City GPS using stereo vision

Next generation driver assistance systems require a precise localization. However, global navigation satellite systems (GNSS) often exhibit a paucity of accuracy due to shadowing effects in street canyon like scenarios rendering this solution insufficient for many tasks. Alternatively 3D laser scanners can be utilized to localize the vehicle within a previously recorded 3D map. These scanners however, are expensive and bulky hampering a wide spread use. Herein we propose to use stereo cameras to localize the ego vehicle within a previously computed visual 3D map. The proposed localization solution is low cost, precise and runs in real time. The map is computed once and kept fixed thereafter using cameras as sole sensors without GPS readings. The presented mapping algorithm is largely inspired by current state of the art simultaneous localization and mapping (SLAM) methods. Moreover, the map merely consists of a sparse set of landmark points keeping the map storage manageably low.