This paper presents a realization of our methodology: to provide support for safety and convenience applications via unsupervised learning algorithms operating on anonymous probe vehicle traces. Our results are good for the limited data sample we collected, but we need to study and improve the robustness and autonomy of our algorithms. Our final goal is to let the centerline refinement and lane clustering processes operate unattended, receiving GPS data from probe vehicles and refining digital maps. However, we need more analysis of different types of driving conditions, such as curved roads, and tools to measure data quality in the absence of labels to have sufficient confidence in our models.
We also plan to make more comprehensive evaluations of our lane models using prototype vehicular applications. These applications will indicate the commercial viability and effectiveness of our approach. Our initial application is a simple lane position task that recognizes the position of the vehicle relative to the lane structure of its current road segment. This application is simple enough for rapid development, but relevant to complete deployable applications such as lanekeeping and lane departure warning.
This study focused on road centerlines and lane models, but data mining over position traces can yield many more types of geospatially specific knowledge, particularly when paired with a geographic information system. Virtually any database with a geographic component, such as records about how often a vehicle comes near different types of locations, can benefit from a suitably large set of position traces. Elsewhere we have reported work on predicting traffic controls [12] and travel times [5]. Since position traces are inherently individual, we are also developing methods to construct personal digital maps, with features such as preferred routes and typical speeds. It is now possible to quickly and cheaply accumulate volumes of position traces that let one annotate objects in a geographic database with real-world behavior. This capability has the potential for impact on many applications areas, from safety to navigation to marketing.