Hi, you should be able to use this method to do the undistortion. I do not quite understand your issue with the detection of particles (which articles?).

http://www.ipol.im/pub/art/2011/ags-alde/revisions/2010-07-28/article.pdf

Regards, Changming Sun

you can try tracking moving objects using optical flow maps from frame to frame, kalman filter (or non parametric camshift), you can find matlab or opencv code for all this by simply searching 

Thanks for the comment dear Wolfgang Reichl. That is helpful.

the exposed ball (which is really a dark blue one) needs to be detected in each frame since it moves through the channel from right to left.

the background subtraction is hard since the distinguishing between the exposed ball and the bed spheres (black and white spheres) is hard. 

could the techniques which you've mentioned rewrite and organized in an algorithm of image correction?

my first problem is the correction of the frames considering having no calibrating images, just know some measures.

thanks

Thanks Dr. Sun and Dr. Ramachandra for your reply.

Dr. Sun thanks for te article and the tool you recommended. I will try it.

Dr. Ramachandra can I track this object with optical flow techniques. I haven't used it before. I thought I could detect the exposed ball (in image attached) by finding a circle around the ball with hough transform method then find the space in different frames so I can find the particle velocity at the end.

Dear Wolfgang. for subtraction, there's a problem of light intensity variation (illumination) in the background frame and other frames (which needed to be studied). when I convert the frame to grayscale the foreground (which is the ball) is not detectable at all. then if I want to binarize it, use thresholding, it's not possible to detect the ball. I also used "decorrstretch" command. which just made it easier for the ball to be detected. but I can't binarize it since the gray scale of the image is also not good to work on it.

Hello Hamed,

converting the frame to grayscale and subsequent thresholding doesn't work because the brightness of the ball is similar to that of the background. However, the color is different, so you could try to do a HSV transform (hue, saturation, value) and perform a segmentation with respect to the hue channel (i.e. detecting the blue of the ball). See the result in the attachment that was achieved using your sample image with a hue interval from 0.62 to 0.78, a minimum saturation of 0.15, and a minimum brightness/value of 0.2. Subsequent morphological operations (opening, closing) and the definition of a region of interest could lead to a quite robust detection of the blue ball.

Good luck and best regards, Ralf

Hello dear Ralf,

Thanks for your answer. I appreciate it.

Yes I agree on what you have mentioned about grayscale image.

I did this to the image by "decorrstretch" command in Matlab (attached). Is the image you have attached grayscale? because it's black&white.

the problem is If I don't convert it to grayscale. Then I can't binarize it and detect it.

Many thanks

take a look ar this morphological operations(imopen/imclose)

I=imread('frame3333.png');

bg=imread('frame1595.png'); %background frame

D=I-bg;

bmat=D(:,:,3);  %blue works

se=strel('disk',5);

I2=imopen(bmat,se);

I3=imclose(bmat,se);

Simply taking the blue channel is not sufficient for a good segmentation, you need to take into account all color channels for calculating the hue/saturation/brightness and filtering out the 'right' blue pixels before applying the morph operators. The image I attached is actually a grayscale image with 255 for the segmented pixels and 0 for the background. I'm not familiar how hue filtering is done in Matlab (I don't use Matlab) but I'm quite sure that it is possible.

Thanks. you're right.  in the attached picture you have sent the ball is not there. is there?

What do you use for image processing?

Yes, the ball is there, see the big white spot in the lower middle of the right half. I took your first sample image for the segmentation. 

I'm using a proprietary software we developed in our research group. But the steps I used are quite generic, so Matlab should provide this type of color segmentation as well.

You could try denoising autoencoders:

https://www.academia.edu/25585807/Autoencoding_Video_Frames

http://eric-yuan.me/dae/

Good comments and suggestions already made and want to add:

1. Can you make more than one image available to work with?  Having two images where the particle / ball is in different positions will allow change detection methods to be applied while working with only one image spectral and spatial methods need to be used rather than change detection methods.

2. If the ball is blue and the background either black or white perhaps using the ratio of blue to green or to red might be helpful.

Pat

Dear Pat S. Chavez, Jr.

Thanks for your response. You want me to add other pictures?

I have 5674 frames from the video. I really believed now that PTV has some big uncertainties because the ball detection in these frames are not quite exact.

in the first frames light, which is there from the start, doesn't let us to have a well detected ball!. maybe I need to do it manually!

in other frames which ball is in the middle, the ball's shadow also detected as an edge.

I attached some of the image processing I've done to detect the ball in different frames.

ِDear Wolfgang Reichl,

That's a lot useful thanks. I have problems with light. in the first frames the ball is surrounded by light so the ball can't properly be detected. and in the middle frames, the shadow of the ball makes some uncertainty in the ball detection.

in some frames, the ball detection is fine like the pic I attached. but in some no I can't detect it with precision.

the detection image I mentioned in the previous response is here

Thanks. That's a nice job you've done. but as you see in the image I attached the white dot you detected is not the whole ball. so I will get the false centeroid.

one more thing there's a problem with shadow, How can I remove it from detection. I tried to normalize it but hadn't worked

The problem with your image is that parts of the ball are in the shadow, too. If you have the chance to install a second light source on the left side for a better illumination the processing based on hue and saturation as shown by Wolfgang Reichl would be quite robust with appropriate threshold parameters.