A high-level road model information fusion framework and its application to multi-lane speed limit inference

We propose a high-level road model information fusion framework to combine regulatory traffic elements, e.g. traffic signs, with lane geometry and digital map information for robust inference of lane-specific traffic rules. In this process, special care is given to adequately consider incomplete, uncertain, and inconsistent information sources with i) spatial, ii) existence, and iii) attribute uncertainties. First, Bayesian networks are employed for logical lane assignment of traffic elements under incorporation of traffic regulation knowledge and soft position relation evidences. The position relations are estimated via Monte Carlo simulations by taking spatial lane geometry and existence uncertainties into account. Second, Dempster-Shafer theory is used not only for fusing simultaneously detected traffic signs based on a novel belief mass transfer over adjacent lanes to recover from false sign classifications but also for traffic situation-dependent, lane-specific fusion of digital map attributes with sensor-inferred attributes. The framework is applied to the task of multi-lane speed limit inference, which gives lane-specific speed limit information in form of belief mass functions and runs in real-time on an experimental vehicle.