Detecting sidewalks in OpenStreetMap using machine learning LATEX Version
Eirik Gjeruldsen · NORA - Norwegian Open Research Archives · 2020
In the last couple of years there has been multiple machine learning projects focusing on detecting a variety of objects in OpenStreetMap.The explanation for this is the improvements done to both computer hardware and the increase of data in the OpenStreetMap database.As previous research done has primarily been done using image analysis and more traditional algorithms analysing spatial data in combination with tag data has been overlooked.This thesis aim to find the possibilities of using such a model to predict sidewalks, and study the differences between developed and developing countries.This is done by using two machine learning algorithms XGBoost and Neural Network in cooperation with features extracted from spatial data and tags contained within the OpenStreetMap database.The best performing model is then used to classify data from a developing country and exported to a custom made windows application to manually analyse images from the predictions.The experiments conducted indicates a connection between different inputs fed to the models and the ground truth, and produce a result with high accuracy.However, when classifying sidewalks on developing country data the image analysis show a big difference in results when testing on a country from Europe and a country from Africa.Despite the variation in results this thesis provides valuable input in regards to the limitations and constraints of the data in OpenStreetMap.