Online Road Model Generation From Evidential Semantic Grids

The knowledge about the local environment is of utmost importance for all robotics and autonomous driving applications. Currently this information is often extracted from high definition maps used in combination with a highly-accurate localization restricting the operation to prior mapped areas and making it vulnerable to changes in the environment. On the other hand, occupancy grids, a well known representation of the static environment, provide a common way to model the environment with online sensor measurement data collected during the operation. However, the approach is limited to static occupancy probabilities without further classification or differentiation. This paper addresses the topic of estimating the local static environment solely from online sensor measurements by using an evidential semantic grid. Based on the Dempster-Shafer theory and a novel frame of discernment, sensor measurements, such as lane markings, point clouds from image-based semantic segmentation, occupancy grids and observed traffic participants are fused into an evidential grid estimating the semantic meaning of each grid cell. Afterwards, an online road model is generated by extraction lane geometries from the evidential grid. Real sensor data from German highways and urban areas is used to show the effectiveness of the proposed approach.