Map change detection using GPS position data

Abstract: Technology advancement in autonomous driving is accelerating. For the technology to be safe it is crucial for the vehicles to have an updated map, meaning all vehicles should have a correct and identical representation of the current road network. This makes change detection in the maps of great importance, in order to continuously understand and recognize the features that need to be updated. This thesis aims to develop and evaluate methods to continuously and automatically update maps using only crowd-sourced Global Positioning System (GPS) data. The approach was to create inferred maps using a method based on an approximate Kernel Density Estimation (KDE). The road network is represented using a mathematical graph. A Hidden Markov Model (HMM) and the Viterbi algorithm are used to map match GPS data to the road network. In the updating routine, new roads are added and old ones are removed. Furthermore, temporary changes are flagged. Three evaluation methods, two set-based and one path-based, are proposed which complement one another by taking different aspects into account—both geometric and topologic. The proposed map-inference method is robust to noise compared to many other map-generation algorithms, however, it is computationally heavy. Therefore, we propose creating geographically smaller maps and fusing multiple maps together. One of the main challenges was the parameter tuning of various thresholds, since the implemented algorithms are sensitive with respect to the accuracy of the data. The path-based evaluation is the only method where parameter tuning is not needed. Evaluation results show successful map updates on road level, where the accuracy was further increased when using an OpenStreetMap (OSM) as base map. However, results show that the methodology is not appropriate to obtain lane-level accuracy.