Ego-Motion Estimation Using Affine Correspondences

Abstract

Ego-motion estimation of a moving vehicle is an essential task in modern robotics, autonomous vehicles (self-driving cars), and unmanned aerial drones. There are many ways to accomplish ego-motion estimation, for example by using different sensors such as radar, laser, or LiDAR. However, motion estimation can be accurately achieved using video input. Cameras become more available, affordable, and accurate. Most vision-based solutions for ego-motion estimation depend on matching algorithms that use Point Correspondences (PCs) between images to estimate the motion. However, motion estimation using Affine Correspondences (ACs) requires fewer correspondences; therefore, it takes fewer iterations to converge in the sample-and-test algorithms, such as random sample analysis (RANSAC). Affine-based solutions are faster while possessing similar accuracy to their point-based counterparts. Using affine correspondences to solve geometric computer vision problems is a relatively new practice. This paper is the first to survey the use of ACs for motion and relative pose estimation, as well as to provide an analysis of the proposed affine correspondences-based technique. The experimental results show that ego-motion estimation with ACs is faster than PCs due to the fewer required matches. Using ACs has the potential to solve various computer vision problems since they provide additional valuable information from the scene.