A Real‐Time and Fast LiDAR–IMU–GNSS SLAM System with Point Cloud Semantic Graph Descriptor Loop‐Closure Detection

Abstract

Herein, a real‐time, fast, tightly coupled simultaneous localization and mapping (SLAM) system is proposed. The deep neural network is used to segment the point cloud semantically to construct the point cloud semantic map descriptor, and the global navigation satellite system real‐time kinematic position is used to detect the loop closure. Finally, a factor‐graph model is used for global optimization. The working principle of the SLAM system is introduced in detail, including the construction of the factor graph, the generation of the point cloud semantic graph descriptor, the generation of the ring graph, the loop‐closure detection process, and the global optimization. Indoor and outdoor experiments are conducted to validate the performance of overall trajectory estimation and loop‐closure detection. Compared to traditional methods, the proposed method offers improved accuracy and real‐time performance in trajectory estimation. It effectively addresses issues such as limited feature constraints, large computational consumption, and subpar real‐time performance, allowing for quick and accurate loop‐closure detection. Moreover, the method demonstrates a time consumption reduction of two‐thirds compared to traditional approaches while exhibiting superior overall loop‐closure detection performance.