Maps of roads are expensive to build and maintain. Although modern cartography involves using satellite and aerial imagery along with GPS traces, using this data to update maps involves having humans analyze the data in a time consuming process, and thus maps of rapidly growing cities (where infrastructure is constantly under construction) are still often inaccurate and incomplete outside the urban core. Current approaches to map inference, however, do not provide consistent results under different levels of GPS noise and sparsity, and almost always fail to capture complex road features such as highway intersections.
We aim to apply machine learning techniques on a combination of GPS data and satellite imagery to produce more accurate maps. Using convolutional neural networks will allow the encoding of a prior on typical road geometries found in actual cities; algorithmic approaches often produce jagged roads with many abrupt turns that are unrealistic. Additionally, we believe that using both GPS data and satellite imagery to infer maps will yield more complete and accurate maps: in regions where roads are clearly visible in satellite images, we can extract very accurate geometry of the road from the image, while GPS data can help to correct errors where roads are obstructed by trees, shadows, or overpasses.
Robust map inference approaches will enable automatic identification of errors in existing maps, and possibly even automatic application of updates and corrections. Our eventual goal is to use map inference approaches to improve the OpenStreetMap dataset.