Video analysis: problem finding the vertical mean of edges

I have a video file which shows an object that is fluctuating up and down along its horizontal length. I want to divide each frame into regions of 50 horizontal pixels so that I can find a single point in each region that is horizontally in the centre of the region and vertically in the centre of the upper and lower edges of the object. I have used edge detection to find the top and bottom edges and I am using a loop to find the centres of the object in the regions. The problem: I have used “mean” to find the vertical centroid of the upper and lower object edges in each region, y_m. However y_m for some bizarre reason always outputs a value very close to the bottom edge of the object, usually below the bottom edge. I have no idea why as there are only 1s (1s for the edges, 0s for everything else) along the edges in each region so the mean should at least be between the upper and lower edges. Could anyone help? Thank you in advance.

Source = vision.VideoFileReader('video.mp4',... 'ImageColorSpace', 'Intensity');

Edge = vision.EdgeDetector('Method', 'Canny', ... 'ThresholdSource', 'Auto',... 'NonEdgePixelsPercentage', 98.3);

VideoEdges = vision.VideoPlayer('Name', 'Edges'); VideoEdges.Position = [700 50 600 400];

F=1; while ~isDone(Source)

    frame     = step(Source);      
    edges     = step(Edge, frame);     

    w=50; %width of regions, 60 pixels.
    N=1920/w; %number of regions of width w.
    for n=1:N
        for s=1:w
            for j=300:800
                for Rni=w*(n-1)+s
                    if edges(j,Rni)==1

                        y_m(F,n)=mean(j);

                    end
                end
            end
            x_m(F,n)=Rni+w/2;     %centre of each individual region
        end
        F=F+1; %for every frame individually
    end
step(VideoEdges, edges);       
end

release(Source);

[Image Analyst]

It's hard to visualize the problem. Can you attach a screenshot or image or something?

Answer 1

[Image Analyst]

What are you trying to find and measure? The black snake-like thing? If so, why are you using edge detection???? And if so, what do you want to find out about it? Its centroid, length, area, perimeter, mean intensity, what??? Also why do you have what looks like bad reflections? Why not get rid of those with a better image capture setup (e.g. crossed polarizers or whatever)?

[Image Analyst]

Well you definitely can improve your image capture situation with the right geometry, specifically, where the light is located. And by polarizing the incident light and using a rotating polarizer in front of the lens, and by turning off all other room lights, you can eliminate the spurious reflections.

Now I haven't seen any other images but that one but it looks like that leaf is darker than the background. If that is always the case then there is no reason at all to do edge detection . That's the kind of thing I see promoted by novices who don't know any better. Edge detection can present problems, like broken edges, edges that don't align with the edge of the object, double edge (edges on both sides of the edge), etc. Unless you have the situation were the leaf turns almost invisible or transparent so that it is the same intensity as that of the background, then I see no reason to use it. You can use simple intensity thresholding to find the leaf object that I see. It will be more accurate and reliable.

binaryImage = grayImage < threshold;
labeledImage = bwlabel(binaryImage);
measurements = regionprops(labeledImage, grayImage, 'all');

See my Image Segmentation Tutorial http://www.mathworks.com/matlabcentral/fileexchange/?term=authorid%3A31862 for a full demo.

[Image Analyst]

I think that looks right. What release of MATLAB do you have? If it's earlier than 2008a you can't use grayImage in the regionprops call.

[Image Analyst]

You don't need a loop over threshold. Just set it equal to 100 - no for loop needed. I don't know why it doesn't like all. Try getting rid of it and see what happens.

[Image Analyst]

Attach your m-file and whatever data I need to run it.

[Image Analyst]

Upload just one frame as a PNG file then.

[Image Analyst]

A video is just a series of frames (individual images) so why wouldn't it work for the video?