GenPa-SLAM: Using a General Panoptic Segmentation for a Real-Time Semantic Landmark SLAM

Alongside a detailed knowledge about the current environment, the ego position is a major aspect in autonomous driving. Solving the task of localization and mapping typically consists of a front-end and a back-end. While the front-end extracts features and solves the data association problem, the back-end performs the localization and the mapping of the extracted features. In this work, we present a novel LiDAR-based SLAM algorithm: In the front-end, a general panoptic segmentation algorithm is used for extracting clustered and classified objects. Knowing the class of each individual cluster gives us the opportunity to efficiently extract various geometric primitives, such as cylinders, planes, lines or bushes. Furthermore, potentially dynamic things, such as cars and trucks, are neglected. A graph-based back-end optimizes the vehicle trajectory and the landmarks position online. Our approach is evaluated on the KITTI odometry benchmark and further experiments using two different LiDAR sensors on our testing vehicles.