# LPIR-GSeg: LiDAR Projection Image-Based and Region-Wise Ground Segmentation in Urban Environment

## Abstract

Fast and accurate ground segmentation is an essential task for various applications, such as point cloud clustering, 3-D object segmentation, and simultaneous localization and mapping (SLAM). To enhance both accuracy and efficiency, our method avoids the traditional plane-fitting approach for 3-D point clouds. We propose a light detection and ranging (LiDAR) projection image-based and region-wise ground segmentation method in urban road scenarios named LPIR-GSeg. We calibrate the mounted angle between the LiDAR and the horizontal plane to correct the heights of 3-D points. For coarse segmentation, we encode the point cloud into a concentric zone model representation and find the corresponding grid for each 3-D point. The heights of the grids will be used for fast ground segmentation, which is used as an initial value for the fine segmentation. To improve the accuracy, we project the point cloud into various cylindrical images based on the LiDAR projection model. Then, a region-growing algorithm based on the projection images is applied to achieve refined ground segmentation. Comprehensive experiments were conducted with a real-world platform and the SemanticKITTI datasets. Our proposed method achieves a precision of 95.93% and a recall of 92.20% on the SemanticKITTI dataset.

## Authors

Yang Sun *School of Instrument Science and Engineering, Southeast University, Nanjing, China* [ORCID: 0009-0002-3106-0164](https://orcid.org/0009-0002-3106-0164)

Shengyi Liu *School of Instrument Science and Engineering, Southeast University, Nanjing, China* [ORCID: 0000-0002-6639-372X](https://orcid.org/0000-0002-6639-372X)

Chao Yan *School of Electrical Engineering and Automation, Suzhou University of Technology, Changshu, China* [ORCID: 0000-0002-2365-3122](https://orcid.org/0000-0002-2365-3122)

Kaiwei Tang *School of Instrument Science and Engineering, Southeast University, Nanjing, China* [ORCID: 0000-0003-4441-7468](https://orcid.org/0000-0003-4441-7468)

Qing Wang *School of Instrument Science and Engineering, Southeast University, Nanjing, China* [ORCID: 0000-0002-6940-6805](https://orcid.org/0000-0002-6940-6805)

Xiaolin Meng *School of Instrument Science and Engineering, Southeast University, Nanjing, China* [ORCID: 0000-0003-2440-8054](https://orcid.org/0000-0003-2440-8054)

## Publication Information

**Journal:** IEEE Sensors Journal **Year:** 2025 **Volume:** 25 **Issue:** 16 **Pages:** 31357-31369 **DOI:** [10.1109/JSEN.2025.3579686](https://doi.org/10.1109/JSEN.2025.3579686) **Article Number:** 11045759 **ISSN:** Print ISSN: 1530-437X, Electronic ISSN: 1558-1748, CD: 2379-9153

## Metrics

**Total Downloads:** 113

## Funding

- China Civil Aviation Safety Capability Project (Grant: 2023-73)
- Natural Science Foundation of Jiangsu Higher Education Institutions of China (Grant: 24KJB420001)
- National Natural Science Foundation of China (Grant: 42374029)

---

## Keywords

**IEEE Keywords:** Point cloud compression, Laser radar, Image segmentation, Accuracy, Three-dimensional displays, Sensors, Fitting, Roads, Deep learning, Simultaneous localization and mapping

**Index Terms:** Urban Environments, Light Detection And Ranging, Ground Segment, Point Cloud, Segmentation Method, Projection Images, Segmentation Accuracy, Urban Road, Final Segmentation, Simultaneous Localization And Mapping, Image-based Methods, Urban Scenarios, Fast Segmentation, Performance Of Algorithm, Image Pixels, Flat Surface, Precision And Recall, Road Surface, Learning-based Methods, Ground Plane, Ground Points, Height Threshold, Vertical Resolution, Point Cloud Segmentation, Random Sample Consensus, Plane Fitting, Segmentation Results, Ground Slope, Current Point, Tilt Angle

**Author Keywords:** 3-D light detection and ranging (LiDAR) simultaneous localization and mapping (SLAM), autonomous driving, coarse-to-fine segmentation, ground segmentation, LiDAR projection images

undefined
## SECTION I. Introduction

Ground segmentation, a critical preprocessing step in point cloud perception tasks, is fundamental for autonomous driving systems with structured representations of drivable areas. By segmenting ground from 3-D light detection and ranging (LiDAR) point clouds, this step supports essential functions such as robotic autonomous navigation, collision-free path planning, and simultaneous localization and mapping (SLAM), while directly impacting the perception systems in dynamic object detection. Compared to point cloud detection, segmentation methods based on cameras are more mature and efficient for understanding complex scenes, utilizing hierarchical feature extraction and parallel computing. However, 3-D LiDAR remains the most important sensor for complex scene understanding due to its all-weather functionality, high-precision ranging, and panoramic coverage, compared with other sensors such as cameras and 4-D millimeter-wave radars.

The accuracy-efficiency tradeoff is the biggest challenge in LiDAR-based ground segmentation. For example, Velodyne HDL-64 generates millions of point cloud points per second, embedding rich environmental semantic information with significant computational cost. Current approaches typically utilize the partition strategy to divide the point cloud into a few grids, enhancing feature extraction and segmentation accuracy, such as [^1] and [^2]. However, excessive region partitioning and fitting significantly increase computational complexity and become difficult to maintain the same accuracy in various scenarios. Besides, the flat ground assumption is not always satisfied in complex terrains, such as roads with inclines or grass areas, especially for traditional methods relying on ground height estimation [^3] or plane fitting [^4]. For embedded platforms, this accuracy-efficiency tradeoff is particularly critical, and the presence of hyperparameters further complicates engineering deployment. Notably, the computer vision field has established a mature framework for image segmentation tasks [^5], offering valuable insights for point cloud segmentation. For example, the mechanical 3-D LiDAR can produce ordered point clouds and project sparse point clouds onto a 2-D plane to form LiDAR projection images with the advantages of feature representation and computational optimization. However, research on ground segmentation using such projection images remains limited, a striking contrast to its theoretical promise.

As deep learning advances, the perception becomes more and more accurate and fast [^6], [^7]. However, our design does not use any deep learning models as ground segmentation is the front end of the entire perception task. Ground segmentation is a module with low computational requirements and low algorithmic complexity, which traditional methods without deep learning are sufficient. We hope to design a simple, plug-and-play, and nonlearning segment module for our LiDAR SLAM system.

Addressing the inherent tradeoff between accuracy and efficiency in existing ground segmentation methods, this study innovatively integrates image processing techniques with region partitioning strategies, proposing a fast ground point cloud segmentation algorithm based on a “coarse-to-fine” paradigm. The framework significantly reduces computational overhead while maximizing segmentation accuracy by avoiding complex spatial fitting and optimization processes. The core contributions of this research are summarized as follows.

1. A new calibration method for the LiDAR tilt angle is introduced to compensate for the height of 3-D points. This method significantly reduces the computational complexity of traditional height-based segmentation and requires only flat scenarios, such as underground parking lots or plazas, for rapid calibration, demonstrating excellent engineering practicality.
2. We design an adaptive ground segmentation algorithm combining grid processing with LiDAR projection maps. The algorithm divides LiDAR point clouds into multiscale concentric sector regions, establishing adaptive thresholds through neighborhood height relationships for coarse segmentation. To further address false negative issues in coarse segmentation, we innovatively introduce a region-growing algorithm based on LiDAR projection maps, achieving precise fine segmentation.
3. Experimental validation shows that our method demonstrates excellent performance on both the SemanticKITTI [^8] and BotanicGarden [^9] benchmark datasets and real-world scenes, particularly exhibiting superior robustness in nonplanar terrains such as inclined roads. In terms of time efficiency, our method performs comparably to state-of-the-art algorithms.
4. To promote academic exchange and technical verification, we have implemented the complete algorithm framework in C++ and made the source code publicly available at https://github.com/sychina/LPIR-GSeg

## SECTION II. Related Works

### A. Learning-Based Method

Learning-based ground segmentation methods can be categorized into two main types: traditional machine learning methods and deep learning methods. In traditional machine learning, the Markov random field (MRF)-based segmentation algorithm has been most studied. Rummelhard et al. [^10] presented a 3-D point cloud ground segmentation system, based on a dynamic estimation of local ground elevation and slope, by modeling the ground as a spatio-temporal conditional random field. Golovinskiy and Funkhouser [^11] first proposed a point cloud segmentation framework using min-cut. They built a *k*-nearest neighbor graph and used foreground/background seed points to achieve binary segmentation of point clouds, which relied on the selection of seed points. Later, Huang et al. [^12] and Guo et al. [^13] set ground and nonground seed points in advance based on MRF to reduce computation with hierarchical segmentation, which improved running speed but lowered accuracy. Overall, MRF-based methods need prior knowledge to set background and foreground points, but the computation cost is too high for real-time use. In contrast, deep learning methods run very efficiently during inference. However, training these models requires tens of thousands of labeled point cloud frames. Deep learning methods [^14] can learn ground features from training data, such as [^4] and [^15]. Paigwar et al. [^14] proposed an end-to-end ground segmentation method in a grid-based representation. It uses points in the pillars to learn pillar-wise features that can be scattered back to a 2-D pseudo image, and it performs well in public datasets. LiDAR projection images are also used in models to improve segmentation accuracy [^16]. Most autonomous vehicles are embedded with limited processors, which makes it difficult to deploy deep learning ground segmentation models with high computation costs for those platforms. In addition, the learning-based algorithm does not have scene generalization capabilities resulting in poor performance in other scenes and sensors, requiring retraining.

### B. Fitting- and Partition-Based Method

Fitting- and partition-based methods are often used together. In 3-D space, plane fitting commonly uses the random sample consensus (RANSAC) algorithm [^17], as in [^18]. The principal component analysis (PCA) algorithm [^19] is another option to fit ground planes. RANSAC-based methods give better plane fitting accuracy than PCA-based methods when handling noisy point clouds. However, PCA is faster than RANSAC in computation speed, such as [^20], [^21], and [^22] with global or local fitting. They fit the ground as one whole or multiple planes and segment the ground with height or variance thresholds of the distance between the points and the plane. However, these methods reduce fitting accuracy on nonflat ground and are easily influenced by point clouds in the vertical direction. Furthermore, multiple fitting iterations are needed to get results from point clouds, which leads to high computational costs. Most real road surfaces are nonflat curves. To avoid curve fitting, partition-based methods divide point clouds into many small grids and fit ground in each grid [^23], [^24]. Alternatively, Lim et al. [^1] and Lee et al. [^2] measured point cloud traits in grids as ground features, such as flatness, elevation, and uprightness, and can be deployed on field-programmable gate array (FPGA) platforms [^25]. An inertial measurement unit (IMU) sensor can also be used to measure the pitch and roll angles of LiDAR, rotating the raw point cloud into an upright orientation for ground segment [^26]. This approach works well for complex terrain and ground-to-nonground transition zones but still limits parameter tuning, data reliance, and real-time performance. For example, overpartition leads to segmentation errors in edges with sudden height changes, such as curbs or tire contact zones. In addition, the selection of height threshold struggles with complex terrain shifts, such as vegetation-covered areas, leading to mistaken for low obstacles. These drawbacks directly reduce the recall and precision of downstream detection modules.

### C. Ordered Point Cloud and Image-Based Method

Ordered point clouds refer to point clouds in which each point follows a well-defined order, similar to the arrangement of pixels in an image, which directly corresponds to the physical scanning pattern of mechanical spinning LiDAR. Leveraging ordered point clouds, the geometric relationships between scan lines can be determined based on the ground slope [^27], [^28], enabling dynamic adjustment of height thresholds to segment nonground points that do not meet the required spacing. Some approaches comprehensively set distance and height thresholds by considering factors such as varying point cloud density at different distances [^29], the slope of the ground plane [^30], and the relationship between the ground plane’s normal vector and the LiDAR tilt angle [^31]. Although these methods exhibit excellent filtering performance in structured scenarios, like urban roads with clear obstacle boundaries and flat surfaces, and are well-suited for deployment in autonomous driving systems, they have limitations. In bumpy or inclined road conditions, the uneven distribution of scan lines can significantly reduce segmentation accuracy. Moreover, the effectiveness of these algorithms is strictly dependent on sensor installation parameters (i.e., the LiDAR must be mounted horizontally on the vehicle roof). Additionally, 3-D point clouds can be projected onto a 2-D image based on their vertical and horizontal resolutions (commonly referred to as LiDAR projection images), allowing for ground segmentation through image processing techniques. This LiDAR projection image contains a depth image, height image, or slope image [^32]. Subsequently, pixel thresholds can be set according to the geometric relationships among the point clouds [^33], and segmentation can be performed using image processing operations, such as dilation, erosion, closing, opening [^34], and convolution [^35]. However, due to pixel and projection errors, this approach often yields erroneous segmentation results along image boundaries, necessitating additional compensatory methods.

In summary, research on image-based point cloud segmentation remains relatively underexplored. This study addresses challenges such as restricted accuracy on the edge and achieving performance that is comparable to state-of-the-art methods.

## SECTION III. System Introduction

### A. Framework of LPIR-GSeg

The pipeline of the proposed work LPIR-GSeg is shown in Fig. 1. The following paragraphs highlight the description and reasoning behind each module of LPIR-GSeg. LPIR-GSeg mainly consists of three parts: preprocessing, coarse segmentation, and fine segmentation.

![Figure 1](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng1-3579686-large.gif)

*Fig. 1. Overview of LPIR-Seg. In the preprocessing module, some noise points under the actual ground are removed to calibrate the LiDAR tilt angle, and then, we project the point cloud into the concentric zone model. A fast coarse segmentation module divides point cloud into high-confidence ground points ${P}_{g}$ and nonground points ${P}_{\textit {ng}}$ through adaptive height thresholding and PCA. A fine segmentation module extracts undetermined boundary points ${P}_{\textit {un}}$ with multithreshold of LiDAR projection images and refines point cloud through ground region growing algorithms.*

### B. Preprocessing

The preprocessing module employs a spherical projection-based geometric correction framework, estimating LiDAR mounting inclination in real time through ground trend fitting and achieving point cloud geometric correction. The processing pipeline comprises some critical steps.

#### 1) Ground Reflection Noise Filtering:

Reflective materials prevalent in urban environments (e.g., glass curtain walls and metal surfaces) generate anomalous noise points, typically exhibiting negative height values (below actual ground level), severely affecting the accuracy of minimum point elevation threshold calculation in the coarse segmentation module, as shown in Fig. 2. To address this, it is common practice to set a custom height threshold $h_{\text {noise}}$ and intensity threshold $i_{\text {noise}}$ to filter these noise points with small incidence angles and low intensity.

![Figure 2](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng2-3579686-large.gif)

*Fig. 2. Diagram of ground reflection noise. Demonstrating typical ground reflection noise phenomena using urban road point cloud data captured by RS-Ruby-Lite LiDAR. Red rectangles highlight the point clouds that needed to be removed caused by reflection.*

#### 2) LiDAR Projection Model:

Mechanical spinning LiDAR exhibits deterministic geometric characteristics in their scanning patterns, with vertical resolution and horizontal resolution forming a regular spherical sampling grid. Based on this, we adopt a spherical projection model to map 3-D point clouds onto a 2-D image plane, generating image-based feature representations. For any LiDAR point $\mathbf {p_{i}}=(x_{i},y_{i},z_{i})$ in the LiDAR point cloud $p = \{ {\mathbf {p}_{1},\mathbf {p}_{2},\mathbf {p}_{3},\ldots ,\mathbf {p}_{\mathbf {i}},\ldots ,\mathbf {p}_{\mathbf {N}}} \}$, the projection image $\mathbf {I}_{u_{i},v_{i}}$ can be obtained by the following equation:

$$
\begin{align*} \begin{pmatrix} u_{i} \\ v_{i} \end{pmatrix} = \begin{pmatrix} {1}/{2} \times \left [{{ 1 - \tan ^{-1}\left ({{ {y_{i}}/{x_{i}} }}\right ) \cdot \pi ^{-1} }}\right ] \cdot I_{\text {width}} \\ \left [{{ 1 - \left ({{ \sin ^{-1}\left ({{ {z_{i}}/{d_{i}} }}\right ) + v_{\text {fov}}^{\text {up}} }}\right ) \cdot 1/{v_{\text {fov}}^{\text {up} + \text {down}}} }}\right ] \cdot I_{\text {height}} \end{pmatrix} \tag {1}\end{align*}
$$

where $I_{\text {width}} \times I_{\text {height}}$ represents the image resolution, image width $I_{\text {width}} = h_{\text {fov}}/h_{\text {res}}$ is calculated by horizontal field of view (FOV) and horizontal resolution $h_{\text {res}}$, and image height $I_{\text {height}} = v_{\text {fov}}/v_{\text {res}}$ is calculated by vertical FOV $v_{\text {fov}} = v_{\text {fov}}^{\text {up}} + v_{\text {fov}}^{\text {down}}$ and vertical resolution $v_{\text {res}}$. Image height is usually consistent with the number of LiDAR lines. $d_{i} = (x_{i}^{2} + y_{i}^{2})^{1/2}$ is the distance from the LiDAR point to the origin of the LiDAR on the *XOY* plane.

#### 3) LiDAR Tilt Angle Calibration:

Nonhorizontal installation may be caused by LiDAR mounting errors, support structure deformation, and uneven tire pressure. We propose a real-time tilt angle calibration method based on ground trend fitting shown in Fig. 3. The tilt angle calibration improves the coarse segmentation performance by compensating for the point cloud height, and a before-and-after comparison image is shown in Fig. 4. Calibration of the tilt angle is not required for each execution of the code. As part of the sensor extrinsic parameters, it requires calibration only once for the same acquisition hardware and sensors.

![Figure 3](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng3-3579686-large.gif)

*Fig. 3. Diagram of the inclination angle of LiDAR installation. A projection of the angle of inclination in the ZOX and ZOY planes, where $\rightarrow {V}$ indicates that the vehicle’s direction of travel is to the right and $\otimes {V}$ indicates that the vehicle’s direction of travel is perpendicular to the paper.*

![Figure 4](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng4-3579686-large.gif)

*Fig. 4. Before-and-after tilt angle calibration. The green points indicate the point cloud before calibration, while the blue points indicate that after calibration.*

First, the vehicle is driven to a flat and open road surface, e.g., an underground parking lot, assuming that the local road surface is an ideal plane. Generate a LiDAR depth image $\mathbf {I}_{\text {range}}$ from the current frame point cloud and extract the point cloud from four main directions (front, rear, left, and right). To ensure fitting accuracy, sample point clouds within a 3-pixel width on both sides of each fitting line. Then, valid fitting points are selected by setting ground height thresholds $h_{\text {thr}}$ and distance thresholds $d_{\text {thr}}$. By RANSAC, the fitting lines in *XOZ* and *YOZ* ground planes can be obtained by the following equation:

$$
\begin{align*} \widehat {z_{x}} = k_{xz}x + b_{x} \\ \widehat {z_{y}} = k_{yz}y + b_{y}. \tag {2}\end{align*}
$$

where $k_{xz}$ and $k_{yz}$ are the slope of the ground planes and $\widehat {z_{x}}$ and $\widehat {z_{y}}$ are the compensation coefficients of *z*. $b_{x}$ and $b_{y}$ indicate the sensor height at the origin in the LiDAR coordinate, and the error between them should be less than 1 cm. If the error is too large, recalibration is required on a flatter surface. Finally, the compensated LiDAR point cloud $z_{t}$ is obtained through the observed measurement with the following equation:

$$
\begin{equation*} z_{t} = z - \left ({{k_{xz}x + k_{yz}y}}\right ). \tag {3}\end{equation*}
$$

#### 4) Concentric Grid Model:

In the complex terrain of urban environments, slope variations, curvature changes, and single plane or surface assumptions are inadequate for accurate modeling. Therefore, we propose a region segmentation strategy based on nonuniform sector grids. This design is based on two key observations: First, the ability of ground detection decreases with distance that the near regions exist most of the ground points. Second, point cloud density follows a radial decay distribution, with near density being a few times that of far.

Thus, an adaptive grid partitioning strategy is adopted: high-resolution small grids for near regions and low-resolution large grids for far regions. The number of grids is determined by the radius resolution $\text {res}_{r}$ and the azimuth resolution $\text {res}_{\theta }$. For a specific corrected LiDAR point $Z_{r,\theta }^{t}$, belonging to the *i*th circle, *j*th grid $Z_{r,\theta }^{i,j}$, as shown in the figure, *i* and *j* are defined as follows:

$$
\begin{align*} & \begin{cases} \displaystyle i = \text {floor}\left ({{ \frac {r}{{\text {res}}_{r}^{c}} }}\right ) \\ \displaystyle j = \text {floor}\left ({{ \frac {{1}/{2}\lbrack {1 - \theta \cdot \pi ^{- 1}} \rbrack \cdot I_{\text {width}}}{{\text {res}}_{\theta }^{c}} }}\right ) \end{cases}~,\text {if}~r \lt = r_{\text {thr}}^{\text {distance}} \\ & \begin{cases} \displaystyle i = \text {floor}\left ({{ \frac {r}{{\text {res}}_{r}^{c}} }}\right ) + \text {floor}\left ({{ \frac {r - r_{\text {thr}}^{\text {distance}}}{{\text {res}}_{r}^{f}} }}\right ) \\ \displaystyle j = \text {floor}\left ({{ \frac {{1}/{2}\lbrack {1 - \theta \cdot \pi ^{- 1}} \rbrack \cdot I_{\text {width}}}{{\text {res}}_{\theta }^{f}} }}\right ) \end{cases},\text {if}~r \gt r_{\text {thr}}^{\text {distance}}\!\!. \tag {4}\end{align*}
$$

In (4), *i* and *j* start from 0, and the function $\text {floor}(\cdot)$ indicates that the result is rounded down. $r = (x^{2} + y^{2})^{1/2}$, $\theta = {\tan ^{- 1}({y,x})}$, $\text {res}_{r}^{c}$ represents the radial resolution of the near grid, $\text {res}_{\theta }^{c}$ represents the near grid orientation angle resolution, $r_{\text {thr}}^{\text {distance}}$ represents the distance threshold in the vicinity, $\text {res}_{r}^{f}$ represents the radial resolution of the distant grid, and $\text {res}_{\theta }^{f}$ represents the angular resolution of the distant grid direction.

### C. Coarse Segmentation

In urban road environments, ground points typically appear as the lowest points within local regions. For rapid segmentation, this study employs an adaptive height threshold strategy, combined with a multithreaded accelerated PCA algorithm for ground segmentation.

#### 1) Ground Seed Point Selection Based on Adaptive Height Threshold:

First, the point cloud of the current frame is traversed, and the maximum height $h_{\max }^{i,j}$, minimum height $h_{\min }^{i,j}$, and the average height $h_{a_{n}}^{i,j}$ within the surrounding $3\times 3$ neighborhood grids are calculated for each sector grid. The point cloud within each sector grid $Z_{r,\theta }^{i,j}$ uses (5) to distinguish ground point $P_{g}$ from nonground point $P_{ng}$

$$
\begin{align*} \begin{cases}P_g=\left\{P_i \in Z_{r, \theta}^{i, j}| | h_i-h_{\min }^{i, j} \mid \leq h_{\mathrm{thr}}^{i, j}\right. \\\text { or } \left.\left|h_{\max }^{i, j}-h_{n_{\text {amin }}}^{i, j}\right| \leq h_{\mathrm{athr}}^{i, j}\right\} \\\mid P_{n g}=\left\{P_i \in Z_{r, \theta}^{i, j} \mid P_i \notin P_g\right\} .\end{cases} \tag {5}\end{align*}
$$

In (5), $h_{\text {thr}}^{i,j} = h_{\min }^{i,j} + \lambda$ is the height threshold of grid $(i,j)$ and $\lambda = {Z_{r,\theta }^{i,j}}_{\text {size}} \times \xi _{\text {slope}}$ is the compensation coefficient for the ground slope factor, which is obtained by multiplying the size of grid ${Z_{r,\theta }^{i,j}}_{\text {size}}$. We set the maximum slope $\xi _{\text {slope}}$ of ground 5%$\thicksim 8$%. $h_{a_{\text {thr}}}^{i,j}$ is the height threshold of the adjacent grids of grid $(i,j)$. The selection of ground seed points comprises two aspects. First, after eliminating ground reflection noise, the lowest point within a grid is typically a ground point; Second, if the highest point in the current grid is lower than the lowest point in the neighboring grids and the trend variation within the current grid is minimal, it confirms that the entire grid represents ground. This method allows for fast segmentation of flat ground.

#### 2) Ground Feature Calculation Based on Fast PCA:

Ground segmentation can be treated as the issue of binary classification, and its goal is to segment a 3-D point cloud into two sets. However, for inclined road surfaces, the seed points may include oversegmented points false positive (FP) and undersegmented points false negative (FN), as illustrated in Fig. 5. FP often appears at the boundary between ground and nonground points, such as tires, street lights, and the base of trees. FN typically occurs on steep slopes and uneven grassy areas. The FN points in the segmentation results can be reduced by decreasing the grid size, but FP points cannot be mitigated by increasing the grid resolution.

![Figure 5](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng5-3579686-large.gif)

*Fig. 5. Schematic of adaptive height threshold segmentation. Green represents TP, red represents TN, blue represents FP, and yellow represents FN. The green parallelogram represents the plane fit by PCA.*

Fig. 5 shows the schematic of adaptive height threshold segmentation. Green represents true positive (TP), red represents true nagetive (TN), blue represents FP, and yellow represents FN. The green parallelogram represents the plane fit by PCA. In this article, the ground trend is estimated by using PCA to calculate the ground features in the grid, and the height threshold is adjusted to reduce the number of FP and FN points. The ground seed point $P_{g}$ in grid ${Z_{r,\theta }^{i,j}}_{\text {size}}$ is selected for plane estimation using the linear method with the following equation:

$$
\begin{align*} ax + by + cz & = {-} d \\ \mathbf {n}^{T}\mathbf {x} & = - d. \tag {6}\end{align*}
$$

In (6), $\mathbf {n} = \lbrack {a,b,c} \rbrack ^{T}$ represents the plane normal vector and $\mathbf {x} = \lbrack {x,y,z} \rbrack ^{T}$. The multithreaded accelerated PCA module is used to calculate **n**, and the covariance matrix of ground seed point $P_{g}$ in the current grid can be obtained with the following equation:

$$
\begin{equation*} \mathbf {C}_{g} = \frac {1}{| P_{g} |}{\sum _{p_{i} \in P_{g}}{\left ({{p_{i} - \overline {p}}}\right )\left ({{p_{i} - \overline {p}}}\right )^{T}}} \tag {7}\end{equation*}
$$

where $\overline {p}$ is the average of nonground points. The eigenvalues and eigenvectors corresponding to the three main directions of the point cloud distribution can be obtained by singular value decomposition (SVD). The three eigenvalues of the covariance matrix $\mathbf {{C}_{g}}$ are sorted as $\lambda _{1} \gt \lambda _{2} \gt \lambda _{3}$, corresponding to the eigenvector $\boldsymbol {\xi }_{1}$–$\boldsymbol {\xi }_{3}$. For planar distribution, the eigenvector $\boldsymbol {\xi }_{3}$ perpendicular to the plane has the smallest eigenvalue $\lambda _{3}$, corresponding to the normal vector **n** of the plane. After that, the plane parameter *d* is obtained. By calculating the average distance of seed points from the plane as a threshold, the number of FP and FN points can be effectively reduced.

### D. Fine Segmentation

While the coarse segmentation module effectively handles flat urban road scenarios, its performance significantly degrades in unstructured terrains (e.g., slopes and grassy areas). To address this, we propose a cascaded region-growing algorithm based on LiDAR projection maps to enhance ground modeling accuracy in complex environments through secondary refinement.

#### 1) LiDAR Projection Image-Based Region Segmentation:

Based on the LiDAR projection model, LiDAR point clouds can be projected onto a fixed-size image to determine pixel positions. The pixel values of the projection image are determined by the position, distance, and spatial relationships of the LiDAR points with their surrounding points. Depending on the type of projection image, subsequent modules utilize geometric information such as contours, depth, and positional relationships of the surrounding environment. Additionally, computer vision image processing techniques, such as erosion, dilation, opening, and closing operations, are applied to segment the LiDAR point clouds. The projected image corresponding to the LiDAR point $\mathbf {p_{i}}$ includes the seven following categories, as shown in Fig. 6.

1. *Range Image*$\mathbf {I}_{\text {range}}$*:* The pixel value is $\mathbf {I}_{\text {range}}({u_{i},v_{i}}) = \mathbf {p}_{\mathbf {i}}^{d}$.
2. *Intensity Image*$\mathbf {I}_{\text {intensity}}$*:* The pixel value is $\mathbf {I}_{\text {intensity}}({u_{i},v_{i}}) = \mathbf {p}_{\mathbf {i}}^{\text {intensity}}$.
3. *Height Image*$\mathbf {I}_{z}$*:* The pixel value is $\mathbf {I}_{z}({u_{i},v_{i}}) = \mathbf {p}_{\mathbf {i}}^{z}$.
4. *Slope Image*$\mathbf {I}_{\text {slope}}$*:* The pixel value is the ratio of the height change rate and depth change rate of each LiDAR point and its adjacent pixels. $\mathbf {I}_{\text {slope}}({u_{i},v_{i}}) = (| {\mathbf {I}_{z}(u_{i} + 1,v_{i})-\mathbf {I}_{z}(u_{i},v_{i})} |/$) $| {\mathbf {I}_{\text {range}}({u_{i} + 1,v_{i}}) - \mathbf {I}_{\text {range}}({u_{i},v_{i}})} |$ reflects the smoothness of the road surface.
5. *Edge Image*$\mathbf {I}_{\text {edge}}$*:* The pixel value is the height difference between the pixel corresponding to the current LiDAR point and the left and right adjacent pixels corresponding to the LiDAR point. $\mathbf {I}_{\text {edge}}^{\text {right}}({u_{i},v_{i}}) = | \mathbf {I}_{z}({u_{i+1},v_{i}}) -\mathbf {I}_{z}({u_{i},v_{i}})|$ and $\mathbf {I}_{\text {edge}}^{\text {left}}({u_{i},v_{i}}) = | \mathbf {I}_{z}({u_{i},v_{i}}) -\mathbf {I}_{z}({u_{i-1},v_{i}})|$. Pixels in the image are classified into edge pixels and nonedge pixels. Edge pixels represent the boundaries of objects, while nonedge pixels may correspond to flat surfaces or buildings.
6. *Elevation Standard Deviation Image*$\mathbf {I}_{\text {stdz}}$*:* The pixel value is $\mathbf {I}_{\text {zstd}}({u_{i},v_{i}}) = \overline {{\mathbf {I}_{z}({u_{i},v_{i}})}^{2}} - {\overline {\mathbf {I}_{z}({u_{i},v_{i}})}}^{2}$, and elevation variogram reflects the rate of elevation change in the image and is used to find the ground plane and object boundaries.
7. *Intensity Standard Deviation Image*$\mathbf {I}_{\text {stdi}}$*:* The pixel value is $\mathbf {I}_{\text {zsti}}({u_{i},v_{i}}) = \overline {{\mathbf {I}_{z}({u_{i},v_{i}})}^{2}} - {\overline {\mathbf {I}_{z}({u_{i},v_{i}})}}^{2}$. The point cloud intensity variance map reflects the rate of change of intensity in the image, and the reflection intensity of each object is different, which can be used to find the boundary of each object.

![Figure 6](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng6-3579686-large.gif)

*Fig. 6. LiDAR projection false-color image.*

According to different LiDAR projection images, different threshold intervals are set to classify image pixels. The classification rules are given as follows:

$$
\begin{align*} f_{\text {edge}}\left ({{u_{i},v_{i}}}\right ) & = \begin{cases} \displaystyle 1, \text {if}~ \mathbf {I}_{\text {edge}}^{\text {right}}\left ({{u_{i},v_{i}}}\right ) \geq T_{\text {edge}} \text {or} \\ \displaystyle \qquad \mathbf {I}_{\text {edge}}^{\text {left}}\left ({{u_{i},v_{i}}}\right ) \geq T_{\text {edge}} \\ \displaystyle 0, \text {else} \end{cases} \tag {8}\\ f_{\text {slope}}\left ({{u_{i},v_{i}}}\right ) & = \begin{cases} \displaystyle -1, & \text {if}~ \mathbf {I}_{\text {slope}}\left ({{u_{i},v_{i}}}\right ) \leq T_{\text {slope}}^{\text {low}} \\ \displaystyle 0, & \text {if}~ T_{\text {slope}}^{\text {low}} \lt \mathbf {I}_{\text {slope}}\left ({{u_{i},v_{i}}}\right ) \lt T_{\text {slope}}^{\text {high}} \\ \displaystyle 1, & \text {if}~ \mathbf {I}_{\text {slope}}\left ({{u_{i},v_{i}}}\right ) \geq T_{\text {slope}}^{\text {high}} \end{cases} \tag {9}\\ f_{\text {stdz}}\left ({{u_{i},v_{i}}}\right ) & = \begin{cases} \displaystyle -1, & \text {if}~ \mathbf {I}_{\text {stdz}}\left ({{u_{i},v_{i}}}\right ) \leq T_{\text {stdz}}^{\text {low}} \\ \displaystyle 0, & \text {if}~ T_{\text {stdz}}^{\text {low}} \lt \mathbf {I}_{\text {stdz}}\left ({{u_{i},v_{i}}}\right ) \lt T_{\text {stdz}}^{\text {high}} \\ \displaystyle 1, & \text {if}~ \mathbf {I}_{\text {stdz}}\left ({{u_{i},v_{i}}}\right ) \geq T_{\text {stdz}}^{\text {high}} \end{cases} \tag {10}\\ f_{\text {stdi}}\left ({{u_{i},v_{i}}}\right ) & = \begin{cases} \displaystyle -1, & \text {if}~ \mathbf {I}_{\text {stdi}}\left ({{u_{i},v_{i}}}\right ) \leq T_{\text {stdi}}^{\text {low}} \\ \displaystyle 0, & \text {if}~ T_{\text {stdi}}^{\text {low}} \lt \mathbf {I}_{\text {stdi}}\left ({{u_{i},v_{i}}}\right ) \lt T_{\text {stdi}}^{\text {high}} \\ \displaystyle 1, & \text {if}~ \mathbf {I}_{\text {stdi}}\left ({{u_{i},v_{i}}}\right ) \geq T_{\text {stdi}}^{\text {high}} \end{cases} \tag {11}\end{align*}
$$

where $T_{\text {edge}}$, $T_{\text {slope}}$, $T_{\text {stdz}}$, and $T_{\text {stdi}}$ and their superscript low and high constitute the threshold of pixel-level segmentation. The left and right edge images can screen out the boundary of the object’s height change according to the pixel, but the point may be located on the ground or the object, so it is classified as an unknown classification point. The segmentation motivation of $\mathbf {I}_{\text {slope}}$, $\mathbf {I}_{\text {stdz}}$, and $\mathbf {I}_{\text {stdi}}$ is to use the lower threshold of the feature to find the ground pixel from the nonedge pixel and the higher threshold to reconfirm the nonground pixel from the edge pixel. According to the segmentation result of $f_{\text {edge}}$, $f_{\text {slope}}$, $f_{\text {stdz}}$, and $f_{\text {stdi}}$, the classification result figure $\mathbf {I}_{\text {label}}$ is generated. Finally, the region growth algorithm is used to further refine the unknown classification points with (12). Physically, they represent the thresholds for the relationship between adjacent pixels in the LiDAR projection images. The initial parameters are obtained by calculating the variance of the pixels and then fine-tuned based on the segmentation results empirically

$$
\begin{align*} \mathbf {I}_{\text {label}}\left ({{u_{i},v_{i}}}\right ) = \begin{cases} \displaystyle l_{\text {ground}}, \text {if}~ f_{\text {edge}} = 0 \text {and} \\ \displaystyle \qquad -3 \leq f_{\text {slope}} + f_{\text {stdz}} + f_{\text {stdi}} \lt 0 \\ \displaystyle l_{\text {unknown}}, \text {if}~ f_{\text {edge}} = 1 \text {and} \\ \displaystyle \qquad f_{\text {slope}} + f_{\text {stdz}} + f_{\text {stdi}} = 0 \\ \displaystyle l_{\text {noground}}, \text {if}~ f_{\text {edge}} = 0 \text {and} \\ \displaystyle \qquad 0 \lt f_{\text {slope}} + f_{\text {stdz}} + f_{\text {stdi}} \leq 3. \end{cases} \tag {12}\end{align*}
$$

#### 2) Fine Segmentation Based on Region Growth Algorithm:

After the aforementioned process, both ground points and nonground points are high-confidence points, and the next task is to distinguish unknown classification points. First, the point cloud in the rough segmentation classification results $P_{g}$ and $P_{ng}$ that is different from the region segmentation result is set as the unknown classification point. Due to projection errors and noise, region segmentation and rough segmentation may include misclassified points, which are often surrounded by correctly classified points. Additionally, for urban roads, the results of ground segmentation are typically continuous, extending from nearby areas to distant ones. Therefore, a ground region growing algorithm is applied to the LiDAR projection map under the assumption that the ground region is connected. The ground-classified points from both the rough segmentation and regional segmentation are used as seed points, and the algorithm iteratively traverses the unknown classified points in the 8-neighborhood. By determining whether the change in LiDAR intensity between consecutive points exceeds a predefined threshold, the algorithm distinguishes whether the current point is a ground point, thereby generating the classified result map $\widehat {\mathbf {I}_{\text {label}}}$. Finally, the classification ground point cloud $\widehat {P_{g}}$ and nonground point cloud $\widehat {P_{ng}}$ are obtained by inverse projection. The algorithm flowchart is shown in Algorithm 1.

![Figure 7](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng16-3579686-large.gif)

*Algorithm 1 GroundRegionGrowing*

## SECTION IV. Experiment

In this article, the accuracy and real-time performance of the proposed algorithm will be evaluated qualitatively and quantitatively in two parts: a public dataset and a self-collected dataset. The SemanticKITTI [^8], an extension of the KITTI dataset with additional manually annotated semantic labels for LiDAR point clouds, serves as the primary dataset for quantitative analysis. BotanicGarden [^9] collected unstructured scenes for a wheeled robot and did not provide semantic annotations for LiDAR point clouds, we utilize this dataset for qualitative analysis in unstructured environments. The self-collected dataset captures urban street environments in Hefei City, Anhui Province, China, encompassing scenarios such as flat roads, uphill and downhill roads, tunnels, and overpasses. To demonstrate the superiority of the proposed algorithm in terms of segmentation accuracy and computational efficiency, this section conducts a comparative analysis against several existing open-source methods. All algorithms are executed on a personal computer with 16-GB RAM, AMD Ryzen7 5800H CPU, and NVIDIA GEFORCE RTX 3050TI GPU.

This section employs precision and recall as evaluation metrics to assess the accuracy of ground segmentation as well as to identify potential undersegmentation and overfitting issues. The $F_{1}$-score integrates both precision and recall and is utilized to comprehensively evaluate the overall performance of the algorithm. The definitions of these metrics are as follows:

$$
\begin{align*} \text {Precision} & = \frac {\text {TP}}{\text {TP} + \text {FP}} \\ \text {Recall} & = \frac {\text {TP}}{\text {TP} + \text {FN}} \\ F_{1}\text {-}\text {score} & = 2 \times \frac {\text {Precision} \cdot \text {Recall}}{\text {Precision} + \text {Recall}} \tag {13}\end{align*}
$$

where TP, FP, and FN correspond to the true positive, false positive, and false negative ground segmentation results of the “ground” category, respectively. Since the primary objective of ground segmentation is to facilitate ground constraints, object detection, or drivable area extraction, in urban road environments, categories such as “pedestrians,” “cars,” “bicycles,” “trucks,” and “vegetation” are also considered nonground points. Conversely, areas like “terrain” and “sidewalks,” which are accessible to vehicles or robots, are classified as ground points.

### A. SemanticKITTI Experiment

The SemanticKITTI dataset extends the KITTI odometry benchmark by incorporating semantic annotations. This dataset was acquired using a vehicle equipped with a Velodyne HDL-64E S2 LiDAR sensor with a 360° field of view, capturing 10-Hz semantic point clouds. It provides comprehensive annotations for both dynamic and static objects in road environments, including road surfaces, buildings, vehicles, vegetation, pedestrians, and traffic signs. The dataset comprises ten training sequences covering urban, highway, and rural scenarios, making it particularly suitable for evaluating the performance of ground segmentation algorithms across diverse environments. This study conducted a comprehensive comparative analysis of state-of-the-art algorithms and various open-source methods, including Patchwork and Patchwork++. Quantitative evaluation results of different segmentation algorithms are presented in Table I. The visualization of segmentation results is illustrated in the accompanying Figs. 7 and 8, where green and red point clouds represent TP and TN, respectively. Blue (FP) and yellow (FN) point clouds indicate false positives and false negatives. Specifically, blue points denote erroneously segmented ground points, reflecting overfitting issues in the algorithm’s segmentation process, while yellow points represent undetected ground points, indicating underfitting problems that affect segmentation accuracy.

![Figure 8](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng.t1-3579686-large.gif)

*TABLE I*

![Figure 9](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng7-3579686-large.gif)

*Fig. 7. Qualitative comparison of ground segmentation on the SemanticKITTI Dataset. The point cloud shown in the figure is extracted from the SemanticKITTI 00 sequence. The leftmost column displays the segmentation results from DipG-Seg, the middle column shows the results from Patchwork++, and the rightmost column presents the results from our proposed method. In the figure, green points indicate TP points, red points denote TN points, blue points represent FP points, and yellow points indicate FN points.*

![Figure 10](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng8abcd-3579686-large.gif)

*Fig. 8. Qualitative results of the ablation experiments on different stages of the segmentation algorithm for the SemanticKITTI 00 and 07 sequences. The left column shows the segmentation results for modules A+B, while the right column displays the results for modules A+B+C+D. (a) 00 A + B. (b) 00 A + B + C + D. (c) 07 A + B. (d) 07 A + B + C + D.*

All metrics in this article obtained from SemanticKITTI were generated on our personal computer. Both the codes and parameters were downloaded online from their open-source platforms. Based on our experimental results, neither Patchwork nor Patchwork++ achieved the results reported in this article. Additionally, after studying the code of these open-source methods, we found that they handle the vegetation category in SemanticKITTI differently. Since vegetation cannot be clearly distinguished as either ground points or nonground points, this leads to varying results across different sequences. In this article, we uniformly treat vegetation as nonground points and obtain the metric evaluation results reported in this article under this consistent standard.

DipG-Seg is a point cloud segmentation method based on LiDAR projection mapping. It projects point clouds onto images, where pixel values are determined by the elevation or distance of corresponding mapped points. While this approach utilizes image segmentation techniques for ground point cloud extraction, it suffers from oversegmentation issues at object boundaries due to projection errors between point clouds and pixels, as well as morphological opening and closing operations. This method inevitably misclassifies boundary points as nonground points, even when these points are actually on the ground surface, as illustrated by the yellow points in Fig. 7.

Patchwork++, a grid-based ground plane fitting method, demonstrates superior performance on flat terrains by dividing point clouds into multiple sectors. However, its effectiveness diminishes in rugged terrains. As the ground segmentation algorithm with the highest recall rate in recent years, it attempts to extract ground points from every partition to maximize recall. This strategy inevitably leads to the misclassification of points at object lower boundaries and wall bases as nonground points, consequently sacrificing precision, as shown in Fig. 7.

In addition, we compared it with other partition-based ground segmentation methods. GroundGrid [^29] divides the LiDAR points into many grids to maintain a grid map and uses the maximum and minimum height of each grid for segmentation. In KITTI 00, it got 92.42% precision, 98.70% recall, and 95.46% $F1$-score, which shows good performance. GroundGrid exhibits high segmentation accuracy, especially for the distant road and terrain. GroundGrid suffers from oversegmentation on the car roof and the distant corners of the wall, which produces higher recall with lower precision. For autonomous vehicles in urban environments, oversegmentation is more unacceptable than undersegmentation because incorrect ground height estimation may lead to bad ground constraints or ground reconstruction. GroundGrid needs global poses of each frame as input and costs over 25 ms on average to maintain the memory-consuming grid map.

As demonstrated in Fig. 7, our method achieves accurate ground segmentation on slopes and rugged terrains. According to the results in Table I, compared with other methods, our approach demonstrates strong competitiveness in most precision and recall metrics, achieving the highest average $F1$-score among all evaluated algorithms. Although the computational complexity is slightly higher than DipG-Seg, it is lower than Patchwork++.

Our method achieves lower recall compared to Patchwork++ and GroundGrid because we employ ground connectivity constraints, treating small ground points as nonground points, which may lead to undersegmentation. However, we achieve higher precision and the $F1$-score, which represents the overall performance of an algorithm. Such a design is intended to obtain 100% ground points, aiming to improve downstream modules, like ground reconstruction and ground constraint. For ground segmentation tasks, oversegmentation is far more harmful than undersegmentation. Unextracted ground points do not have a significant impact on downstream applications. In summary, although our method does not outperform Patchwork++ and GroundGrid in all metrics, it achieves the best overall performance and has the advantage of being easy to deploy.

### B. Parameter Studies

To balance the efficiency and accuracy of segmentation, we supplemented the parameter studies, as shown in Table II. Fig. 9 shows that different thresholds have a minor impact on the performance of the algorithm and do not affect the application of the algorithm. For different LiDAR, we recommend adjusting thresholds based on the various number of scan lines and vertical resolution. Although deep learning indeed has advantages in learning these parameters, the primary purpose of this work is to provide a lightweight, computing-efficient, and reliable front-end ground segmentation module for LiDAR SLAM.

![Figure 11](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng.t2-3579686-large.gif)

*TABLE II*

![Figure 12](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng9-3579686-large.gif)

*Fig. 9. Precision, recall, ${F}1$ -score, and accuracy vary with parameters.*

### C. Ablation Studies

To verify the contribution of each submodule, the ablation study designed a progressive experiment to evaluate the impact of each module on segmentation performance. The experimental results are shown in Table III.

![Figure 13](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng.t3-3579686-large.gif)

*TABLE III*

In A, the preprocessing module estimated the LiDAR tilt angle and removed reflection noise, which significantly improved segmentation accuracy. After the preprocessing module, the precision increased by 3.36% and 2.81% in the SemanticKITTI 00 and 07 sequences, respectively. The recall increased by 0.04% and 0.24%, while the overall performance, as measured by the $F1$-score, rose by 1.86% and 1.82%. Since the LiDAR point cloud becomes more aligned with the ground, this leads to more accurate height thresholds. Based on the ablation experiments, the tilt angle is more beneficial for autonomous vehicles driving on flat terrain, primarily improving the accuracy and speed of coarse segmentation, but it does not significantly impact the final results. However, for wheeled robots operating on uneven terrain, the sensor’s horizontal plane is rarely parallel to the ground, so the tilt angle does not need to be calibrated frequently.

In B, the coarse segmentation module provides an initial segmentation of the ground points and successfully extracts the majority of ground points. However, due to the limitations of the fixed height threshold, oversegmentation errors occur in boundary areas. In the SemanticKITTI 00 and 07 sequences, the coarse segmentation module achieves the highest recall of 96.47% and 95.38%, respectively, identifying correctly the majority of ground points. However, its precision is generally insufficient, at only 84.04% and 88.08%. As a result, the overall $F1$-score for the coarse segmentation module is 89.83% and 90.04%. This finding aligns with the qualitative experiment results. As shown in Fig. 8(a) and (c), our method with only modules A + B can segment the majority of ground points. However, the overfitting still occurs with a high FP ratio in boundary areas, as shown by the blue points. Overfitting is mainly observed in regions such as the bottom of walls, the connection between vehicle tires and the ground, as well as grassy areas. Also, only very few underfitting FN points were observed.

In C and D, by projection image threshold and ground region growing, the algorithm achieves the most balanced and accurate results. The visualization of projection processing is shown in Fig. 10.

![Figure 14](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng10-3579686-large.gif)

*Fig. 10. Visualization of LiDAR projection image processing (raw $\rightarrow $ morphology $\rightarrow $ segmentation).*

The last and most important claim is that our method runs fast as shown in Fig. 11. It only takes around 9.62 out of 14.19 ms to fine-segment the repaired image and 4.41 ms to coarse segment. It proves that our algorithm has good real-time performance and can be used as a front-end module such as SLAM and ground reconstruction.

![Figure 15](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng11-3579686-large.gif)

*Fig. 11. Runtime of our method on SemanticKITTI 07.*

In the SemanticKITTI 00 and 07 sequences, the algorithm achieved the highest segmentation accuracy, improving to 97.27% and 97.84%, with a 9.87% and 9.76% increase, respectively. Although the recall decreased slightly by 1.66% and 0.7%, the overall algorithm reached the most balanced and accurate result, with the $F1$-score improving by 4.31% and 4.25%. While the overfitting FP points were restored, a small number of underfitting FN points were introduced due to fuzzy object boundaries and projection errors. As shown in Fig. 8(b) and (d), the complete algorithm still misclassifies points when identifying objects like grass but achieves higher classification accuracy at road boundaries. These results fully validate the effectiveness of the proposed algorithm and its advantage in handling ground-to-nonground transition areas.

### D. BotanicGarden Experiment

BotanicGarden [^3] builds complex field environment datasets in a botanic garden of more than 48000 m2 to validate algorithms in unstructured environments. It works in a four-wheel-drive differential mechanism, carrying the gray and color stereo cameras, a mechanic-spinning LiDAR, a solid-state MEMS LiDAR, a GNSS/IMU system, and a wheel odometry.

As we can see from Fig. 12. GndNet lacks generalization ability in unstructured scenes. According to the qualitative analysis, GndNet has the lowest performance among the three algorithms. It tends to oversegment and cannot adequately estimate ground height. The performance of GroundGrid is comparable to our proposed method. Qualitative analysis shows that our method achieves higher segmentation accuracy. GroundGrid tends to overfit at close distances (within 1 m) and long distances, as shown in Fig. 12(b). This is due to the lack of connectivity constraints for the ground. In contrast, the algorithm proposed in this article yields good qualitative experimental results in unstructured scenarios. Our approach does not require a neural network, and parameter tuning is fast and efficient without additional prior information such as robot pose as input. However, due to the lack of fine-tuning in image threshold selection, there is an undersegmentation phenomenon. In summary, the qualitative experiments on this dataset demonstrate that our method is not only suitable for ground segmentation in urban road scenes but also has the capability for segmenting uneven surfaces in unstructured environments. In addition, the drawback of the GroundGrid is that it requires pose as input instead of just point clouds. Since the system is highly coupled with gridmap, it cannot be flexibly and quickly deployed for datasets except KITTI. It requires additional coding to adapt the parameters and data interface. Furthermore, GroundGrid needs to specify the odometry coordinate system and the relationship between different coordinates to perform the transformation.

![Figure 16](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng12abc-3579686-large.gif)

*Fig. 12. Qualitative results of (a) GndNet, (b) GroundGrid, and (c) ours in BotanicGarden.*

### E. Real-World Experiment

The real experimental results, presented in Fig. 13, serve as a qualitative analysis of the ground segmentation algorithm, which was constructed on a modified electric passenger vehicle, as illustrated in Fig. 14. The top of the vehicle is equipped with three LiDARs, mainly using the LiDAR RS-Ruby-Lite-80 LiDAR in the middle with a 360° field of view, serving as the primary data source; The CGI-430 GNSS/IMU dual-antenna navigation system at the rear provides centimeter-level RTK positioning and heading. Wheel speed sensors are mounted on the chassis for real-time four-wheel rotational velocity measurement. Only raw LiDAR point cloud data in the middle were utilized in this experimental validation.

![Figure 17](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng13abc-3579686-large.gif)

*Fig. 13. Qualitative comparison of ground segmentation algorithms in urban scenarios. The blue rectangle indicates oversegmentation, and the yellow rectangle indicates undersegmentation. (a) DipG-Seg. (b) Patchwork++. (c) Ours.*

![Figure 18](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng14-3579686-large.gif)

*Fig. 14. Experimental platform: the autonomous electric vehicle equipped with RS-Ruby-Lite LiDAR, CGI-430 GNSS/IMU integrated navigation system, and wheel speed sensors.*

To evaluate algorithm adaptability in complex urban environments, multiscenario data were collected on elevated expressways in Heighfei City, Anhui Province, China (Fig. 15). In Fig. 13, the ground truth of segmentation labels of real-world datasets is hard to obtain. Therefore, the test environment included ramp transition zones with significant grade changes that were only used to qualitatively evaluate the algorithm.

![Figure 19](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/7361/11126918/11045759/meng15-3579686-large.gif)

*Fig. 15. Real-world dataset in highway environment.*

The experimental results indicated that the DipG-Seg, shown in Fig. 13(a), exhibits pronounced oversegmentation when processing 80-line LiDAR data, particularly generating anomalous noise points near guardrails and vegetation. This suggests that the method struggles to adapt to high-line LiDAR. Despite this, it still achieves the highest ground recall rate, indicating that image-based segmentation methods need to be optimized with height threshold constraints.

The results show that Patchwork++, shown in Fig. 13(b), achieved higher accuracy compared to digpseg. However, it suffered from characteristic undersegmentation in ramp transition zones, shown in red rectangle areas. Besides, the experimental analysis revealed its sensitivity to parameters like grid resolution, which needs platform-specific parameter tuning.

In comparison, our method successfully maintains high segmentation accuracy even on roads with significant elevation changes, as shown in Fig. 13(c). Notably, in the rectangles marked steep slope section, our method achieved highly accurate continuous ground segmentation. However, there is still oversegmentation noise at vegetation boundaries. These findings demonstrate the effectiveness of our ground segmentation approach in processing real-world point clouds, and the high accuracy further implies enhanced driving safety for mobile platforms employing the algorithm.

## SECTION V. Conclusion

In this article, we propose a rapid and accurate ground segmentation method for urban environments, termed LPIR-Seg. This approach leverages LiDAR projection images and a coarse-to-fine segmentation strategy. It fully exploits the horizontal mounting configuration of the LiDAR sensor and the distinctive features of urban road surfaces by applying a height thresholding and block-fitting strategy for coarse segmentation. Subsequently, we integrate an improved image segmentation technique based on LiDAR depth maps, which fuses the coarse segmentation results to classify pixels into three categories: ground, nonground, and unknown. Finally, the unknown pixels are refined using a fine segmentation module, significantly enhancing the overall accuracy. We conducted extensive experiments on both the public dataset and real-world datasets, which encompass diverse scenarios and multiple LiDAR sensor models, to validate our approach. The results on SemanticKITTI demonstrate that our method outperforms other feature-based algorithms in terms of ground segmentation accuracy. The experiments on BotanicGarden indicate that our method also works well on unstructured scenes. Additional experiments further confirm the robustness of our approach under various conditions.

However, our comparative analysis also indicates that improvements are still needed in recall rate and execution efficiency. In particular, the extraction accuracy suffers for ambiguous boundary pixels, especially in regions with uneven grass or shrubbery. Furthermore, the algorithm’s ability to capture spatial semantic information is limited. In future work, we plan to integrate our approach with deep learning-based point cloud semantic segmentation methods to enhance semantic accuracy and to pursue parallel optimization techniques to further improve efficiency.

## References

[^1]: H. Lim, M. Oh, and H. Myung, “Patchwork: Concentric zone-based region-wise ground segmentation with ground likelihood estimation using a 3D LiDAR sensor,” IEEE Robot. Autom. Lett., vol. 6, no. 4, pp. 6458–6465, Oct. 2021. [IEEE](https://ieeexplore.ieee.org/document/9466396) [Google Scholar](https://scholar.google.com/scholar?as_q=Patchwork%3A+Concentric+zone-based+region-wise+ground+segmentation+with+ground+likelihood+estimation+using+a+3D+LiDAR+sensor&as_occt=title&hl=en&as_sdt=0%2C31)

[^2]: S. Lee, H. Lim, and H. Myung, “Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3D point cloud,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Oct. 2022, pp. 13276–13283. [IEEE](https://ieeexplore.ieee.org/document/9981561) [Google Scholar](https://scholar.google.com/scholar?as_q=Patchwork%2B%2B%3A+Fast+and+robust+ground+segmentation+solving+partial+under-segmentation+using+3D+point+cloud&as_occt=title&hl=en&as_sdt=0%2C31)

[^3]: Y. Pan, P. Xiao, Y. He, Z. Shao, and Z. Li, “MULLS: Versatile LiDAR SLAM via multi-metric linear least square,” in Proc. IEEE Int. Conf. Robot. Autom. (ICRA), May 2021, pp. 11633–11640. [IEEE](https://ieeexplore.ieee.org/document/9561364) [Google Scholar](https://scholar.google.com/scholar?as_q=MULLS%3A+Versatile+LiDAR+SLAM+via+multi-metric+linear+least+square&as_occt=title&hl=en&as_sdt=0%2C31)

[^4]: Y. Qian, X. Wang, Z. Chen, C. Wang, and M. Yang, “Hy-Seg: A hybrid method for ground segmentation using point clouds,” IEEE Trans. Intell. Vehicles, vol. 8, no. 2, pp. 1597–1606, Feb. 2023. [IEEE](https://ieeexplore.ieee.org/document/9810973) [Google Scholar](https://scholar.google.com/scholar?as_q=Hy-Seg%3A+A+hybrid+method+for+ground+segmentation+using+point+clouds&as_occt=title&hl=en&as_sdt=0%2C31)

[^5]: S. B. Damelin and N. S. Hoang, “On surface completion and image inpainting by biharmonic functions: Numerical aspects,” Int. J. Math. Math. Sci., vol. 2018, pp. 1–8, 2018. [DOI](https://doi.org/10.1155/2018/3950312) [Google Scholar](https://scholar.google.com/scholar?as_q=On+surface+completion+and+image+inpainting+by+biharmonic+functions%3A+Numerical+aspects&as_occt=title&hl=en&as_sdt=0%2C31)

[^6]: D. Kong, X. Li, Q. Xu, Y. Hu, and P. Ni, “SC\_LPR: Semantically consistent LiDAR place recognition based on chained cascade network in long-term dynamic environments,” IEEE Trans. Image Process., vol. 33, pp. 2145–2157, 2024. [IEEE](https://ieeexplore.ieee.org/document/10444711) [Google Scholar](https://scholar.google.com/scholar?as_q=SC%5C_LPR%3A+Semantically+consistent+LiDAR+place+recognition+based+on+chained+cascade+network+in+long-term+dynamic+environments&as_occt=title&hl=en&as_sdt=0%2C31)

[^7]: P. Ni, X. Li, D. Kong, and X. Yin, “Scene-adaptive 3D semantic segmentation based on multi-level boundary-semantic-enhancement for intelligent vehicles,” IEEE Trans. Intell. Vehicles, vol. 9, no. 1, pp. 1722–1732, Jan. 2024. [IEEE](https://ieeexplore.ieee.org/document/10122665) [Google Scholar](https://scholar.google.com/scholar?as_q=Scene-adaptive+3D+semantic+segmentation+based+on+multi-level+boundary-semantic-enhancement+for+intelligent+vehicles&as_occt=title&hl=en&as_sdt=0%2C31)

[^8]: J. Behley, “SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), Oct. 2019, pp. 9296–9306. [IEEE](https://ieeexplore.ieee.org/document/9010727) [Google Scholar](https://scholar.google.com/scholar?as_q=SemanticKITTI%3A+A+dataset+for+semantic+scene+understanding+of+LiDAR+sequences&as_occt=title&hl=en&as_sdt=0%2C31)

[^9]: L. Yuan-zhi, “BotanicGarden: A high-quality dataset for robot navigation in unstructured natural environments,” IEEE Robot. Autom. Lett., vol. 9, no. 3, pp. 2798–2805, Jan. 2024. [IEEE](https://ieeexplore.ieee.org/document/10415477) [Google Scholar](https://scholar.google.com/scholar?as_q=BotanicGarden%3A+A+high-quality+dataset+for+robot+navigation+in+unstructured+natural+environments&as_occt=title&hl=en&as_sdt=0%2C31)

[^10]: L. Rummelhard, A. Paigwar, A. Nègre, and C. Laugier, “Ground estimation and point cloud segmentation using SpatioTemporal conditional random field,” in Proc. IEEE Intell. Vehicles Symp. (IV), Jun. 2017, pp. 1105–1110. [IEEE](https://ieeexplore.ieee.org/document/7995861) [Google Scholar](https://scholar.google.com/scholar?as_q=Ground+estimation+and+point+cloud+segmentation+using+SpatioTemporal+conditional+random+field&as_occt=title&hl=en&as_sdt=0%2C31)

[^11]: A. Golovinskiy and T. Funkhouser, “Min-cut based segmentation of point clouds,” in Proc. IEEE 12th Int. Conf. Comput. Vis. Workshops ICCV Workshops, Sep. 2009, pp. 39–46. [IEEE](https://ieeexplore.ieee.org/document/5457721) [Google Scholar](https://scholar.google.com/scholar?as_q=Min-cut+based+segmentation+of+point+clouds&as_occt=title&hl=en&as_sdt=0%2C31)

[^12]: W. Huang, “A fast point cloud ground segmentation approach based on coarse-to-fine Markov random field,” IEEE Trans. Intell. Transp. Syst., vol. 23, no. 7, pp. 7841–7854, Jul. 2022. [IEEE](https://ieeexplore.ieee.org/document/9410344) [Google Scholar](https://scholar.google.com/scholar?as_q=A+fast+point+cloud+ground+segmentation+approach+based+on+coarse-to-fine+Markov+random+field&as_occt=title&hl=en&as_sdt=0%2C31)

[^13]: C. Guo, W. Sato, L. Han, S. Mita, and D. McAllester, “Graph-based 2D road representation of 3D point clouds for intelligent vehicles,” in Proc. IEEE Intell. Vehicles Symp. (IV), Jun. 2011, pp. 715–721. [IEEE](https://ieeexplore.ieee.org/document/5940502) [Google Scholar](https://scholar.google.com/scholar?as_q=Graph-based+2D+road+representation+of+3D+point+clouds+for+intelligent+vehicles&as_occt=title&hl=en&as_sdt=0%2C31)

[^14]: A. Paigwar, Ö. Erkent, D. Sierra-Gonzalez, and C. Laugier, “GndNet: Fast ground plane estimation and point cloud segmentation for autonomous vehicles,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Oct. 2020, pp. 2150–2156. [IEEE](https://ieeexplore.ieee.org/document/9340979) [Google Scholar](https://scholar.google.com/scholar?as_q=GndNet%3A+Fast+ground+plane+estimation+and+point+cloud+segmentation+for+autonomous+vehicles&as_occt=title&hl=en&as_sdt=0%2C31)

[^15]: W. Dou and Y. Wang, “Point cloud graph for LiDAR segmentation,” Measurement, vol. 242, Jan. 2025, Art. no. 115851. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0263224124017366 [Google Scholar](https://scholar.google.com/scholar?as_q=Point+cloud+graph+for+LiDAR+segmentation&as_occt=title&hl=en&as_sdt=0%2C31)

[^16]: A. Milioto, I. Vizzo, J. Behley, and C. Stachniss, “RangeNet ++: Fast and accurate LiDAR semantic segmentation,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Nov. 2019, pp. 4213–4220, doi: 10.1109/iros40897.2019.8967762. [IEEE](https://ieeexplore.ieee.org/document/8967762) [Google Scholar](https://scholar.google.com/scholar?as_q=RangeNet+%2B%2B%3A+Fast+and+accurate+LiDAR+semantic+segmentation&as_occt=title&hl=en&as_sdt=0%2C31)

[^17]: M. A. Fischler and R. C. Bolles, “Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography,” Commun. ACM, vol. 24, no. 6, pp. 381–395, Jun. 1981, doi: 10.1145/358669.358692. [DOI](https://doi.org/10.1145/358669.358692) [Google Scholar](https://scholar.google.com/scholar?as_q=Random+sample+consensus%3A+A+paradigm+for+model+fitting+with+applications+to+image+analysis+and+automated+cartography&as_occt=title&hl=en&as_sdt=0%2C31)

[^18]: X. Wei, J. Lv, J. Sun, and S. Pu, “Ground-SLAM: Ground constrained LiDAR SLAM for structured multi-floor environments,” 2021, arXiv:2103.03713. [Google Scholar](https://scholar.google.com/scholar?as_q=Ground-SLAM%3A+Ground+constrained+LiDAR+SLAM+for+structured+multi-floor+environments&as_occt=title&hl=en&as_sdt=0%2C31)

[^19]: A. Nurunnabi, D. Belton, and G. West, “Diagnostics based principal component analysis for robust plane fitting in laser data,” in Proc. 16th Int. Conf. Comput. Inf. Technol., Mar. 2014, pp. 484–489. [IEEE](https://ieeexplore.ieee.org/document/6997319) [Google Scholar](https://scholar.google.com/scholar?as_q=Diagnostics+based+principal+component+analysis+for+robust+plane+fitting+in+laser+data&as_occt=title&hl=en&as_sdt=0%2C31)

[^20]: P. Narksri, E. Takeuchi, Y. Ninomiya, Y. Morales, N. Akai, and N. Kawaguchi, “A slope-robust cascaded ground segmentation in 3D point cloud for autonomous vehicles,” in Proc. 21st Int. Conf. Intell. Transp. Syst. (ITSC), Nov. 2018, pp. 497–504. [IEEE](https://ieeexplore.ieee.org/document/8569534) [Google Scholar](https://scholar.google.com/scholar?as_q=A+slope-robust+cascaded+ground+segmentation+in+3D+point+cloud+for+autonomous+vehicles&as_occt=title&hl=en&as_sdt=0%2C31)

[^21]: A. Mukherjee, S. D. Das, J. Ghosh, A. S. Chowdhury, and S. K. Saha, “Fast geometric surface based segmentation of point cloud from LiDAR data,” in Proc. Pattern Recognit. Mach. Intell., Jan. 2019, pp. 415–423. [DOI](https://doi.org/10.1007/978-3-030-34869-4_45) [Google Scholar](https://scholar.google.com/scholar?as_q=Fast+geometric+surface+based+segmentation+of+point+cloud+from+LiDAR+data&as_occt=title&hl=en&as_sdt=0%2C31)

[^22]: H. Lim, S. Hwang, and H. Myung, “ERASOR: Egocentric ratio of pseudo occupancy-based dynamic object removal for static 3D point cloud map building,” IEEE Robot. Autom. Lett., vol. 6, no. 2, pp. 2272–2279, Apr. 2021. [IEEE](https://ieeexplore.ieee.org/document/9361109) [Google Scholar](https://scholar.google.com/scholar?as_q=ERASOR%3A+Egocentric+ratio+of+pseudo+occupancy-based+dynamic+object+removal+for+static+3D+point+cloud+map+building&as_occt=title&hl=en&as_sdt=0%2C31)

[^23]: M. Himmelsbach, F. v. Hundelshausen, and H.-J. Wuensche, “Fast segmentation of 3D point clouds for ground vehicles,” in Proc. IEEE Intell. Vehicles Symp., Jun. 2010, pp. 560–565. [IEEE](https://ieeexplore.ieee.org/document/5548059) [Google Scholar](https://scholar.google.com/scholar?as_q=Fast+segmentation+of+3D+point+clouds+for+ground+vehicles&as_occt=title&hl=en&as_sdt=0%2C31)

[^24]: D. Zermas, I. Izzat, and N. Papanikolopoulos, “Fast segmentation of 3D point clouds: A paradigm on LiDAR data for autonomous vehicle applications,” in Proc. IEEE Int. Conf. Robot. Autom. (ICRA), May 2017, pp. 5067–5073. [IEEE](https://ieeexplore.ieee.org/document/7989591) [Google Scholar](https://scholar.google.com/scholar?as_q=Fast+segmentation+of+3D+point+clouds%3A+A+paradigm+on+LiDAR+data+for+autonomous+vehicle+applications&as_occt=title&hl=en&as_sdt=0%2C31)

[^25]: J. Carvalho, L. Cunha, S. Pinto, and T. Gomes, “FESTA: FPGA-enabled ground segmentation technique for automotive LiDAR,” IEEE Sensors J., vol. 24, no. 22, pp. 38005–38014, Nov. 2024. [IEEE](https://ieeexplore.ieee.org/document/10705951) [Google Scholar](https://scholar.google.com/scholar?as_q=FESTA%3A+FPGA-enabled+ground+segmentation+technique+for+automotive+LiDAR&as_occt=title&hl=en&as_sdt=0%2C31)

[^26]: M. Oh, “TRAVEL: Traversable ground and above-ground object segmentation using graph representation of 3D LiDAR scans,” IEEE Robot. Autom. Lett., vol. 7, no. 3, pp. 7255–7262, Jul. 2022. [IEEE](https://ieeexplore.ieee.org/document/9794594) [Google Scholar](https://scholar.google.com/scholar?as_q=TRAVEL%3A+Traversable+ground+and+above-ground+object+segmentation+using+graph+representation+of+3D+LiDAR+scans&as_occt=title&hl=en&as_sdt=0%2C31)

[^27]: T. Shan and B. Englot, “LeGO-LOAM: Lightweight and ground-optimized LiDAR odometry and mapping on variable terrain,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Oct. 2018, pp. 4758–4765. [IEEE](https://ieeexplore.ieee.org/document/8594299) [Google Scholar](https://scholar.google.com/scholar?as_q=LeGO-LOAM%3A+Lightweight+and+ground-optimized+LiDAR+odometry+and+mapping+on+variable+terrain&as_occt=title&hl=en&as_sdt=0%2C31)

[^28]: Z. Wang and G. Liu, “Improved LeGO-LOAM method based on outlier points elimination,” Measurement, vol. 214, Mar. 2023, Art. no. 112767. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0263224123003317 [Google Scholar](https://scholar.google.com/scholar?as_q=Improved+LeGO-LOAM+method+based+on+outlier+points+elimination&as_occt=title&hl=en&as_sdt=0%2C31)

[^29]: N. Steinke, D. Goehring, and R. Rojas, “GroundGrid: LiDAR point cloud ground segmentation and terrain estimation,” IEEE Robot. Autom. Lett., vol. 9, no. 1, pp. 420–426, Nov. 2023. [IEEE](https://ieeexplore.ieee.org/document/10319084) [Google Scholar](https://scholar.google.com/scholar?as_q=GroundGrid%3A+LiDAR+point+cloud+ground+segmentation+and+terrain+estimation&as_occt=title&hl=en&as_sdt=0%2C31)

[^30]: D. Wei, H. Ran, Z. Li, and T. Gao, “Enhanced low drift LiDAR SLAM with dynamic noise using ground principal plane constraint,” Measurement, vol. 245, Mar. 2025, Art. no. 116549. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0263224124024345 [Google Scholar](https://scholar.google.com/scholar?as_q=Enhanced+low+drift+LiDAR+SLAM+with+dynamic+noise+using+ground+principal+plane+constraint&as_occt=title&hl=en&as_sdt=0%2C31)

### Additional References