The final study combined the centerline refinement and the lane clustering processes. This experiment is most similar to how we expect to actually deploy the system, because the system initially generates lane models with no information beyond the baseline digital map. The procedure was similar to lane clustering alone, except the system computed the offsets of the first trace from the NavTech baseline. After computing the offsets and evaluating the predictions for each trace, the system refined the digital map centerline with the trace. We expected the results to be poor at first because of the inaccurate centerlines, but quickly approach levels in the previous experiment as the centerline improved. Figure 6 plots the average accuracy of the interleaved processes over 50 random orderings of the traces. As expected, the early results were poor, although somewhat above chance.3 Starting at the fifth trace, the combined algorithm performed comparably to clustering on the most accurate centerline.
 |
The results of this experiment show that, starting with baseline geometry that is commercially available, it is possible to generate an accurate road centerline and lane models after a few high-precision GPS passes. Our position recorder is compact and robust enough to operate unattended in any car, and our algorithms make no assumption regarding particular route or lane changing characteristics, so an entire city highway network could be modeled by distributing a number of recording units to vehicles. The vehicles, acting as probes during their normal driving patterns, accumulate information about the highway network. Once a vehicle has sufficiently sampled its driving patterns, it can relay its data to a centralized mapping service. The speed at which coverage and accuracy of the digital map improves is proportional to the number of recording units in operation. This is a low-cost technique for automatically mapping highways with high geometric accuracy.