[Background] Traditional methods, due to their static receptive field design, struggle to adapt to the significant scale differences among cars, pedestrians, and cyclists in urban autonomous driving scenarios. Moreover, cross-scale feature fusion often leads to hierarchical interference. [Methodology] To address the key challenge of cross-scale representation consistency in 3D object detection for multi-class, multi-scale objects in autonomous driving scenarios, this study proposes a novel method named VoxTNT. VoxTNT leverages an equalized receptive field and a local-global collaborative attention mechanism to enhance detection performance. At the local level, a PointSetFormer module is introduced, incorporating an Induced Set Attention Block (ISAB) to aggregate fine-grained geometric features from high-density point clouds through reduced cross-attention. This design overcomes the information loss typically associated with traditional voxel mean pooling. At the global level, a VoxelFormerFFN module is designed, which abstracts non-empty voxels into a super-point set and applies cross-voxel ISAB interactions to capture long-range contextual dependencies. This approach reduces the computational complexity of global feature learning from O(N²) to O(M²) (where M << N, M is the number of non-empty voxels), avoiding the high computational complexity associated with directly applying complex Transformers to raw point clouds. This dual-domain coupled architecture achieves a dynamic balance between local fine-grained perception and global semantic association, effectively mitigating modeling bias caused by fixed receptive fields and multi-scale fusion. [Results] Experiments demonstrate that the proposed method achieves a single-stage detection Average Precision (AP) of 59.56% for moderate-level pedestrian detection on the KITTI dataset, an improvement of approximately 12.4% over the SECOND baseline. For two-stage detection, it achieves a mean Average Precision (mAP) of 66.54%, outperforming the second-best method, BSAODet, which achieves 66.10%. Validation on the WOD dataset further confirms the method’s effectiveness, achieving 66.09% mAP, which outperforms the SECOND and PointPillars baselines by 7.7% and 8.5%, respectively. Ablation studies demonstrate that the proposed equalized local-global receptive field mechanism significantly improves detection accuracy for small objects. For example, on the KITTI dataset, full component ablation resulted in a 10.8% and 10.0% drop in AP for moderate-level pedestrian and cyclist detection, respectively, while maintaining stable performance for large-object detection. [Conclusions] This study presents a novel approach to tackling the challenges of multi-scale object detection in autonomous driving scenarios. Future work will focus on optimizing the model architecture to further enhance efficiency.

[Objectives] The 3D model reconstruction of buildings based on a model-driven approach using airborne LiDAR building point clouds relies on fitting the building point cloud to predefined geometric primitives. However, due to the uneven density and noise in the building point cloud, errors often arise in structural details during the primitive fitting process, leading to reduced reconstruction accuracy. To address this issue, this study proposes a 3D model reconstruction method for airborne LiDAR building point clouds based on sequential quadratic programming and elevation step correction. [Methods] First, a primitive library containing classical roof structures is established, including simple roofs, complex roofs, and steep roofs. An adjacency matrix is constructed by incorporating the adjacency relationships and ridge properties between roof patches. The best-matching primitives are then selected from the primitive library based on the adjacency matrix. Next, the shape parameters of the selected primitives are optimized using the sequential quadratic programming algorithm to achieve a globally optimal fitting state. The initial 3D model is then generated. To further enhance accuracy, the relative position of the building models and the roof point clouds in 3D space is refined through translation and rotation, reducing the relative distance deviation and improving the fitting precision. Finally, the City Geography Markup Language (CityGML) is used to store the reconstructed 3D building models, ensuring clear structure and correct topology, which facilitates the visual representation of reconstruction results. [Results] Ten sets of classical building point clouds from the 3D Building dataset were selected for the 3D model reconstruction experiment. The proposed method was compared with existing reconstruction approaches based on the same model-driven framework, and classical accuracy evaluation matrics were used for quantitative analysis. The average objective function value for the selected experimental data was 0.32 m, which is 0.03 m higher than the comparison method, indicating improved accuracy. The horizontal average deviation between the reconstructed building elements and the building point cloud was 0.10 m, while the vertical average deviation was 0.04 m. [Conclusions] In summary, the optimal shape parameters, obtained through the sequential quadratic programming algorithm, enable the construction of 3D building models with complete topology and regular shapes. Additionally, the elevation step correction, which utilizes the average point spacing of the roof point cloud as the step length, effectively enhances the reconstruction accuracy of 3D building models.

[Objectives] Simultaneous Localization and Mapping (SLAM) using LiDAR technology is a pivotal technology in the fields of positioning and navigation and spatial data acquisition. It offers distinct advantages, such as reliable performance and the ability to generate intuitive maps. However, in environments lacking geometric features or with high feature similarity, traditional LiDAR SLAM techniques often yield inaccurate results, because the current LiDAR degenerate environment detection methods require heuristic thresholding, calculation of indirect evaluation indexes, and have low detection efficiency. In order to predict such anomalies and address this issue, it is crucial to effectively detect LiDAR degenerate environments. [Methods] This study proposed a LiDAR degenerate environment detection method based on improved XGBoost algorithms. It achieved direct degradation detection of individual-frame point clouds by analyzing the geometric structure of the environment. Specifically, a classification feature system was established utilizing LiDAR point cloud data, followed by the construction of XGBoost decision trees. Subsequently, the Fuzzy Comprehensive Evaluation (FCE) algorithm was employed to compute a comprehensive importance metric for each feature, thereby facilitating the formation of an effective subset of features which leads to enhanced detection accuracy. Meanwhile, a bidirectional feature screening strategy based on Spearman's rank Correlation Coefficient (SCC) was implemented to construct the feature subset and increase model training efficiency. Moreover, by integrating Sliding Window (SW) and Boyer-Moore Voting (BMV) strategies, we performed secondary correction on the initial XGBoost detection outcomes, further enhancing the precision of LiDAR degenerate environment detection. The validity of the proposed method and its effectiveness in detecting LiDAR degenerate environments were verified through the collection of real scene data. [Results] Our results demonstrated the efficacy of components of the method in this paper. The success rate of LiDAR degenerate environment detection reached 94.41%, and the false detection rate of well-conditioned environment reached 1.24%. Compared to the LOAM degradation detection module, the detection success rate was improved by 10.91%, the false detection rate was reduced by 95.26%, and the detection efficiency was improved by 56.97%. [Conclusions] These findings highlight the achievement of efficient and precise LiDAR degenerate environment detection. The improved XGBoost algorithm in this paper is also applicable to other fields for improving data processing efficiency and accuracy. The LiDAR degenerate environment detection method proposed in this paper can not only be effectively applied to the LiDAR SLAM process but also play an important role in the fields of Geographic Information System (GIS) and Remote Sensing (RS).

[Objectives] Scene understanding based on 3D laser point clouds plays a core role in many applications such as object detection, 3D reconstruction, cultural relic protection, and autonomous driving. The semantic classification of 3D point clouds is an important task in scene understanding, but due to the large amount of data, diverse targets, and large-scale differences, as well as the occlusion of buildings and trees, this task still poses challenges. The existing deep learning models for point cloud classification face several challenges due to the unstructured and disordered nature of point clouds. These challenges include inadequate extraction of local and global features and the absence of an efficient mechanism for context feature integration, making it challenging to achieve fine-grained classification of ground objects. Therefore, this study introduces a novel point cloud feature classification approach that incorporates a multi-scale convolutional attention network for both local and global features. [Methods] To address the lack of structure in point clouds, we construct a local weighted graph to model the positional relationships between central points and their neighboring points. This graph facilitates dynamic adjustments of kernel weights, enabling the extraction of more representative local features. Simultaneously, we introduce a global graph attention module to account for the overall spatial distribution of points, address the disorder of point clouds, and effectively capture global contextual features, thereby integrating information at different scales. Furthermore, we design an adaptive weighted pooling module to facilitate the seamless fusion of local and global features, thus maximizing the network's classification performance. [Results] The proposed method is evaluated using the publicly available Toronto-3D point cloud dataset and a campus point cloud dataset obtained from real measurements. We compare its performance against various network models, including Pointnet++, DGCNN, RandLA-Net, BAAF-Net, and BAF-LAC, The experimental results show that the OA and MIoU of our method in the Toronto-3D dataset are 97.21% and 85.46%, respectively. Compared with network models such as Pointnet++, DGCNN, RandLA Net, BAAF Net, and BAF-LAC, OA has improved by 1.99% to 8.21%, and MIoU has improved by 3.23% to 35.86%. In the campus dataset, the OA and MIoU of our method in this paper are 97.38% and 85.70%, respectively. OA has improved by 0.58% to 10.53%, and MIoU has improved by 2.01% to 32.01%. [Conclusions] These results surpass those achieved by the comparison networks and effectively overcome problems such as large changes in target scale and building occlusion, establishing our method's capability to achieve high-precision and efficient fine classification of ground objects in complex road scenes.

Addressing the issue of Ultra-Wideband (UWB) signal obstruction by obstacles in indoor environments, which leads to Non Line of Sight (NLOS) errors, this paper presents a fusion positioning method based on Light Detection And Ranging (LiDAR) point cloud for identifying UWB NLOS. This method utilizes LiDAR point cloud information to assist in the identification of UWB NLOS and leverages UWB Line of Sight (LOS) ranging values to eliminate cumulative errors in the LiDAR Simultaneous Localization and Mapping (SLAM) positioning process, thereby enhancing the accuracy and robustness of indoor fusion positioning. Initially, the method processes the LiDAR point cloud using an octree, constructs the ranging direction based on the location information of UWB base stations, and extracts the relevant point cloud data in the ranging direction from the LiDAR point cloud. Subsequently, the 3D Alpha Shape algorithm is employed to extract contours of obstacles that may hinder UWB signal propagation within the extracted point clouds. Furthermore, by analyzing the spatial relationship between the extracted obstacle contours and the UWB ranging direction, the presence of NLOS conditions in UWB signals is effectively determined. Finally, NLOS ranging values identified during the UWB ranging process are excluded, and a tight integration approach is used with an Extended Kalman Filter (EKF) to fuse UWB LOS ranging values with LiDAR SLAM positioning data, eliminating cumulative errors in the LiDAR SLAM positioning outcomes, thereby enhancing the precision and robustness of fusion positioning. Experimental results demonstrate that this method significantly improves positioning accuracy in indoor environments, increasing the positioning accuracy by 96.13% compared to the positioning method that tightly combines the original UWB ranging values with LiDAR SLAM using EKF, with a positioning error of 0.067 m, achieving sub-meter level indoor positioning accuracy.

The accurate extraction of tunnel sections is a pivotal step in tunnel deformation analysis. However, due to inadequate illumination, the reflection and occlusion caused by dust and structural elements generate noise and erroneous points in the point cloud data, complicating data processing. Additionally, the intricate geometry of tunnel features, such as curved surfaces, corners, and cracks, renders traditional section extraction algorithms ineffective for point cloud data. Consequently, there is an urgent need for more efficient and robust algorithms. To address this issue, this paper proposes a method for continuous tunnel section extraction based on laser point cloud data. First, a combination filtering method is introduced, integrating Random Sample Consensus (RANSAC) cylindrical fitting and radius filtering to effectively remove scattered outliers and noise points adhering to the tunnel walls with sparse density. Next, the tunnel central axis is obtained via bidirectional projection, and a mathematical function model is established in line with the principle that ‘a straight line intersecting the tunnel central axis intersects the tunnel wall’, enabling the continuous extraction of tunnel section point clouds. Finally, the fitting center coordinates of the section points and the tunnel design radius are used as parameters to calculate the radial deviation of the tunnel points, representing the shape variables. The tunnel point cloud is visually rendered using these shape variables to display the overall deformation of the tunnel. In this paper, laser point cloud data from three sections of a subway tunnel in Chengdu are used as the experimental data. The results show that the mean values of Class I error, Class II error, and total error are 1.48%, 1.03%, and 1.21%, respectively, with the Kappa coefficient reaching 97.45% when using this method for noise filtering. Compared to traditional methods such as least squares, density clustering, and normal deviation algorithms, this method reduces cumulative errors by 9.34%, 10.61%, and 4.41%, respectively, while increasing the Kappa coefficient by 5.36%, 6.38%, and 3.65%. This demonstrates the enhanced robustness and accuracy of the proposed method. Moreover, the mean deviation between the tunnel section fitting radius obtained through this method and the design radius is merely 1.36 mm, compared to deviations of 1.60 mm and 6.00 mm with existing methods, achieving reductions of 2.5 mm and 2.7 mm, respectively. The range of the tunnel shape variable is between 0 and 18 mm, and the overall deformation of the tunnel is visually displayed through point cloud rendering. The method provides a reliable foundation and essential support for tunnel safety monitoring.

Simultaneous Localization and Mapping (SLAM) based on Light Detection and Ranging (LiDAR) is one of the core technologies widely used in the field of indoor and outdoor integrated positioning and spatial data acquisition. It has the advantage of strong anti-interference without considering light conditions. However, in degenerate environments lacking geometric features, the traditional LiDAR odometry based on feature registration algorithm or Iterative Closest Point (ICP) algorithm usually produce inaccurate results. To address the above problem, this paper proposes a LiDAR odometer positioning method for degenerate environments using modified laser intensity and an accurate LiDAR degenerate environment detection method. Firstly, an intensity correction method is proposed based on laser intensity properties and the elimination idea, and an unsupervised position correction method is used to jointly optimize the original point cloud. Based on this, four point cloud features are extracted, and the fuzzy comprehensive evaluation method is used to accurately detect the degradation of each frame point cloud. For well-conditioned point clouds, line and surface features are extracted to construct the interest point set. For degenerate point clouds, a " MI word search " method is proposed to extract well-conditioned features based on laser intensity, and the interest point set is constructed with effective line and surface features. Finally, the robust LiDAR odometry is constructed by processing the inter-frame interest point sets through suitable registration algorithms. The experimental results show that the correct rate of laser intensity correction is 93.34 %. The correct detection rate of LiDAR degenerate point cloud is 98.58 %, and the error detection rate of well-conditioned point cloud is only 2.24 %. Compared with the LiDAR odometry based on feature registration algorithm and ICP algorithm, the positioning Root Mean Square Error (RMSE) of the LiDAR odometry positioning method proposed in this paper is reduced by 90.00% and 83.96%, respectively, and the Maximum Error (MAXE) is reduced by 86.23 % and 79.07 %, respectively. The time required for single inter-frame registration is 0.006 9 s, which effectively improves the positioning accuracy of traditional LiDAR odometry in degenerate environment while maintaining data processing efficiency. Compared with the other methods in the same field, our method shows more significant advantages. In addition, since the method in this paper is completely innovated and improved on the basis of traditional LiDAR odometer, it has a low overall complexity and significant advantages in terms of secondary development and portability.

Visibility refers to the spatial extent that an observer can see from a specific location under certain environmental conditions. Existing visibility analysis methods, based on point cloud data, mainly focus on the outdoor perspective and do not fully consider the problem of point cloud voids. To fill the gaps in the current indoor visibility research, this paper proposes an indoor 3D visibility analysis method based on point cloud data from UAVs and ground survey data. Our method reduces the influence of point cloud voids and aims to improve the accuracy of the visibility analysis. The method presented in this paper is divided into three parts: data preprocessing, visual space construction, and visual distance correction. In the data preprocessing stage, UAV point cloud data is used to extract the building roof contour and DEM, which are utilized for subsequent visual distance correction. Ground-based point cloud data is then used to extract the center points of the windows on the building façade, which serve as observation viewpoints. In the viewspace construction stage, the viewpoint is taken as the center, and the theoretical visual distance is taken as the radius. Points within the theoretical visual distance are selected to construct a hemispherical viewspace. The points in the visual field space are projected onto the depth image to form the visual field image. The visual field of view is then calculated, and the visual range of the field of view space is analyzed to simulate the external scene observed through the window in the room. During this process, point cloud voids can lead to errors in the visual range. Specifically, when the positions of the scanning center and the viewpoint differ significantly, missing parts may appear in the visible space, and the pass-through errors may also occur near the viewpoint due to insufficient point cloud density. In the sight distance correction stage, this paper categorizes the viewspace using spatial attributes and proposes a sight distance correction method to attenuate the effect of point cloud voids on the volume of the visible space by considering the feature continuity and occlusion relationship within different categories of the viewspace. Finally, visibility analysis is conducted using the visual space volume index at the viewpoint and is compared with the existing voxel-based and surface-based visibility analysis methods, as well as a manual evaluation. The DTW distances for comparisons are 48 247, and 240, respectively. The results show that the proposed method has better consistency with manual evaluation results. It can analyze the visibility of indoor environments more effectively, and is also applicable to outdoor visibility analysis. This method provides a comprehensive and reliable visibility analysis strategy for urban planning, landscape design, and other related fields.

Urban 3D modeling is indispensable for digital twinning and the development of smart cities. The effective extraction of building outlines is a critical step in achieving high-precision urban modeling and 3D mapping. At present, the extraction of building outlines from airborne point cloud data still faces challenges, such as low efficiency and accuracy with conventional methods and limited calibration samples. In response to these challenges, this paper introduces a deep learning method for extracting building outlines from 3D airborne point clouds. The airborne LiDAR point clouds are the primary data input. First, through vertical projection to the XOY plane, point clouds of buildings with the application of progressive morphological filtering are converted to rasterized elevation that characterizes spatial variation of terrain and visible light raster images that depict texture differences. Then, the deep learning model based on Lines-Convolutional Neural Networks (Line-CNN) is employed to preliminarily extract line features from raster images, encompassing stages of feature extraction, node prediction, route generation, and others. To enhance the quality of the primary straight-line extraction, an optimization strategy is introduced, which incorporates a range of comprehensive trimming and completion operations, aligning with information extracted from both the elevation and visible light raster images. Simultaneously, false line segments are eliminated, and missing lines are added, resulting in the regular and complete building outline features. To verify the proposed model, the airborne point cloud data from NUIST campus and the ISPRS H3D 2019 datasets are utilized in the experiment. Our results show that the proposed method accurately and comprehensively extracts building outline features from LiDAR images, achieving an impressive average accuracy and completeness rate, both up to 90%. Furthermore, the proposed method is highly efficient and effectively addresses the challenge of insufficient 3D calibration samples in traditional methods, making it suitable for various applications, particularly large-scale urban 3D modeling and cadastral surveying. To sum up, the proposed method constitutes a significant stride in advancing urban modeling and 3D mapping. It provides a novel solution to address the challenges associated with building outline extraction, particularly within the context of smart cities and digital twins. Due to the model's high accuracy, completeness, and efficiency, our method is highly helpful for a wide range of applications in the urban planning and geospatial information fields.

The precision of individual tree segmentation is important for survey of forest resources. However, traditional individual tree segmentation algorithms suffer from issues such as near tree confusion and low computational efficiency when processing large-scale point cloud data. To address these issues, this paper introduces an improved K-means clustering method that combines spectral clustering and particle swarm optimization for individual tree segmentation of airborne LiDAR point clouds. The proposed method is designed to overcome the limitations of conventional methods by increasing the accuracy of tree segmentation and optimizing the processing of large and complex point cloud data. By combining advanced techniques in spectral clustering and particle swarm optimization, the proposed method significantly improves the precision and efficiency of individual tree segmentation. Firstly, the voxelization of the point cloud data is performed using the Mean Shift algorithm, where the adaptive bandwidth and Gaussian kernel function compute the similarity between voxels, resulting a Gaussian similarity graph reflecting the properties of voxels. This graph not only encapsulates the space structure of the forest but also improves the accuracy of the subsequent data analysis and processing. After voxelization, the Nyström method is applied to efficiently manage the Gaussian similarity graph. This method uses K-nearest neighbor search to select representative samples, effectively reducing the computational burden associated with spectral clustering when dealing with large-scale datasets. By selecting representative samples, the algorithm ensures that the main features of the data are retained, facilitating a more manageable and accurate clustering process. This method optimizes the processing of large amounts of point cloud data by balancing computational efficiency with the requirement to maintain data integrity and accuracy, thus providing a robust foundation for accurate tree segmentation. Using the Nyström approximation, approximate eigenvectors of the similarity graph are obtained, facilitating an effective mapping from the high-dimensional space to a low-dimensional feature space. Finally, the particle swarm optimization algorithm is introduced to enhance the K-means clustering process. This optimization algorithm first randomly initializes a set of particles, each representing a set of potential cluster centers. In each iteration, the particles update their clustering speed and position based on the best historical position of the individual and the best historical position of the group, adjusting the clustering centers to minimize the internal cluster distance. In this paper, publicly available point cloud data from NEWFOR is selected for experiments. The experimental results show that the segmentation results obtained by the proposed algorithm are 5.3% higher in accuracy and 23 times more efficient than the comparison algorithm.

ICESat-2, a new-generation satellite for spaceborne lidar altimetry, adopts a multi-beam single-photon counting regime. The presence of a large number of noisy photons in its detection data is well known, and the large amount of photonic data poses a challenge to data transmission and processing due to the limited processor performance and satellite storage resources on board the satellite. Therefore, in order to efficiently denoise the raw detection data in orbit and reduce the data volume for transmission from space to ground, this study proposes a histogram denoising algorithm based on the Poisson distribution characteristics of the raw detection photons. The method includes creating vertical and skewed histograms of the photon point cloud. Firstly, the point cloud data segment is divided into a two-dimensional grid to form a vertical histogram based on the photon transmission distance. The thresholds for distinguishing signal and noise photons are calculated from the mean and standard deviation of the number of photons in the histogram box, and signal photons are assigned with low, medium, and high confidence labels to characterize signal reliability. This is the process of vertical histogramming. Secondly, the slope information is obtained by linearly fitting the medium- and high-confidence photons from the first step. The photon transmission distance is projected to the vertical direction along the slope to form a tilted histogram, and the signal is recognized twice, and the confidence labels are merged. Finally, the noise is eliminated to achieve the purpose of compressing the original data for downlinking. This study also compares the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm and Ordering Points to Identify the Clustering Structure (OPTICS) algorithm to evaluate the performance of the histogram denoising algorithm. Experiments were conducted using ATL02 data for eight surface types. The results show that DBSCAN and OPTICS are not suitable for data denoising in urban areas, with F-values of 0.766 and 0.765, respectively. In contrast, the histogram algorithm is robust and adaptable to different landform types, with F-values of more than 0.90 in all the experimental areas. The DBSCAN algorithm results in the loss of signal photons, while the OPTICS algorithm produces spurious signal clusters. In contrast, these problems are effectively avoided due to the inclusion of signal rate constraints and horizontal to overlapping division histogram processing techniques in the histogram algorithm. Among the histogram algorithms, the vertical histogram algorithm achieves a signal photon recall (R) of 1, and the average precision (P) and F-value are more than 0.90. This improves the operation efficiency by 12 times, 3473 times, and 1528 times compared with the vertical/skewed histogram, DBSCAN, and OPTICS algorithms, respectively. In addition, the average operation time of the vertical histogram algorithm is only 0.048 seconds, realizing efficient data denoising and compression. The denoising result of the vertical histogram algorithm initially meets the requirement of on-orbit processing efficiency (<0.25 seconds). This study can provide a technical reference for future on-orbit denoising of spaceborne photon-counting LiDAR data.

The accurate extraction of street trees is of great significance for the construction of ecological garden cities and the promotion of urban intelligence. However, in vehicle LiDAR point cloud data, there are many situations where the street tree and the adjacent ground objects, such as buildings or other vegetation, are occluded or connected to each other. These situations make it difficult to accurately extract street trees. To solve the above problems, a hierarchical dynamic region-growing method for street tree extraction was proposed in this paper. Firstly, the ground points were filtered out using the point cloud rasterization method, and street trees were preliminarily extracted according to the projection features of ground objects. Secondly, the point cloud data was stratified with equal height intervals based on the distribution characteristics of ground objects. A hierarchical point cloud space was constructed to further obtain information about street trees and interfering objects. Then, the dynamic region-growing operation was carried out in the hierarchical point cloud space, and the attribute information of the point cloud between the same layer and the adjacent layer was obtained. Based on the obtained attribute information, point cloud clusters were generated to distinguish street trees from interfering objects. Finally, the interfering objects were filtered out to achieve accurate street tree extraction using the geometric features of the interfering objects and the rod-like features of the street trees. In this paper, competition data provided by LiDAR conference and point cloud data provided by Open DataLab website were selected for our experiment. The street data of Lille and Paris from the Open DataLab website were mainly used. The experimental results demonstrate that the accuracy and completeness of the proposed method for extracting street trees are above 98.69% and 97.73%, respectively. The proposed method can achieve accurate and complete street tree extraction even when street trees and adjacent ground objects are occluded or connected with each other. Overall, the experimental results strongly support the effectiveness and reliability of the proposed method for street tree extraction. Furthermore, the hierarchical dynamic region-growing method proposed in this paper for street tree extraction demonstrates remarkable data applicability. This method exhibits a robust capacity to adapt to different datasets, making it suitable for a wide range of scenarios. Notably, even when the independence of street trees is not immediately evident, the proposed approach can still effectively extract them.

Large outdoor point clouds have rich spatial structures and are one of the important means of obtaining geographic information. They have broad application prospects in fields such as autonomous driving, robot navigation, and 3D reconstruction. Due to its inherent irregularity, complex geometric structural features, and significant changes in land scale, the accuracy of point cloud segmentation remains a huge challenge. At present, most point cloud segmentation methods only extract features based on the original 3D coordinates and color information of point cloud data and have not fully explored the information contained in point cloud data with rich spatial information, especially the problem of insufficient utilization of geometric and color information in large-scale point clouds. In order to effectively address the aforementioned issues, this paper introduces the CMGF-Net, a method for semantic segmentation of point clouds that effectively integrates color information and multi-scale geometric features. In this network, dedicated modules are designed for extracting geometric feature information and semantic feature information. In the geometric feature information extraction path, to fully leverage the geometric characteristics of point cloud data, two feature extraction modules are designed: the Relative Position Feature (RPF) extraction module and the Local Geometry Properties (LGP)extraction module, both focusing on the characteristics of the local neighborhood. In the RPF module, spatial normal information of the 3D point cloud and relative spatial distances are utilized to extract the relative positional relationships between neighboring points and the central point. The LGP module exploits the unique performance characteristics of point cloud geometric properties across different terrains, integrating geometric attribute features from the local region. Subsequently, the designed Local Geometric Feature Fusion module (LGF) combines the extracted feature information from the RPF and LGP modules, yielding fused geometric feature information. Furthermore, to learn multi-scale geometric features from the point cloud, CMGF-Net conducts geometric feature extraction at different scales within the network layers. Eventually, the extracted geometric features are hierarchically fused with semantically extracted features based on color information. By extracting multi-scale geometric features and integrating semantic features, the learning ability of the network is enhanced. The experimental results show that our proposed network model achieves a mean Intersection Over Union (mIoU) of 78.2% and an Overall Accuracy (OA) of 95.0% on the Semantic3D dataset, outperforming KPConv by 3.6% and 2.1%, respectively. On the SensatUrban dataset, it achieves a mIOU of 59.2% and an OA of 93.7%. These findings demonstrate that the proposed network model, CMGF-Net, yields promising results in the segmentation of large-scale outdoor point clouds.

With the development of 3D sensors and 3D reconstruction techniques, the registration and fusion of cross-source point clouds have become a research hotspot. However, traditional registration methods use a single feature as the registration primitive, which leads to problems such as weak spatial geometric constraints. Combining multiple structural features with joint constraints can improve the registration accuracy to a certain extent. In order to fuse cross-source point cloud data with high accuracy and fully express the façade information in the scene, this paper proposes a cross-source point cloud registration method based on the constraints of line and surface features. Firstly, the homonymous line and plane features in the cross-source point cloud are extracted by RANSAC algorithm, which are mainly used to constrain the point cloud model in registration. Then the quaternion method is used to describe the spatial transformation parameters based on the line and surface features. The rotation and transformation in arbitrary 3D space can be realized at a faster calculation speed compared with other representations while also avoiding the gimbal lock phenomenon. The line features are used as the constraints of registration, the spatial transformation objective function is constructed, and the parameters related to the transformation are estimated to complete a coarse registration and solve the scale variability. Based on the coarse registration, the surface features are further used as the constraints to solve the rotation matrix and translation parameters to achieve a fine registration. The use of surface features instead of point features as the registration primitives can avoid the selection of common points from massive point cloud data, reducing the accidental errors selected by human selection, avoiding the accumulation of errors, and further improving the registration accuracy. Finally, experiments are conducted using the image-matched point clouds and LiDAR point cloud data for a small area and a large area to study the feasibility of this paper's method in different scales. Results show that the RMSE values for the small-area single building, multiple buildings, and large-area building clusters are 0.364 7, 0.032, and 0.614 6, respectively. The maximum angle between the homonymous surfaces does not exceed 1.5°, the minimum is less than 0.1°, and the mean value of the angle is within the range of 1°. The coarse registration based on line feature constraints can solve the scaling problem well in different scenarios, and the fine registration based on surface feature constraints can further improve the accuracy of the rotation matrix and translation parameters. These results indicate that the cross-source point cloud registration method based on line and surface feature constraints is feasible at different scales.

Accurate segmentation of building roofs is important for 3D building model reconstruction. However, traditional segmentation methods have some problems such as under-segmentation, over-segmentation, and difficulty in accurately segmenting small surfaces due to the diverse and complex characteristics of building roofs such as varying sizes and shapes, as well as the properties of LiDAR point clouds such as uneven density and large amount of data. To address these problems, this paper proposes a building roof segmentation method using voxel-based region growing to improve the segmentation accuracy from airborne LiDAR point clouds. First, the voxel size is determined based on the density of points derived from the point clouds projected onto the xy-plane, and the voxelization of the point cloud is performed. Then, the normal vector and curvature of each voxel are estimated using the PCA method. This leads to the initial roof segmentation results based on voxel growing. During the segmentation process, the voxel with the minimum curvature value is selected as the initial seed voxel, and the surrounding 26-neighborhood voxels are assigned as the growing voxels. The growth is constrained by the angle of the normal vector between the seed voxel and the growing voxels. The new seed voxels to be grown are determined iteratively based on the absolute difference of the curvature values between the current seed voxel and the growing voxels. The growth continues until no new seed voxels appear. This process is repeated by selecting a new initial seed voxel until the segmentation of all voxels is completed. Finally, the final roof surface is obtained through optimization processes such as merging the over-segmented roof surface into the initial segmentation results, repairing the integrity of the roof surface, and extracting small surfaces for complicated buildings. In this paper, airborne LiDAR point cloud data from two regions, i.e., Vaihingen and Toronto, provided by ISPRS official website, are selected to perform roof segmentation experiments of single buildings and building areas. The results show that the completeness, accuracy, and quality of the point cloud segmentation of complicated building roofs are 95.36%~99.58%, 94.83%~100%, and 90.65%~98.28%, respectively. The proposed method can effectively improve the accuracy of roof segmentation based on LiDAR point clouds without under-segmentation and over-segmentation problems, which provides reliable basic data for the automatic construction of 3D building models based on LiDAR data.

Geometric and semantic integration of 3D building models are important infrastructure data for smart city, they are conducive for promoting the refined management and intelligent application of building facilities. However, most of the existing point cloud-based modeling methods focus on the reconstruction of geometric models with simple roof structure, and semantic and topological relations are ignored. Moreover, these methods are sensitive to noise, which are difficult to assure topological consistency and geometric accuracy. To solve these problems, this paper proposes a 3D primitive fitting algorithm for automatically reconstructing building models with complex roof structure from point clouds. Firstly, a 3D building primitive library is designed, including various 3D building primitives with simple and complex roof types. Secondly, an individual building point cloud input is segmented into multiple planes using RANSAC algorithm. The Roof Topology Graph (RTG) is then generated according to the relationship of roof planes, and the roof type of point cloud is subsequently recognized by comparison of RTG between point cloud and building primitives. Thirdly, the reconstruction is formulated as an optimization problem that minimizes the Point-to-Mesh Distance (PMD) between the point cloud and the candidate meshed building primitive. The sequential quadratic programming optimization algorithm with necessary constraints is adopted to perform holistically primitive fitting, so as to estimate the shape and position parameters of a 3D primitive. Finally, the parameterized model is automatically converted into City Geography Markup Language (CityGML) building model based on the prior 3D building primitive. The generated CityGML LoD2 (second level of detail) models are different from mesh models created by conventional building modeling methods, which are represented with geometric, semantic, and topological information. To evaluate the quality and performance of the proposed approach, airborne lidar and photogrammetric building point clouds with different roof structures are collected from public datasets for test. Several building models with complex roof types are successfully reconstructed by using this approach, and the average PMD of five models is 0.17 m. The proposed algorithm is also compared with three other methods. Experimental results indicate that the proposed method achieves the best geometric accuracy, because the average PMD of each model is less than that of other methods. Moreover, this automatic primitive fitting method is efficient, and it is also robust to noise and local data missing. This study demonstrates that the resulting building models can well fit the input point cloud with topologic integrity and rich semantic. This method provides great potential for accurate and rapid reconstruction of geometric-semantic coherent building models with complex roof condition.

The road boundary extraction from the vehicle-borne LiDAR point cloud can be easily affected by the occlusion of vehicles and pedestrians in urban environment. These occlusion phenomena will cause two problems: one is the generation of pseudo-boundary points, and the other is the discontinuity of boundary lines. We find that the elevation standard deviation constraint can effectively deal with the problems caused by occlusion. First, data pre-processing is carried out, including point cloud subdivision, cloth simulation filtering, and scanline storage. Second, a continuous double window is established based on scanlines. The bidirectional moving window method is adopted to construct elevation difference constraint, angle value constraint, and elevation standard deviation constraint to obtain candidate road points. Then, based on the continuity of road boundary, the DBSCAN algorithm in density clustering is used to generate relatively continuous and accurate road boundary points. Finally, the cumulative curvature value and distance of the boundary breakpoint region are calculated to determine whether a location is a junction. The boundary breakpoint is not considered connected if it is a junction; otherwise, it is regarded as a breakpoint caused by vehicle or pedestrian occlusion. The quadratic polynomial curve is used to smooth the boundary points, and the mathematical parameter model of the boundary is obtained. The experimental results show that the accuracy of road boundary extraction can exceed 80% in an urban environment with more occlusion.

Image geo-localization is a technique that obtains the geographic location information of an image through a series of methods, so as to establish a mapping relationship with the real geographic space. This technique is important for further image information mining and has potential application value in cyberspace surveying and mapping, intelligence acquisition, user outdoor positioning, and augmented reality. Despite the tremendous progress in the field of computer vision, high-precision automatic geo-localization of images still needs to be addressed due to the involvement of multiple fields such as image feature extraction, large-scale data retrieval, large-scale point cloud processing, deep learning, geographic information feature extraction, geometric modeling and reasoning, semantic scene understanding, context-based reasoning, and multiple data fusion. This paper reviews the progress of image geo-localization research, mainly including image geo-localization methods, image geo-localization datasets, image geo-localization evaluation methods, and summary and prospect of image geo-localization. Firstly, image geolocation methods are classified into three categories, i.e., image retrieval, 2D-3D matching, and cross-modal retrieval, according to the relevance of the research content. Secondly, the datasets and evaluation methods used for image geo-localization research are categorized and summarized. The geo-localization datasets include image datasets, cross-view datasets, Structure from Motion (SfM) datasets, and multimodal datasets, etc. The image geo-localization evaluation metrics include Top-k candidates, localization error, position and orientation error per video frame, and accuracy/recall. Finally, the current status of image geo-localization research is analyzed, and the future research directions of image geo-localization are outlined in terms of global geo-localization, natural area geo-localization, multi-method fusion for geo-localization, Point of Interest (POI) data-based geo-localization, and pre-selected location refinement.

The multispectral LiDAR system can simultaneously provide the 3D space and spectral information of the target ground object, which is convenient for ground object recognition, land cover/use classification, and scene understanding. However, most multispectral LiDAR point cloud classification methods cannot fully mine the geometric information of point clouds and achieve poor performance in fine-scale classification. To overcome this limitation, this paper presents a continuous kernel point convolutional network which uses local point cloud geometric information to enhance features. Firstly, the network combines a random sampling with a farthest point sampling to quickly process large-scale multispectral LiDAR point clouds. Then, an enhanced convolution module based on continuous variable convolution is designed to improve the semantic information expression of multispectral LiDAR point cloud data. In order to address the problem that kernel point convolution simply using the distance relationship between the geometric space and feature space of neighboring points and centroids is insufficient to express the local information as a complementary feature of the kernel point convolution network, the local features given to the kernel points are enhanced by using the position relationship between neighboring points and centroids while aggregating the local features to provide richer semantic information for the multispectral LiDAR point cloud classification network. Finally, the weighted label smoothing loss and the Lovasz-Softmax loss are combined to further improve the classification performance. The results on the Titan multispectral LiDAR dataset show that the proposed network achieves an overall accuracy of 96.80%, a macro-F₁ index of 88.51%, and a mIoU value of 83.42%, which is superior to the state-of-the-art (SOTA) multispectral LiDAR data networks. The proposed model uses the combination of grid sampling and KD-Tree to better preserve the geometric features of the original point cloud. In the case of a single batch of 65,536 points, the point cloud sampling time is reduced by 28 261.79 ms compared with similar multispectral LiDAR point cloud classification networks. This Study demonstrates the potential of enhanced feature kernel points convolutional network for multispectral LiDAR point cloud classification tasks.

Airborne LiDAR point clouds are the main data source for obtaining high-quality Digital Elevation Model (DEM), and surface roughness, as the main derivative of DEM, plays an important role in geoscience research. However, there is no clear conclusion about the influence of the airborne LiDAR point cloud data density and interpolation methods on the accuracy of DEMs and surface roughness. Thus, this paper evaluates the performance of five classical interpolation methods including Ordinary Kriging (OK), Radial Basis Function (RBF), Triangulated Irregular Network (TIN), Natural Neighbor (NN), and Inverse Distance Weighting (IDW) for quantifying surface roughness using different LiDAR data density (90%, 70%, 50%, 30%, and 10% of the original data) in three study sites with different terrain characteristics. The results show that: (1) the accuracy of each DEM interpolation algorithm decreases with the decrease of point cloud density, and when the data amount is reduced to 30% of the original data amount, the accuracy of different algorithms is obviously different. Among them, RBF and OK have the highest accuracy, while IDW has the lowest accuracy; (2) the DEM error is positively correlated with surface roughness. With the decrease of data density, the correlation coefficients between DEM error and roughness obtained by OK, RBF, and IDW methods all decrease, and the correlation coefficients between DEM error and roughness obtained by TIN and NN decrease first and then increase at density of 30%; (3) The surface roughness error extracted from DEM based on all interpolation methods increases with the decrease of data density, and the accuracy of IDW derived roughness is the highest when the data density is 90% and 70%. When the data density is reduced by 50%, RBF can capture terrain changes more accurately.

At present, the cross-view matching technology of remote sensing images cannot directly use large-scale satellite images for matching, which is difficult to meet the requirements of large-scale complex scene matching tasks and relies on large-scale datasets, thus lacking a good generalization ability. Aiming at the above problems, this paper proposes a cross-view remote sensing image matching method based on visual transformation using the quality-aware template matching method combined with the multi-scale feature fusion algorithm. In this method, the ground multi-view images are collected by using handheld photographic equipment. The portability and flexibility of the handheld photographic equipment can make it easier for us to collect multi-view images covering the target area. The acquired images are densely matched to generate point cloud data, and principal component analysis is used to fit the best ground plane and perform projection transformation to realize the conversion from the ground side view to the aerial view. Then, a feature fusion module is designed for the VGG19 network. The low, medium and high-level features extracted from remote sensing images are fused to obtain rich spatial and semantic information of remote sensing images. The fusion features of semantic information and spatial information can resist large-scale differences. Finally, the quality-aware template matching method is used. The features extracted from the ground images are matched with the fusion features of the remote sensing images. The matching soft ranking results are obtained, and the non-maximum suppression algorithm is used to select high-quality matching results. The experimental results show that the method proposed in this paper has a high accuracy and strong generalization ability without the need of large-scale datasets. The average matching success rate is 64.6%, and the average center point offset is 5.9 pixels. The matching results are accurate and complete, which provide a new solution for the task of cross-view image matching in large scenes.

The errors of DEM derived from scattered topographic scanning data can be divided into errors propagated from topographic data to raster nodes and errors caused by information loss owing to scattered measuring data being incapable of completely reproducing the terrain surface. This study gave an analytical solution of DEM random errors propagated from data random errors in rasterizing the Triangulated Irregular Network (TIN) model by linear interpolation. Moreover, using the topographic scanning data of gullies recorded by a Terrestrial Laser Scanner (TLS), a method was derived for estimating the systematic and random errors caused by information loss in constructing DEM from the TIN model. The results show that both the mean systematic error and random error resulting from information loss are related to the density of effective measuring points which are located at the vertexes of triangles of the TIN model with at least a node of DEM inside them. Besides, the two kinds of errors are also related to terrain characteristics, among which the mean concavity has an impact on the mean systematic error and the terrain roughness affects the random error. For the DEM with a resolution of 0.1 m×0.1 m, the mean systematic error (H_d, in meter) can be estimated by H_{\mathrm{d}} = 0.0580 \frac{C_{\mathrm{m}}}{D_{\mathrm{e}\mathrm{p}}} + 0.000 \mathrm{} 001, and the random error at a DEM node (σ_t, in meter) can be calculated as {\sigma }_{\mathrm{t}}^2 = {\sigma }_z^2 + ( 0.1593 D_{\mathrm{e}\mathrm{p}}^{-1.049} C_{\mathrm{s}\mathrm{d}}^{0.9811} {)}^2, where C_m is the mean concavity defined as the mean of total height difference (ΔZ_d, in meter) between surrounding eight grids and the middle grid in a 3×3 sliding window, D_ep is the density of effective points (points/dm²), σ_z is the analytical solution of the random error at a DEM node propagated from data random errors, and C_sd is the terrain roughness defined as the standard deviation of the parameter ΔZ_d.