Concept study for vehicle self-localization using neural networks for detection of pole-like landmarks

Abstract

This paper discusses and showcases a software framework for the self-localization of autonomous vehicles in an urban environment. The general concept of this framework is based on the semantic detection and observation of objects in the surrounding environment. For the object detection three different perception approaches are compared; LiDAR based, stereo camera based and mono camera based using a neural net. The investigated objects all share the same geometrical shape; they are vertical with a high aspect ratio. To compute the pose of the vehicle an Adaptive Monte-Carlo Algorithm has been implemented. Hence it is necessary to create a high-precision digital map this is done with a dense map, the detected objects and the LiDAR point cloud. Comparison with an earlier paper have shown that this approach keeps the global positioning accuracy around 0.50 m and leads to more robust results in highly dynamic scenarios where a small amount of objects can be detected.