Classify blobs vs clutter based on motion

Hi Folks,

I am looking to classify blobs representing objects of 'interest'. I would like to detect these objects of interest. These objects are in very regular motion. However they are mixed up with rest of the clutter blobs in the image. I would like to classify all blobs in two categories a) objects of interest b) clutter. They clutter, does not undergo motion like that of objects which I would like to detect. Please look at the attached image.

Is there a way I can take all potential blobs in the image and curve fit them - blobs with close to least error from the curve would be objects of my interest ? What type of curve(s) can I fit ? How do I go about fitting the curve ? Should I just put centroids of all blobs in an array and pass that to fitobject = fit(x,y,fitType) ? How would I classify object ? After curve fit ?

Answer 1

[Image Analyst]

Use optical flow, in the computer Vision System Toolbox.

[Ankit Chourasia]

Hi ImageAnalyst I am not an expert. I am learning up optical flow, upon your advice, could you please write few more lines to give me a start ? Which specific function to try (e.g. should I try Lucas Kanade )?

[Image Analyst]

Sorry - I don't have that toolbox. You'll just have to look at their examples.

http://www.mathworks.com/discovery/optical-flow.html

http://www.mathworks.com/matlabcentral/fileexchange/?sort=downloads_desc&term=tag%3A%22optical+flow%22

Those people have written far more lines on it than I ever could.

Answer 2

[Ankit Chourasia]

Thanks Image Analyst for your pointers. I implemented LK method, on successive frames and got optical flow motion vectors. Please see attached image. However, I am not able to figure out how to use these vectors to reject the "outliers".

Intutively, I feel that lot of these objects of interest follow regular parabolic motion. Wouldn't fitting curve work better ? After I fit a parabolic curve, can I not detect the outliers from that curve ?

[Image Analyst]

Sounds reasonable. Give it a try.

Answer 3

[Ankit Chourasia]

Forgot to attach the image.