Robot Visual Navigation and Homing: Theory and Applications
In this talk I will present the visual homing algorithm for robots. The task is to guide a robot (usually a mobile robot or AGV) to a target position and orientation. We present a user friendly method to perform this task which does not require complicated setup operations. The robot, equipped with a camera is taken once to the target position. There it takes an image. Later when it is required to return to the goal position it takes another image and comparing it to the original image is able to estimate the direction to the goal. The only requirement is that there will be some overlap between the scenes seen in the two images. The algorithm is based in part on contributions we made to the theory of multiple view geometry in computer vision. This theory does not require that the algorithm be given any 3D information about the objects in the scene.
In the first stage of the algorithm, corresponding features in the two images have to be found. The major problem here is that the low level algorithms which recover these matches may yield a very high percentage of incorrect matches, which if not detected yield incorrect estimations of the target position. This is a basic component of many computer vision problems. In the second part of the talk I will describe our new method for separating incorrect matches (outliers) from the correct ones (inliers). This algorithm develops and uses techniques in robust statistics. The resulting algorithm is able to work in cases where the percentage of outliers is very high. In these cases the algorithm achieves a speedup of one to two orders of magnitude with respect to the state of the art algorithms, enabling visual homing algorithms and many other computer vision algorithms to work effectively in these hard cases.
In the last part of the talk I will describe the mobile robot that we are building in the Technion and its current and future capabilities.