input
case 'hog',
interp_factor = 0.02;
kernel.sigma = 0.5;
kernel.poly_a = 1;
kernel.poly_b = 9;
features.hog = true;
features.hog_orientations = 9;
cell_size = 4;
otherwise
error('Unknown feature.')
end
assert(any(strcmp(kernel_type, {'linear', 'polynomial', 'gaussian'})), 'Unknown kernel.')
%在matlab中assert函数用来判断一个expression是否成立
%strcmp是用于做字符串比较的函数
switch video
case 'choose',
%ask the user for the video, then call self with that video name.
video = choose_video(base_path);
if ~isempty(video),
[precision, fps] = run_tracker(video, kernel_type, ...
feature_type, show_visualization, show_plots);
if nargout == 0, %don't output precision as an argument
clear precision
end
end
case 'all',
%all videos, call self with each video name.
%only keep valid directory names
dirs = dir(base_path);
videos = {dirs.name};
videos(strcmp('.', videos) | strcmp('..', videos) | ...
strcmp('anno', videos) | ~[dirs.isdir]) = [];
%the 'Jogging' sequence has 2 targets, create one entry for each.
%we could make this more general if multiple targets per video
%becomes a common occurence.
videos(strcmpi('Jogging', videos)) = [];
videos(end+1:end+2) = {'Jogging.1', 'Jogging.2'};
all_precisions = zeros(numel(videos),1); %to compute averages
all_fps = zeros(numel(videos),1);
if ~exist('matlabpool', 'file'),
%no parallel toolbox, use a simple 'for' to iterate
for k = 1:numel(videos),
[all_precisions(k), all_fps(k)] = run_tracker(videos{k}, ...
kernel_type, feature_type, show_visualization, show_plots);
end
else
%evaluate trackers for all videos in parallel
if matlabpool('size') == 0,
matlabpool open;
end
parfor k = 1:numel(videos),
[all_precisions(k), all_fps(k)] = run_tracker(videos{k}, ...
kernel_type, feature_type, show_visualization, show_plots);
end
end
%compute average precision at 20px, and FPS
mean_precision = mean(all_precisions);
fps = mean(all_fps);
fprintf('\nAverage precision (20px):% 1.3f, Average FPS:% 4.2f\n\n', mean_precision, fps)
if nargout > 0,
precision = mean_precision;
end
case 'benchmark',
%running in benchmark mode - this is meant to interface easily
%with the benchmark's code.
%get information (image file names, initial position, etc) from
%the benchmark's workspace variables
seq = evalin('base', 'subS');
target_sz = seq.init_rect(1,[4,3]);
pos = seq.init_rect(1,[2,1]) + floor(target_sz/2);
img_files = seq.s_frames;
video_path = [];
%call tracker function with all the relevant parameters
positions = tracker(video_path, img_files, pos, target_sz, ...
padding, kernel, lambda, output_sigma_factor, interp_factor, ...
cell_size, features, false);
%return results to benchmark, in a workspace variable
rects = [positions(:,2) - target_sz(2)/2, positions(:,1) - target_sz(1)/2];
rects(:,3) = target_sz(2);
rects(:,4) = target_sz(1);
res.type = 'rect';
res.res = rects;
assignin('base', 'res', res);
otherwise
%we were given the name of a single video to process.
%get image file names, initial state, and ground truth for evaluation
[img_files, pos, target_sz, ground_truth, video_path] = load_video_info(base_path, video);
%call tracker function with all the relevant parameters
[positions, time] = tracker(video_path, img_files, pos, target_sz, ...
padding, kernel, lambda, output_sigma_factor, interp_factor, ...
cell_size, features, show_visualization);
%calculate and show precision plot, as well as frames-per-second
precisions = precision_plot(positions, ground_truth, video, show_plots);
fps = numel(img_files) / time;
fprintf('%12s - Precision (20px):% 1.3f, FPS:% 4.2f\n', video, precisions(20), fps)
if nargout > 0,
%return precisions at a 20 pixels threshold
precision = precisions(20);
end
end
end
2 选择序列choose_video.m
输入base_path即benchmark序列库的路径，输出你选择的序列的名称（如’boy’，’Car4’等）
利用listdlg函数生成了一个列表GUI，方便进行选择
function video_name = choose_video(base_path)
%CHOOSE_VIDEO
% Allows the user to choose a video (sub-folder in the given path).
%
% Joao F. Henriques, 2014
% http://www.isr.uc.pt/~henriques/
%---------------------------------------------------------------------------------------------
%ispc用来判断当前电脑是不是Windows系统，是返回1，不是返回0
%process path to make sure it's uniform
if ispc(), base_path = strrep(base_path, '\', '/'); end
if base_path(end) ~= '/', base_path(end+1) = '/'; end
%list all sub-folders 列出所有子文件夹
contents = dir(base_path); %dir('G:\Matlab')列出指定目录下所有子文件夹和文件
names = {};
for k = 1:numel(contents),%numel()：返回数组或者向量中所含元素的总数。
name = contents(k).name;
if isdir([base_path name]) && ~any(strcmp(name, {'.', '..'})),%isdir用于判断输入是否表示一个文件夹
%any函数作用：判断元素是否为非零元素any(v),如果v是非零元素返回true(即1)否则返回flase(即0)
names{end+1} = name; %#ok
end
end
%no sub-folders found 没有找到子文件夹
if isempty(names), video_name = []; return; end
%isempty(names)????判断names是否为空，如果为空，结果为1，否则为0.
%choice GUI
choice = listdlg('ListString',names, 'Name','Choose video', 'SelectionMode','single');
%'Name','Choose video',代表的是选择框的名字是Choose video
%listdlg列表选择对话框
if isempty(choice), %user cancelled
video_name = [];
else
video_name = names{choice};
end
end
3 提取序列中目标的参数load_video_info.m
- 输入base_path和通过choose_video选择得到的序列名video_name，得到所选序列的存储路径。
video_path = [base_path video '/'];
- 并且打开benchmark中预留的groundtruth_rect.txt文件，将序列中目标的参数存储到ground_truth中。
filename = [video_path 'groundtruth_rect' suffix '.txt'];
f = fopen(filename);
- ground_truth是一个矩阵，行数为序列的帧数。开始时列数为4，第三列和第四列为目标跟踪框的宽和高，第一列和第二列为目标跟踪框左上角的列坐标和行坐标（这样我猜大概是因为imcrop函数中行和列的位置是反过来的吧）。跟踪框左上角坐标再加上其二分之一的边长后由4列变为2列，表示目标的中心位置轨迹。
将第一帧的参数作为初始化时的参数target_sz和pos（有4个序列作者应该是出于跟踪精度的需要修改了参数，暂且不表）。
target_sz = [ground_truth(1,4), ground_truth(1,3)];%目标大小(高，宽）
pos = [ground_truth(1,2), ground_truth(1,1)] + floor(target_sz/2);%中心点坐标
ground_truth = ground_truth(:,[2,1]) + ground_truth(:,[4,3]) / 2;%4列变为2列
- img_files按顺序存储了所有的序列图片的文件名，如’0001.jpg’（图片不必设置为0001.jpg，0002.jpg…0030.jpg这种格式，即使是从一个完整序列中截取的片段，也可以顺利读取，比如0031.jpg,0033.jpg…0080.jpg），长度为图片数。
如果要加入自己采集的序列，最好将名称都改成“xxxx.jpg”的格式，并且要放在一个叫img的子文件夹下，比如这个路径：“F:\track\Benchmark\IRA\img\0001.jpg”。
video_path = [video_path 'img/'];%图片应放在video_path路径下的img子文件夹中
img_files = dir([video_path '*.jpg']);
- 可以自己仿照benchmark里的groundtruth_rect.txt进行标注，
比如“F:\track\Benchmark\IRA\groundtruth_rect.txt”。不标注的话就需要对代码做一些修改，使其可以顺利读入无标注的序列，并通过imcrop的方式初始化跟踪窗口。
详细代码如下：
function [img_files, pos, target_sz, ground_truth, video_path] = load_video_info(base_path, video)
%LOAD_VIDEO_INFO
% Loads all the relevant information for the video in the given path:
% the list of image files (cell array of strings), initial position
% (1x2), target size (1x2), the ground truth information for precision
% calculations (Nx2, for N frames), and the path where the images are
% located. The ordering of coordinates and sizes is always [y, x].
%
% Joao F. Henriques, 2014
% http://www.isr.uc.pt/~henriques/
%see if there's a suffix, specifying one of multiple targets, for
%example the dot and number in 'Jogging.1' or 'Jogging.2'.
if numel(video) >= 2 && video(end-1) == '.' && ~isnan(str2double(video(end))),
suffix = video(end-1:end); %remember the suffix
video = video(1:end-2); %remove it from the video name
else
suffix = '';
end
%full path to the video's files
if base_path(end) ~= '/' && base_path(end) ~= '\',
base_path(end+1) = '/';
end
video_path = [base_path video '/'];
%try to load ground truth from text file (Benchmark's format)
filename = [video_path 'groundtruth_rect' suffix '.txt'];
f = fopen(filename);
assert(f ~= -1, ['No initial position or ground truth to load ("' filename '").'])
%the format is [x, y, width, height]
try
ground_truth = textscan(f, '%f,%f,%f,%f', 'ReturnOnError',false);
catch %#ok, try different format (no commas)
frewind(f);
ground_truth = textscan(f, '%f %f %f %f');
end
ground_truth = cat(2, ground_truth{:});
fclose(f);
%set initial position and size
target_sz = [ground_truth(1,4), ground_truth(1,3)];
pos = [ground_truth(1,2), ground_truth(1,1)] + floor(target_sz/2);
if size(ground_truth,1) == 1,
%we have ground truth for the first frame only (initial position)
ground_truth = [];
else
%store positions instead of boxes
ground_truth = ground_truth(:,[2,1]) + ground_truth(:,[4,3]) / 2;
end
%from now on, work in the subfolder where all the images are
video_path = [video_path 'img/'];
%for these sequences, we must limit ourselves to a range of frames.
%for all others, we just load all png/jpg files in the folder.
frames = {'David', 300, 770;
'Football1', 1, 74;
'Freeman3', 1, 460;
'Freeman4', 1, 283};
idx = find(strcmpi(video, frames(:,1)));
if isempty(idx),
%general case, just list all images
img_files = dir([video_path '*.png']);
if isempty(img_files),
img_files = dir([video_path '*.jpg']);
assert(~isempty(img_files), 'No image files to load.')
end
img_files = sort({img_files.name});
else
%list specified frames. try png first, then jpg.
if exist(sprintf('%s%04i.png', video_path, frames{idx,2}), 'file'),
img_files = num2str((frames{idx,2} : frames{idx,3})', '%04i.png');
elseif exist(sprintf('%s%04i.jpg', video_path, frames{idx,2}), 'file'),
img_files = num2str((frames{idx,2} : frames{idx,3})', '%04i.jpg');
else
error('No image files to load.')
end
img_files = cellstr(img_files);
end
end