Communicationless navigation through robust visual odometry

GPS navigation is often found undependable in urban situations where tall structures occlude large parts of the sky. To keep accurate position in these situations, we need an alternative method. We propose a novel visual odometry method that is shown to provide reliable relative motion estimation in typical urban road driving using a single camera. While the short-term accuracy is good, relative motion estimation by itself is susceptible to drift and therefore insufficient to provide good long-term absolute position estimates. To overcome this drift, we propose to warp the visual odometry output to a stored map. This warping must be able to cope with temporary discrepancies between visual odometry and map data. The proposed mapping does not make a hard decision about road position, but instead entertains all plausible hypotheses about the followed trajectory and their associated warping costs. Evaluation on real test sequences proves that the method succesfully eliminates drift, and on average stays within 7 metres of simultaneously recorded GPS data. This proves that the combined visual odometry and mapping are sufficient to provide positioning with comparable accuracy to GPS in those situations when GPS is unavailable.