Enhancing Digital Maps with AI-Derived Confidence Information

Digital maps function as a crucial sensor in the realm of self-driving cars. Beyond their use in Lane Keeping Assistance (LKA) and delivering mission statements to autonomous vehicles, their pivotal role extends to initiating takeover requests well in advance. Similar to the expectations from other sensors, it is imperative for maps to convey a confidence level. This paper introduces three innovative strategies to instill such information. The first encourages map providers to offer timestamp information reflecting when map features were last verified. The second approach advocates for the inclusion of variance information in crowd-sourced maps. By providing not only mean values for map features but also standard deviation information, subsequent fusion modules can decide whether they trust the map, or the information provided by other sensors more. In the third approach, we will leverage Variational Auto Encoders (VAE) for finding map outliers. If a self-driving vehicle encounters a map situation it has not seen before, the system should trigger a takeover request, as it likely hasn’t encountered such a situation during training and might not know how to handle it. Map outliers and map features with high uncertainty are handled in a similar way by self-driving cars and are therefore presented together in this paper.