Robust Optical Flow with Combined Lucas-Kanade/Horn-Schunck and Automatic Neighborhood Selection

Abstract

Differential optical flow methods are widely used within the computer vision community. They are classified as being either local, as in the Lucas-Kanade method, or global, such as in the Horn-Schunck technique. Local differential techniques are known to have robustness under noise, whilst global techniques are able to produce dense optical flow fields. We will show that the Horn-Schunck Technique, when combined with Lucas-Kanade, can yield the advantage of having both robust and dense optical flow fields. Selection of neighborhood size is an important tuning parameter for the combined Lucas-Kanade/Horn-Schunck technique. Choosing the optimal neighborhood is a difficult task and greatly effects the performance of optical flow results. We outline a method for the automatic selection of neighborhood size based on Stein's unbiased risk estimator (SURE). Algorithms are derived for a combined Lucas-Kanade/Horn-Schunck technique with automatic neighborhood selection. The performance of SURE neighborhood selection for the combined optical flow technique is simulated via Matlab, providing an illustration of the performance that is attainable.