Localization of vehicle in adverse conditions such as in dense traffic conditions is a challenging problem and state-of-the-art techniques often make the vehicle get lost, requiring a re-localization technique to correctly reset the vehicle pose. The visual place recognition and loop-closure based re-localization techniques need to store a very large map and take a lot of time to re-localize the vehicle. We solve the problem by making a semantic map which is used to re-localize the vehicle, if and once it gets lost by conventional localization techniques. The semantic map is created using a test vehicle with sophisticated sensors, and the map can be used by any vehicle with a stereo camera for re-localization. It is assumed that the test vehicle has a budget stereo camera which produces numerous false positives to be rejected by the re-localizer; while the vehicle also misses many key landmarks during the run due to heavy traffic. These are the challenges which are overcome by the designed re-localization algorithm. The vehicle is tested on a highway scenario in Bengaluru, India for multiple runs in a highway segment. Results confirm accurate re-localization on a semantic map generated from road-signs.
Re-localization for Self-Driving Cars using Semantic Maps
2020-11-01
349414 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Springer Verlag | 2023
|