Accurately perceiving drivable surfaces and ground obstacles from 3D LiDAR point clouds is a core task in the environmental perception of autonomous driving systems. Therefore, developing efficient and reliable point cloud segmentation algorithms is of paramount importance. Compared to structured environments, unstructured environments typically exhibit characteristics such as rough terrain and complex, unevenly distributed obstacles. These traits pose significant challenges to most existing point cloud segmentation algorithms, leading to issues such as low accuracy and poor real-time performance, which in turn severely affect the safety of navigation decisions made by autonomous vehicles in complex environments. To address these challenges, we propose a cascaded point cloud segmentation algorithm aimed at enhancing both the accuracy and efficiency of point cloud segmentation and providing reliable support for the safe operation of autonomous vehicles in unstructured environments.

To address the challenges in real-time performance and accuracy of LiDAR point cloud segmentation algorithms in complex, rugged, and unstructured environments, we propose a cascaded point cloud segmentation algorithm tailored for such environments. The algorithm consists of two main processes: ground point cloud segmentation and above-ground object segmentation. Given the characteristics of unstructured environments, such as complex terrain and unordered obstacle distribution, we first introduce inertial measurement unit (IMU) pre-integration to correct distortions in the LiDAR point clouds. The corrected point clouds are then projected into the dynamic concentric zone model (DCZM), and the principal component analysis (PCA) algorithm is used to classify the point clouds into ground and non-ground points. Based on the spatial distribution characteristics of LiDAR point clouds, a hierarchical clustering algorithm is proposed to segment above-ground objects. For high-density point clouds near the sensor, a 3D voxel clustering algorithm based on spherical coordinates is applied, which fully exploits the spatial continuity of the point cloud data to achieve efficient and accurate clustering. For sparse point clouds at greater distances from the sensor, a point cloud clustering algorithm is utilized based on key eigenvalues and normal vector angle constraints. Compared to traditional clustering algorithms, this method better characterizes the local geometric structure of the point clouds, thus enhancing segmentation performance in sparse regions.

In this study, we select sequences 01 (highway scenario), 09 (rural occlusion scenario), and 10 (undulating terrain scenario) from the Semantic KITTI dataset as experimental data for point cloud segmentation. These sequences reflect the effect of dynamic objects, occlusion, sparsity, and ground undulation on point cloud data in unstructured environments, thereby providing representative scenarios for algorithm validation. To quantitatively evaluate the effectiveness of the proposed algorithm, we use five key metrics for ground point cloud segmentation: precision (P), recall (R), F1 score (F₁), accuracy (A), and time (T). For above-ground object segmentation, we use three key metrics: segmentation entropy (SE) and scene accuracy (ACCscene). For ground point cloud segmentation, we conduct both qualitative (Fig. 9) and quantitative (Table 3) comparisons between the proposed algorithm and existing open-source algorithms, including GPF, LineFit, Patchwork, and Patchwork++. The results indicate that the proposed method outperforms the others in overall performance, particularly excelling in time. Compared to Patchwork++, the proposed method reduces time by 11.57 ms, which demonstrates significant improvement in real-time performance. For above-ground object segmentation, we compare the proposed algorithm with existing open-source algorithms, such as CVC, Travel, and Adaptive Clustering, and conduct both qualitative (Fig. 10) and quantitative (Table 4) analyses. The results show that the proposed method outperforms the others in overall performance, especially in terms of scene accuracy and time. Compared to the Travel method, the proposed algorithm improves these two metrics by 23.01% and 14.09 ms, respectively, which indicates its substantial advantage in segmentation accuracy and processing efficiency. To comprehensively evaluate the performance of the algorithm, we also conduct qualitative validation of ground point cloud segmentation (Fig. 11) and above-ground object segmentation (Fig. 12) using an all-terrain unmanned vehicle in a complex forest environment with significant slopes. The experimental results indicate that the proposed method still exhibits considerable advantages over the aforementioned algorithms in complex environments, which further demonstrates its effectiveness and reliability in unstructured environments.

We address the problem of point cloud segmentation in complex and dynamic unstructured environments, proposing a cascaded point cloud segmentation algorithm that processes ground and above-ground object point clouds separately. The algorithm corrects radar point cloud motion distortions through IMU pre-integration and introduces a DCZM for dynamically partitioning radar perception areas. Subsequently, we propose a hierarchical clustering algorithm for above-ground object segmentation, considering factors such as point cloud density, angular resolution, and sparsity, based on the spatial distribution characteristics of 3D point clouds. Experimental results on the Semantic KITTI dataset and all-terrain autonomous vehicle tests demonstrate that the proposed algorithm meets real-time requirements while achieving high segmentation accuracy and robustness, which shows strong practical value. This work focuses on enabling real-time navigation in unknown, unstructured environments by recognizing relevant objects. In the future, the algorithm will be combined with advanced deep learning models to further optimize its segmentation capability in complex scenarios and enhance its practical application value.