Robust Perception for Formula Student Driverless Racing

Abstract: Building an autonomous system for race cars requires robust and highly accurate perception running in real time. This thesis proposes a novel ground removal strategy for 3D LiDAR perception, modelling the ground as several planes, and a novel clustering method for LiDARs that sweep the scene in a predefined pattern resulting in a 20-fold performance increase over clustering methods commonly used for this problem. To weed out spurious information produced by relying only on LiDAR perception, two sensor fusion methods using a camera are introduced and evaluated, one using the immensely popular YOLO network. All algorithms were evaluated in real world scenarios using a fully functioning Formula Student driverless vehicle built by the Lund Formula Student 2021 team.