Better Features to Track by Estimating the Tracking Convergence Region

Reliably tracking key points and textured patches from frame to frame is the basic requirement for many bottomup computer vision algorithms. The problem of selecting the features that can be tracked well is addressed here. The Lucas-Kanade tracking procedure is commonly used. We propose a method to estimate the size of the tracking procedure convergence region for each feature. The features that have a wider convergence region around them should be tracked better by the tracker. The size of the convergence region as a new feature goodness measure is compared with the widely accepted Shi-Tomasi feature selection criteria.