|M.Sc Student||Cherevatsky Boris|
|Subject||Estimating Epipolar Geometry for Mobile Robots|
|Department||Department of Computer Science||Supervisors||PROF. Ilan Shimshoni|
|PROF. Ehud Rivlin|
|Full Thesis text|
This thesis addresses the problem of robot navigation using natural visual features in a planar environment. Suppose we have a mobile robot equipped with a camera.
The robot can move on the floor and capture images with the camera, and save them on a storage. The robot acts inside a room which is surrounded by walls, which are orthogonal to the floor. Many visual features appear on the walls, and the robots goal is to go from one pose to another by using only those visual features. The robot can use only images for navigation.
Suppose a human operator defined a set of targets in advance. Each target is a position on a plane (X,Z coordinates) and a rotation (around Y-axis) in some fixed coordinate system. This position and orientation pair is denoted as pose. In every target an image is captured to represent the target. When the robot reaches a target the image it should see is exactly the same image captured in advance by an imaginary human operator.
The robot is given as an input a set of images, representing a set of targets, and a source image, captured from the starting position. The robot should make an accurate move towards each target in a sequence which can be decomposed as a rotation and translation on the plane.
This algorithm can be used as a step in a higher level homing algorithm for visual navigation. This problem was already addressed in the general case, but in this special case when all the visual features are located on the walls and the robot moves on the floor plane, there is a special structure for the essential Matrix and the homography, and a novel algorithm is proposed for finding both of them from only two points correspondence which gives a higher precision and solves numerical stability issues. A theoretical connection between those entities is given, and a fast algorithm for finding one from the other is proposed.
Experiments, comparing the suggested algorithm to the well known 3-point and 8-point methods, were carried out in a simulated robotic environment. The advantage of the suggested algorithm in accuracy and runtime is described here.
Finally a navigation experiment of a mobile robot, with a single camera, was carried out. The robot successfully reached the targets with the proposed algorithm.