Learning of lane information reliability for intelligent vehicles

Automated riving is becoming the focus of various research institutions and companies. In this context, road estimation is one of the most important tasks. Many works propose to realize this task by employing one or multiple of the following orthogonal information sources: road markings from optical lane recognition, leading vehicle, digital map. etc. Each of them has its own strength and drawbacks in different situations. However, many existing approaches assume that the sources are equally reliable. Incorporating reliability estimates into the fusion of theses sources can significantly increase the availability of automated driving in most scenarios. In this work, we propose a novel concept to define, measure, learn and integrate reliabilities into the road estimation task. We introduce a new error metric in which the reliability is defined as the angle discrepancy between the estimated road course and the manually driven trajectory. Based on a large database containing sensor and context information from different situations, a Bayesian Network and Random Forests are trained to learn the reliabilities. The estimated reliabilities are used to discard unreliable sources in the fusion process. Experimental results prove our concept.