LLama-SLAM: Learning High-Quality Visual Landmarks for Long-Term Mapping and Localization

The precise localization of vehicles is an important requirement for autonomous driving or advanced driver assistance systems. Using common GNSS the ego position can be measured but not with the reliability and precision necessary. An alternative approach to achieve precise localization is the usage of visual landmarks observed by a camera mounted in the vehicle. However, this raises the necessity of reliable visual landmarks that are easily recognizable and persistent. We propose a novel SLAM algorithm that focuses on learning and mapping such visual long-term landmarks (LLamas). The algorithm therefore processes stereo image streams from several recording sessions in the same spatial area. The key part within LLama-SLAM is the assessment of the landmarks with quality values that are inferred as viewpoint dependent probabilities from observation statistics. By adding solely landmarks of high quality to the final LLama Map, it can be kept compact while still allowing reliable localization. Due to the long-term evaluation of the GNSS measurement during the sessions, the landmarks can be positioned precisely in a global referenced coordinate system. For a first assessment of the algorithm's capabilities, we present some experimental results from the mapping process combining three sessions recorded over two months on the same route.