Semantic Assisted LiDAR Odometry with Loop Closure in Large Scale Urban Environment

Compared to the vision-based approach, LiDAR-based SLAM has shown a great advantage in depicting geometric characteristics but still suffers from accumulated localization errors during long-term operation in large-scale scenarios. Introducing semantic information to the current system helps to discover higher-level features and establish a stronger association of features in different frames. In this paper, we utilize semantic information to present an integral LiDAR odometry that combines adaptive downsampling feature with label-specified registration to boost the performance of odometry estimation, together with Scan Context as the loop closure module to constrain the amplification of cumulative errors. Experiments are conducted based on the well-known KITTI dataset, which reveals that the proposed framework achieves higher accuracy with an average RTE of 0.97% in real-time and shows great robustness toward various scenarios.