Finding map feature correspondences in heterogeneous geospatial datasets

Abstract

In an automated map making process, map features like lane-markings, traffic-signs, poles, stop-lines and similar other features are extracted using deep learning methods from various sources of imagery or sensor data. These sources come with their own positional errors due to which the map features extracted from these sources are always misaligned with respect to each other, making the conflation of map features a difficult task. We propose a novel method to find map feature correspondences between 2 sets of map feature datasets obtained from different sources by first converting them into a heterogeneous geospatial graph and then doing node representation learning using a graph neural network that can generate vector embeddings that encode information of morphology, attributes, and absolute and relative positions of the map feature with respect to its neighbours along with aggregated information from its neighbours. This process can be employed to generate embeddings of map feature nodes, which are amicable to identifying spatially similar and corresponding map feature nodes across disparate sources with varying degree of similarity scores. When applied aptly, these map feature correspondences between two sources can be used as anchor points to perform spatial alignment with linear or nonlinear transforms, leading to a better conflation.