Real-time localization is a key to autonomous vehicle systems. This paper utilizes a camera-based visual simultaneous localization and mapping (SLAM) called ORB-SLAM for the localization of autonomous vehicles. In this work, we present a dynamic location transformation (DLT) method to correct the errors when applying relocalization using the SLAM of previously produced maps. Our experiments show that the localization accuracy can reach less than 1 meters when the simulated scenes with 1 km-long road contain sufficient object features. However, accuracy is decreased to several meters with similar conditions when using real data. Once the DLT method is applied, accuracy can be improved, that is, accuracy can be as good as those of the simulated data.
- IA-09-0009 (2M)