查找两个视频之间匹配的 SIFT 特征

Question

我从两个视频中提取了 SIFT 特征进行匹配。 我需要将第二个视频的每个特征与存储在第一个视频的一系列特征中的特征进行比较。我在设置代码时遇到了麻烦，因此当我有对应关系时，我可以获得该功能所在的框架。我能怎么做？谁能给我一个代码示例？

这是我的代码：

obj = VideoReader('video2.avi');
lastFrame = read(obj, inf);
numFrames = obj.NumberOfFrames;

%estrazione frame e sift
for k = 1 : 3 % numFrames / 5
    disp(['Processing frame #', num2str(k)]);
    this_frame = read(obj, k * 5); % leggi solo un fotogramma ogni 5 per velocizzarle la cosa
    this_frame = imresize(this_frame, 0.5); % rimpiccioliamolo per questioni di efficienza!
    I = single(rgb2gray(this_frame)) ;
    [f,d] = vl_sift(I);  % estrazione feature

    features{k} = f;     % salviamo le feautre e i relativi descrittori in delle celle
    descriptors{k} = d;
end
save('feature_input', 'features');
save('descrittori_input', 'descriptors');

%%% un esempio di come ripescare i dati...
pippo = load('feature_input');
newfeat = pippo.features;
pippo = load('descrittori_input');
newdesc = pippo.descriptors;

for k = 1 : 3
    disp(['Le feature del fotogramma', num2str(k), ' sono: ']);
    f = cell2mat( newfeat(k) );
    f(:, 1:10) % ne mostriamo solo un pezzetto... le posizioni delle prime 10 features
end


obj2 = VideoReader('video2u.avi');
lastFrame = read(obj2, inf);
numFrames = obj2.NumberOfFrames;



%estrazione frame e sift video2
for k2 = 1 : 3 % numFrames / 5
    disp(['Processing frame #', num2str(k2)]);
    this_frame2 = read(obj2, k2 * 5); % leggi solo un fotogramma ogni 5 per velocizzarle la cosa
    this_frame2 = imresize(this_frame2, 0.5); % rimpiccioliamolo per questioni di efficienza!
    K = single(rgb2gray(this_frame2)) ;
    [f2,d2] = vl_sift(K);  % estrazione feature

    features2{k2} = f2;     % salviamo le feautre e i relativi descrittori in delle celle
    descriptors2{k2} = d2;
end




save('feature2_input', 'features2');
save('descrittori2_input', 'descriptors2');

%%% un esempio di come ripescare i dati...
pippo2 = load('feature2_input');
newfeat2 = pippo2.features2;
pippo2 = load('descrittori2_input');
newdesc2 = pippo2.descriptors2;

for k2 = 1 : 3
    disp(['Le feature del fotogramma', num2str(k2), ' sono: ']);
    f2 = cell2mat( newfeat2(k2) );
    f2(:, 1:10) % ne mostriamo solo un pezzetto... le posizioni delle prime 10 features


end


[matches, scores] = vl_ubcmatch(d, d2, 1.5) ;

% sift points plot

    subplot(1,2,1);
    imshow(uint8(I));
    hold on;
    plot(f(1,matches(1,:)),f(2,matches(1,:)),'b*');


    subplot(1,2,2);
    imshow(uint8(K));
    hold on;
    plot(f2(1,matches(2,:)),f2(2,matches(2,:)),'r*');


    figure;

     %-------------  

 % RANSAC

X1 = f(1:2,matches(1,:)) ; X1(3,:) = 1 ;
X2 = f2(1:2,matches(2,:)) ; X2(3,:) = 1 ;


numMatches = size(matches,2) ;

for t = 1:100
  % estimate homograpyh
  subset = vl_colsubset(1:numMatches, 4) ;
  A = [] ;
  for i = subset
    A = cat(1, A, kron(X1(:,i)', vl_hat(X2(:,i)))) ;
  end


  [U,S,V] = svd(A) ;


H{t} = reshape(V(:,9),3,3) ;

  % score homography
  X2_ = H{t} * X1 ;
  du = X2_(1,:)./X2_(3,:) - X2(1,:)./X2(3,:) ;
  dv = X2_(2,:)./X2_(3,:) - X2(2,:)./X2(3,:) ;
  ok{t} = (du.*du + dv.*dv) < 6*6 ;
  score(t) = sum(ok{t}) ;
end



[score, best] = max(score) ;
H = H{best};
ok = ok{best};


% sift feature matching 

   dh1 = max(size(K,1)-size(I,1),0) ;
   dh2 = max(size(I,1)-size(K,1),0) ;


subplot(2,1,1) ;
imagesc([padarray(I,dh1,'post') padarray(K,dh2,'post')]) ;
 colormap (gray);
o = size(I,2) ;
line([f(1,matches(1,:));f2(1,matches(2,:))+o], ...
     [f(2,matches(1,:));f2(2,matches(2,:))]) ;


axis image off ;

subplot(2,1,2) ;
imagesc([padarray(I,dh1,'post') padarray(K,dh2,'post')]) ;
 colormap (gray);
o = size(I,2) ;
line([f(1,matches(1,ok));f2(1,matches(2,ok))+o], ...
     [f(2,matches(1,ok));f2(2,matches(2,ok))]) ;
title(sprintf('%d (%.2f%%) inliner matches out of %d', ...
              sum(ok), ...
              100*sum(ok)/numMatches, ...
              numMatches)) ;
axis image off ;

drawnow ;

end

Answer 1

您正在使用[matches, scores] = vl_ubcmatch(d, d2, 1.5) ; 进行匹配。 d 仅包含最新帧的描述符。你应该这样做：

for nFrames=1:3
   [matches{nFrames}, scores{nFrames}] = vl_ubcmatch(descriptors{nFrames}, descriptors2{nFrames}, 1.5);
end

从这里，您应该能够获得帧之间的匹配。