Map inference in the face of noise and disparity

This paper describes a process for automatically inferring maps from large collections of opportunistically collected GPS traces. In this type of dataset, there is often a great disparity in terms of coverage. For example, a freeway may be represented by thousands of trips, whereas a residential road may only have a handful of observations. Additionally, while modern GPS receivers typically produce high-quality location estimates, errors over 100 meters are not uncommon, especially near tall buildings or under dense tree coverage. Combined, GPS trace disparity and error present a formidable challenge for the current state of the art in map inference. By tuning the parameters of existing algorithms, a user may choose to remove spurious roads created by GPS noise, or admit less-frequently traveled roads, but not both.

In this paper, we present an extensible map inference pipeline, designed to mitigate GPS error, admit less-frequently traveled roads, and scale to large datasets. We demonstrate and compare the performance of our proposed pipeline against existing methods, both qualitatively and quantitatively, using a real-world dataset that includes both high disparity and noise. Our results show significant improvements over the current state of the art.