Make computer to see the things the way you see.Reduce the dimensions.
To extract finger tips from hand, various techniques can be used like convex hull points, template matching etc..  One of the most difficult task it to extract hand from background. Kinect make it easy for us. The approach I used is somewhat  different to extract finger points.Check out the video. Green points show finger tips and red points show valley points.

Extracting hand is very tedious task but with kinnect hand point and depth image it is easy to do.lets see how  to do it.

Get kinect.Set up OpenNI and Nite. Download c# version of OpenCV. Play with PointViewer sample.Now you know  how to get HandPoint.. here is the step you can follow.

Background subtraction:

  • Get the hand point.
  • Get DepthMap.
  • Take a frame of   M x M pixels(say 200×200 px) from the DepthMap around hand point.
  • To subtract background  ,get depth of HandPoint from DepthMap. Make all pixel to value 0 whose depth is more or less than 5 cm and make rest of pixels to 1 in frame.

HandExtraction:

Here OpenCv save  a lot time. Extract contours using cvFindContour function and select the one contour which contains HandPoint. This contour is  our hand.

FingerTips Extraction:

So we have HandPoint and Hand contour. You can use convex hull  and convexity defect to extract desired  points and then try to identify finger tips.But it has many problems ,like you get more than  convex hull point near corner. So you need to add  filter to reduce it to one point for further processing.  and even more we are reading depthmap that is generated by IR projection.So in realtime video feed data we are getting changes over time ,in few frame point may be present and in few point may not be there.So our algorithme should able to cope this variation in image. Expirement your self you will know what I am talking about.

Polor cordinates:

We have image of lets say 200×200 px. what we will do is  make HandPoint as center of Cartesian system. Then we will find r and theta of each pixel.

r=sqrt(point.x*point.x+point.y*point.y).

theta=atan(y/x).

Now   The idea is  values of r for each point is one dimention curve. find the local maxima and mimima of the curve.bingo.. local maxima is probable finger tips and mimima is propable valley points.