学步园

matlab版源程序：http://download.csdn.net/detail/cv_richie/4982673

C程序：http://download.csdn.net/detail/cv_richie/4982678

晚点我会把我注释的版本发上来。

2. tldInit.m

function tld = tldInit(opt,tld)

lk(0);

if ~isempty(tld);
    handle = tld.handle;
    tld = opt;
    tld.handle = handle;
else
    tld = opt;
end


% INITIALIZE DETECTOR =====================================================

% Scanning grid 创建扫描网格 tld.grid保存网格坐标 tld.scales保存网格尺度
[tld.grid tld.scales] = bb_scan(tld.source.bb,size(tld.source.im0.input),tld.model.min_win); %6*所有不同尺度下的网格数 1-4行为坐标 5尺度标号 6尺度下的网格列数标号

% Features 
tld.nGrid     = size(tld.grid,2);   %tld.nGrid保存网格数量
tld.features  = tldGenerateFeatures(tld.model.num_trees,tld.model.num_features,0);  %初始化FERN中的pixel comparison特征对tld.featrue.x为52*10 tld.feature.type='forest'

% Initialize Detector
fern(0); % cleanup 清除FERN分类器
fern(1,tld.source.im0.input,tld.grid,tld.features,tld.scales); % allocate structures 初始化ensemble分类器

% Temporal structures
tld.tmp.conf = zeros(1,tld.nGrid);
tld.tmp.patt = zeros(tld.model.num_trees,tld.nGrid);  %10*网格数

% RESULTS =================================================================

% Initialize Trajectory   初始化轨迹信息空间
tld.img             =   cell(1,length(tld.source.idx));  %CELL生成的是元胞
tld.snapshot     =   cell(1,length(tld.source.idx));
tld.dt                =   cell(1,length(tld.source.idx));
tld.bb               =   nan(4,length(tld.source.idx));  %nan生成的是空数组
tld.conf            =   nan(1,length(tld.source.idx));
tld.valid            =   nan(1,length(tld.source.idx));
tld.size              =   nan(1,length(tld.source.idx));
tld.trackerfailure = nan(1,length(tld.source.idx));
tld.draw    = zeros(2,0);
tld.pts     = zeros(2,0);
% Fill first fields
tld.img{1}  = tld.source.im0;
tld.bb(:,1) = tld.source.bb;
tld.conf(1) = 1;
tld.valid(1)= 1;
tld.size(1) = 1;

% TRAIN DETECTOR ==========================================================

% Initialize structures
tld.imgsize = size(tld.source.im0.input);  %tld.imgsize保存图像大小
tld.X       = cell(1,length(tld.source.idx)); % training data for fern 训练ensemble的样本
tld.Y       = cell(1,length(tld.source.idx)); 
tld.pEx     = cell(1,length(tld.source.idx)); % training data for NN 训练最近邻的样本
tld.nEx     = cell(1,length(tld.source.idx));
overlap     = bb_overlap(tld.source.bb,tld.grid); % bottleneck 返回与bb与每个网格的重叠比

% Target (display only)
tld.target = img_patch(tld.img{1}.input,tld.bb(:,1));%分割出目标区域的图像

% Generate Positive Examples产生正样本
[pX,pEx,bbP] = tldGeneratePositiveData(tld,overlap,tld.img{1},tld.p_par_init); %PX输出前十网格的特征，每行是一个网格，10列是一层随机；pex最大网格归一化图像 bbp最大网格的坐标
pY = ones(1,size(pX,2));%正样本的正标号
% disp(['# P patterns: ' num2str(size(pX,2))]);
% disp(['# P patches : ' num2str(size(pEx,2))]);

% Correct initial bbox
tld.bb(:,1) = bbP(1:4,:);%最大重叠度的网格 程序中把制定框转化为重叠度最大的网格

% Variance threshold
tld.var = var(pEx(:,1))/2; %方差分类器阈值为最大重叠度网格方差的的一半
% disp(['Variance : ' num2str(tld.var)]);

% Generate Negative Examples
[nX,nEx] = tldGenerateNegativeData(tld,overlap,tld.img{1});%重叠度小于0.2的网格中产生
% disp(['# N patterns: ' num2str(size(nX,2))]);
% disp(['# N patches : ' num2str(size(nEx,2))]);

% Split Negative Data to Training set and Validation set
[nX1,nX2,nEx1,nEx2] = tldSplitNegativeData(nX,nEx);%负样本随机平均分为两组
nY1  = zeros(1,size(nX1,2));%负样本的负标号

% Generate Apriori Negative Examples
%[anX,anEx] = tldGenerateAprioriData(tld);
%anY = zeros(1,size(anX,2));
% disp(['# apriori N patterns: ' num2str(size(anX,2))]);
% disp(['# apriori N patches : ' num2str(size(anEx,2))]);

tld.pEx{1}  = pEx; % save positive patches for later
tld.nEx{1}  = nEx; % save negative patches for later
tld.X{1}    = [pX nX1]; 
tld.Y{1}    = [pY nY1]; 
idx         = randperm(size(tld.X{1},2));
tld.X{1}    = tld.X{1}(:,idx);%随机化后的初始样本集
tld.Y{1}    = tld.Y{1}(:,idx);%随机化后的初始标号集

% Train using training set ------------------------------------------------

% Fern
bootstrap = 2;
fern(2,tld.X{1},tld.Y{1},tld.model.thr_fern,bootstrap);  %thr_fern为0.5训练初始随机厥分类器

% Nearest Neightbour 

tld.pex = [];
tld.nex = [];

tld = tldTrainNN(pEx,nEx1,tld);%训练初始NN分类器
tld.model.num_init = size(tld.pex,2);%训练后的正样本数

% Estimate thresholds on validation set  ----------------------------------

% Fern
conf_fern = fern(3,nX2);%输出为输入样本的分类结果即正样本可信度,可信度越高,分类效果不好,需要更新
tld.model.thr_fern = max(max(conf_fern)/tld.model.num_trees,tld.model.thr_fern);%该阈值影响后面的学习,越大则需要更新的样本越多

% Nearest neighbor
conf_nn = tldNN(nEx2,tld);
tld.model.thr_nn = max(tld.model.thr_nn,max(conf_nn));%该阈值越大，则需要被添加的模型正样本越多
tld.model.thr_nn_valid = max(tld.model.thr_nn_valid,tld.model.thr_nn);%跟踪时用到的有效性阈值