Optical flow for obstacle detection in mobile robots

Abstract

The Problem: We want to let a mobile robot explore its environment, detecting and avoiding obstacles as it goes. This will allow it to create a map of that environment, and then find its way around in it. We would like it to be able to do this using visual input. It should need as little human intervention, in the way of pre-supplied maps or other input, as possible. The system should also be robust with respect to sensor noise and sloppy actuation. Motivation: Robots currently have a hard time finding their way around on their own. This is OK if the robot is being closely watched by a person, but in some situations, this isn’t practical. A mobile robot wandering the surface of Mars could not be guided remotely because of the large time-delay involved. A small army of robots moving around a factory floor, making and delivering things, would require an unreasonably large number of operators. The reason we will focus on vision as opposed to other sensors (for example, sonar sensors), is that cameras are relatively cheap, and images from a camera give much more information than distance measurements that would be given by sonar or laser range finders. Vision provides the best balance in several factors such as sensor noise, amount of information, and cost, and the only reason it is not more widely used is because it is difficult to extract the relevant data fromcapturedimages. Optical flow is one useful piece of information that can be extracted from sequential images captured by a mobile robot [2]. The optical flow describes how the different parts of the image are moving. As the robot advances, objects