A density clustering method, system, device and medium based on a graph neural network

Abstract

The application discloses a density clustering method and system based on a graph neural network, a device, and a medium. The method is based on three-dimensional point cloud data to be processed, and a multi-scale hybrid sparse graph including multiple nodes and multiple scales is constructed. Node features, edge features, scale features, and graph structure reliability input features are extracted from the multi-scale hybrid sparse graph to be input into a multi-task density relationship field prediction network to obtain a predicted density relationship field. Based on the predicted density relationship field, a low-confidence point cloud node set is determined. According to the low-confidence point cloud node set, selective accurate correction of the predicted density relationship field is performed to form a corrected density relationship field. Based on the corrected density relationship field, point cloud adaptive clustering is performed to generate initial clustering labels. According to the initial clustering labels and the corrected density relationship field, point cluster level closed-loop correction is performed to obtain a final clustering result. The method can adapt to complex density dynamic distribution of point clouds and significantly improve point cloud clustering accuracy and robustness.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and data processing technology, and in particular to a density clustering method, system, device and medium based on graph neural networks. Background Technology

[0002] 3D point cloud data is widely used in computer vision tasks such as autonomous driving, robot navigation, 3D reconstruction, augmented reality, industrial measurement, and intelligent transportation. Unlike regular 2D images, 3D point clouds typically consist of a large number of spatially discrete points, each containing features such as 3D coordinates, laser reflection intensity, color, normal vector, and curvature. By performing cluster analysis on 3D point cloud data, roads, vehicles, pedestrians, buildings, obstacles, and background structures can be divided into different point clusters, and object boundaries, occlusion areas, and noise points can be further identified, thus providing crucial data support for high-level scene understanding and spatial perception.

[0003] However, 3D point cloud data is characterized by irregular sampling, uneven local density, significant differences in scale between near and far, complex occlusion relationships, and the presence of noise points. In practical data acquisition scenarios such as LiDAR, the point cloud is dense in areas close to the sensor, while it is sparse in distant or occluded areas. Existing clustering techniques use a uniform distance metric or the assumption of regular cluster shapes for partitioning. This single-scale and static approach is ill-suited to the complex dynamic distribution of point cloud density. It easily leads to small target clusters at a distance being swallowed up by large background areas, misclassification of adjacent target boundary points, and noise points being incorrectly assigned to normal clusters. Consequently, the clustering accuracy and robustness fail to meet the requirements of practical applications. Summary of the Invention

[0004] The main objective of this disclosure is to propose a density clustering method, system, device, and storage medium based on graph neural networks, which can adapt to the complex dynamic density distribution of point clouds and significantly improve the accuracy and robustness of point cloud clustering.

[0005] A first aspect of this application provides a density clustering method based on graph neural networks, the method comprising: Based on the 3D point cloud data to be processed, a multi-scale hybrid sparse graph is constructed; the multi-scale hybrid sparse graph includes multiple nodes and multiple scales. Extract node features, edge features, scale features, and graph structure reliability input features from the multi-scale hybrid sparse graph; The node features, edge features, scale features, and graph structure reliability input features are input into a multi-task density relation field prediction network to obtain the predicted density relation field. Based on the predicted density relationship field, determine the set of low-confidence point cloud nodes; Based on the set of low-confidence point cloud nodes, selective precise correction is performed on the predicted density relation field to form a corrected density relation field; Based on the corrected density relation field, perform adaptive clustering of the point cloud to generate initial cluster labels; Based on the initial clustering labels and the corrected density relationship field, point cluster-level closed-loop correction is performed to obtain the final clustering result.

[0006] The first aspect of this application provides a density clustering method based on graph neural networks. This method constructs a multi-scale hybrid sparse graph, including multiple nodes and multiple scales, based on the 3D point cloud data to be processed. Node features, edge features, scale features, and graph structure reliability input features are extracted from the multi-scale hybrid sparse graph and input to a multi-task density relation field prediction network to obtain a predicted density relation field. Based on the predicted density relation field, a set of low-confidence point cloud nodes is determined. Selective and precise correction is performed on the predicted density relation field according to the low-confidence point cloud node set to form a corrected density relation field. Adaptive clustering of the point cloud is performed based on the corrected density relation field to generate initial cluster labels. Point cluster-level closed-loop correction is performed based on the initial cluster labels and the corrected density relation field to obtain the final clustering result. This method can adapt to the complex dynamic density distribution of point clouds and significantly improves the accuracy and robustness of point cloud clustering.

[0007] To achieve the above objectives, a second aspect of the present invention provides a density clustering system based on graph neural networks, the system comprising: A construction module is used to construct a multi-scale hybrid sparse graph based on the 3D point cloud data to be processed; the multi-scale hybrid sparse graph includes multiple nodes and multiple scales; The extraction module is used to extract node features, edge features, scale features, and graph structure reliability input features from the multi-scale hybrid sparse graph; The input module is used to input the node features, edge features, scale features, and graph structure reliability input features into the multi-task density relation field prediction network to obtain the predicted density relation field; The determination module is used to determine the set of low-confidence point cloud nodes based on the predicted density relationship field; The correction module is used to perform selective and precise correction on the predicted density relation field based on the set of low-confidence point cloud nodes to form a corrected density relation field. The execution module is used to perform adaptive clustering of the point cloud based on the corrected density relation field and generate initial cluster labels; The clustering module is used to perform point cluster-level closed-loop correction based on the initial clustering labels and the corrected density relationship field to obtain the final clustering result.

[0008] To achieve the above objectives, a third aspect of the present invention provides an electronic device, comprising: at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, the instructions being executed by the at least one control processor to enable the at least one control processor to perform the above-described density clustering method based on a graph neural network.

[0009] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions for causing a computer to execute the above-described density clustering method based on a graph neural network.

[0010] It is understood that the beneficial effects of the second to fourth aspects compared with the related technologies are the same as the beneficial effects of the first aspect compared with the related technologies. Please refer to the relevant description in the first aspect above, which will not be repeated here. Attached Figure Description

[0011] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a schematic flowchart of a density clustering method based on graph neural networks provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a density clustering system based on a graph neural network provided in an embodiment of this application; Figure 3 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0012] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0013] In the description of this application, the use of terms such as "first," "second," etc., is for the purpose of distinguishing technical features only and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated.

[0014] In the description of this application, it should be understood that the orientation descriptions, such as up, down, etc., are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0015] In the description of this application, it should be noted that, unless otherwise explicitly defined, terms such as "setup," "installation," and "connection" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this application in conjunction with the specific content of the technical solution.

[0016] 3D point cloud data is widely used in computer vision tasks such as autonomous driving, robot navigation, 3D reconstruction, augmented reality, industrial measurement, intelligent transportation, and scene understanding. Unlike regular 2D images, 3D point clouds typically consist of a large number of spatially discrete points, each of which can contain features such as 3D coordinates, laser reflection intensity, color, normal vector, curvature, height, or local geometric attributes. By performing cluster analysis on 3D point cloud data, roads, vehicles, pedestrians, buildings, vegetation, obstacles, background structures, and isolated anomalies can be divided into different point clusters, and object boundaries, occluded areas, and noise points can be further identified.

[0017] However, 3D point cloud data is characterized by irregular sampling, uneven local density, significant differences in scale between near and far, complex occlusion relationships, and the presence of mixed noise points. In LiDAR acquisition scenarios, point clouds are typically denser in areas closer to the sensor and sparser in areas farther from the sensor or occluded. Large areas of roads, walls, and building facades tend to form high-density point clusters, while pedestrians, traffic signs, small obstacles, object edges, and distant targets may form low-density small clusters, boundary points, or isolated points. If a uniform distance metric or assumption of regular cluster shapes is still used for segmentation, it can easily lead to small target point clusters being swallowed up by background areas, misclassification of boundary points, or noise points being incorrectly classified into normal point clusters.

[0018] Based on this, embodiments of this application provide a density clustering method, system, electronic device, and medium based on graph neural networks, which aims to adapt to the complex dynamic density distribution of point clouds and significantly improve the accuracy and robustness of point cloud clustering.

[0019] The density clustering method, system, electronic device, and medium based on graph neural networks provided in this application are specifically described through the following embodiments. First, the density clustering method based on graph neural networks in this application embodiment is described.

[0020] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0021] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0022] The density clustering method based on graph neural networks provided in this application relates to the fields of artificial intelligence and data processing technology. This density clustering method can be applied to terminals, servers, or software running on either a terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the density clustering method based on graph neural networks, but is not limited to the above forms.

[0023] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0024] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirection to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0025] Therefore, referring to Figure 1 This application provides a density clustering method based on graph neural networks. This method is applied to a central controller, which can be a server, an electronic device, or a mobile terminal, etc. There are no specific limitations here. The method includes the following steps S110 to S170.

[0026] S110. Based on the 3D point cloud data to be processed, construct a multi-scale hybrid sparse graph; the multi-scale hybrid sparse graph includes multiple nodes and multiple scales.

[0027] In this step, the 3D point cloud data to be processed is acquired, and a multi-scale hybrid sparse graph is constructed based on this data. The aforementioned 3D point cloud data to be processed refers to a set of 3D point cloud data that requires clustering processing. This data can be point cloud data collected by LiDAR, point cloud data acquired by a 3D scanner, point cloud data output by a depth camera, or point cloud data collected by other 3D sensing devices. Each point in the 3D point cloud data to be processed typically contains 3D coordinate information and may also contain attribute information such as laser reflection intensity, color, normal vector, curvature, and height. For example, in an autonomous driving scenario, the 3D point cloud data to be processed can be a single frame of road point cloud data collected by an onboard LiDAR. This single frame contains 3D points of targets such as vehicles, pedestrians, guardrails, buildings, and vegetation on the road ahead of the vehicle.

[0028] The aforementioned multi-scale hybrid sparse graph refers to a sparse graph data structure that represents the connectivity between nodes in point cloud data at different scales in the form of a graph structure. A multi-scale hybrid sparse graph includes multiple nodes and multiple scales. Nodes are the basic units of the graph structure; each node corresponds to a point in the 3D point cloud data to be processed or a superpoint region aggregated from neighboring points. Scales refer to different granular levels used to describe the neighborhood range of a node; different scales correspond to different numbers of nearest neighbors and local distance ranges. By setting multiple scales, the graph structure can simultaneously reflect the density distribution characteristics of the point cloud in different spatial ranges. For example, for a vehicle-mounted LiDAR point cloud, three scales can be set, corresponding to neighborhood ranges with 8, 16, and 32 nearest neighbors respectively, thereby simultaneously capturing the local structural features of densely populated areas and sparsely populated areas.

[0029] The process of constructing a multi-scale hybrid sparse graph includes preprocessing the 3D point cloud data to be processed, node mapping, scale setting, and construction of multiple types of candidate edges. Preprocessing may include coordinate normalization, missing value imputation, and feature standardization to eliminate data differences caused by different acquisition devices or conditions. Node mapping can directly map each original point cloud point to a node, or it can aggregate adjacent points falling within the same voxel unit into a superpoint region according to a preset spatial voxel size and map it to a node. Scale setting involves setting multiple scales arranged from smallest to largest on the same node set, with each scale including the number of nearest neighbors, local distance range, and candidate connection rules. Construction of multiple types of candidate edges involves forming candidate neighborhoods at each scale and constructing various types of candidate edges within these neighborhoods, including nearest neighbor edges, mutually nearest neighbor edges, higher-density dependent candidate edges, and cross-scale bridging edges.

[0030] In a specific application example, a single frame of point cloud data from a city road vehicle-mounted LiDAR point cloud dataset is used as the 3D point cloud data to be processed. This single frame contains approximately 120,000 points, from which 512 points are selected as the point cloud to be processed. After coordinate normalization and feature standardization, each point is mapped to a node, forming a node set. Three scales are set on the node set, with the number of nearest neighbors corresponding to the three scales being 8, 16, and 32, respectively. Based on these three scales, approximately 26,000 candidate connections are constructed, thus forming a multi-scale hybrid sparse graph.

[0031] S120. Extract node features, edge features, scale features, and graph structure reliability input features from a multi-scale hybrid sparse graph.

[0032] In this step, based on the multi-scale hybrid sparse graph constructed in step S110, multiple types of features are extracted for input to the prediction network, including node features, edge features, scale features, and graph structure reliability input features. The aforementioned node features refer to feature vectors describing the attributes of a node itself. They originate from the node feature matrix formed in step S110 and may include one or more of the following: region centroid coordinates, number of points within the region, average reflection intensity, average color, average normal vector, curvature statistics, height statistics, local scale information, and variance or extreme values ​​of the aforementioned attributes. For example, for a node in the point cloud of a vehicle-mounted LiDAR, its node features may include the three-dimensional coordinates of the corresponding point, the reflection intensity value, the height value, and the point count statistics of the local region where the point is located.

[0033] The aforementioned edge features refer to the feature information used to describe the relationship between the two nodes connected by a candidate edge. These features include at least the node distance, forward nearest neighbor ranking, reverse nearest neighbor ranking, mutual nearest neighbor identifier, edge type encoding, and scale edge reliability. The node distance refers to the Euclidean distance or the combined spatial and geometric distance between two nodes in three-dimensional space. Forward and reverse nearest neighbor rankings represent the position of a node in another node's nearest neighbor list, respectively. The mutual nearest neighbor identifier indicates whether two nodes are neighbors. The edge type encoding distinguishes different edge types such as nearest neighbor edges, mutual nearest neighbor edges, higher-density dependent candidate edges, and cross-scale bridging edges. Scale edge reliability measures the credibility of a candidate edge at the current scale.

[0034] The scale features mentioned above refer to the feature information used to describe the configuration parameters of each scale. They are encoded by the number of nearest neighbors, the local distance range, and the candidate connection rules corresponding to the scale. For example, for a scale with 16 nearest neighbors, a local distance range of 2 meters, and a candidate connection rule of nearest neighbor connection, the scale features can be formed by encoding the above three parameters into numerical vectors and concatenating them. The graph structure reliability input features mentioned above refer to the feature information used to characterize the stability of the local candidate edge structure of a node. These include average scale edge reliability, cross-scale neighborhood consistency, and the proportion of mutual nearest neighbor connections. Average scale edge reliability is the average level of the reliability of candidate edges associated with a node; cross-scale neighborhood consistency measures whether the candidate neighborhood of the same node is stable at adjacent scales; and the proportion of mutual nearest neighbor connections is the proportion of mutual nearest neighbor connections among the candidate edges associated with a node.

[0035] For example, in the specific application example of the urban road vehicle-mounted LiDAR point cloud above, based on the road point cloud nodes and three neighborhood scales obtained in step S110, the node features of each road point cloud node, the edge features of each candidate edge, the scale features of each scale, and the graph structure reliability input features of each node are extracted to obtain multi-type features for inputting the prediction network.

[0036] S130. Input the node features, edge features, scale features, and graph structure reliability input features into the multi-task density relation field prediction network to obtain the predicted density relation field.

[0037] In this step, the node features, edge features, scale features, and graph structure reliability input features extracted in step S120 are input into a pre-trained multi-task density relation field prediction network, which then outputs a predicted density relation field. This multi-task density relation field prediction network is a network model based on a graph neural network message propagation structure. It simultaneously outputs prediction results related to various density clustering decisions through multi-scale neighborhood information aggregation and scale fusion. This network obtains hidden node representations through node encoding, edge encoding, and multi-scale message propagation, and aggregates neighborhood information on sparse edges at different scales, enabling subsequent predictions to utilize node attributes, edge relationships, and scale differences simultaneously.

[0038] The aforementioned predicted density relation field refers to the set of various prediction quantities output by the network to characterize the density state, dependencies, and prediction reliability of nodes in a multi-scale neighborhood. The predicted density relation field includes multi-scale node density, fused node density, adaptive neighborhood radius, higher-density dependencies, center candidate score, boundary probability, noise probability, graph structure reliability prediction score, and prediction uncertainty. Specifically, higher-density dependencies represent the dependencies of the current node on nodes with higher density within the candidate connection range; the adaptive neighborhood radius represents the local neighborhood range used when performing density comparisons or cluster expansion for this node; the graph structure reliability prediction score represents the network's prediction result on the reliability of the candidate connection structure around the current node; and prediction uncertainty represents the stability of the prediction result for this node.

[0039] For example, in the specific application example of the urban road vehicle-mounted LiDAR point cloud described above, the features of the road point cloud nodes, edge features, scale features, etc., obtained in step S120 are input into the trained multi-task density relation field prediction network to obtain the predicted density relation field of the road point cloud. This predicted density relation field is used to characterize the density state, dependency relationship, and prediction reliability of the road point cloud nodes in the multi-scale neighborhood.

[0040] In one training example, 60 frames of point cloud samples from similar urban road vehicle-mounted LiDAR point clouds were used for training, accumulating approximately 39,600 training points, which were then aggregated to obtain approximately 21,000 training nodes. The network hidden dimension was set to 128, the number of training epochs was set to 50, and the learning rate was set to 0.001. During training, the results in the predicted density relation field were used as training objectives, and a density direction consistency constraint was introduced to ensure that the predicted density direction on candidate connections was consistent with the density direction in the training objective.

[0041] S140. Based on the predicted density relation field, determine the set of low-confidence point cloud nodes.

[0042] In this step, based on the predicted density relation field output in step S130, nodes with low prediction reliability are identified from all nodes according to indicators such as prediction uncertainty, density gradient conflict, and local neighborhood stability, forming a low-confidence point cloud node set. This low-confidence point cloud node set refers to the set of nodes with low prediction reliability in the predicted density relation field. Nodes in this set require selective precise correction in subsequent steps to improve prediction accuracy. Low-confidence point cloud nodes are defined as nodes with low prediction reliability in the predicted density relation field. A node is identified as a low-confidence node when it exhibits high prediction uncertainty, high boundary probability, low graph structure reliability prediction score, large density gradient conflict, or low local neighborhood stability.

[0043] Specifically, for each node, the node confidence is calculated based on the prediction uncertainty in the predicted density relation field and the graph structure reliability prediction score. The lower the prediction uncertainty and the higher the graph structure reliability prediction score, the higher the node confidence, indicating that the predicted relation field of that node is more reliable. At the same time, the neighborhood distance of each node is determined based on the multi-scale hybrid sparse graph, and the local estimated density is calculated based on the neighborhood distance. Then, by combining the node confidence, neighborhood distance, and local estimated density, low-confidence point cloud nodes are screened from all nodes.

[0044] For example, in the specific application example of the urban road vehicle-mounted LiDAR point cloud mentioned above, 18 out of approximately 450 road point cloud nodes were identified as low-confidence nodes. In urban road vehicle-mounted LiDAR point clouds, boundary points between adjacent vehicles, vehicle-guardrail intersections, sparse pedestrian points, distant building facades, and isolated anomalous reflection points are generally more likely to be identified as low-confidence nodes.

[0045] S150. Based on the set of low-confidence point cloud nodes, perform selective precise correction on the predicted density relation field to form the corrected density relation field.

[0046] In this step, the low-confidence point cloud node set obtained in step S140 and the predicted density relation field obtained in step S130 are used as inputs to perform selective exact correction on the low-confidence point cloud nodes, forming a corrected density relation field. The aforementioned selective exact correction refers to performing local recalculation and neighborhood re-search only on low-confidence point cloud nodes, while retaining the prediction results obtained in step S130 for nodes not identified as low-confidence point cloud nodes. This improves the prediction accuracy of low-confidence regions while reducing the overall computational load. The corrected density relation field refers to the density relation field obtained after selective exact correction, which includes the corrected multi-scale node density, the fused node density, the adaptive neighborhood radius, higher density dependencies, center candidate scores, boundary probabilities, noise probabilities, graph structure reliability prediction scores, and prediction uncertainties.

[0047] Specifically, for nodes belonging to the low-confidence point cloud node set, local precise recalculation and extended neighborhood re-retrieval are performed. Within the extended neighborhood, the precise spatial distance or spatial-geometric combination distance is recalculated, and candidate nodes with closer distances, higher edge reliability, or more consistent geometric attributes are preferentially retained to compensate for errors in the fast prediction within the low-confidence region. For nodes not belonging to the low-confidence point cloud node set, local recalculation is not performed; the fast prediction from step S130 is directly used.

[0048] For example, in the specific application example of the urban road vehicle-mounted LiDAR point cloud described above, selective precise correction is performed on 18 low-confidence nodes to redetermine the higher-density dependencies of these nodes, while the remaining approximately 432 nodes retain the prediction results obtained in step S130. This correction is used to reduce the risk that the dependencies of distant sparse vehicles, small pedestrian targets, and object boundary areas are determined solely by the rapid prediction results.

[0049] S160. Perform adaptive clustering of the point cloud based on the corrected density relation field to generate initial cluster labels.

[0050] In this step, adaptive point cloud clustering is performed based on the corrected density relation field output in step S150 to generate initial cluster labels. The aforementioned adaptive point cloud clustering refers to the clustering process that adaptively divides nodes into different clusters based on information such as the density, dependencies, and candidate center scores of each node in the corrected density relation field. The initial cluster labels refer to the node-level labels generated by the adaptive point cloud clustering, used to initially identify the cluster to which each node belongs. The generation process of the initial cluster labels includes operations such as selecting candidate center nodes, calculating dependency confidence, label propagation, and generating candidate boundary points, noise points, and outliers.

[0051] Specifically, a set of candidate centers is selected based on the center candidate score, noise probability, boundary probability, and graph structure reliability prediction score in the corrected density relation field, and initial cluster labels are assigned to the candidate center nodes. Then, for non-center nodes with dependency mappings, the dependency confidence is calculated. When the dependency confidence reaches a preset threshold, the node inherits the label of its dependent node; when the dependency confidence does not reach the preset threshold or the dependency mapping is empty, the node does not inherit the dependent node label and is treated as a boundary point candidate or a pending node for subsequent processing.

[0052] For example, in the specific application example of the urban road vehicle-mounted LiDAR point cloud mentioned above, adaptive clustering is performed based on the corrected density relation field to initially divide the road point cloud nodes into different point clusters such as vehicle point clusters, guardrail point clusters, and building point clusters, and generate initial clustering labels for each node.

[0053] S170. Perform point cluster-level closed-loop correction based on the initial cluster labels and the corrected density relationship field to obtain the final clustering result.

[0054] In this step, the initial clustering labels generated in step S160 and the corrected density relation field output in step S150 are used as inputs to perform point cluster-level closed-loop correction, obtaining the final clustering result. The aforementioned point cluster-level closed-loop correction refers to performing correction operations at the point cluster level, such as boundary point reassignment, noise point and outlier re-determination, reliable connectivity component splitting, point cluster merging, and dependent path multicenter determination, based on the initial clustering result, to eliminate possible misclassification and omissions in the initial clustering. The aforementioned final clustering result refers to the final clustering label, boundary point identifier, noise point identifier, outlier identifier, and node confidence information output for each node after point cluster-level closed-loop correction.

[0055] Specifically, reliable connectivity and node attributes of each node are determined based on the corrected density relation field. Point clustering is then performed on each node according to its attributes and reliable connectivity, resulting in multiple point clusters. For candidate boundary points, if their dependency mapping is not empty, their dependency confidence reaches a preset threshold, and the dependent node already has an initial label, then that dependency label is preferentially considered; otherwise, the number of reliable neighbors of the boundary point connected to each initial point cluster through reliable candidate edges is compared, and the boundary point is reassigned to the point cluster with the most reliable neighbors. For candidate noise points or abnormal points, if there are neighbors in their corrected neighborhood connected through reliable candidate edges, already having obtained the same initial cluster label, and whose number reaches a preset threshold, then they are restored as nodes in the corresponding point cluster.

[0056] For example, in the above-mentioned urban road vehicle-mounted lidar point cloud clustering scenario, compared with methods that only cluster based on fixed neighborhoods or single-scale density relationships, this method can improve the stability and accuracy of clustering results for distant sparse targets, adjacent target boundaries, and occlusion transition regions.

[0057] In this embodiment, a multi-scale hybrid sparse graph is constructed based on the 3D point cloud data to be processed. Node features, edge features, scale features, and graph structure reliability input features are extracted from the multi-scale hybrid sparse graph. These features are then input into a multi-task density relation field prediction network to obtain a predicted density relation field. A low-confidence point cloud node set is determined based on the predicted density relation field. Selective and precise correction is performed on the predicted density relation field based on the low-confidence point cloud node set to form a corrected density relation field. Point cloud adaptive clustering is performed based on the corrected density relation field to generate initial cluster labels. Finally, point cluster-level closed-loop correction is performed based on the initial cluster labels and the corrected density relation field to obtain the final clustering result. Thus, on the one hand, by simultaneously capturing the density distribution characteristics of point clouds in different spatial ranges through multi-scale hybrid sparse graphs, the problem that a single scale is difficult to adapt to the complex dynamic density distribution of point clouds in existing technologies is overcome, and the problem of small target point clusters at a distance being swallowed up by large background areas is avoided. On the other hand, by selective and accurate correction of low-confidence point cloud nodes and point cluster-level closed-loop correction, local accurate recalculation is performed on nodes with low reliability of prediction results, and closed-loop correction is performed on the initial clustering results at the point cluster level to avoid misclassification of adjacent target boundary points and incorrect classification of noise points into normal point clusters, thereby improving clustering accuracy and robustness and meeting the needs of practical applications.

[0058] In some embodiments, the above-described construction of a multi-scale hybrid sparse graph based on the three-dimensional point cloud data to be processed includes steps S201 to S203.

[0059] S201. Preprocess each 3D point cloud in the 3D point cloud data to be processed to obtain the corresponding point cloud nodes and node feature vectors, so as to form a node set and a node feature matrix.

[0060] In this step, each 3D point cloud in the 3D point cloud data to be processed is preprocessed to obtain the corresponding point cloud node and node feature vector, forming a node set and a node feature matrix. The preprocessing mentioned above refers to the operation of standardizing the original point cloud data, which may include coordinate normalization, missing value imputation, and feature standardization. Coordinate normalization is used to unify the 3D coordinates of the point cloud to a preset coordinate range, eliminating coordinate differences caused by different acquisition devices or acquisition locations; missing value imputation is used to complete point cloud points with missing attribute information, such as using the neighborhood mean to imput points lacking reflection intensity information; feature standardization is used to unify feature values ​​of different dimensions to a similar numerical range, preventing a certain dimension feature from dominating subsequent calculations due to excessively large values.

[0061] The aforementioned point cloud node refers to the basic unit in the graph structure. It can correspond to an original point cloud point in the 3D point cloud data to be processed, or it can correspond to a superpoint region formed by aggregating adjacent original point cloud points. For cases where superpoints are not enabled, each original point cloud point is directly mapped to a point cloud node. For cases where superpoints are enabled, adjacent original point cloud points falling within the same voxel unit are aggregated into a superpoint region according to a preset spatial voxel size, and this superpoint region is mapped to a point cloud node. The construction process of the superpoint region is as follows: The voxel unit to which the original point cloud point belongs is calculated based on its 3D coordinates; original points with the same voxel index are grouped into the same superpoint region; for each superpoint region, its mean characteristics, coordinate standard deviation, spatial range, and number of points are statistically analyzed, serving as the feature components of the corresponding point cloud node.

[0062] The node feature vectors mentioned above refer to feature vectors used to describe node attributes, which are statistically obtained based on the original point or superpoint region corresponding to the node. Node feature vectors may include one or more of the following: region centroid coordinates, number of points within the region, average reflection intensity, average color, average normal vector, curvature statistics, height statistics, local scale information, and variance or extreme values ​​of the aforementioned attributes. The node set mentioned above refers to the set composed of all point cloud nodes, and the node feature matrix mentioned above refers to a matrix formed by arranging the node feature vectors of all nodes in node order. For example, for a point cloud containing 512 points to be processed, if superpoints are not enabled, a node set containing 512 nodes is formed. The node feature vector of each node contains features in five dimensions: three-dimensional coordinates, reflection intensity, and height. The node feature matrix has a dimension of 512 rows and 5 columns.

[0063] S202. Set a scale set on the node set. Each scale in the scale set includes the number of nearest neighbors, the local distance range, and the candidate connection rules.

[0064] In this step, a scale set is set on the node set formed in step S201. Each scale in the scale set includes the number of nearest neighbors, the local distance range, and candidate connection rules. The aforementioned scale set refers to a set composed of multiple scales, arranged in ascending order of neighborhood range, with at least two scales in the scale set. Each scale corresponds to a neighborhood range granularity. By setting multiple scales, the graph structure can simultaneously reflect the density distribution characteristics of the point cloud in different spatial ranges.

[0065] The number of nearest neighbors mentioned above refers to a parameter used to limit the number of neighbors that each node has to participate in candidate connections at this scale. For example, the number of nearest neighbors can be set to values ​​such as 8, 16, or 32. The larger the number of nearest neighbors, the larger the neighborhood range of the node at this scale. The local distance range mentioned above refers to a parameter used to limit the spatial distance or spatial and geometric combination distance centered on the node, which is used to constrain the maximum connection distance of candidate edges.

[0066] For example, the local distance range can be set to values ​​such as 1 meter, 2 meters, or 5 meters. No candidate edges are established between node pairs exceeding this distance range. The aforementioned candidate connection rule refers to the rule used to describe the connection method adopted at this scale. It can employ one or more of the following connection methods: nearest neighbor connection, mutual nearest neighbor connection, density-dependent candidate connection, or cross-scale bridging. For example, for a scale with 8 nearest neighbors, the candidate connection rule can adopt the nearest neighbor connection method, that is, selecting the 8 nearest neighbors for each node to establish candidate edges.

[0067] For example, in the specific application example of the urban road vehicle-mounted LiDAR point cloud mentioned above, three scales are set on the node set containing 512 nodes. The number of nearest neighbors for the three scales are 8, 16, and 32, respectively, and each scale is arranged in ascending order of neighborhood range. The number of nearest neighbors for the first scale is 8, the local distance range is 1 meter, and the candidate connection rule is nearest neighbor connection; the number of nearest neighbors for the second scale is 16, the local distance range is 2 meters, and the candidate connection rule is a combination of nearest neighbor connection and mutual nearest neighbor connection; the number of nearest neighbors for the third scale is 32, the local distance range is 5 meters, and the candidate connection rule is a combination of nearest neighbor connection and cross-scale bridging.

[0068] S203. Based on the scale set, node set, and node feature matrix, a multi-scale hybrid sparse graph is constructed.

[0069] In this step, a multi-scale hybrid sparse graph is constructed based on the scale set set set in step S202, the node set formed in step S201, and the node feature matrix. Specifically, at each scale, a scale candidate neighborhood is formed based on the node set and the node feature matrix, and various types of candidate edges are constructed within this scale candidate neighborhood, including nearest neighbor edges, mutually nearest neighbor edges, higher-density dependent candidate edges, and cross-scale bridging edges. Nearest neighbor edges provide basic adjacency relationships; mutually nearest neighbor edges indicate that two nodes are located within each other's nearest neighbor range, used to reflect connection stability; cross-scale bridging edges are used to retain candidate connections in sparse regions, occluded transition regions, and regions with insufficient local connectivity.

[0070] Higher-density dependency candidate edges serve as the input range and constraints for subsequent dependency probability prediction. They are constructed as follows: first, a preliminary density trend is estimated based on the local number of nodes, local kernel density, neighborhood distance, or existing prior density. Then, candidate edges that may point to higher-density nodes are retained within the candidate neighborhood. For each candidate edge, the deviation of the distance between nodes along the candidate edge from the typical distance in the current scale neighborhood is calculated. The reliability of the scale edge is then calculated by combining forward nearest neighbor ranking, reverse nearest neighbor ranking, mutual nearest neighbor identifiers, and edge type. A higher reliability is achieved if the candidate edge distance is closer to the typical distance in the current local neighborhood, ranks higher in both forward and reverse nearest neighbor rankings, the two nodes are mutual neighbors, the candidate edge is a cross-scale bridging edge, or the same node pair appears repeatedly in adjacent scale candidate neighborhoods.

[0071] Candidate edges at each scale are merged, deduplicated, and scale identifiers and edge types are retained to obtain a multi-scale hybrid sparse graph. The multi-scale hybrid sparse graph includes information such as node set, node feature matrix, scale set, candidate neighborhood at each scale, predicted candidate neighborhood, cross-scale bridging neighborhood, candidate edge set, and scale edge reliability.

[0072] For example, in the specific application example of the urban road vehicle-mounted LiDAR point cloud mentioned above, about 26,000 candidate connections were constructed based on 512 nodes and three neighborhood scales, forming a multi-scale hybrid sparse graph.

[0073] In this embodiment, each 3D point cloud in the 3D point cloud data to be processed is preprocessed to obtain the corresponding point cloud node and node feature vector. A scale set including the number of nearest neighbors, local distance range and candidate connection rules is set on the node set. Then, a multi-scale hybrid sparse graph is constructed based on the scale set, the node set and the node feature matrix. In this way, the graph structure can simultaneously reflect the density distribution characteristics of the point cloud in different spatial ranges, providing rich multi-scale structural information for subsequent network prediction and overcoming the problem that a single scale is difficult to adapt to the complex dynamic density distribution of the point cloud.

[0074] In some embodiments, the above-mentioned inputting node features, edge features, scale features and graph structure reliability input features into a multi-task density relation field prediction network to obtain a predicted density relation field includes steps S301 to S303.

[0075] S301. Input the node features, edge features, scale features, and graph structure reliability features into the multi-task density relation field prediction network for encoding and multi-scale neighborhood information aggregation to obtain the hidden representations of multiple nodes and density clustering decision parameters.

[0076] In this step, the node features, edge features, scale features, and graph structure reliability input features extracted in step S120 are input into a multi-task density relation field prediction network. This network performs encoding and multi-scale neighborhood information aggregation to obtain hidden representations of multiple nodes and density clustering decision parameters. The encoding mentioned above refers to the process by which the network transforms and abstracts the input features, including node encoding and edge encoding. Node encoding maps node features to a high-dimensional hidden space to obtain the initial hidden representation of the node; edge encoding fuses edge features with the hidden representations of the connected nodes to obtain edge-enhanced node representations.

[0077] The aforementioned multi-scale neighborhood information aggregation refers to the process by which the network aggregates neighborhood information across sparse edges at different scales, enabling subsequent predictions to simultaneously utilize node attributes, edge relationships, and scale differences. At each scale, the network interacts with its neighboring nodes within the candidate neighborhood of that scale, allowing the node's hidden representation to incorporate neighborhood information at that scale. By aggregating information at multiple scales, the node's hidden representation can simultaneously reflect local structural features across different spatial ranges.

[0078] The aforementioned hidden representation refers to the high-dimensional feature representation of nodes obtained after encoding and aggregation of multi-scale neighborhood information, which integrates node attributes, edge relationships, and multi-scale neighborhood information. The aforementioned density clustering decision parameters refer to various prediction parameters output by the network for density clustering decisions, including prediction adaptive radius, prediction higher-density dependency probability, center candidate score, boundary probability, noise probability, graph structure reliability prediction score, and prediction uncertainty.

[0079] For example, in the specific application example of the point cloud of vehicle-mounted lidar on urban roads mentioned above, the network hidden dimension is set to 128, that is, the hidden representation of each node is a 128-dimensional vector. The network obtains the hidden representation of 512 nodes and the corresponding density clustering decision parameters through multi-scale neighborhood information aggregation.

[0080] S302. Based on the hidden representations and scale features of multiple nodes, determine the fusion weights of each node in the multiple nodes for each scale in the multiple scales, so as to calculate the fusion density of each node in the multiple nodes.

[0081] In this step, based on the hidden representations and scale features of multiple nodes obtained in step S301, the fusion weights of each node across multiple scales are determined to calculate the fusion density of each node. The aforementioned fusion weights refer to the weight values ​​obtained after normalizing the node's adaptation to each scale, and the sum of the fusion weights for each scale is 1. The network first calculates the node's adaptation to each scale based on its hidden representations and scale features, and then normalizes the adaptation to each scale into scale-gated weights.

[0082] The aforementioned fusion density refers to the node density value obtained by weighting and fusing the predicted densities at each scale according to their corresponding fusion weights. Specifically, for a node, its fusion density is equal to the weighted sum of the predicted densities at each scale and the corresponding scale gating weights. This process allows dense regions, sparse regions, boundary regions, and occlusion transition regions to utilize information contributions at different scales. For example, dense regions can rely more on the fine density information of small-scale neighborhoods, while sparse regions can rely more on the coarse density information of large-scale neighborhoods.

[0083] For example, in the specific application example of the urban road vehicle-mounted LiDAR point cloud mentioned above, for a certain road point cloud node, the network calculates the fusion weights of the node for the three scales as 0.5, 0.3, and 0.2 based on its hidden representation and scale features at three scales, and the predicted densities at the three scales are 0.8, 0.6, and 0.4, respectively. Therefore, the fusion density of the node is 0.5 multiplied by 0.8 plus 0.3 multiplied by 0.6 plus 0.2 multiplied by 0.4, resulting in 0.66. For nodes located in densely populated vehicle areas, the network may assign higher fusion weights at smaller scales to capture fine local density changes; for nodes located in sparsely populated areas at greater distances, the network may assign higher fusion weights at larger scales to utilize density information over a wider range.

[0084] S303. Based on the fusion density of each node in the multiple nodes and the density clustering decision parameters of the multiple nodes, the predicted density relationship field is obtained.

[0085] In this step, the predicted density relation field is obtained based on the fusion density of each node in the multiple nodes obtained in step S302 and the density clustering decision parameters of the multiple nodes obtained in step S301. Specifically, the multi-task prediction head outputs the predicted adaptive radius, the predicted higher density dependency probability, the center candidate score, the boundary probability, the noise probability, the graph structure reliability prediction score, and the prediction uncertainty for each node, which together with the fusion density constitute the predicted density relation field.

[0086] The aforementioned adaptive prediction radius refers to the local neighborhood range used by the network when performing density comparisons or cluster expansion for predicted nodes. The aforementioned higher-density dependency probability refers to the probability that a node points to a higher-density node within its candidate neighborhood; this prediction is only performed within the candidate neighborhood, and the edge features of the candidate edges are used as network input to help determine whether a candidate node is suitable as a higher-density dependency node. During prediction, if the fusion density of a candidate node is higher than that of the current node, the candidate node can participate in the normalization of the current node's higher-density dependency probability; if the candidate node's density is not higher than that of the current node, it will not participate in the normalization as a higher-density dependency candidate. The aforementioned center candidate score refers to the probability score of a node predicted by the network as a cluster center candidate. The aforementioned boundary probability refers to the probability that a node predicted by the network is located at the boundary of a node cluster. The aforementioned noise probability refers to the probability that a node predicted by the network is a noise point. The aforementioned graph structure reliability prediction score refers to the network's prediction result regarding the reliability of the candidate connection structure around the current node. The aforementioned prediction uncertainty refers to the stability of the network's prediction results.

[0087] For example, in the specific application example of the urban road vehicle-mounted LiDAR point cloud described above, the network outputs a predicted adaptive radius, a predicted higher density dependency probability, a center candidate score, a boundary probability, a noise probability, a graph structure reliability prediction score, and a prediction uncertainty for each of the 512 road point cloud nodes. These outputs, together with the fused density, constitute the predicted density relation field of the road point cloud. This predicted density relation field is used in subsequent steps to identify low-confidence nodes and serves as a prediction benchmark for selective accurate correction.

[0088] In this embodiment, node features, edge features, scale features, and graph structure reliability input features are input into a multi-task density relation field prediction network for encoding and multi-scale neighborhood information aggregation to obtain the hidden representation of nodes and density clustering decision parameters. Based on the hidden representation and scale features, the fusion weight of each node for each scale is determined to calculate the fusion density. Then, the predicted density relation field is obtained based on the fusion density and density clustering decision parameters. In this way, the network can simultaneously use node attributes, edge relationships, and scale differences for prediction, and scale gating weights can be used to make different density regions adopt information contributions at different scales, thereby improving the adaptability of the predicted density relation field to complex density distributions of point clouds.

[0089] In some embodiments, the above-described determination of the low-confidence point cloud node set based on the predicted density relation field includes steps S401 to S405.

[0090] S401. Extract the graph structure reliability prediction score and prediction uncertainty of multiple nodes from density clustering decision parameters.

[0091] In this step, the graph structure reliability prediction score and prediction uncertainty of multiple nodes are extracted from the density clustering decision parameters obtained in step S303. The graph structure reliability prediction score refers to the network's prediction of the credibility of candidate connection structures around the current node; its value ranges from 0 to 1, with a higher score indicating a more credible candidate connection structure around the node. The prediction uncertainty refers to the stability of the network's prediction results; its value ranges from 0 to 1, with lower uncertainty indicating a more stable prediction result.

[0092] Specifically, the graph structure reliability prediction score and prediction uncertainty of each node are read from the predicted density relation field to form two corresponding numerical sequences.

[0093] For example, in the specific application example of the urban road vehicle-mounted lidar point cloud mentioned above, the graph structure reliability prediction score and prediction uncertainty of each node are extracted from the density clustering decision parameters of 512 road point cloud nodes, forming a graph structure reliability prediction score sequence containing 512 elements and a prediction uncertainty sequence containing 512 elements.

[0094] S402. Calculate the node confidence level of the corresponding node based on the graph structure reliability prediction score and prediction uncertainty of multiple nodes.

[0095] In this step, based on the graph structure reliability prediction scores and prediction uncertainties of multiple nodes extracted in step S401, the node confidence score of the corresponding node is calculated. The node confidence score is an indicator used to measure the credibility of the node's predicted relationship field, calculated based on the prediction uncertainty and the graph structure reliability prediction score. Specifically, the node confidence score equals 1 minus the prediction uncertainty multiplied by the graph structure reliability prediction score; if necessary, the result is limited to the range of 0 to 1. The lower the prediction uncertainty and the higher the graph structure reliability prediction score, the higher the node confidence score, indicating a more credible predicted relationship field for that node, and subsequent steps are more inclined to retain the network's rapid prediction results.

[0096] For example, for a road point cloud node, if its prediction uncertainty is 0.2 and its graph structure reliability prediction score is 0.8, then the node confidence score of this node is 1 minus 0.2 multiplied by 0.8, resulting in 0.64. For another node, if its prediction uncertainty is 0.05 and its graph structure reliability prediction score is 0.95, then the node confidence score of this node is 1 minus 0.05 multiplied by 0.95, resulting in approximately 0.9, indicating that the prediction result for this node is relatively reliable.

[0097] S403. Based on a multi-scale hybrid sparse graph, determine the neighborhood distance of each node.

[0098] In this step, based on the multi-scale hybrid sparse graph constructed in step S110, the neighborhood distance of each node is determined. The neighborhood distance refers to the distance information between a node and its candidate neighbors at the base scale. It can be the set of distances between the node and each candidate neighbor at the base scale, or the distance between the node and its k-th nearest neighbor at the base scale. The base scale refers to the scale with the smallest neighborhood range in the scale set, i.e., the first scale.

[0099] Specifically, candidate neighborhoods of each node at the base scale are read from the multi-scale hybrid sparse graph, and the distances between the node and each candidate neighbor are calculated to form neighborhood distance information.

[0100] For example, in the specific application example of the urban road vehicle-mounted LiDAR point cloud mentioned above, for a certain road point cloud node, the node's 8 candidate neighbors at the basic scale are read from the multi-scale hybrid sparse graph, and the Euclidean distance between the node and the 8 candidate neighbors is calculated to form neighborhood distance information containing 8 distance values.

[0101] S404. Calculate the local estimated density of the corresponding node based on the neighborhood distance of each node.

[0102] In this step, the local estimated density of the corresponding node is calculated based on the neighborhood distances of each node determined in step S403. The aforementioned local estimated density refers to the local density value of the node estimated based on neighborhood distances. This can be achieved using kernel density-based estimation, nearest neighbor distance-based estimation, or other neighborhood distance-based density estimation methods. For example, the local estimated density can be the reciprocal of the average neighborhood distance; a smaller average neighborhood distance indicates a denser point cloud around the node, resulting in a higher local estimated density.

[0103] For example, for a road point cloud node, if the average neighborhood distance of its 8 candidate neighbors at the base scale is 0.5 meters, then the local estimated density of this node can be taken as the reciprocal of 0.5, i.e., 2.0. For another node located in a distant sparse region, if the average neighborhood distance of its 8 candidate neighbors at the base scale is 2.0 meters, then the local estimated density of this node can be taken as the reciprocal of 2.0, i.e., 0.5, indicating that the point cloud around this node is relatively sparse.

[0104] S405. Based on the node confidence, neighborhood distance, and local estimated density of each node, identify low-confidence point cloud nodes from multiple nodes to form a set of low-confidence point cloud nodes.

[0105] In this step, based on the node confidence calculated in step S402, the neighborhood distance determined in step S403, and the local estimated density calculated in step S404, low-confidence point cloud nodes are identified from multiple nodes to form a low-confidence point cloud node set. Specifically, whether a node is a low-confidence node is determined by comprehensively considering indicators such as node confidence, density gradient conflict degree, and local neighborhood stability. Density gradient conflict degree refers to the degree of inconsistency between the density direction predicted by the network and the density direction estimated locally. It is calculated by comparing and fusing the predicted density direction with the locally estimated density direction for candidate neighbors connected to the node and whose edge reliability is not lower than a preset reliable edge threshold. Local neighborhood stability is a weighted comprehensive value of the node's average scale edge reliability, cross-scale neighborhood consistency, and the proportion of mutual nearest neighbor connections.

[0106] A node is added to the low-confidence point cloud node set if it meets any of the following conditions: prediction uncertainty is not less than a preset uncertainty threshold; node confidence is not greater than a preset confidence threshold; density gradient conflict is not less than a preset conflict threshold; local neighborhood stability is not greater than a preset stability threshold; or boundary probability is not less than a low-confidence boundary threshold. Nodes that do not meet the above conditions retain the network's fast prediction results.

[0107] For example, in the specific application example of the above-mentioned urban road vehicle-mounted lidar point cloud, 18 out of approximately 450 road point cloud nodes were identified as low-confidence nodes. These 18 nodes are mainly distributed in areas such as boundary points between adjacent vehicles, intersection points between vehicles and guardrails, sparse pedestrian points, distant building facade points, and isolated anomalous reflection points.

[0108] In this embodiment, the graph structure reliability prediction score and prediction uncertainty are extracted from the density clustering decision parameters, the node confidence is calculated, the neighborhood distance is determined based on the multi-scale hybrid sparse graph and the local estimated density is calculated, and then the node confidence, neighborhood distance and local estimated density are combined to determine the set of low-confidence point cloud nodes. In this way, nodes with low prediction reliability can be accurately identified, providing an accurate target node set for subsequent selective and precise correction, avoiding the computational overhead of performing precise correction on all nodes, and ensuring that nodes with low reliability can be corrected in a targeted manner.

[0109] In some embodiments, the above-described selective precise correction of the predicted density relation field based on the low-confidence point cloud node set to form a corrected density relation field includes steps S501 to S506.

[0110] S501. Based on the set of low-confidence point cloud nodes, determine the low-confidence point cloud nodes and non-low-confidence point cloud nodes in the predicted density relation field.

[0111] In this step, based on the low-confidence point cloud node set determined in step S405, all nodes in the predicted density relation field are divided into two categories: low-confidence point cloud nodes and non-low-confidence point cloud nodes. The aforementioned low-confidence point cloud nodes refer to nodes belonging to the low-confidence point cloud node set; these nodes require local exact recalculation and neighborhood re-search in subsequent steps. The aforementioned non-low-confidence point cloud nodes refer to nodes not belonging to the low-confidence point cloud node set; these nodes do not undergo local exact recalculation in subsequent steps and directly use the network's fast prediction results.

[0112] For example, in the specific application example of the urban road vehicle-mounted LiDAR point cloud mentioned above, approximately 450 road point cloud nodes were divided into 18 low-confidence point cloud nodes and approximately 432 non-low-confidence point cloud nodes. The 18 low-confidence point cloud nodes are mainly distributed in areas such as boundary points between adjacent vehicles, vehicle-guardrail intersections, sparse pedestrian areas, distant building facades, and isolated anomalous reflection points. The prediction results in these areas have lower reliability and require accurate correction.

[0113] S502. For low-confidence point cloud nodes in the predicted density relation field, perform extended neighborhood re-retrieval to obtain the extended neighborhood of the corresponding node.

[0114] In this step, for low-confidence point cloud nodes in the predicted density relation field, an extended neighborhood re-search is performed to obtain the extended neighborhood of the corresponding node. This extended neighborhood re-search refers to the operation of re-searching for candidate nodes within an extended neighborhood range larger than the base scale nearest neighbor range, used to compensate for errors in the fast prediction in low-confidence regions. The extended neighborhood is formed by combining the predicted candidate neighborhood, the cross-scale bridging neighborhood, and the candidate set obtained from the re-search within the extended neighborhood range; it is the union of this extended neighborhood with the predicted candidate neighborhood and the cross-scale bridging neighborhood.

[0115] Specifically, for low-confidence point cloud nodes, the precise spatial distance or spatial and geometric combined distance is recalculated within the expanded neighborhood, and candidate nodes with closer distances, higher edge reliability, or more consistent geometric attributes are prioritized. If the number of candidate nodes obtained from the expanded retrieval is insufficient, the currently available predicted candidate neighborhood and cross-scale bridging neighborhood are retained; if there are many candidate nodes, they are filtered based on distance, edge reliability, and geometric consistency. For example, for a low-confidence point cloud node with 8 basic-scale nearest neighbors, the expanded neighborhood re-retrieval can re-search for candidate nodes within a range of 32 nearest neighbors, resulting in an expanded neighborhood containing more candidate nodes, thereby improving the accuracy of the node's dependency determination.

[0116] S503. Determine the candidate distances for the corresponding nodes based on the extended neighborhood.

[0117] In this step, candidate distances for corresponding nodes are determined based on the extended neighborhood obtained in step S502. These candidate distances are representative distance values ​​determined based on the distances between candidate nodes in the extended neighborhood and the current node. These distances can be quantiles of the distances between candidate nodes in the extended neighborhood, such as the median, quartile, or other preset quantiles. The candidate distances are used to calculate the adaptive radius after subsequent correction, to compensate for potential radius estimation errors that may exist in low-confidence regions during rapid prediction.

[0118] For example, for a low-confidence point cloud node, its extended neighborhood contains 20 candidate nodes. Calculate the distance between this node and the 20 candidate nodes and sort them in ascending order. Take the 75th percentile as the candidate distance. If the distance between this node and the 15th node after sorting is 1.2 meters, then the candidate distance is 1.2 meters.

[0119] S504. Based on the confidence level and candidate distance of low-confidence point cloud nodes in the predicted density relation field, perform corresponding node correction to obtain the first processing result.

[0120] In this step, based on the confidence level of low-confidence point cloud nodes in the predicted density relation field and the candidate distance determined in step S503, corresponding node correction is performed to obtain the first processing result. The node correction includes operations such as calculating the corrected adaptive radius, determining the corrected neighborhood, estimating the corrected density, updating the corrected edge reliability, and determining the corrected dependency probability and dependency mapping. The first processing result refers to the density relation field correlation quantities obtained after the low-confidence point cloud nodes have been corrected, including the corrected adaptive radius, corrected neighborhood, corrected density, corrected edge reliability, corrected dependency probability, and dependency mapping.

[0121] Specifically, for low-confidence point cloud nodes, the corrected adaptive radius is obtained by weighting the node confidence between the predicted adaptive radius and the candidate distance quantile. The higher the node confidence, the closer the corrected adaptive radius is to the network's predicted radius; the lower the node confidence, the more the corrected adaptive radius depends on the candidate distance. Subsequently, candidate nodes within the corrected adaptive radius, the predicted candidate neighborhood, and the cross-scale bridging neighborhood are merged to obtain the corrected neighborhood, and the corrected density is re-estimated based on this neighborhood. Higher-density dependencies are re-evaluated in the corrected neighborhood, and only candidate nodes with a corrected density higher than the current node participate in dependency probability normalization. The candidate node with the highest corrected dependency probability is selected as the dependency mapping.

[0122] For example, for a low-confidence point cloud node, if its node confidence is 0.3, the predicted adaptive radius is 1.0 meter, and the candidate distance is 1.2 meters, then the corrected adaptive radius is 0.3 multiplied by 1.0 plus 0.7 multiplied by 1.2, resulting in 1.14 meters. Because the node confidence is low, the corrected adaptive radius depends more on the candidate distance, thus compensating for potential radius estimation errors that may exist in the fast prediction within this low-confidence region.

[0123] S505. Extract the prediction results of non-low confidence point cloud nodes from the predicted density relation field.

[0124] In this step, the prediction results for non-low-confidence point cloud nodes are extracted from the predicted density relation field. The prediction results for these non-low-confidence point cloud nodes refer to the prediction quantities output by the network in step S130, including the predicted adaptive radius, predicted density, predicted dependency probability, center candidate score, boundary probability, noise probability, graph structure reliability prediction score, and prediction uncertainty. For non-low-confidence point cloud nodes, local recalculation is not performed; the network's fast prediction results are directly used to reduce the overall computational load.

[0125] For example, in the specific application example of the urban road vehicle-mounted lidar point cloud mentioned above, the prediction results of about 432 non-low confidence point cloud nodes are extracted from the prediction density relation field. The prediction results of these nodes are relatively reliable, and no precise correction is required. The fast prediction of the network is directly retained.

[0126] S506. Based on the first processing result and the prediction results of the non-low confidence point cloud nodes, the corrected density relation field is obtained.

[0127] In this step, based on the first processing result obtained in step S504 and the prediction results of the non-low-confidence point cloud nodes extracted in step S505, the corrected density relation field is obtained. The aforementioned corrected density relation field refers to the density relation field composed of the corrected results of the low-confidence point cloud nodes and the prediction results of the non-low-confidence point cloud nodes. It includes corrected density, corrected radius, corrected neighborhood, corrected edge reliability, corrected higher density dependency probability, dependency mapping, and center candidate score, boundary probability, noise probability, graph structure reliability prediction score, prediction uncertainty, and node confidence, etc., which are retained or updated by the network.

[0128] For example, in the specific application example of the urban road vehicle-mounted LiDAR point cloud mentioned above, the correction results of 18 low-confidence point cloud nodes and the prediction results of approximately 432 non-low-confidence point cloud nodes are combined to obtain a corrected density relation field of approximately 450 road point cloud nodes. This corrected density relation field is used to reduce the risk that the dependencies of distant sparse vehicles, small pedestrian targets, and object boundary areas are determined solely by the rapid prediction results, providing more accurate density relation information for subsequent clustering.

[0129] In this embodiment, low-confidence point cloud nodes and non-low-confidence point cloud nodes in the predicted density relation field are determined based on the set of low-confidence point cloud nodes. An extended neighborhood re-search is performed on the low-confidence point cloud nodes to obtain an extended neighborhood. Candidate distances are determined based on the extended neighborhoods. Node correction is performed based on the confidence level of the low-confidence point cloud nodes and the candidate distances to obtain the first processing result. The prediction results of non-low-confidence point cloud nodes are extracted from the predicted density relation field. Then, the corrected density relation field is obtained based on the first processing result and the prediction results of non-low-confidence point cloud nodes. This can improve the prediction accuracy of low-confidence regions while reducing the overall computational load, avoiding the computational overhead of performing precise correction on all nodes, and ensuring that nodes with low reliability can be corrected in a targeted manner.

[0130] In some embodiments, the above-described point cloud adaptive clustering based on the corrected density relation field to generate initial cluster labels includes steps S601 to S606.

[0131] S601. Extract the center candidate score, noise probability, boundary probability and graph structure reliability prediction score of each node from the corrected density relation field.

[0132] In this step, the center candidate score, noise probability, boundary probability, and graph structure reliability prediction score for each node are extracted from the corrected density relation field obtained in step S506. The center candidate score refers to the probability score of a node predicted by the network as a cluster center candidate; a higher score indicates that the node is more suitable as a cluster center. The noise probability refers to the probability that a node is a noise point predicted by the network; a higher probability indicates that the node is more likely to be a noise point. The boundary probability refers to the probability that a node is located at the boundary of a node cluster predicted by the network; a higher probability indicates that the node is more likely to be located at the boundary of a node cluster. The graph structure reliability prediction score refers to the network's prediction of the credibility of the candidate connection structures around the current node; a higher score indicates that the candidate connection structures around the node are more reliable.

[0133] For example, in the specific application example of the urban road vehicle-mounted lidar point cloud mentioned above, the center candidate score, noise probability, boundary probability and graph structure reliability prediction score of each node are extracted from the corrected density relation field of about 450 road point cloud nodes to form four corresponding numerical sequences for subsequent center candidate node selection.

[0134] S602. Select central candidate nodes and non-central candidate nodes from multiple nodes based on the central candidate score, noise probability, boundary probability and graph structure reliability prediction score.

[0135] In this step, based on the center candidate score, noise probability, boundary probability, and graph structure reliability prediction score extracted in step S601, center candidate nodes and non-center candidate nodes are selected from multiple nodes. The aforementioned center candidate nodes refer to nodes that meet the center candidate criteria, and they serve as cluster centers for subsequent label propagation. The aforementioned non-center candidate nodes refer to nodes that do not meet the center candidate criteria, and their respective clusters need to be determined through dependency mapping.

[0136] Specifically, nodes whose center candidate score reaches the center threshold, whose noise probability is not higher than the noise threshold, whose boundary probability is not higher than the boundary threshold, and whose graph structure reliability prediction score is not lower than the reliability threshold can be used as center candidate nodes. For example, for a road point cloud node, if its center candidate score is 0.85 and the center threshold is 0.6, its noise probability is 0.05 and the noise threshold is 0.3, its boundary probability is 0.1 and the boundary threshold is 0.4, and its graph structure reliability prediction score is 0.9 and the reliability threshold is 0.5, then this node meets all the center candidate conditions and can be used as a center candidate node. For another node, if its boundary probability is 0.6 and the boundary threshold is 0.4, then this node does not meet the center candidate conditions and is used as a non-center candidate node.

[0137] S603. Initialize the initial cluster label of the central candidate node.

[0138] In this step, the initial cluster labels of the central candidate nodes selected in step S602 are initialized. The aforementioned initial cluster label refers to the label used to identify the node's node cluster; each central candidate node receives a unique initial cluster label. For example, if 10 central candidate nodes are selected, labels 1 to 10 are assigned to these 10 nodes as the initial cluster labels of their respective node clusters.

[0139] For example, in the specific application example of the urban road vehicle-mounted LiDAR point cloud mentioned above, several central candidate nodes are selected from about 450 road point cloud nodes. These central candidate nodes may correspond to the center point of vehicles, the center point of guardrails, the center point of buildings, etc. on the road, and a unique initial cluster label is assigned to each central candidate node.

[0140] S604. Determine the dependency mapping of non-central candidate nodes.

[0141] In this step, the dependency mappings of the non-central candidate nodes selected in step S602 are determined. These dependency mappings refer to the mapping relationships between non-central candidate nodes and their higher-density dependent nodes, determined based on the corrected higher-density dependency probabilities in the corrected density relation field. Specifically, for non-central candidate nodes with a corrected density higher than the current node, the candidate node with the highest corrected dependency probability is selected as the dependency mapping; if no candidate node has a corrected density higher than the current node, the dependency mapping is left empty.

[0142] For example, if there are three candidate nodes with higher corrected densities in its corrected neighborhood, with corresponding corrected dependency probabilities of 0.5, 0.3 and 0.2, then the candidate node with the highest corrected dependency probability is selected as the dependency mapping of the non-central candidate node.

[0143] S605. Based on the dependency mapping of non-central candidate nodes, determine the node label of non-central candidate nodes.

[0144] In this step, based on the dependency mapping of the non-central candidate nodes determined in step S604, the node labels of the non-central candidate nodes are determined. The node labels refer to the label affiliation status of the non-central candidate nodes, which can be inherited dependency node labels, boundary point candidates, noise point candidates, or outlier candidate candidates, etc. Specifically, for non-central candidate nodes with dependency mappings, the dependency confidence is calculated. The dependency confidence is obtained by multiplying the corrected higher-density dependency probability, the corrected edge reliability, and the node confidence. When the dependency confidence is not less than a preset dependency confidence threshold, the node inherits the labels of its dependent nodes; when the dependency confidence is less than the preset dependency confidence threshold or the dependency mapping is empty, the node does not inherit the dependency node labels and is treated as a boundary point candidate or a pending node.

[0145] Simultaneously, for each node, the number of reliable neighbors, the degree of mixing of neighborhood labels, and the deviation of node features are calculated, and these are combined with boundary probability, noise probability, corrected density, and dependency confidence to form the basis for candidate judgment. When a node's boundary probability is not lower than a preset boundary candidate threshold, its dependency confidence is lower than a preset dependency confidence threshold, or its degree of mixing of neighborhood labels is not lower than a preset mixing threshold, it is considered a boundary point candidate. When a node's noise probability is not lower than a preset noise candidate threshold, its corrected density is lower than a preset low density threshold, and its number of reliable neighbors is lower than a preset reliable neighbor number threshold, it is considered a noise point candidate. When a node's anomaly probability is not lower than a preset anomaly threshold, it is considered an anomaly point candidate.

[0146] For example, for a non-central candidate node, if its dependency mapping is not empty and its dependency confidence is 0.8, and the preset dependency confidence threshold is 0.5, then the node inherits the label of its dependent node. For another non-central candidate node, if its dependency mapping is empty, then the node does not inherit the label of its dependent node for the time being, and is regarded as a boundary point candidate or a pending node.

[0147] S606. Generate initial cluster labels based on the initial cluster labels of the central candidate nodes and the node labels of the non-central candidate nodes.

[0148] In this step, initial cluster labels are generated based on the initial cluster labels of the central candidate nodes initialized in step S603 and the node labels of the non-central candidate nodes determined in step S605. The aforementioned initial cluster labels refer to the set of labels for all nodes. Central candidate nodes have their own initial cluster labels, while non-central candidate nodes whose dependency confidence reaches a preset threshold inherit the labels of their dependent nodes. The remaining nodes, as boundary point candidates, noise point candidates, or outlier candidate nodes, are not assigned labels at this time.

[0149] For example, in the specific application example of the urban road vehicle-mounted lidar point cloud mentioned above, adaptive clustering is performed based on the corrected density relation field to initially divide the road point cloud nodes into different point clusters such as vehicle point clusters, guardrail point clusters, and building point clusters. The central candidate nodes obtain their respective initial cluster labels, and the nodes among the non-central candidate nodes whose dependency confidence reaches a preset threshold inherit the labels of their dependent nodes. Boundary point candidates, noise point candidates, and abnormal point candidates are not assigned labels for the time being, and are handed over to the subsequent point cluster-level closed-loop correction steps for processing.

[0150] In this embodiment, by extracting the center candidate score, noise probability, boundary probability, and graph structure reliability prediction score of each node from the corrected density relation field, center candidate nodes and non-center candidate nodes are screened out. The initial cluster labels of the center candidate nodes are initialized, the dependency mapping of the non-center candidate nodes is determined, the node label of the non-center candidate nodes is determined based on the dependency mapping, and the initial cluster label is generated according to the initial cluster label of the center candidate nodes and the node label of the non-center candidate nodes. In this way, the nodes can be adaptively divided into different point clusters based on the corrected density relation field, while identifying boundary point candidates, noise point candidates, and abnormal point candidates, providing a basis for subsequent point cluster-level closed-loop correction.

[0151] In some embodiments, the above-mentioned point cluster level closed-loop correction based on the initial clustering labels and the corrected density relationship field is used to obtain the final clustering result, including steps S701 to S703.

[0152] S701. Based on the corrected density relationship field, determine the reliable connectivity and the node attributes of each node.

[0153] In this step, reliable connectivity and node attributes of each node are determined based on the corrected density relation field obtained in step S506. The aforementioned reliable connectivity refers to the connectivity formed by candidate edges whose corrected edge reliability is not lower than a preset reliable edge threshold. Only candidate edges with corrected edge reliability not lower than the preset reliable edge threshold are included in the closed-loop correction. The aforementioned node attributes refer to the node-related attribute information used for point cluster partitioning and closed-loop correction, which may include the node's corrected density, dependency mapping, dependency confidence, center candidate score, boundary probability, noise probability, anomaly probability, node feature deviation, number of reliable neighbors, and neighborhood label mixing degree, etc.

[0154] Specifically, the corrected edge reliability of each candidate edge is read from the corrected density relation field. Candidate edges with corrected edge reliability not lower than a preset reliable edge threshold are added to the reliable candidate edge set to form reliable connections. Simultaneously, node attributes of each node are read from the corrected density relation field and the initial clustering labels for subsequent point cluster partitioning and loop closure correction. For example, for a road point cloud node, its node attributes may include a corrected density of 0.7, a dependency mapping pointing to node number 35, a dependency confidence of 0.8, a center candidate score of 0.2, a boundary probability of 0.6, and 5 reliable neighbors.

[0155] S702. Based on the node attributes and reliable connection relationships of each node, perform point clustering on each node to obtain multiple point clusters.

[0156] In this step, based on the node attributes and reliable connectivity relationships of each node determined in step S701, point clustering is performed on each node to obtain multiple point clusters. The aforementioned point clustering refers to the operation of re-dividing nodes based on the initial clustering labels and in conjunction with reliable connectivity relationships. It includes operations such as boundary point redistribution, re-determination of noise points and outliers, splitting of reliable connected components, merging of point clusters, and determination of multi-center dependency paths.

[0157] Specifically, for boundary point candidates, if their dependency mapping is not empty, their dependency confidence is not less than a preset dependency confidence threshold, and the dependent node already has an initial label, then the dependency label is preferentially referenced; otherwise, the number of reliable neighbors of the boundary point connected to each initial point cluster through reliable candidate edges is compared. When the number of reliable neighbors of an initial point cluster reaches a preset reliable neighbor number threshold and the relatively second-best cluster exceeds a preset redistribution interval, the boundary point is redistributed to that initial point cluster. For noise point candidates or abnormal point candidates, if there are neighbors in their corrected neighborhood that are connected through reliable candidate edges, have obtained the same initial cluster label, and whose number reaches a preset reliable neighbor number threshold, and the dependency confidence of each node in the dependency path is not less than a preset dependency confidence threshold and is eventually connected to the central candidate node, and the node feature deviation does not exceed a preset feature deviation threshold, then it is restored to the corresponding point cluster node.

[0158] If multiple disconnected reliable connected components are formed within the same initial point cluster based on reliable candidate edges, and the size of the connected components meets the minimum point cluster requirement, then the initial point cluster is split into multiple candidate point clusters. For two different candidate point clusters, if the number or proportion of reliable cross-cluster dependencies between them reaches the point cluster merging threshold, and the proportion of noisy or abnormal point candidates after merging does not exceed a preset conflict threshold, the proportion of reliable candidate edges within the merged point cluster is not lower than a preset connectivity threshold, the average feature deviation does not exceed a preset deviation threshold, and the dependency path consistency reaches a preset consistency threshold, then the two point clusters are merged. For the dependency path of each node, it is traced step by step along the dependency mapping. If multiple dependency paths within the same point cluster converge to different central candidate nodes, then label propagation is re-executed within the point cluster starting from these central candidate nodes.

[0159] For example, in the specific application example of the above-mentioned urban road vehicle-mounted lidar point cloud, according to the node attributes and reliable connection relationships of each node, the boundary points between adjacent vehicles are redistributed, and the boundary points that are misassigned to adjacent point clusters are reassigned to the correct point clusters; isolated noise points are re-determined, and noise points that meet the recovery conditions are restored to the corresponding point cluster nodes; point clusters are merged for the broken connectivity components caused by the occlusion transition area, and point clusters that should belong to the same target are re-merged.

[0160] S703. Based on multiple point clusters, generate the final clustering results for each node.

[0161] In this step, based on the multiple point clusters obtained in step S702, the final clustering result for each node is generated. The aforementioned final clustering result refers to the clustering-related information of each node output after point cluster-level closed-loop correction, including the final cluster label, boundary point identifier, noise point identifier, outlier identifier, and node confidence score. When a node corresponds to an over-point region, the final result of that node is written back to all the original point cloud points it contains, according to the mapping relationship between the node and the original point cloud points.

[0162] For example, in the above-mentioned urban road vehicle-mounted LiDAR point cloud clustering scenario, the final clustering label, boundary point identifier, noise point identifier, outlier identifier, and node confidence score of each road point cloud node are generated based on multiple point clusters. Compared with methods that only cluster based on fixed neighborhoods or single-scale density relationships, this method can improve the stability and accuracy of clustering results in distant sparse targets, adjacent target boundaries, and occlusion transition regions, avoiding the swallowing of distant small target point clusters by large background areas, misclassification of adjacent target boundary points, and incorrect classification of noise points into normal point clusters.

[0163] In this embodiment, reliable connectivity and node attributes of each node are determined based on the corrected density relation field. Based on the node attributes and reliable connectivity, each node is divided into multiple point clusters. Then, the final clustering result of each node is generated based on the multiple point clusters. In this way, correction operations such as boundary point redistribution, noise point and abnormal point re-determination, reliable connected component splitting, point cluster merging, and dependent path multi-center determination can be performed at the point cluster level to eliminate possible misclassification and omission in the initial clustering and improve the clustering accuracy and robustness.

[0164] like Figure 2 As shown in some embodiments of this application, a density clustering system based on graph neural networks is provided. The system includes a construction module 210, an extraction module 220, an input module 230, a determination module 240, a correction module 250, an execution module 260, and a clustering module 270. Specifically: Module 210 is used to construct a multi-scale hybrid sparse graph based on the 3D point cloud data to be processed; the multi-scale hybrid sparse graph includes multiple nodes and multiple scales; Extraction module 220 is used to extract node features, edge features, scale features, and graph structure reliability input features from multi-scale hybrid sparse graphs; Input module 230 is used to input node features, edge features, scale features and graph structure reliability input features into the multi-task density relation field prediction network to obtain the predicted density relation field; Module 240 is used to determine the set of low-confidence point cloud nodes based on the predicted density relation field; The correction module 250 is used to perform selective and precise correction on the predicted density relation field based on the low-confidence point cloud node set, so as to form the corrected density relation field. Execution module 260 is used to perform adaptive clustering of point clouds based on the corrected density relation field and generate initial cluster labels; Clustering module 270 is used to perform point cluster-level closed-loop correction based on the initial cluster labels and the corrected density relationship field to obtain the final clustering result.

[0165] It should be noted that the density clustering system based on graph neural networks provided in this embodiment is based on the same inventive concept as the density clustering method based on graph neural networks described above. Therefore, the relevant content of the density clustering method based on graph neural networks described above is also applicable to the content of the density clustering system based on graph neural networks. Therefore, it will not be repeated here.

[0166] The system constructs a multi-scale hybrid sparse graph, including multiple nodes and multiple scales, based on the 3D point cloud data to be processed. Node features, edge features, scale features, and graph structure reliability input features are extracted from the multi-scale hybrid sparse graph and input to a multi-task density relation field prediction network to obtain a predicted density relation field. Based on the predicted density relation field, a set of low-confidence point cloud nodes is determined. Selective and precise correction is performed on the predicted density relation field according to the low-confidence point cloud node set, forming a corrected density relation field. Adaptive point cloud clustering is performed based on the corrected density relation field to generate initial cluster labels. Finally, point cluster-level closed-loop correction is performed based on the initial cluster labels and the corrected density relation field to obtain the final clustering result. This approach can adapt to the complex dynamic density distribution of point clouds, significantly improving the accuracy and robustness of point cloud clustering.

[0167] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described density clustering method based on graph neural networks.

[0168] like Figure 3 As shown, Figure 3 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. The electronic device includes: At least one battery; At least one memory; At least one processor; At least one program; The program is stored in memory, and the processor executes at least one program to implement the density clustering method based on graph neural networks described above in this disclosure.

[0169] This electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0170] The electronic devices according to embodiments of this application will now be described in detail.

[0171] The processor 1600 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this disclosure. The memory 1700 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1700 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1700 and is called and executed by the processor 1600 to implement a density clustering method based on a graph neural network according to an embodiment of this disclosure.

[0172] The input / output interface 1800 is used to implement information input and output. The communication interface 1900 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 2000 transmits information between various components of the device (e.g., processor 1600, memory 1700, input / output interface 1800, and communication interface 1900); The processor 1600, memory 1700, input / output interface 1800 and communication interface 1900 are connected to each other within the device via bus 2000.

[0173] This disclosure also provides a storage medium, which is a computer-readable storage medium storing computer-executable instructions for causing a computer to execute the aforementioned density clustering method based on a graph neural network.

[0174] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0175] The embodiments described in this disclosure are for the purpose of more clearly illustrating the technical solutions of this disclosure and do not constitute a limitation on the technical solutions provided by this disclosure. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by this disclosure are also applicable to similar technical problems.

[0176] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this disclosure, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0177] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0178] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0179] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any related variations, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0180] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0181] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0182] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0183] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0184] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0185] The above is a detailed description of the preferred embodiments of this application. However, the embodiments of this application are not limited to the above-described implementation methods. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the embodiments of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of the embodiments of this application.

[0186] The embodiments of this application have been described in detail above with reference to the accompanying drawings. However, this application is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of this application.

Claims

1. A density clustering method based on graph neural networks, characterized in that, The method includes: Based on the 3D point cloud data to be processed, a multi-scale hybrid sparse graph is constructed; the multi-scale hybrid sparse graph includes multiple nodes and multiple scales. Extract node features, edge features, scale features, and graph structure reliability input features from the multi-scale hybrid sparse graph; The node features, edge features, scale features, and graph structure reliability input features are input into a multi-task density relation field prediction network to obtain the predicted density relation field. Based on the predicted density relationship field, determine the set of low-confidence point cloud nodes; Based on the set of low-confidence point cloud nodes, selective precise correction is performed on the predicted density relation field to form a corrected density relation field; Based on the corrected density relation field, perform adaptive clustering of the point cloud to generate initial cluster labels; Based on the initial clustering labels and the corrected density relationship field, point cluster-level closed-loop correction is performed to obtain the final clustering result; The step of inputting the node features, edge features, scale features, and graph structure reliability input features into a multi-task density relation field prediction network to obtain a predicted density relation field includes: The node features, edge features, scale features, and graph structure reliability input features are input into the multi-task density relation field prediction network for encoding and multi-scale neighborhood information aggregation to obtain the hidden representations and density clustering decision parameters of the multiple nodes. Based on the hidden representations of the plurality of nodes and the scale features, the fusion weights of each node in the plurality of nodes with respect to each scale in the plurality of scales are determined, so as to calculate the fusion density of each node in the plurality of nodes. The predicted density relationship field is obtained based on the fusion density of each node in the plurality of nodes and the density clustering decision parameters of the plurality of nodes; The density clustering decision parameters include a graph structure reliability prediction score and prediction uncertainty. The step of determining the low-confidence point cloud node set based on the predicted density relationship field includes: Extract the graph structure reliability prediction score and prediction uncertainty of the multiple nodes from the density clustering decision parameters; Based on the graph structure reliability prediction score and prediction uncertainty of the multiple nodes, the node confidence of the corresponding node is calculated. Based on the multi-scale hybrid sparse graph, the neighborhood distance of each node is determined; Calculate the local estimated density of the corresponding node based on the neighborhood distance of each node; Based on the node confidence of each node, the neighborhood distance, and the local estimated density, low-confidence point cloud nodes are determined from the plurality of nodes to form the low-confidence point cloud node set. The step of performing selective precise correction on the predicted density relation field based on the low-confidence point cloud node set to form a corrected density relation field includes: Based on the set of low-confidence point cloud nodes, determine the low-confidence point cloud nodes and non-low-confidence point cloud nodes in the predicted density relation field; For low-confidence point cloud nodes in the predicted density relation field, perform extended neighborhood re-retrieval to obtain the extended neighborhood of the corresponding node; Based on the extended neighborhood, determine the candidate distance for the corresponding node; Based on the confidence level of the low-confidence point cloud nodes in the predicted density relation field and the candidate distance, corresponding node correction is performed to obtain the first processing result; The prediction results of the non-low confidence point cloud nodes are extracted from the predicted density relation field; Based on the first processing result and the prediction result of the non-low confidence point cloud node, the corrected density relation field is obtained.

2. The density clustering method based on graph neural networks according to claim 1, characterized in that, The construction of a multi-scale hybrid sparse map based on the 3D point cloud data to be processed includes: Each 3D point cloud in the 3D point cloud data to be processed is preprocessed to obtain the corresponding point cloud node and node feature vector, so as to form a node set and a node feature matrix. A scale set is defined on the set of nodes, and each scale in the scale set includes the number of nearest neighbors, the local distance range, and the candidate connection rules; The multi-scale hybrid sparse graph is constructed based on the scale set, the node set, and the node feature matrix.

3. The density clustering method based on graph neural networks according to claim 1, characterized in that, The step of performing adaptive clustering of the point cloud based on the corrected density relation field to generate initial cluster labels includes: The center candidate score, noise probability, boundary probability, and graph structure reliability prediction score of each node are extracted from the corrected density relation field. Based on the center candidate score, the noise probability, the boundary probability, and the graph structure reliability prediction score, center candidate nodes and non-center candidate nodes are selected from the plurality of nodes; Initialize the initial cluster label of the central candidate node; Determine the dependency mapping of the non-central candidate nodes; Based on the dependency mapping of the non-central candidate nodes, determine the node label of the non-central candidate nodes; The initial cluster label is generated based on the initial cluster label of the central candidate node and the node label of the non-central candidate node.

4. The density clustering method based on graph neural networks according to claim 1, characterized in that, The step of performing point cluster-level closed-loop correction based on the initial clustering labels and the corrected density relationship field to obtain the final clustering result includes: Based on the corrected density relationship field, reliable connectivity relationships and node attributes of each node are determined; Based on the node attributes and reliable connection relationships of each node, point clustering is performed on each node to obtain multiple point clusters; Based on the multiple point clusters, the final clustering result for each node is generated.

5. A density clustering system based on graph neural networks, characterized in that, The system includes: A construction module is used to construct a multi-scale hybrid sparse graph based on the 3D point cloud data to be processed; the multi-scale hybrid sparse graph includes multiple nodes and multiple scales; The extraction module is used to extract node features, edge features, scale features, and graph structure reliability input features from the multi-scale hybrid sparse graph; The input module is used to input the node features, edge features, scale features, and graph structure reliability input features into the multi-task density relation field prediction network to obtain the predicted density relation field; The determination module is used to determine the set of low-confidence point cloud nodes based on the predicted density relationship field; The correction module is used to perform selective and precise correction on the predicted density relation field based on the set of low-confidence point cloud nodes to form a corrected density relation field. The execution module is used to perform adaptive clustering of the point cloud based on the corrected density relation field and generate initial cluster labels; The clustering module is used to perform point cluster-level closed-loop correction based on the initial clustering labels and the corrected density relationship field to obtain the final clustering result; The step of inputting the node features, edge features, scale features, and graph structure reliability input features into a multi-task density relation field prediction network to obtain a predicted density relation field includes: The node features, edge features, scale features, and graph structure reliability input features are input into the multi-task density relation field prediction network for encoding and multi-scale neighborhood information aggregation to obtain the hidden representations and density clustering decision parameters of the multiple nodes. Based on the hidden representations of the plurality of nodes and the scale features, the fusion weights of each node in the plurality of nodes with respect to each scale in the plurality of scales are determined, so as to calculate the fusion density of each node in the plurality of nodes. The predicted density relationship field is obtained based on the fusion density of each node in the plurality of nodes and the density clustering decision parameters of the plurality of nodes; The density clustering decision parameters include a graph structure reliability prediction score and prediction uncertainty. The step of determining the low-confidence point cloud node set based on the predicted density relationship field includes: Extract the graph structure reliability prediction score and prediction uncertainty of the multiple nodes from the density clustering decision parameters; Based on the graph structure reliability prediction score and prediction uncertainty of the multiple nodes, the node confidence of the corresponding node is calculated. Based on the multi-scale hybrid sparse graph, the neighborhood distance of each node is determined; Calculate the local estimated density of the corresponding node based on the neighborhood distance of each node; Based on the node confidence of each node, the neighborhood distance, and the local estimated density, low-confidence point cloud nodes are determined from the plurality of nodes to form the low-confidence point cloud node set. The step of performing selective precise correction on the predicted density relation field based on the low-confidence point cloud node set to form a corrected density relation field includes: Based on the set of low-confidence point cloud nodes, determine the low-confidence point cloud nodes and non-low-confidence point cloud nodes in the predicted density relation field; For low-confidence point cloud nodes in the predicted density relation field, perform extended neighborhood re-retrieval to obtain the extended neighborhood of the corresponding node; Based on the extended neighborhood, determine the candidate distance for the corresponding node; Based on the confidence level of the low-confidence point cloud nodes in the predicted density relation field and the candidate distance, corresponding node correction is performed to obtain the first processing result; The prediction results of the non-low confidence point cloud nodes are extracted from the predicted density relation field; Based on the first processing result and the prediction result of the non-low confidence point cloud node, the corrected density relation field is obtained.

6. An electronic device, characterized in that, It includes at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, which, when executed by the at least one control processor, enable the at least one control processor to perform a density clustering method based on a graph neural network as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform a density clustering method based on a graph neural network as described in any one of claims 1 to 4.

Images