Visual detection and tracking of poorly structured dirt roads

Outdoor mobile robots are often faced with the problem of trying to navigate through an unknown environment. Areas that appear simple to humans can be very difficult for a robot to accurately and consistently describe. Poorly structured dirt roads, such as fire-access tracks and bush-walking tracks, are often overlooked in research but are highly important passageways for emergency support crews, such as search-and-rescue teams and firefighters tackling bush fires. This paper presents a method of autonomously detecting and hence tracking such roads using colour vision. Central to the process is a method of characterising the road surface through a statistical colour description, which makes minimal assumptions about the road. A highly simplified and generalised road model is used to ignore the background and contain the road, and weighted control points are used to generate a spline-based trajectory along the road, which is intended to be used for motion-control of a robot trying to traverse these tracks. Inherent to the system is the avoidance or safe traversal of certain types of obstacles. The combination of simple modelling and efficient processing algorithms has resulted in a usable average processing speed of approximately eight frames per second on the 1.7 GHz Pentium-4 test machine