Since lane prediction involves two concurrent processes, we first tested each process in isolation. Centerline refinement is the most difficult algorithm to evaluate. The only objective evaluation is a comparison with the true centerline, but there are no means of measuring this centerline. Traditional surveying is impractical for busy public highways. Geo-referenced aerial or satellite photographs are alternative sources of raw data, but the data may be noisy, and the vision processing algorithms may not be reliable. Construction blueprints are available, but there is no guarantee that the road is actually built according to the plan. Additionally, all of these alternatives measure the center of the pavement, whereas our technique produces a centerline ``weighted'' toward the most common lane sampled. So even if the independent centerline measurement is very different from our own centerline, it is not clear if that makes a difference in the performance of the overall task.
Besides the final accuracy of the centerline, the rate of convergence is also of interest. Since the system is incremental, it can measure the difference between the centerline at each iteration and a reference centerline. If we plot the average difference between the current and reference centerlines for each iteration, the learning curve describes how quickly the centerline approaches the reference. This is of interest because it lets us estimate how many passes are necessary over a given segment before the centerline stops changing significantly. Ideally the reference centerline would be the true road centerline, but since the true centerline is unavailable we need an approximation. The best approximation available is the final centerline after the system has processed all traces. The rate of convergence in this case describes how quickly the centerline approaches the final result.
Figure 4 plots the difference between the centerline before each trace and the final centerline for a representative highway segment. The original map database was provided by NavTech, Inc., and had an average error of about 7 meters. The plot shows that the major adjustment occurs on the first pass, where the baseline estimate is corrected by a GPS estimate with 1 to 2 meter accuracy. Processing the successive traces slowly improves accuracy by averaging out the noise in the GPS readings. Although there is no ground truth to measure the final accuracy, later processing critically relies on an accurate centerline, so good results in those evaluations imply a sufficiently accurate centerline.
 |