Improving Accuracy of Maps Using Anonymized GPS Traces

However, if our map data has an error, like missing the segment depicted in Figure 7, this sequence will include abnormally low transition probabilities, indicating that the vehicle could not travel on a segment or transit between certain segments within the map data context, even though the vehicle actually did in the real world. Each of the copies has identical map data but a different and evenly distributed set of trips, as illustrated in Figure 9, below:

This approach eliminates bottlenecks caused by high-density cells and leads to more accurate results because each cell is still big enough to contain both the full map data context and GPS points for map matching and error detection. The philosophy behind aggregating results from a large number of trips is that if we see consistency in abnormal probabilities at a given place reported by trips, the root cause of this disparity is more likely to be a map data error than illegal driving behavior or noisy GPS signals.

Source: eng.uber.com