I need to recognize some elements drawn with user’s finger on device’s screen. For the start say they are circle, horizontal line, diagonal line, box, and a few more complicated, like lightning sign.
Not only I need to say what element is more likely to be, but also I need to recognize if element is drawn wrong, so it’s not one of those I recognize.
I wonder what algorithm is best for that, and also what is the best way to collect data for it.
For now I am using OpenCV, algorithm is Support vector machine. I collect data by counting distance from from center coordinate to 10 points of image. It doesn’t works well.
Sounds like you need to extract a feature vector to describe your “gesture”. Ideally the features must be scale and translation invariant. That is: they distinguish the gesture irrespective of the size or angle it is created at.
You might like to consider using Velocity Moments as the features to use if the motion itself is important.
Or, if you treat with only the finished shape of the gesture, you could just use the image based methods provided by OpenCV. You might be particularly interested in the matchShapes() method – based on the Hu Moments.
If you are already using a SVM then that is your supervised classifier into which you feed your extracted feature vector for each gesture (with a label) to train it. Apart from the choice of classifier (PCA, Naive Bayes, ANN’s, etc, etc.) – the crux of the matter is in the quality of your training data.
When I have done something similar in the past I have simply collected training data using a small script to capture motion of the mouse (when the button is held down). I seem to remember at the time using the UNIX command ‘xev’ to capture the window mouse events.