Extracting road information from recorded GPS data using snap-drift neural network

Abstract

Research towards an innovative solution to the problem of automated updating of road network databases is presented. It moves away from existing methods where vendors of road network databases either go through the time consuming and logistically challenging process of driving along roads to register changes or use update methods that rely on remote sensing images. The solution presented here would allow users of road network dependent applications (e.g. in-car navigation system or Sat Nav) to passively collect characteristics of any “unknown route” (departure from the known roads in the database) on behalf of the provider. These data would be processed either by an on-board artificial neural network (ANN) or transferred back to the Sat Nav provider and input into their ANN along with similar track data provided by other service users, to decide whether or not to automatically update (add) the “unknown road” to the road database. The solution presented here addresses the feasibility of identifying roads and assigning them into classes from recorded global positioning system (GPS) trajectory data. GPS trajectory data collected in London are analysed using a snap-drift neural network (SDNN) which categorises them into their strongest natural groupings, by combining clustering with feature detection in a single ANN. The key variables required are discussed. We have demonstrated that the SDNN offers a fast method of learning that preserves feature discovery. Using only GPS trajectory information, the SDNN is able to group collected points to reveal travelled road segments. The results showed that relying only on the winning node, a grouping accuracy of about 71% is achieved compared to 51% from learning vector quantisation (LVQ). On analysis and further experimentation with the SDNN d-nodes a grouping accuracy of nearly 100% was achieved, but with a high count of unique d-node combinations.