There are many potential uses for an in-car system that can determine position on the road relative to lane markings. These can be roughly grouped into safety applications, such as lane departure warnings [17,1], and convenience applications, such as lane-changing advice to improve traffic flow [8]. Although Differential Global Positioning System (DGPS) receivers are often accurate enough to locate a vehicle to within a lane,1 there is no reliable source of lane locations.
In this paper, we present and evaluate the first system, to our knowledge, that enhances digital road maps with descriptions of lane structure, including number of lanes and their locations. Our approach involves mining massive amounts of DGPS traces from floating probe vehicles to augment the digital maps with lane information, creating a resource usable by any lane-related automotive application. Our system does not require special vehicles or expensive hardware to collect data, unlike some GPS mapping methods [4]. Our approach to data collection is to unobtrusively and indiscriminately gather as much data as possible from multiple drivers going about their ordinary business and to mine the resultant traces for knowledge about the road network.
Previous work on lane boundary finding has focused on directly performing tasks, such as lanekeeping, using machine vision to find lane markings related to the vehicle position [2,11]. This approach is limited in several ways. First, the vision system must be correctly calibrated to the lane markings it will sense. Our reliance on DGPS traces effectively lets us use the driver's lanekeeping ability to identify the center of the lane. The absolute nature of the data also provides information on upcoming terrain not directly sensible from the vehicle. Second, it is difficult, if not impossible, to build an accurate database of lane models with machine vision, or any other relative sensing method, alone. This is because the straightforward approach to building such a database is to store the lane structures in a spatially absolute reference system, and vehicles without an absolute sensing method, such as GPS, have no way to register the data spatially. An advantage to using machine vision techniques is that GPS accuracy suffers when the satellite view is partially or totally obstructed. In fact, deployed systems will probably use a combination of both technologies. We are currently developing a positioning system that combines GPS and local sensors to compensate for satellite visibility problems [15].
The next section motivates this work in more detail by describing some uses of a lane-sensitive vehicle. We then discuss some related work on problems similar in spirit to our own and some possible approaches to the problem. After this, we describe our solution to the problem, a system that creates an accurate description of road centerlines from a commercially-available map with relatively low accuracy and induces lane models by unsupervised learning. We evaluate the lane models against manual lane labels on highways. Finally, we describe some plans for future extensions to the work.