Robust and real-time multi-cue map verification for the road ahead

Digital maps are a key information for road course estimation in large distances, especially for automated driving with higher speeds or during nighttime. However, digital maps may contain errors caused by road changes or inaccurate recordings. As these errors may lead to severe application level failures, early and reliable detection of these errors, generally termed map verification, is a must. This work presents an online approach to verification of digital maps for automotive use based on neural multi-sensor fusion. Unlike previous works, the system is able to verify the digital map integrity ahead of the vehicle, before the error position is reached. Thus, appropriate countermeasures to mitigate possible effects can be initiated upfront. On the basis of a large database of map geometry and sensor information, a neural network is trained to classify the digital map integrity by optimally fusing sensor information at different abstraction levels. Using this technique, our system is able to verify the map integrity 100 meters ahead of the vehicle with an error rate of less than 3.4 %, while being real-time capable at the same time. Experiments validate the performance and robustness of this approach.