This paper presents an online feature selection algorithm
for video object tracking. Using the object and background
pixels from the previous frame as training samples, we model the
feature selection problem as finding a good subset of features to
better classify object from background in current frame.