Point Tracking using Optical Flow - c#

Hello I'm trying to apply point tracking to a scene.
Now I want to get the points only moving in horizontally. Anyone have any thoughts on this?
The arrays "Actual" and "nextfeature" contain the relevant x,y coordinates. I tried to get the difference from the two arrays, it did not work. I tried to get the optical flow using Farneback but it didn't gave me a satisfying result. I would really appreciate if anyone can give me any thoughts on how to get the points only moving in horizontal line.
thanks.
Here is the code.
private void ProcessFrame(object sender, EventArgs arg)
{
PointF[][] Actual = new PointF[0][];
if (Frame == null)
{
Frame = _capture.RetrieveBgrFrame();
Previous_Frame = Frame.Copy();
}
else
{
Image<Gray, byte> grayf = Frame.Convert<Gray, Byte>();
Actual = grayf.GoodFeaturesToTrack(300, 0.01d, 0.01d, 5);
Image<Gray, byte> frame1 = Frame.Convert<Gray, Byte>();
Image<Gray, byte> prev = Previous_Frame.Convert<Gray, Byte>();
Image<Gray, float> velx = new Image<Gray, float>(Frame.Size);
Image<Gray, float> vely = new Image<Gray, float>(Previous_Frame.Size);
Frame = _capture.RetrieveBgrFrame().Resize(300,300,Emgu.CV.CvEnum.INTER.CV_INTER_AREA);
Byte []status;
Single[] trer;
PointF[][] feature = Actual;
PointF[] nextFeature = new PointF[300];
Image<Gray, Byte> buf1 = new Image<Gray, Byte>(Frame.Size);
Image<Gray, Byte> buf2 = new Image<Gray, Byte>(Frame.Size);
opticalFlowFrame = new Image<Bgr, Byte>(prev.Size);
Image<Bgr, Byte> FlowFrame = new Image<Bgr, Byte>(prev.Size);
OpticalFlow.PyrLK(prev, frame1, Actual[0], new System.Drawing.Size(10, 10), 0, new MCvTermCriteria(20, 0.03d),
out nextFeature, out status, out trer);
for (int x = 0; x < Actual[0].Length ; x++)
{
opticalFlowFrame.Draw(new CircleF(new PointF(nextFeature[x].X, nextFeature[x].Y), 1f), new Bgr(Color.Blue), 2);
}
new1 = old;
old = nextFeature;
Actual[0] = nextFeature;
Previous_Frame = Frame.Copy();
captureImageBox.Image = Frame;
grayscaleImageBox.Image = opticalFlowFrame;
//cannyImageBox.Image = velx;
//smoothedGrayscaleImageBox.Image = vely;
}
}

First... I can only give you a general idea about this, not a code snippet...
Here's how you may do this:
(One of the many possible approaches of tackling this problem)
Take the zero-th frame and pass it through goodFeaturesToTrack. Collect the points in an array ...say, initialPoints.
Grab the (zero + one) -th frame. With respect to the points grabbed from step 1, run it through calcOpticalFlowPyrLK. Store the next points in another array ...say, nextPoints. Also keep track of status and error vectors.
Now, with initialPoints and nextPoints in tow, we leave the comfort of openCV and do things our way. For every feature in initialPoints and nextPoints (with status set to 1 and error below an acceptable threshold), we calculate the gradient between the points.
Accept only those point for horizontal motion whose angle of slope is either around 0 degrees or 180 degrees. Now... vector directions won't lie perfectly at 0 or 180... so take into account a bit of +/- threshold.
Repeat step 1 to 4 for all frames.
Going through the code you posted... it seems like you've almost nailed steps 1 and 2.
However, once you get the vector nextFeature, it seems like you're drawing circles around it. Interesting ...but not what we need.
Check if you can go about implementing the gradient calculation and filtering.

Related

Calculate and write DFT, discrete Fourier transformation with Emgu.CV Cuda

I want very basically to run a Fourier transformation on an image, using the Emgu.CV (4.4) library.
For the none-Cuda code, it seams to do something, going into the right direction with this code:
using Emgu.CV;
using Emgu.CV.Structure;
using Emgu.CV.Cuda;
private static void StartDefault()
{
Image<Gray, float> image = new Image<Gray, float>(#"C:\temp\emgucv\original_4096.jpg");
var forward = new Image<Gray, float>(image.Rows, image.Cols);
var inverse = new Image<Gray, float>(image.Rows, image.Cols);
CvInvoke.Dft(image, forward, Emgu.CV.CvEnum.DxtType.Forward, image.Rows);
CvInvoke.Dft(forward, inverse, Emgu.CV.CvEnum.DxtType.Inverse, image.Rows);
forward.Save(#"C:\temp\emgucv\original_4096_dft_fw.jpg");
inverse.Save(#"C:\temp\emgucv\original_4096_dft_in.jpg");
}
Now I tried to run the same using Cuda (in this case only forward to keep things simple).
private static void Start1Cuda()
{
Image<Gray, float> image = new Image<Gray, float>(#"C:\trash\emgucv\original_128.jpg");
CudaImage<Gray, float> cudaImage = new CudaImage<Gray, float>(image.Height, image.Width);
cudaImage.Upload(image);
// channels 2 here, else it crashes on .Dft with "is_complex_input || is_complex_output"
GpuMat resultMatForward = new GpuMat(cudaImage.Size.Height, cudaImage.Size.Width, cudaImage.Depth, channels: 2, true);
CudaInvoke.Dft(image, resultMatForward, image.Size, Emgu.CV.CvEnum.DxtType.Forward);
Image<Gray, float> forward = new Image<Gray, float>(image.Height, image.Width);
resultMatForward.Download(forward);
forward.Save(#"C:\trash\emgucv\original_128_dft.jpg");
}
But this crashes at the very last line, trying to save the result. Exeption: Emgu.CV.Util.CvException: 'OpenCV: image.channels() == 1 || image.channels() == 3 || image.channels() == 4'
As mentioned in comment, and only thing I see here as an issue, is the number of channels.
What has to be done, to avoid the exception and get the dft transformated image saved.

Emgucv turn the black background transparent

I'm very new to Emgucv, so need a little help?
The code below is mainly taken from various places from Google. It will take a jpg file (which has a green background) and allow, from a separate form to change the values of h1 and h2 settings so as to create (reveal) a mask.
Now what I want to be able to do with this mask is to turn it transparent.
At the moment it will just display a black background around a person (for example), and then saves to file.
I need to know how to turn the black background transparent, if this is the correct way to approach this?
Thanks in advance.
What I have so far is in C# :
imgInput = new Image<Bgr, byte>(FileName);
Image<Hsv, Byte> hsvimg = imgInput.Convert<Hsv, Byte>();
//extract the hue and value channels
Image<Gray, Byte>[] channels = hsvimg.Split(); // split into components
Image<Gray, Byte> imghue = channels[0]; // hsv, so channels[0] is hue.
Image<Gray, Byte> imgval = channels[2]; // hsv, so channels[2] is value.
//filter out all but "the color you want"...seems to be 0 to 128 (64, 72) ?
Image<Gray, Byte> huefilter = imghue.InRange(new Gray(h1), new Gray(h2));
// TURN IT TRANSPARENT somewhere around here?
pictureBox2.Image = imgInput.Copy(mask).Bitmap;
imgInput.Copy(mask).Save("changedImage.png");
I am not sure I really understand what you are trying to do. But a mask is a binary object. A mask is usually black for what you do not want and white for what you do. As far as I know, there is no transparent mask, as to me that makes no sense. Masks are used to extract parts of an image by masking out the rest.
Maybe you could elaborate on what it is you want to do?
Doug
I think I may have the solution I was looking for. I found some code on stackoverflow which I've tweaked a little :
public Image<Bgra, Byte> MakeTransparent(Image<Bgr, Byte> image, double r1, double r2)
{
Mat imageMat = image.Mat;
Mat finalMat = new Mat(imageMat.Rows, imageMat.Cols, DepthType.Cv8U, 4);
Mat tmp = new Mat(imageMat.Rows, imageMat.Cols, DepthType.Cv8U, 1);
Mat alpha = new Mat(imageMat.Rows, imageMat.Cols, DepthType.Cv8U, 1);
CvInvoke.CvtColor(imageMat, tmp, ColorConversion.Bgr2Gray);
CvInvoke.Threshold(tmp, alpha, (int)r1, (int)r2, ThresholdType.Binary);
VectorOfMat rgb = new VectorOfMat(3);
CvInvoke.Split(imageMat, rgb);
Mat[] rgba = { rgb[0], rgb[1], rgb[2], alpha };
VectorOfMat vector = new VectorOfMat(rgba);
CvInvoke.Merge(vector, finalMat);
return finalMat.ToImage<Bgra, Byte>();
}
I'm now looking at adding SmoothGaussian to the mask to create a kind on blend, where the two images are layered, rather than a sharp cut-out.

How to detect number of objects from image?

I have a windows forms application and i want to count number of objects from a medical images. For instance
I used an algorithm which can take the contours of every object from the image.
private void findContoursToolStripMenuItem_Click(object sender, EventArgs e)
{
Image<Gray, byte> imgOutput = imgInput.Convert<Gray, byte>().ThresholdBinary(new Gray(100), new Gray(255));
Emgu.CV.Util.VectorOfVectorOfPoint contours = new Emgu.CV.Util.VectorOfVectorOfPoint();
Mat hier = new Mat();
Image<Gray, byte> imgout = new Image<Gray, byte>(imgInput.Width, imgInput.Height, new Gray(100));
CvInvoke.FindContours(imgOutput, contours, hier, Emgu.CV.CvEnum.RetrType.External, Emgu.CV.CvEnum.ChainApproxMethod.ChainApproxSimple);
CvInvoke.DrawContours(imgout, contours, -1, new MCvScalar(123, 0, 0));
pictureBox2.Image = imgout.Bitmap;
}
But I can't find out the number of cells from the image. Any advice or algorithm I have to use ?
I search within EMGU documentation but I don't find any function which does somethink like I want.
Any advice or answer will be rewarded.
If you consider that is too broad, I don't want the implemented algorithm. I just need some ideas or a suggestion of algorithm i have to use.
It probably a bit basic and brute force, but how about selecting a random point on the image that is close to the green colour, then effectively search for 'matching' colours (with a tolerance for the same colour. As you visit each pixel, colour it black so you don't look at it again and count how many pixels you have coloured in. Each time you select a pixel, make sure it's not black. Once you can't find any more points, if the number of black pixels is greater than a tolerance (so you only find 'big' polygons), then count it in the number of cells.

Emgu image conversion from Image<Gray,float> to Image<Gray,Byte> results in intensity loss?

We are performing image sharpening of a gray scale image of type Image by subtracting the Laplacian of the image from the original image. The result, if saved as a JPEG, has well defined edges and contrast. However, if the resultant image is converted to Bitmap OR "Image<Gray, Byte>"
and saved as JPEG, the intensity is reduced and the sharpening effect is lost. I suspected that converting to Bitmap may be causing this problem. So, I saved some of the intermediate images and also converted the image to "Image<Gray,Byte>". This did not help. I also tried to scale the image using a simple method. This too did not help.
The above behaviour is also true when we perform Laplace and subtract the resultant image from the original image. Illustrations are below (code has been modified for simplicity):
...
Image<Gray, Byte> sharpenedImage = Sharpen(filter, originalprocessedImage);
ProcessedImage = sharpenedImage.ToBitmap(); // Or ProcessedImage.Bitmap;
ProcessedImage.Save("ProcessedImage.jpg"); // results in intensity loss
...
public Image<Gray, Byte> Sharpen(Image<Gray, Byte> inputFrame)
{
ConvolutionKernelF Sharpen1Kernel = new ConvolutionKernelF (new float[,] { { -1,-1,-1 }, { -1, 8,-1 }, { -1,-1,-1 } });
Image<Gray, float> newFloatImage = inputFrame.Convert<Gray, float>();
Image<Gray, float> newConvolutedImage = newFloatImage.Convolution(Sharpen1Kernel);
Image<Gray, float> convolutedScaledShiftedImage = newFloatImage.AddWeighted(newConvolutedImage, 1.0, 1.0, 0);
// added for testing
convolutedScaledShiftedImage .Save("ConvolutedScaledShiftedImage .jpg");
//Now try to scale and save:
Image<Gray, float> scaledImageFloat = convolutedScaledAddedImage.Clone();
Image<Gray, float> scaledImageFloat2 = ScaleImage(scaledImageFloat);
// added for testing
scaledImageFloat.Save("ScaledImage.jpg");
// added for testing
scaledImageFloat2.Convert<Gray,Byte>().Save("ScaledImage-8Bits.jpg");
// both of these return the images of lower intensity
return scaledImageFloat2.Convert<Gray,Byte>();
return convolutedScaledShiftedImage.Convert<gray,Byte>();
}
While the ConvolutedScaledShifteImage.jpeg is brighter and with better contrast, "ScaledImage.jpeg" and "ScaledImage-8Bits.jpeg" have lost the intensity levels as compared to ConvolutedScaledShifteImage.jpeg. The same is true for ProcessedImage.jpeg.
The ScaleImage is below. This was not really necessary. As the Convert was losing intensity, I tried to do the conversion and check:
Image<Gray, float> ScaleImage(Image<Gray, float> inputImage)
{
double[] minValue;
double[] maxValue;
Point[] minLocation;
Point[] maxLocation;
Image<Gray, float> scaledImage = inputImage.Clone();
scaledImage.MinMax(out minValue, out maxValue, out minLocation, out maxLocation);
double midValue = (minValue[0] + maxValue[0] ) / 2;
double rangeValue = (maxValue[0]) - (minValue[0]);
double scaleFactor = 1 / rangeValue;
double shiftFactor = midValue;
Image<Gray, float> scaledImage1 = scaledImage.ConvertScale<float>(1.0, Math.Abs(minValue[0]));
Image<Gray, float> scaledImage2 = scaledImage1.ConvertScale<float>(scaleFactor * 255, 0);
return scaledImage2;
}
Would anybody be able to suggest what could be going wrong and why the intensities are lost in the above operations? Thanks.
Edit: fixed the formatting issue... conversion was from Image<Gray, float> to Image<Gray, Byte>
Edit 12-Jan: I further dug into OpenCV code and as I understand, when you save an image of type Image<Gray,float> to JPEG, imwrite() first converts the image to an 8-bit image image.convertTo( )temp, CV_8U ); and writes to the file. When the same operation is performed with Convert<Gray,Byte>(), the intensities are not the same. So, It is not clear what is the difference between the two.

Measure difference of two images using emgucv

i need to compare two images and identify differences on them as percentage. "Absdiff" function on emgucv doesn't help with that. i already done that compare example on emgucv wiki. what i exactly want is how to get two image difference in numerical format?
//emgucv wiki compare example
//acquire the frame
Frame = capture.RetrieveBgrFrame(); //aquire a frame
Difference = Previous_Frame.AbsDiff(Frame);
//what i want is
double differenceValue=Previous_Frame."SOMETHING";
if you need more detail plz ask.
Thanks in advance.
EmguCV MatchTemplate based comparison
Bitmap inputMap = //bitmap source image
Image<Gray, Byte> sourceImage = new Image<Gray, Byte>(inputMap);
Bitmap tempBitmap = //Bitmap template image
Image<Gray, Byte> templateImage = new Image<Gray, Byte>(tempBitmap);
Image<Gray, float> resultImage = sourceImage.MatchTemplate(templateImage, Emgu.CV.CvEnum.TemplateMatchingType.CcoeffNormed);
double[] minValues, maxValues;
Point[] minLocations, maxLocations;
resultImage.MinMax(out minValues, out maxValues, out minLocations, out maxLocations);
double percentage = maxValues[0] * 100; //this will be percentage of difference of two images
The two images need to have the same width and height or MatchTemplate will throw an exception. In case if we want to have an exact match.
Or
The template image should be smaller than the source image to get a number of occurrence of the template image on the source image
EmguCV AbsDiff based comparison
Bitmap inputMap = //bitmap source image
Image<Gray, Byte> sourceImage = new Image<Gray, Byte>(inputMap);
Bitmap tempBitmap = //Bitmap template image
Image<Gray, Byte> templateImage = new Image<Gray, Byte>(tempBitmap);
Image<Gray, byte> resultImage = new Image<Gray, byte>(templateImage.Width,
templateImage.Height);
CvInvoke.AbsDiff(sourceImage, templateImage, resultImage);
double diff = CvInvoke.CountNonZero(resultImage);
diff = (diff / (templateImage.Width * templateImage.Height)) * 100; // this will give you the difference in percentage
As per my experience, this is the best method compared to MatchTemplate based comparison. Match template failed to capture very minimal changes in two images.
But AbsDiff will be able to capture very small difference as well

Categories

Resources