Removing outliers from the Lucas-Kanade method with a weighted median filter

Abstract

The definition of optical flow is stated as a brightness pattern of apparent motion of objects, through surfaces and edges in a visual scene. This technique is used in motion detection and segmentation, video compression and robot navigation. The Lucas-Kanade method uses information from the image structure to compose a gradient based solution to estimate velocities, also known as movement of X- and Y-direction in a scene. The goal is to obtain an accurate pixel motion from an image sequence

The objective of this thesis is to implement a post processing step with a weighted median lter to a well known optical flow method; the Lucas-Kanade. The purpose is to use the weighted median lter to remove outliers, vectors that are lost due to illumination changes and partial occlusions.

The median filer will replace velocities that are under represented in neighbourhoods. A moving object will have corners not just edges, and these vectors have to be preserved. A weighted median filter is introduced to ensure that the under represented vectors is preserved. Error is measured through angular and endpoint error, describing accuracy of the vector field. The iterative and hierarchical LK method have been studied. The iterative estimation struggles less with single error. Because of this the weighted median filter did not improve the iterative LK-method. The hierarchical estimation is improved by the weighted median and reduced the average error of both angular and endpoint error.