Vision-GPS fusion for high precision vehicle localization in urban environment

In this paper, we proposed a method for vehicle localization with low-cost GPS and a monocular camera. Traditionally, the positioning accuracy of a single GPS can merely be reached at the metric-level both laterally and longitudinally. By integrating vision and GPS with the prior position database, which was generated by the bag-of-words method and improved by lane change detection, our method can achieve effective improvement of positioning accuracy, which is able to acquire the location of lane laterally and more precise vehicle coordinates than raw GPS data. In addition, by leveraging the complementary global information, our GPS-aided vision method has the advantages of avoiding inaccurate pose initialization of SLAM systems and obtaining absolute coordinates which can be deployed to navigation directly. The proposal was described by pseudocode in detail and was evaluated by simulation and a dataset collected by our team in an urban environment. The improvement and validity was verified by experiment based on our proposed evaluation criteria.