Improving Self-Localization Using CNN-based Monocular Landmark Detection and Distance Estimation in Virtual Testbeds

Mobile robots often require knowledge of their precise position. However, in many cases the integration of an outside-in tracking system is not feasible. In the exemplary case of autonomous vehicles, Global Navigation Satellite Systems (GNSS) are available but do not fulfill the precision requirements sufficiently strictly. Monocular cameras are another technology which is already built into many commercially available vehicles to enable additional safety and comfort features. Their image streams can be utilized to enhance localization precision using prior knowledge of landmarks along the roads.In this paper, an integrated architecture for landmark detection, classification, position estimation and landmark-based localization is presented. The system is developed within a virtual testbed allowing for rapid development and evaluation of the employed components. In testing, an accuracy in the order of 10 cm is achieved for localization, constituting a significant improvement compared to systems using positional data from navigation satellites only.