ViSLAB Update

Different methods of background subtraction

I have been researching into different ways to get very effective background subtraction. After talking to Masa this week, I will be looking into a number of different methods. I will be testing these methods using a number of different backgrounds, and seeing how well they hold up when the background is similar in color to the hand placed in front of the screen.

Texture Based subtraction

Uses feature images instead of raw intensity images to separate foreground and background. Paper mentions that foreground and background must "exhibit sufficiently different statistics". May be undone by flesh-colored backgrounds. I have found an OpenCV implementation of this, it should work after a few modifications.

Texture and Motion Pattern Fusion

Finds the probability that a pixel belongs to the background using both spatial and temporal patterns. These probabilities are combined using a user defined weighting.

Region Based Subtraction

Split the images up into small squares of 4x3 pixels, compute the histogram of each square. Combine the values of the squares to make larger squares. Continue until you have squares roughly 100 pixels across. Compare equivalent squares of forground and background, moving in smaller and smaller until you get to the original 4x3 squares.Claims to be very efficient, does not depend on a temporal model.

Over the next two weeks

I will try to implement or find implementations of each method, then test them for reliability and speed using a number of different backgrounds. Masa also briefly mentioned to me a few other ideas he had, so I will clarify the specifics of those and test them as well.