如何使用 for 循环在 Matlab 中处理大型视频且没有内存错误

Question

我是 Matlab 处理的新手，我想在“for 循环”（或不使用它）内读取和处理大型视频（超过 200k 帧）。特别是，我想：

1. 使用 VideoReader 阅读视频，
2. 将视频细分为 n-epoch，每帧 1000 帧，
3. 处理每个 epoch 1000 帧，读取：epoch 的第一帧，跳过两帧，读取帧，跳过两帧，以此类推（例如 i=1:3:nFrames），
4. 考虑到我需要将读取的每个“RGB 帧”转换为 im2bw 的每个时期
5. 转换后，我需要考虑第一个视频帧（“mov(1,1).cdata”）和在该时期内读取的每个帧，进行“corr2”2D 互相关，
6. 将“corr2”的结果存储到向量中。

总之，这就是我需要做的。谢谢大家

---

这是我目前所拥有的，关于“corr2”：

for frame_ind = 1 : nFrames
  mov(frame_ind).cdata = im2bw(rgb2gray(read(xyloObj,frame_ind)),0.20);      
end

%% Corr2 to compare BW video frames
for frame_ind2 = 1:(frame_ind-1)
    R(frame_ind2)=corr2(mov(1,frame_ind2).cdata,mov(1,frame_ind2+1).cdata);
end

 TF= isnan(R); 
 g=sum(TF);
 f=(length(R)-g);


if (g~=(length(R))) 
  %%If Part has errors 
  disp('"Part_1" has video interferences/noise/problems, see "Testresult.txt" for more information.');
 else 
  %%If Part has not errors  
  displ=strcat('"Part_1" has not video interferences/noise/problems.');
 end

Answer 1

这是我的版本：

mov = VideoReader('movie.avi');
nFrames = mov.NumberOfFrames;

len = 1000;     %# epoch length
step = 3;       %# step size

%# indices of each epoch
indices = bsxfun(@plus, 1:step:len, (0:ceil(nFrames/len-1))'*len);   %#'
indices = num2cell(indices,2);
indices{end}(indices{end}>nFrames) = [];

%# loop over each epoch
corr_coef = cell(size(indices));
for e=1:numel(indices)
    %# read first image in epoch
    img1 = read(mov, indices{e}(1));
    img1 = rgb2gray(img1);            %# instead of im2bw(img1, graythresh(img1))

    %# read rest of images in epoch
    corr_coef{e} = zeros(1,numel(indices{e})-1);
    for f=2:numel(indices{e})
        img2 = read(mov, indices{e}(f));
        img2 = rgb2gray(img2);

        %# compute corr2 between the two images
        corr_coef{e}(f-1) = corr2(img1,img2);
    end
end

元胞数组corr_coef包含每个时期的相关系数，其中每个元胞在第一帧和第(i+1)帧之间包含corr2的向量corr_coef{e}(i)。

请注意，如果其中一帧是恒定的（例如全黑），则 2D 相关系数只是 NaN（formula‌​ 中的零除以零）

Answer 2

% create video handle and get number of frames
vidObj = VideoReader(video_file);
nFrames = get(vidObj, 'NumberOfFrames');

blocksize = 1000;
BlocksIDs = 1:blocksize:nFrames;
nBlocks = numel(BlocksIDs);

frame_step = 3;

% cell array with all correlations values grouped by block
xcorrs_all_blocks = cell(1, nBlocks);

for j = 1 : nBlocks

    % if this is the last block, process until the last frame
    if j == nBlocks
        last_frame = nFrames;
    % otherwise, read until next block
    else
        last_frame = BlocksIDs(j + 1);
    end
    % compute the frame numbers that we want to look at
    frame_indices = BlocksIDs(j) : frame_step : last_frame;
    nFrames = numel(frame_indices);

    first_frame = [];
    % pre-allocate array holding the in-block corr2 values.
    xcorrs_this_block = nan(1, nFrames-1);

    for k = 1 : nFrames

        % read-in raw frame from video file.
        raw_frame = read(vidObj, frame_indices(k));            
        % determine level for bw conversion - this might help.
        level = graythresh(raw_frame);
        raw_frame = im2bw(raw_frame, level);

        if k == 1
            % save current frame as first frame for later corr2 processing
            first_frame = raw_frame;
        else
            % calc the correlation between the first frame in the block and each successive frame
            xcorrs_this_block(k-1) = corr2(first_frame, raw_frame);
        end


    end

    % save all xcorr values into global cell array.
    xcorrs_all_blocks{j} = xcorrs_this_block;

end