Since all the approaches covered in the previous section are inadequate, this section describes a new approach to road-modeling that we did implement and evaluate. Although the centerline in the digital map is not accurate enough to compute constant lane offsets, any line parallel to the true lanes is, by definition, a constant distance from the lanes. We have devised a procedure to bring the centerline from the original digital map into alignment with the traces. The procedure computes the ``average'' between the current centerline and a new trace, weighted by the confidence in the centerline and in the trace. As the system incrementally incorporates more traces, it averages out errors in the traces to find a centerline more accurate than any of the traces that went into it. With this centerline refinement procedure, our approach to finding a lane model for a target segment S, covered by a set of position traces P, is to decompose the task into first finding an accurate centerline for S using P, then clustering P's offsets from the centerline into lanes. Next we describe how the system combines new traces into segment centerlines and clusters offsets into lanes.