Chessboard recognition system using signature, principal component analysis and color information
Object tracking is a challenging problem in realtime computer vision, especially when the circumstance is unstable due to variations of lighting, pose, and view-point. This paper presents an online feature selection mechanism by extracting and evaluating multiple color features. Given a tracking image, we use clustering method to segment the object according to different color, and generate Gaussian model for each segment respectively to extract the color feature. Then we judge the discrimination of the features and select an appropriate feature subset, by which the object can be distinguished from the background at the highest SNR(signal noise ratio). This feature selection mechanism is embedded in a mean-shift tracking system that updating the feature set adaptively. Examples are presented to show that our method is robust to complicated object and changing background.