Airport clearance ultrahigh ground feature extraction method fusing LiDAR point cloud and panoramic segmentation model

Abstract

The invention belongs to the technical field of airport clearance safety, and discloses an airport clearance ultrahigh ground feature extraction method fusing LiDAR point cloud and a panoramic segmentation model, and the method comprises the steps: collecting point cloud data in an airport clearance protection region through employing an unmanned plane airborne LiDAR, and carrying out the noise filtering, outlier elimination and datum point-based elevation standardization; constructing a data set with semantic tags and instance numbers, inputting the data set into a SuperCluster panoramic segmentation model, sequentially executing feature extraction, adjacency relation construction, edge weight distribution, local supervision and super-point aggregation, and outputting a segmentation result with semantics and instances; extracting a highest point in each instance point set, and performing spatial comparison with an inner horizontal plane, an approach plane and a transition plane limiting plane established according to an ICAO standard to obtain an instance category, a highest point coordinate, an intrusion depth and a spatial relationship; and performing three-dimensional visual display on the judgment result through the clearance supervision platform.

Description

Technical Field

[0001] This invention belongs to the field of airport airspace safety technology, specifically a method for extracting ultra-high ground features in airport airspace by integrating LiDAR point cloud and panoramic segmentation model. Background Technology

[0002] Airport airspace protection zone is a crucial spatial area for ensuring the safety of aircraft takeoff and landing. Its spatial boundaries are usually determined based on Obstacle Limitation Surfaces (OLS) proposed by the International Civil Aviation Organization (ICAO). Obstacle Limitation Surfaces are used to prevent buildings, trees, communication towers, and other ground features from encroaching on the flight path, ensuring that aircraft have the necessary safety space during takeoff, climb, approach, and landing. As urban construction continues to expand around airports, the number of illegal constructions and tall ground features is gradually increasing. Overly tall ground features have become a major hidden danger affecting airport airspace safety.

[0003] Existing methods for detecting airspace obstacles mainly rely on surveying techniques such as GPS-RTK, total stations, and stereo mapping. While these methods offer high measurement accuracy, they suffer from low operational efficiency and long detection cycles under large-scale and complex terrain conditions, making it difficult to meet the dynamic monitoring requirements of airport airspace. In recent years, with the development of airborne lidar (LiDAR) and deep learning technologies, point cloud data has been introduced into airport airspace detection. LiDAR technology can acquire high-density 3D point clouds, providing a rich information foundation for modeling surface buildings, vegetation, and other features. Deep learning models have also provided new tools for semantic segmentation and instance recognition of point clouds.

[0004] However, existing point cloud detection methods still have limitations. Some methods only perform elevation comparison within the overall point cloud, lacking instance-level recognition of independent targets, resulting in redundant calculations and insufficient discrimination accuracy. They also have limited ability to identify slender or sparse targets such as rod-shaped objects and complex vegetation, which can easily lead to missed detections or misjudgments. At the same time, existing technologies are mostly based on abstract geometric planes for clearance discrimination, which do not directly correspond to the obstacle limit surface parameters of the ICAO standard, and lack engineering applicability and standardization. Summary of the Invention

[0005] The purpose of this invention is to provide a method for extracting ultra-high ground features in airport airspace by integrating LiDAR point clouds and panoramic segmentation models, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for extracting airport airspace high-altitude features by integrating LiDAR point clouds and panoramic segmentation models, wherein the steps of this method include: Point cloud data acquisition: Within the airport's airspace protection zone, three-dimensional point cloud data of the airport's airspace protection zone is acquired. The point cloud data is obtained by an airborne LiDAR carried by a UAV and is acquired under the conditions of 80% forward overlap, 75% lateral overlap, and a flight altitude not exceeding 120 meters. The acquisition results are stored in the standard LiDAR storage format (.las). Point cloud preprocessing: Using 3D point cloud processing software, noise filtering, outlier removal, and elevation standardization are performed on the data obtained in step one, and point cloud data with geometric accuracy correction and unified elevation benchmark are output. Sample construction and annotation: Based on the point cloud data, a sample set is constructed, and semantic categories are performed on buildings, vegetation and rod-shaped objects in the point cloud. Each instance is assigned a unique number to form a training set, a validation set and a test set for supervised learning of the segmentation model. Panoramic segmentation modeling: The point cloud data is input into the SuperCluster panoramic segmentation model. Point-level features are extracted using the semantic feature extraction unit. Instance clustering is achieved by combining point adjacency graph construction and Euclidean distance metric. The model is then optimized using the edge weight optimization unit, local supervision unit, and superpoint aggregation unit to output segmentation results with semantic labels and instance numbers. Instance highest point extraction: In the segmentation results, the highest point of each instance point set is determined, and the highest point is used as a candidate point for airspace intrusion analysis; Obstacle limiting surface modeling: An airport obstacle limiting surface model is established according to the ICAO (International Civil Aviation Organization) standard. The model includes obstacle limiting surface models of the inner horizontal plane, approach plane, and transition plane. Each limiting surface is defined by a reference point and a normal vector and is represented by a point normal equation. Spatial discrimination and output: The candidate points are spatially compared with the obstacle restriction surface. If the candidate point exceeds the restriction surface, it is determined to be an over-height feature; otherwise, it is determined to be a compliant feature. The output results include the instance category, candidate point coordinates, intrusion depth, and relative restriction surface position, and are displayed in three-dimensional visualization on the airport airspace monitoring platform. Over-height instances are marked in red, and compliant instances are marked in green.

[0007] Preferably, point cloud data acquisition is performed within the boundary of the airport's airspace protection zone, and the acquisition route is generated based on the geometric contour of the airspace protection zone, so that the scanning path can cover the entire airspace area and avoid missing areas. The data acquisition platform is an unmanned aerial vehicle (UAV) equipped with an airborne lidar to perform aerial surveys. The flight altitude is set to no more than 120 meters to ensure that the echo points have the spatial resolution required for airspace discrimination, while also complying with the airport's flight altitude management requirements. To ensure data continuity and splicability, the forward overlap was set to 80% and the lateral overlap to 75%, thereby creating sufficient redundant point clouds in the flight strip intersection area. After acquisition, all point cloud data is stored in a unified standard format, specifically .las files. This format supports storing attribute information such as 3D coordinates, intensity values, and timestamps, and is directly compatible with subsequent point cloud preprocessing workflows.

[0008] Preferably, the collected .las format point cloud data needs to be preprocessed before entering the segmentation model to ensure the consistency of the data in terms of spatial scale and elevation datum; The preprocessing process is performed using CloudCompare 3D point cloud processing software. The specific steps are as follows: First, noise points caused by abnormal laser echoes or environmental interference during the scanning process are removed. These points are usually obviously abnormal in intensity value or neighborhood distribution. Secondly, the spatial distance between each point and its neighboring point set is calculated, and isolated points exceeding the set threshold are deleted to eliminate local anomalies caused by outliers. Finally, the airport control point is selected as the elevation benchmark to establish a unified reference elevation surface, and the elevation values ​​of all point cloud data are converted to this reference system to achieve the unification of data from different batches and different flight segments in the same elevation system. After the above processing, the point cloud data has completed geometric accuracy correction and elevation benchmark alignment, and can be directly used as input for subsequent sample construction and panoramic segmentation modeling.

[0009] Preferably, after completing the point cloud data preprocessing, a sample set with semantic labels and instance numbers needs to be constructed for supervised training of the segmentation model. The specific process is as follows: import the preprocessed .las format point cloud data into CloudCompare 3D point cloud processing software, and manually label and verify the point cloud within the airport airspace protection area. The annotation categories are limited to three types: buildings, vegetation, and poles. The category classification criteria are determined based on the geometric features and reflectance intensity characteristics of the ground objects. Within the same category, different independent targets are further distinguished, and each independent target is assigned a unique instance number. The number is generated sequentially and cannot be repeated throughout the entire sample set, so as to uniquely identify each instance during training. After completing semantic annotation and instance numbering, the dataset is divided into training set, validation set, and test set according to the following proportions: training set accounts for 70%, validation set accounts for 20%, and test set accounts for 10%. This allows the model to learn sufficiently during the training phase and to conduct independent performance evaluations during the validation and testing phases. The sample set obtained after the above processing is used as input for panoramic segmentation modeling and has semantic and instance-level information that matches the airport airspace application scenario.

[0010] Preferably, after obtaining preprocessed point cloud data in .las format with semantic category annotation and instance numbering, it is fed into the SuperCluster panoramic segmentation model. This model is adapted to the complex point cloud features of airport airspace protection areas and mainly includes five functional units: semantic feature extraction, adjacency relationship construction, edge weight allocation, local supervision, and superpoint aggregation. Semantic feature extraction unit: The input point cloud is encoded using a Transformer structure. Through a multi-head attention mechanism, the remote dependencies between points are captured, thereby maintaining discrimination accuracy in distinguishing large building blocks, irregular vegetation shapes, and slender and sparse rod-like structures. The adjacency relationship construction unit establishes adjacency relationships based on the Euclidean distance between points. Specifically, within the neighborhood of each point, a set of adjacent points is determined according to a set distance threshold, and a point adjacency graph is generated. This adjacency relationship is used for subsequent clustering and edge weight optimization, and is a key step in organizing the original point cloud data into a graph structure.

[0011] In the formula: , : The three-dimensional coordinates of two points in a point cloud; Spatial distance between two points; when ( When the distance threshold is set, the point is considered to be... and Connected in the adjacency graph; Edge weight allocation unit: Distributes shearing costs on edges in a vertex adjacency graph, defined as follows:

[0012] In the formula: Point-to-point The shear cost; :point With point The probability value of belonging to the same instance, with a range of [0, 1]; when When it increases, The corresponding reduction helps to keep points of the same instance connected. This mechanism is especially critical for sparse targets such as rods, preventing them from being incorrectly segmented during the segmentation process. The allocation rule is based on the probability value of point pairs belonging to the same instance. As the probability value increases, the shearing cost decreases accordingly, thereby encouraging points belonging to the same instance to maintain connectivity. This mechanism is particularly important when dealing with sparse targets such as rods, to avoid them being incorrectly segmented into multiple fragmented instances. Local supervision unit: During model training, independent supervision signals are applied to the semantic prediction results and the object consistency prediction results respectively. Semantic supervision ensures the correctness of point clouds at the category level (buildings, vegetation, poles), while consistency supervision ensures the aggregation of points of the same instance. This dual supervision mechanism effectively reduces the error caused by the conflict between semantic and instance levels. Superpoint aggregation unit: In a local continuous region, multiple adjacent points are merged into a superpoint. Each superpoint inherits the features of its constituent points and performs feature embedding and prediction at the superpoint level. This aggregation strategy can reduce computational complexity while maintaining boundary integrity, so that the segmentation results have both global consistency and can preserve detailed edges. After processing by the above model, the output is point cloud segmentation data with semantic labels and instance numbers, which is directly used as input data for subsequent extraction of the highest point of the instance.

[0013] Preferably, in the segmentation result, each instance point set contains several point cloud data, and each point has three-dimensional coordinates. To determine the highest point of this instance, within the instance's point set, use elevation coordinates. To serve as a benchmark, the point with the highest elevation value is selected as the highest point of this instance, and its elevation is plotted in three-dimensional coordinates. Formatted output; The formula for calculating the highest point is as follows:

[0014] In the formula: The set of points of an instance point set; gather midpoint Elevation coordinates; :gather Maximum elevation value within; Corresponding to The plane coordinates of the point; The coordinates of the highest point obtained through this calculation process It is directly used as input data for obstacle constraint surface modeling and discrimination; Unlike existing common methods that use the overall point cloud elevation range for discrimination, this embodiment limits the extraction of the highest point to each instance point set independently, thereby ensuring that each independent target corresponds to a unique discrimination point and avoiding data redundancy and computational complexity caused by overall comparison.

[0015] Preferably, the obstacle limitation surface modeling is strictly carried out in accordance with the ICAO airspace standards. Three types of obstacle limitation surfaces, namely the inner horizontal plane, the approach plane, and the transition plane, are established within the airport airspace protection area. Each obstacle limitation surface is uniquely determined by a reference point coordinate and a normal vector perpendicular to the plane. Its mathematical expression adopts the point normal equation form. Specifically, as shown in the formula:

[0016] In the formula: The coordinates of any point in space; : The coordinates of the reference point used as a plane reference; : The normal vector component perpendicular to the limiting surface; When the result on the left side of the equation is zero, the point Located on this limiting surface; By setting the reference point coordinates and normal vector respectively, three types of restricted surfaces under the ICAO standard can be obtained: Inner horizontal plane: established from the runway reference point, with the normal vector pointing vertically upwards; Approach plane: defined by the runway endpoints and approach direction, with the normal vector determined by the approach slope; Transition surface: Established from the runway edge points, formed by the combination of the horizontal and vertical components of the normal vector; After modeling is completed, the obstacle constraint surface model serves as the geometric reference for subsequent spatial discrimination of the highest point of the instance, and is used to calculate whether the candidate point is located outside the constraint surface.

[0017] Preferably, in the spatial discrimination and output stage, candidate points extracted from the highest point of the instance are first obtained, and their three-dimensional coordinates are represented as (Xc, Yc, Zc). The horizontal position coordinates (Xc, Yc) of the candidate points are substituted into the plane equation obtained from the obstacle constraint surface modeling to calculate the corresponding elevation value of the point on the constraint surface, denoted as . Compare the elevations Zc and Zlim of the candidate points; When Zc > Zlim, the instance is determined to be an ultra-high ground feature; When Zc≤Zlim, the instance is determined to be a compliant feature; The intrusion depth of a candidate point is defined as:

[0018] In the formula: D: The depth of the candidate point's intrusion relative to the constraint surface; Zc: Elevation coordinates of the candidate point; Zlim: Elevation value of the same horizontal position on the obstacle-limiting surface; When D>0, it means that the candidate point exceeds the constraint surface; When D≤0, it means that the candidate point is inside or coincides with the constraint surface; The judgment result includes the following four items: Instance category (building, vegetation, or stalk); The three-dimensional coordinates (Xc, Yc, Zc) of the candidate point; The spatial relationship of the candidate point relative to the obstacle constraint surface (whether it is above or below the constraint surface); The intrusion depth D of the candidate point; The judgment results are presented in a three-dimensional visualization through the airport airspace monitoring platform, with instances of excessively high ground features marked in red and instances of compliant ground features marked in green, so as to achieve intuitive display and rapid identification in the monitoring scenario.

[0019] An airport airspace over-altitude feature extraction system, based on the above-mentioned method, and comprising: The point cloud acquisition module is used to control the UAV equipped with an airborne LiDAR to perform scanning within the airport's airspace protection area, according to a preset flight path generated based on the area's geometric contour. During the acquisition process, the flight altitude is set to no more than 120 meters, the forward overlap is set to 80%, and the lateral overlap is set to 75%. The acquired point cloud data is uniformly stored as a .las file and used as input for subsequent processing. The point cloud preprocessing module is used to process the .las format point cloud data and performs the following steps in sequence: removing noise points caused by echo anomalies; removing outliers whose distance from the neighboring point set exceeds a set threshold; and standardizing the elevation of all point cloud data based on a unified reference elevation datum established by the airport benchmark, thereby outputting point cloud data that has been geometrically corrected and has a unified elevation datum. The sample construction module is used to import preprocessed point cloud data into point cloud annotation software, and perform manual annotation and verification within the airport's airspace protection area; classify the point cloud according to three semantic categories: buildings, vegetation, and poles, and distinguish different independent targets within the same category; assign a unique instance number to each independent target, ensuring that the number is not repeated throughout the entire sample set; after annotation is completed, the dataset is divided according to the proportion of training set, validation set, and test set to generate a sample set for supervised learning of the panoramic segmentation model; The segmentation analysis module is used to input the labeled data into the SuperCluster panoramic segmentation model and perform semantic feature extraction, adjacency relationship construction, edge weight allocation, local supervision, and superpoint aggregation. Specifically, the semantic feature extraction unit generates point-level feature representations based on the Transformer structure; the adjacency relationship construction unit generates a point adjacency graph based on the Euclidean distance threshold; the edge weight allocation unit adjusts the shearing cost based on the probability that point pairs belong to the same instance; the local supervision unit applies supervision signals to semantic prediction and consistency prediction respectively; and the superpoint aggregation unit merges point sets within local continuous regions and generates superpoint features. Finally, it outputs point cloud segmentation results with semantic labels and instance numbers. The highest point extraction module is used to select the point corresponding to the maximum elevation value as the highest point of the instance in each instance point set of the point cloud segmentation result, with the elevation coordinate value as the comparison benchmark, and output the highest point in three-dimensional coordinate form as the input for spatial discrimination. The obstacle restriction surface modeling module is used to establish three types of obstacle restriction surfaces—inner horizontal plane, approach plane, and transition plane—within the airport's airspace protection area, based on the ICAO airspace clearance standards. Each obstacle restriction surface is uniquely determined by the coordinates of a set reference point and the normal vector perpendicular to that plane, and is represented by a point-normal equation to generate an obstacle restriction surface model for spatial discrimination. The spatial discrimination and visualization module is used to compare the coordinates of the highest point of the instance with the obstacle restriction surface equation, and calculate the spatial positional relationship and intrusion depth of the highest point relative to the restriction surface. When the elevation of the highest point is greater than the height of the corresponding restriction surface, it is judged as an over-height feature; otherwise, it is judged as a compliant feature. The discrimination result includes the instance category, the coordinates of the highest point, the spatial relationship and the intrusion depth information, and is presented in a three-dimensional visualization in the airport airspace supervision platform, where over-height feature instances are marked in red and compliant feature instances are marked in green.

[0020] The beneficial effects of this invention are as follows: 1. This invention utilizes an unmanned aerial vehicle (UAV) equipped with an airborne LiDAR to collect point clouds within the airport's airspace protection area. By limiting parameters such as flight altitude, forward overlap, and lateral overlap, the invention ensures complete spatial coverage and consistent resolution of the acquired point clouds. The collected results are uniformly stored as .las files, and a standardization method based on airport benchmarks is used to establish a unified benchmark across different data batches. This data acquisition and processing method overcomes the shortcomings of existing GPS-RTK and total station detection methods, which suffer from low efficiency and long cycles under large-scale and complex terrain conditions. It provides a reliable and consistent data foundation for subsequent point cloud analysis.

[0021] 2. This invention utilizes the SuperCluster panoramic segmentation model to perform semantic and instance-level processing on point clouds. During the segmentation process, it introduces a Transformer-based feature extraction mechanism, an Euclidean distance adjacency graph construction method, a probability-driven edge weight allocation strategy, a dual supervision mechanism for semantics and consistency, and a superpoint aggregation calculation method. Through the combination of the above multi-units, it can maintain stable performance in the segmentation of three typical targets: buildings, vegetation, and rod-shaped objects. In particular, it significantly improves the problem of missed detection or misclassification in the identification of sparse and slender structures of rod-shaped objects. The resulting segmentation results have semantic labels and instance numbers, laying the data conditions for subsequent instance-level spatial discrimination.

[0022] 3. This invention achieves a complete closed loop from data acquisition and instance identification to spatial discrimination by independently extracting the highest point of each instance point set based on segmentation and spatially comparing it with the obstacle restriction surface established according to ICAO airspace standards. The discrimination output includes information such as target category, three-dimensional coordinates of the highest point, positional relationship with the restriction surface, and intrusion depth, and is presented in a three-dimensional visualization form in the airspace supervision platform. Unlike existing discrimination methods based on the overall point cloud elevation range or abstract geometric plane, this invention can perform standardized comparison of independent ground objects, ensuring that the results correspond to international civil aviation standards and meeting the actual requirements of airport airspace supervision for accuracy and standardization. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the overall technical process of this invention. Figure 2 This is the original point cloud data diagram of the present invention; Figure 3 This invention provides a point cloud data map containing semantic tags; Figure 4 This is a point cloud data map containing instance labels for the present invention; Figure 5 This is a graph showing the loss values ​​during the training and validation phases of this invention. Figure 6 This is a graph showing the changes in the evaluation metrics for panoramic segmentation on the validation set of this invention. Figure 7 This is the confusion matrix diagram of the present invention; Figure 8 This is a graph showing the qualitative results on the test set of this invention; Figure 9 This is a diagram showing the detection results at the highest point of this invention; Figure 10 This is a diagram showing the results of extracting superelevation points for each semantic category in this invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] like Figures 1 to 10 As shown, this embodiment of the invention provides a method for extracting airport airspace high-altitude features by fusing LiDAR point clouds and panoramic segmentation models. The steps of this method include: Point cloud data acquisition: Within the airport's airspace protection zone, three-dimensional point cloud data of the airport's airspace protection zone is acquired. The point cloud data is obtained by an airborne LiDAR carried by a UAV and is acquired under the conditions of 80% forward overlap, 75% lateral overlap, and a flight altitude not exceeding 120 meters. The acquisition results are stored in the standard LiDAR storage format (.las). Point cloud preprocessing: Using 3D point cloud processing software, noise filtering, outlier removal, and elevation standardization are performed on the data obtained in step one, and point cloud data with geometric accuracy correction and unified elevation benchmark are output. Sample construction and annotation: Based on point cloud data, a sample set is constructed, and semantic categories are performed on buildings, vegetation and rod-shaped objects in the point cloud. Each instance is assigned a unique number to form a training set, a validation set and a test set for supervised learning of the segmentation model. Panoramic segmentation modeling: Point cloud data is input into the SuperCluster panoramic segmentation model. Point-level features are extracted using the semantic feature extraction unit. Instance clustering is achieved by combining point adjacency graph construction and Euclidean distance metric. The model is then optimized using edge weight optimization unit, local supervision unit, and superpoint aggregation unit, and the segmentation result with semantic labels and instance numbers is output. Instance highest point extraction: In the segmentation results, the highest point of each instance point set is determined and used as a candidate point for airspace intrusion analysis; Obstacle limiting surface modeling: An airport obstacle limiting surface model is established according to the ICAO (International Civil Aviation Organization) standard. The model includes obstacle limiting surface models for the inner horizontal plane, approach plane, and transition plane. Each limiting surface is defined by a reference point and a normal vector and is represented by a point normal equation. Spatial discrimination and output: The candidate point is spatially compared with the obstacle restriction surface. If the candidate point exceeds the restriction surface, it is judged as an over-height feature; otherwise, it is judged as a compliant feature. The output results include the instance category, candidate point coordinates, intrusion depth and relative restriction surface position, and are displayed in three-dimensional visualization on the airport airspace supervision platform. Over-height instances are marked in red, and compliant instances are marked in green.

[0026] Among them, point cloud data acquisition is performed within the boundary of the airport's airspace protection zone, and the acquisition route is generated according to the geometric contour of the airspace protection zone, so that the scanning path can cover the entire airspace area and avoid missing areas. The data acquisition platform is an unmanned aerial vehicle (UAV) equipped with an airborne lidar to perform aerial surveys. The flight altitude is set to no more than 120 meters to ensure that the echo points have the spatial resolution required for airspace discrimination, while also complying with the airport's flight altitude management requirements. To ensure data continuity and splicability, the forward overlap was set to 80% and the lateral overlap to 75%, thereby creating sufficient redundant point clouds in the flight strip intersection area. After the data acquisition is completed, all point cloud data are stored in a unified standard format, specifically .las format files. This format supports storing attribute information such as three-dimensional coordinates, intensity values, and timestamps, and is directly compatible with subsequent point cloud preprocessing workflows. Compared to existing general point cloud acquisition methods that do not limit the acquisition area and storage format, this embodiment combines the boundary limitation of the airport airspace protection area, the route geometry generation rules, the overlap parameter constraints, and the unified data format to form a specific acquisition specification for the airport airspace scenario, and provides standardized input for subsequent point cloud preprocessing and segmentation modeling.

[0027] The collected .las format point cloud data needs to be preprocessed before entering the segmentation model to ensure the consistency of the data in terms of spatial accuracy and elevation datum. The preprocessing process is performed using CloudCompare 3D point cloud processing software. The specific steps are as follows: First, noise points caused by abnormal laser echoes or environmental interference during the scanning process are removed. These points are usually obviously abnormal in intensity value or neighborhood distribution. Secondly, the spatial distance between each point and its neighboring point set is calculated, and isolated points exceeding the set threshold are deleted to eliminate local anomalies caused by outliers. Finally, the airport control point is selected as the elevation benchmark to establish a unified reference elevation surface, and the elevation values ​​of all point cloud data are converted to this reference system to achieve the unification of data from different batches and different flight segments in the same elevation system. After the above processing, the point cloud data has completed geometric accuracy correction and elevation benchmark alignment, and can be directly used as input for subsequent sample construction and panoramic segmentation modeling; Unlike existing processing methods that only use general filtering or average plane correction, this embodiment introduces airport-specific benchmarks for elevation unification, ensuring the comparability and consistency of point cloud data across flight segments and batches in airport airspace scenarios.

[0028] After completing the point cloud data preprocessing, a sample set with semantic labels and instance numbers needs to be constructed for supervised training of the segmentation model. The specific process is as follows: import the preprocessed .las format point cloud data into CloudCompare 3D point cloud processing software, and manually label and verify the point cloud within the airport's airspace protection area. The annotation categories are limited to three types: buildings, vegetation, and poles. The category classification criteria are determined based on the geometric features and reflectance intensity characteristics of the ground objects. Within the same category, different independent targets are further distinguished. Each independent target is assigned a unique instance number, which is generated sequentially and cannot be repeated throughout the entire sample set, so as to uniquely identify each instance during training. After completing semantic annotation and instance numbering, the dataset is divided into training set, validation set, and test set according to the following proportions: 70% for the training set, 20% for the validation set, and 10% for the test set. This allows the model to learn sufficiently during the training phase and to conduct independent performance evaluations during the validation and testing phases. The sample set obtained after the above processing is used as input for panoramic segmentation modeling and has semantic and instance-level information that matches the airport airspace application scenario. Unlike existing general sample sets built on urban point clouds, this embodiment has made clear limitations on category settings, instance numbering rules and proportional division, thus forming a specific sample system applicable to airport airspace scenarios.

[0029] After obtaining preprocessed point cloud data in .las format with semantic category annotation and instance numbering, it is fed into the SuperCluster panoramic segmentation model. This model is adapted to the complex point cloud features of airport airspace protection areas and mainly includes five functional units: semantic feature extraction, adjacency relationship construction, edge weight allocation, local supervision, and superpoint aggregation. Semantic feature extraction unit: The input point cloud is encoded using a Transformer structure. Through a multi-head attention mechanism, the remote dependencies between points are captured, thereby maintaining discrimination accuracy in distinguishing large building blocks, irregular vegetation shapes, and slender and sparse rod-like structures. The adjacency relationship construction unit establishes adjacency relationships based on the Euclidean distance between points. Specifically, within the neighborhood of each point, a set of adjacent points is determined according to a set distance threshold, and a point adjacency graph is generated. This adjacency relationship is used for subsequent clustering and edge weight optimization, and is a key step in organizing the original point cloud data into a graph structure.

[0030] In the formula: , : The three-dimensional coordinates of two points in a point cloud; Spatial distance between two points; when ( When the distance threshold is set, the point is considered to be... and Connected in the adjacency graph; Edge weight allocation unit: Distributes shearing costs on edges in a vertex adjacency graph, defined as follows:

[0031] In the formula: Point-to-point The shear cost; :point With point The probability value of belonging to the same instance, with a range of [0, 1]; when When it increases, The corresponding reduction helps to keep points of the same instance connected. This mechanism is especially critical for sparse targets such as rods, preventing them from being incorrectly segmented during the segmentation process. The allocation rule is based on the probability value of point pairs belonging to the same instance. As the probability value increases, the shearing cost decreases accordingly, thereby encouraging points belonging to the same instance to maintain connectivity. This mechanism is particularly important when dealing with sparse targets such as rods, to avoid them being incorrectly segmented into multiple fragmented instances. Local supervision unit: During model training, independent supervision signals are applied to the semantic prediction results and the object consistency prediction results respectively. Semantic supervision ensures the correctness of point clouds at the category level (buildings, vegetation, poles), while consistency supervision ensures the aggregation of points of the same instance. This dual supervision mechanism effectively reduces the error caused by the conflict between semantic and instance levels. Superpoint aggregation unit: In a local continuous region, multiple adjacent points are merged into a superpoint. Each superpoint inherits the features of its constituent points and performs feature embedding and prediction at the superpoint level. This aggregation strategy can reduce computational complexity while maintaining boundary integrity, so that the segmentation results have both global consistency and can preserve detailed edges. After processing by the above model, the output is point cloud segmentation data with semantic labels and instance numbers, which is directly used as input data for subsequent extraction of the highest point of the instance. Unlike existing point cloud segmentation techniques that only perform semantic classification or rely on a single clustering method, this embodiment combines multiple units such as Transformer feature extraction, probabilistic edge weight allocation, and superpoint aggregation to form a panoramic segmentation model optimized for airport airspace scenarios. It has particular adaptability and stability, especially in the segmentation of slender and sparse targets such as rods. Figure 5 The evolution curves of training loss and validation loss are shown. Throughout the training process, both curves decrease rapidly in the early stage and then converge smoothly, indicating that the model is trained sufficiently and effectively. A key and positive phenomenon is that the training loss is always higher than the validation loss, which is a typical sign that the strong regularization strategy is working effectively. These strategies improve the generalization ability of the model by increasing the difficulty of the training phase. This is also confirmed by the fact that the validation loss curve remains stable in the later stage of training and does not rebound. Figure 6 This demonstrates the evolution of panoptic segmentation performance metrics on the validation set. Panoptic Quality (PQ), Recognition Quality (RQ), and Segmentation Quality (SQ) all showed rapid performance improvements in the early stages of training, subsequently entering a stable plateau. Specifically, Segmentation Quality (SQ) first reached and stabilized at a peak of approximately 75%, indicating excellent model performance in instance segmentation accuracy. Recognition Quality (RQ), as a comprehensive metric measuring detection and classification capabilities, stabilized at approximately 57%. Panoptic Quality (PQ), the final evaluation metric resulting from the product of the two, converged to approximately 43% and remained stable in the later stages of training. like Figure 7 As shown in the figure, the diagonal elements of the matrix clearly display the classification accuracy of each category, with darker colors indicating higher accuracy. The off-diagonal elements reveal the confusion between categories. As can be seen from the figure, the model exhibits excellent classification performance across all categories. For common categories such as buildings and vegetation, the classification accuracy reaches as high as 95.6% and 93.2%, respectively. This result strongly demonstrates the effectiveness of our training strategy in solving the problem of small and sparse target recognition. The model demonstrates excellent classification ability across all categories, especially achieving a high accuracy of 88.7% for the key ultra-tall target such as rod-shaped objects, and significantly suppresses confusion between categories, verifying its effectiveness and robustness in the recognition of small and sparse targets. Figure 8 The figure shows qualitative visualization results of some test scenarios. From the perspective of overall and local magnification, the figure comprehensively compares the model input, the real category and the predicted category, and the real instance and the predicted instance. Through qualitative visualization analysis, the model proposed in this invention shows high accuracy in multi-class target segmentation. In particular, it successfully identifies sparse rod-shaped objects in occluded environments, which verifies its strong identification ability for key ultra-high ground objects. At the same time, it also reveals that there is still room for optimization in fine-grained distinction in other categories.

[0032] In the segmentation results, each instance point set contains several point cloud data, and each point has three-dimensional coordinates. To determine the highest point of this instance, within the instance's point set, use elevation coordinates. To serve as a benchmark, the point with the highest elevation value is selected as the highest point of this instance, and its elevation is plotted in three-dimensional coordinates. Formatted output; The formula for calculating the highest point is as follows:

[0033] In the formula: The set of points of an instance point set; :gather midpoint Elevation coordinates; :gather Maximum elevation value within; Corresponding to The plane coordinates of the point; The coordinates of the highest point obtained through this calculation process It is directly used as input data for obstacle constraint surface modeling and discrimination; Unlike existing common methods that use the overall point cloud elevation range for discrimination, this embodiment limits the extraction of the highest point to each instance point set independently, thereby ensuring that each independent target corresponds to a unique discrimination point and avoiding data redundancy and computational complexity caused by overall comparison. Figure 9 The detection results of the highest points of all instances are displayed intuitively. Red marks indicate that their elevations exceed the limit surface, posing a potential threat to airspace safety, while green marks indicate that they meet the airspace requirements. By superimposing the results on the inclined 3D model and the 3D vector of the airspace obstacle limit surface, the precise spatial positioning of each potential obstacle is achieved.

[0034] Among them, the obstacle limitation surface modeling is strictly implemented in accordance with the ICAO airspace standards. Three types of obstacle limitation surfaces are established in the airport airspace protection area: inner horizontal surface, approach surface and transition surface. Each obstacle limitation surface is uniquely determined by a reference point coordinate and a normal vector perpendicular to the plane. Its mathematical expression adopts the point normal equation form. Specifically, as shown in the formula:

[0035] In the formula: The coordinates of any point in space; : The coordinates of the reference point used as a plane reference; : The normal vector component perpendicular to the limiting surface; When the result on the left side of the equation is zero, the point Located on this limiting surface; By setting the reference point coordinates and normal vector respectively, three types of restricted surfaces under the ICAO standard can be obtained: Inner horizontal plane: established from the runway reference point, with the normal vector pointing vertically upwards; Approach plane: defined by the runway endpoints and approach direction, with the normal vector determined by the approach slope; Transition surface: Established from the runway edge points, formed by the combination of the horizontal and vertical components of the normal vector; After the modeling is completed, the obstacle constraint surface model serves as the geometric reference for subsequent spatial discrimination of the highest point of the instance, and is used to calculate whether the candidate point is located outside the constraint surface; Unlike existing abstract geometric planes that are only publicly disclosed, this embodiment defines the types of limiting surfaces (inner horizontal plane, approach plane, transition plane), the geometric definition method (reference point + normal vector), and the regulatory background that complies with ICAO airspace standards, forming modeling constraints unique to airport airspace and providing an engineering-applicable reference benchmark for subsequent judgment.

[0036] In the spatial discrimination and output stage, candidate points extracted from the highest point of the instance are first obtained, with their three-dimensional coordinates represented as (Xc, Yc, Zc). The horizontal coordinates (Xc, Yc) of the candidate points are then substituted into the plane equation obtained from the obstacle constraint surface modeling to calculate the corresponding elevation value of the point on the constraint surface, denoted as [equation missing]. Compare the elevations Zc and Zlim of the candidate points; When Zc > Zlim, the instance is determined to be an ultra-high ground feature; When Zc≤Zlim, the instance is determined to be a compliant feature; The intrusion depth of a candidate point is defined as:

[0037] In the formula: D: The depth of the candidate point's intrusion relative to the constraint surface; Zc: Elevation coordinates of the candidate point; Zlim: Elevation value of the same horizontal position on the obstacle-limiting surface; When D>0, it means that the candidate point exceeds the constraint surface; When D≤0, it means that the candidate point is inside or coincides with the constraint surface; The judgment result includes the following four items: Instance category (building, vegetation, or stalk); The three-dimensional coordinates (Xc, Yc, Zc) of the candidate point; The spatial relationship of the candidate point relative to the obstacle constraint surface (whether it is above or below the constraint surface); The intrusion depth D of the candidate point; The judgment results are presented in a three-dimensional visualization through the airport airspace supervision platform, with instances of excessively tall features marked in red and instances of compliant features marked in green, so as to achieve intuitive display and rapid identification in the supervision scenario; Unlike existing schemes that only make judgments based on the overall point cloud elevation range, this embodiment compares candidate points one by one at the instance level and within the range, and outputs structured result information including intrusion depth, ensuring that the judgment process directly corresponds to the ICAO limit surface model, forming a complete airport airspace monitoring closed loop. Figure 10 As can be seen, the test set statistically analyzed and semantically categorized the excessive height of the highest points of each instance, revealing the main sources of ground features threatening airport airspace. The results showed that 38.5% of the ground feature instances in the area exhibited excessive height, with trees being the primary source, detecting 26 instances of excessive height, exceeding the compliant limit (18 instances). Buildings were the second most common source, with 9 instances of excessive height. However, most buildings remained at safe heights. The fundamental reason for this situation lies in the differences in ground feature attributes: tree heights change dynamically and are prone to exceeding airspace limits when lacking maintenance, resulting in continuous and gradual risks; building heights are fixed, and excessive height issues are mostly historical legacies or due to insufficient planning. Based on this, targeted and dynamic management measures should be established to periodically review vegetation growth and control dynamic risks; for super high-rise buildings, legal review should be conducted and measures such as installing or removing warning lights should be taken to eliminate long-term safety hazards.

[0038] An airport airspace over-altitude feature extraction system, based on the above-mentioned method, and comprising: The point cloud acquisition module is used to control the UAV equipped with an airborne LiDAR to perform scanning within the airport's airspace protection area, according to a preset flight path generated based on the area's geometric contour. During the acquisition process, the flight altitude is set to no more than 120 meters, the forward overlap is set to 80%, and the lateral overlap is set to 75%. The acquired point cloud data is uniformly stored as a .las file and used as input for subsequent processing. The point cloud preprocessing module is used to process .las format point cloud data and performs the following steps in sequence: remove noise points caused by echo anomalies; remove outliers whose distance from the neighboring point set exceeds a set threshold; and perform elevation standardization on all point cloud data based on a unified reference elevation datum established by the airport benchmark, thereby outputting point cloud data with geometric accuracy correction and a unified elevation datum. The sample construction module is used to import preprocessed point cloud data into point cloud annotation software, and perform manual annotation and verification within the airport's airspace protection area; classify the point cloud according to three semantic categories: buildings, vegetation, and poles, and distinguish different independent targets within the same category; assign a unique instance number to each independent target, ensuring that the number is not repeated throughout the entire sample set; after annotation is completed, the dataset is divided according to the proportion of training set, validation set, and test set to generate a sample set for supervised learning of the panoramic segmentation model; The segmentation analysis module is used to input labeled data into the SuperCluster panoramic segmentation model and perform semantic feature extraction, adjacency relation construction, edge weight allocation, local supervision, and superpoint aggregation. Specifically, the semantic feature extraction unit generates point-level feature representations based on the Transformer structure; the adjacency relation construction unit generates a point adjacency graph based on the Euclidean distance threshold; the edge weight allocation unit adjusts the shearing cost based on the probability that point pairs belong to the same instance; the local supervision unit applies supervision signals to semantic prediction and consistency prediction respectively; and the superpoint aggregation unit merges point sets within local continuous regions and generates superpoint features. Finally, it outputs point cloud segmentation results with semantic labels and instance numbers. The highest point extraction module is used to select the point corresponding to the maximum elevation value as the highest point of the instance in each instance point set of the point cloud segmentation result, with the elevation coordinate value as the comparison benchmark, and output the highest point in three-dimensional coordinate form as the input for spatial discrimination. The obstacle restriction surface modeling module is used to establish three types of obstacle restriction surfaces—inner horizontal plane, approach plane, and transition plane—within the airport's airspace protection area, based on the ICAO airspace clearance standards. Each obstacle restriction surface is uniquely determined by the coordinates of a set reference point and the normal vector perpendicular to that plane, and is represented by a point-normal equation to generate an obstacle restriction surface model for spatial discrimination. The spatial discrimination and visualization module is used to compare the coordinates of the highest point of an instance with the obstacle restriction surface equation, and calculate the spatial positional relationship and intrusion depth of the highest point relative to the restriction surface. When the elevation of the highest point is greater than the height of the corresponding restriction surface, it is judged as an over-height feature; otherwise, it is judged as a compliant feature. The discrimination results include the instance category, the coordinates of the highest point, the spatial relationship and the intrusion depth information, and are presented in a three-dimensional visualization in the airport airspace supervision platform. Over-height feature instances are marked in red, and compliant feature instances are marked in green.

[0039] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for extracting ultra-high ground features in airport airspace by integrating LiDAR point clouds and panoramic segmentation models, characterized in that: The steps of this method for extracting ultra-high ground features in airport airspace by integrating LiDAR point clouds and panoramic segmentation models include: Point cloud data acquisition: Within the airport's airspace protection zone, three-dimensional point cloud data of the airport's airspace protection zone is acquired. The point cloud data is obtained by an airborne LiDAR carried by a UAV and is acquired under the conditions of 80% forward overlap, 75% lateral overlap, and a flight altitude not exceeding 120 meters. The acquisition results are stored in the standard LiDAR storage format (.las). Point cloud preprocessing: Using 3D point cloud processing software, noise filtering, outlier removal, and elevation standardization are performed on the data obtained in step one, and point cloud data with geometric accuracy correction and unified elevation benchmark are output. Sample construction and annotation: Based on the point cloud data, a sample set is constructed, and semantic categories are performed on buildings, vegetation and rod-shaped objects in the point cloud. Each instance is assigned a unique number to form a training set, a validation set and a test set. Panoramic segmentation modeling: The point cloud data is input into the SuperCluster panoramic segmentation model. Point-level features are extracted using the semantic feature extraction unit. Instance clustering is achieved by combining point adjacency graph construction and Euclidean distance metric. The model is then optimized using the edge weight optimization unit, local supervision unit, and superpoint aggregation unit to output segmentation results with semantic labels and instance numbers. Instance highest point extraction: In the segmentation results, the highest point of each instance point set is determined, and the highest point is used as a candidate point for airspace intrusion analysis; Obstacle limiting surface modeling: An airport obstacle limiting surface model is established according to the ICAO (International Civil Aviation Organization) standard. The model includes obstacle limiting surface models of the inner horizontal plane, approach plane, and transition plane. Each limiting surface is defined by a reference point and a normal vector and is represented by a point normal equation. Spatial discrimination and output: The candidate points are spatially compared with the obstacle restriction surface. If the candidate point exceeds the restriction surface, it is determined to be an over-height feature; otherwise, it is determined to be a compliant feature. The output results include the instance category, candidate point coordinates, intrusion depth, and relative restriction surface position, and are displayed in three-dimensional visualization on the airport airspace monitoring platform. Over-height instances are marked in red, and compliant instances are marked in green.

2. The method for extracting airport airspace hyper-elevation features by fusing LiDAR point clouds and panoramic segmentation models according to claim 1, characterized in that: The point cloud data acquisition is performed along a preset route within the boundary of the airport's airspace protection zone. The route is generated based on the geometric contour of the airport's airspace protection zone to ensure that the point cloud covers the entire airspace protection zone. The data acquisition was carried out by an UAV equipped with an airborne lidar. The flight altitude was set to no more than 120 meters, the forward overlap was set to 80%, and the lateral overlap was set to 75%. The acquired point cloud data was uniformly stored in .las format files as input for subsequent point cloud preprocessing and panoramic segmentation.

3. The method for extracting airport airspace hyper-elevation features by fusing LiDAR point clouds and panoramic segmentation models according to claim 2, characterized in that: The point cloud preprocessing uses CloudCompare 3D point cloud processing software to process the acquired .las format point cloud data; The processing includes, in sequence: removing noise points generated during the scanning process, removing outliers whose distance from the neighboring point set exceeds a set threshold, and standardizing the elevation of all point clouds according to the unified reference elevation datum established by the airport benchmark, thereby generating point cloud data that has been geometrically corrected and adapted to the input of the panoramic segmentation model.

4. The method for extracting airport airspace hyper-elevation features by fusing LiDAR point clouds and panoramic segmentation models according to claim 3, characterized in that: The sample construction and annotation include: The preprocessed .las format point cloud data was imported into CloudCompare software, and the point cloud within the airport's airspace protection area was manually labeled and verified. The points in the point cloud are classified according to three preset semantic categories: buildings, vegetation, and poles, and different independent targets are further distinguished within the same category. Each independent target is assigned a unique instance number, which is not repeated throughout the entire sample set; The data with completed semantic labels and instance numbers were divided into training, validation, and test sets in a 7:2:1 ratio to generate a dataset that can be used for supervised learning of panoptic segmentation models.

5. The method for extracting airport airspace hyper-elevation features by fusing LiDAR point clouds and panoramic segmentation models according to claim 4, characterized in that: The panoramic segmentation modeling specifically includes: The preprocessed point cloud data in .las format, which has been semantically categorized and assigned instance numbers, is fed into the SuperCluster panoramic segmentation model. The model includes the following units: The semantic feature extraction unit uses a Transformer structure to represent the features of each point in the input point cloud. The adjacency relationship construction unit establishes a point adjacency graph based on the Euclidean distance threshold between points; The edge weight allocation unit allocates the shearing cost of the edge according to the probability value that the point pair belongs to the same instance, and stipulates that the shearing cost decreases accordingly when the probability value increases. Local supervision units apply independent supervision signals to the semantic prediction results and the object consistency prediction results, respectively. The superpoint aggregation unit aggregates multiple points into a superpoint within a local continuous region and performs feature embedding and prediction at the superpoint level. After processing by the model, a point cloud segmentation result with semantic labels and instance numbers is output. The segmentation result serves as the direct input data for extracting the highest point of the instance.

6. The method for extracting airport airspace hyper-elevation features by fusing LiDAR point clouds and panoramic segmentation models according to claim 5, characterized in that: The extraction of the highest point of the instance is limited to: in each instance point set obtained from the segmentation result, the elevation coordinate Z value of each point in the point cloud is used as the comparison benchmark, and the single point with the largest Z value is selected as the highest point of the instance, and the output is in the form of the three-dimensional coordinates (X, Y, Z) of the highest point. The output is directly used as the input data for obstacle constraint surface modeling and discrimination.

7. The method for extracting airport airspace hyper-elevation features by fusing LiDAR point clouds and panoramic segmentation models according to claim 6, characterized in that: The obstacle constraint surface modeling includes: According to the ICAO (International Civil Aviation Organization) airspace standards, three types of obstacle restriction surfaces are established within the airport's airspace protection area: the inner horizontal surface, the approach surface, and the transition surface. Each obstacle constraint surface is uniquely determined by the preset reference point coordinates and the normal vector perpendicular to the constraint surface, and is represented in the form of a point normal equation. The obstacle constraint surface model serves as the geometric reference for spatial discrimination of the highest point in subsequent instances.

8. The method for extracting airport airspace hyper-elevation features by fusing LiDAR point clouds and panoramic segmentation models according to claim 7, characterized in that: The spatial discrimination and output include: Based on the three-dimensional coordinates (X, Y, Z) of the candidate point extracted from the highest point of the instance, the coordinates are substituted into the constraint surface equation obtained by modeling the obstacle constraint surface to calculate the constraint surface height threshold of the candidate point at that location, and the Z coordinate of the candidate point is compared with the height threshold. When the Z coordinate is greater than the height threshold of the restricted surface, the corresponding instance is determined to be an overly tall feature; when the Z coordinate is less than or equal to the height threshold, the corresponding instance is determined to be a compliant feature. The discrimination results include instance category, three-dimensional coordinates of candidate points, spatial relationship of candidate points relative to the restriction surface, and intrusion depth information, and are presented in a three-dimensional visualization through the airport airspace monitoring platform. Instances of ultra-high ground features are marked in red, while instances of compliant ground features are marked in green.

9. A system for extracting ultra-high ground features in airport airspace, characterized in that: The airport airspace over-altitude feature extraction system is based on claim 8 above, and the system includes: The point cloud acquisition module is used for UAV-borne LiDAR scanning of airport airspace protection areas, setting flight altitude and overlap parameters, and outputting point cloud data in .las format; The point cloud preprocessing module performs noise filtering, outlier removal, and elevation standardization, outputting data corrected for geometric accuracy. The sample construction module is used to assign semantic labels and instance numbers to point clouds based on building, vegetation and rod categories, and generate training and validation samples. The segmentation analysis module is used to run the SuperCluster panoramic segmentation model, which combines semantic feature extraction, adjacency clustering, edge weight optimization, local supervision and superpoint aggregation to output semantic labels and instance numbers. The highest point extraction module is used to determine the highest point in each instance point set and use the highest point as spatial discrimination input; The constraint surface modeling module is used to create obstacle constraint surface models according to ICAO standards and generate point-normal equations. The spatial discrimination and visualization module is used to compare the highest point of an instance with the restricted surface, output the superelevation feature category, coordinates and intrusion depth, and display it in three dimensions on the airport airspace monitoring platform with red or green markers.

Images