A Solid-State LiDAR SLAM System With Dual Verification Mechanism for Loop Closure Detection

As an essential component of simultaneous localization and mapping (SLAM) technology, loop closure detection can enhance the overall accuracy and consistency of the map. However, existing loop closure detection algorithm often perform poorly in solid-state LiDAR SLAM system. Solid-state LiDARs typically have a narrow field of view (FOV), which limits their ability to capture the surrounding environment, resulting in less environmental feature information available for loop closure detection and consequently impacting detection accuracy. To address this issue, this article designs a solid-state LiDAR SLAM system with a dual verification mechanism for loop closure detection. The system utilizes FAST-LIO2 as the LiDAR-inertial odometry (LIO) module, providing the loop closure detection module with registered scans and estimated odometry data. In the loop closure detection module, a stable triangle descriptor (STD) is generated, and then a quick historical descriptors matching is carried out in a hash table which serves as the descriptor database. Subsequently, an initial verification is carried out using normal vectors and curvature to generate a larger set of loop candidates. Next, a rigorous secondary verification of loop closure is performed using normal distributions transform (NDT) registration which leverages global optimization to achieve the optimal registration result, effectively avoiding incorrect loop closure detection and thus maintaining high accuracy and robustness of loop closure detection. Finally, the loop closure factors are integrated into graph optimization as global factors, thereby improving solid-state LiDAR SLAM accuracy in localization and mapping. This article extensively evaluates the proposed system on public datasets and datasets collected on campus. The experimental results show that compared to the state-of-the-art system FAST-LIO2, our proposed system improves pose accuracy by an average of 37% (root mean square value), indicating smaller overall errors and superior overall localization accuracy. On the other hand, the system proposed in this article implemented strict loop frame matching, which prevented incorrect loop closure detection, resulting in more accurate mapping.