A Real-Time Mapping System With Temporal Scan Context via Odometry Local Trajectories

Simultaneous localization and mapping (SLAM) directly impacts the quality of maps used for downstream tasks such as localization and path planning. Cumulative errors from interframe matching are eliminated in most state-of-the-art LiDAR SLAM algorithms through loop closure detection. However, the benefits of loop closure detection for trajectory correction are emphasized in these methods, while the reciprocal contribution of odometry to loop closure detection is overlooked. In this article, a real-time SLAM system leveraging the interaction between a loop closure detection and a front-end odometry method is proposed. Initially, multiple point clouds (from the neighbor of the retrieval frame or historical frame) are transformed into a unified coordinate system using short-term poses from the front-end odometry, and local scenario maps are created. Then, long-term place similarity is identified by encoding these local maps from neighboring frames into 2-D descriptors for loop closure detection. Finally, poses between loop closure pairs are fed into a graph structure to correct and update the odometry trajectory. The experimental results on the KITTI and JLU datasets demonstrate that our loop closure detection method achieves an average F1-score of 0.8445 with retrieval efficiency in under 5 ms. Moreover, the proposed SLAM system runs in real-time on our self-built Tiguan platform, and effectively corrects accumulated errors during long-distance mapping. The code is available at https://github.com/TrisMask/TSC.