Practical ego-motion estimation for mobile robots

Abstract

Accurate ego-motion estimation is a difficult problem that humans perform with relative ease. This paper describes two methods that are used in conjunction to estimate the ego motion of an intelligent autonomous vehicle from vision alone. First, a cross-correlation method is used to select a promising patch in the image. The optical flow information for this patch is used to determine linear and angular velocity of the intelligent autonomous vehicle. Lines in the image are then used to provide an estimate of the ego motion of the vehicle. The gradient of the line as well as the distance to the line allow the computation of current wheel velocities. Both methods have been implemented on real robots and have been tested in a treasure hunt competition. These methods greatly improved the exploration as well as accuracy of the generated maps of the environment.