As a crucial aspect of automatic driving, unmanned vehicle technology has garnered extensive attention and research in both academia and industry. Autonomous vehicles require robust perception and decision-making systems for autonomous navigation, with simultaneous localization and mapping (SLAM) being one of the core components. While many advanced SLAM algorithms have achieved stable and high-precision positioning and mapping, challenges persist. For example, non-smooth and uneven roads can distort collected data, making it difficult to establish reliable feature correspondence between frames, leading to significant map drift and positioning errors. Given the inaccuracy of existing laser SLAM algorithms in ground segmentation, low feature-matching efficiency, and the high computational demands of traditional loop closure detection methods based on Euclidean distance, we propose a lidar SLAM algorithm with optimized ground segmentation and closed-loop detection strategy.

We first introduce a more reliable ground segmentation method for non-smooth roads during the preprocessing stage. By establishing a concentric region model for each point cloud frame and using principal component analysis (PCA) to extract statistical characteristics, we design a ground likelihood estimation binary classification method to remove unstable ground points. This approach addresses the issue of misclassifying non-ground points as ground due to small slopes between adjacent laser points, achieving more accurate segmentation of non-smooth road surfaces. Additionally, we introduce a sphere feature point extraction method alongside the standard edge and plane feature points. This enhances the point cloud’s useful information for matching between consecutive frames, improving both efficiency and robustness while reducing the influence of redundant point clouds. In addition, we propose a robust global alignment strategy based on Mahalanobis distance to replace the traditional iterative closest point (ICP) matching method using Euclidean distance. Mahalanobis distance, which measures covariance, can more effectively calculate the similarity between two unknown sample sets, thereby improving the accuracy of ICP closed-loop matching without excessive computational overhead.

We evaluate the positioning accuracy and trajectory in both urban and rural scenes using the KITTI dataset. The proposed method is compared with LOAM, LeGO-LOAM, and FAST-LIO algorithms for quantitative analysis. The absolute pose error (APE) results (Table 1) show that, based on the root mean square error-index, our algorithm significantly improves positioning accuracy and stability. In the ablation experiment (Table 2), our algorithm significantly improves positioning accuracy and stability. Through EVO trajectory visualization (Fig. 8), the proposed algorithm shows superior consistency with the ground truth trajectory in terms of trajectory deviation and closed-loop integrity, verifying the algorithm’s deployability. The algorithm’s time complexity and ability to process large datasets are also evaluated (Table 3). Compared with the LeGO-LOAM algorithm, our method improves the average loop frame matching calculation by 16.56%, meeting both high accuracy and real-time deployment requirements for unmanned vehicles. Finally, the algorithm’s robustness and generalization are validated using the M2DGR dataset and real-world mining environments in Shandong province (Figs. 10 and 11). The results confirm that our algorithm meets practical application needs.

Addressing the inaccuracies of existing laser SLAM algorithms in ground segmentation, low feature-matching progress, and high computational cost of traditional closed-loop detection methods based on Euclidean distance, we propose a laser SLAM algorithm based on optimized ground segmentation and closed-loop detection. This method uses the LeGO-LOAM framework to extract regional statistical features of the point cloud based on a concentric region model, incorporating a ground likelihood estimation binary classification method to accurately segment non-ground points and remove unstable ground points. Additionally, the extraction of sphere features enhances the accuracy of inter-frame matching. Finally, we optimize the closed-loop strategy with a robust decoupling global alignment based on Mahalanobis distance, effectively correcting cumulative errors and improving overall positioning and mapping accuracy. Comparative experiments using the KITTI public dataset demonstrate the advantages of our algorithm in positioning accuracy, and the M2DGR dataset further verifies its applicability in real-world scenarios. Our algorithm successfully constructs a globally consistent 3D map with high precision.