A bidirectional dynamic constraint euclidean-db scan point cloud clustering segmentation method and device

Abstract

The application discloses a bidirectional dynamic constraint Euclidean-DBSCAN point cloud clustering and segmentation method and device, which comprises the following steps: acquiring original three-dimensional point clouds around a low-speed unmanned vehicle, uniformly downsampling and filtering the original three-dimensional point clouds through a voxel grid filter with adaptive resolution parameters to obtain an effective detection point cloud set; outputting a first-level rough clustering cluster set for the effective detection point cloud set; initializing a final target point cloud cluster set for storing a final clustering result, and the initial state is an empty set; serially traversing each first-level rough clustering cluster in the first-level rough clustering cluster set, and performing bidirectional dynamic constraint decision; solving the geometric center of each point cloud cluster in the final target point cloud cluster set F by using principal component analysis, performing multi-angle scanning calculation of the best heading angle in a preset yaw angle search space, accurately solving the length, width and height of the directional enclosing sum, and publishing the result after coordinate system mirror mapping conversion. The application mainly applies to the technical field of vehicles.

Description

Technical Field

[0001] This invention relates to the field of vehicle technology, specifically to a bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation method and apparatus. Background Technology

[0002] In the field of low-speed autonomous driving (such as park logistics and community cleaning), unmanned vehicles frequently operate in high-density near-field environments such as narrow passages and mixed human-machine traffic. Due to the limited operating space, the perception system faces more stringent requirements regarding the integrity of the edge contours of obstacles around the low-speed vehicle (to avoid collision hazards caused by target contour shrinkage) and the real-time processing capabilities of the algorithm on a low-power embedded platform. Non-deep learning point cloud clustering techniques applied to vehicle environmental perception mainly rely on single-stage global retrieval schemes, including pure global Euclidean clustering and pure global DBSCAN clustering. These two schemes not only fail to solve the edge feature loss, encirclement, and shrinkage problems caused by density clustering, but also incur huge unnecessary computational overhead. Therefore, how to solve the edge feature loss caused by density clustering and improve the utilization rate of computing power is a technical problem that urgently needs to be studied in the industry. Summary of the Invention

[0003] This invention provides a bidirectional dynamic constraint Euclidean-DBSCAN point cloud clustering and segmentation method and apparatus to solve the technical problem of poor adaptation effect in the prior art, and at least provides a beneficial option or creates conditions.

[0004] This invention provides a bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation method, including: Step 1, acquiring the original 3D point cloud surrounding a low-speed unmanned vehicle. Adaptive resolution parameters using a voxel grid filter. Uniform downsampling is performed, and spatial pass-through filtering is applied based on preset vertical, horizontal, and height axis filtering intervals to extract the effective detection point set within the near-field region of interest. ;

[0005] Step 2: For the valid detection point set Construct a global first search space and apply large distance tolerance. Execute the global first-level Euclidean coarse clustering function Produces a set of first-level coarse clusters. And initialize a final target point cloud cluster set to store the final clustering results; The initial state is an empty set;

[0006] Step 3: Serially traverse each first-level coarse cluster in the first-level coarse cluster set. The two-way dynamic constraint decision-making is performed, which includes: the first sub-step and the second sub-step.

[0007] The first sub-step includes: extracting the current first-level coarse clusters. Point cloud quantity ;

[0008] If the number of point clouds < This triggers a positive skip constraint, causing the first-level coarse clusters to be... Save directly to the final target point cloud cluster set F; if ≥ Then proceed to the second sub-step, where, The preset adaptive traffic splitting threshold;

[0009] The second sub-step includes: extracting the current first-level coarse clusters. The internal points construct locally bounded Utilizing a narrow sensing radius and the minimum number of neighborhood points of the core point Execute local DBSCAN clustering function Output the set of second-level sub-clusters ; where, definition And based on the M value, perform the following feedback decision in reverse: If Then discard the first-level coarse cluster. Store the M secondary sub-clusters into the final target point cloud cluster set F; if M=1, discard the secondary clustering results. Revert and fully inherit the first-level coarse cluster Store the final target point cloud cluster set F; if M=0, then determine the first-level coarse cluster. Invalid; it is rejected from being included in the final target point cloud cluster set F.

[0010] Step 4: Use principal component analysis to solve for the geometric center of each point cloud cluster in the final target point cloud cluster set F. Perform multi-angle scanning within the preset yaw angle search space to calculate the optimal heading angle, accurately calculate the length, width, and height of the directed enclosing sum, and then perform coordinate system mirror mapping transformation before publishing it externally.

[0011] Furthermore, in step 1, the vertical filtering interval is... The horizontal filtering interval satisfies

[0012] The height axis filtering range is: removing items that meet the following conditions. ground point or High altitude point.

[0013] Furthermore, in step 2, the global first-order Euclidean coarse clustering function By utilizing the large radius connectivity property, all sparse edge points extending around the object are completely enclosed in each independent first-order coarse cluster. middle.

[0014] Furthermore, in the first sub-step, during the positive skip constraint, when < If the current primary coarse cluster is determined to be a low-density small target, its secondary search is waived to avoid the low-density small target being filtered out as a whole.

[0015] Furthermore, in the second sub-step, when M=1, it is determined that the current first-level coarse cluster is continuous and has no multi-objective adhesion, the reverse constraint mechanism is activated, and the current second-level clustering result is discarded. Completely inherits the first-level coarse cluster. This is to ensure that its original three-dimensional physical dimensions are not distorted.

[0016] Furthermore, the adaptive resolution parameters of the voxel grid filter The narrow sensing radius of the local DBSCAN And the large distance tolerance of the global first-level Euclidean coarse clustering Satisfying the spatial scale hierarchical progression inequality: < < .

[0017] Furthermore, in step 4, the method of external publication is to package and publish the three-dimensional oriented enclosure and state of the obstacle through a ROS topic.

[0018] On the other hand, a bidirectional dynamic constraint Euclidean-DBSCAN point cloud clustering and segmentation device is provided, including: an acquisition and filtering module, a output construction module, a decision execution module and a fitting output module;

[0019] The acquisition and filtering module is used to: acquire the original 3D point cloud surrounding the low-speed unmanned vehicle. Adaptive resolution parameters using a voxel grid filter. Uniform downsampling is performed, and spatial pass-through filtering is applied based on preset vertical, horizontal, and height axis filtering intervals to extract the effective detection point set within the near-field region of interest. ;

[0020] The output module is used for: targeting the effective detection point set Construct a global first search space and apply large distance tolerance. Execute the global first-level Euclidean coarse clustering function Produces a set of first-level coarse clusters. And initialize a final target point cloud cluster set to store the final clustering results; The initial state is an empty set;

[0021] The decision execution module is used to: serially traverse each first-level coarse cluster in the first-level coarse cluster set. The two-way dynamic constraint decision-making is performed, which includes: the first sub-step and the second sub-step.

[0022] The first sub-step includes: extracting the current first-level coarse clusters. Point cloud quantity ;

[0023] If the number of point clouds < This triggers a positive skip constraint, causing the first-level coarse clusters to be... Save directly to the final target point cloud cluster set F; if ≥ Then proceed to the second sub-step, where, The preset adaptive traffic splitting threshold;

[0024] The second sub-step includes: extracting the current first-level coarse clusters. The internal points construct locally bounded Utilizing a narrow sensing radius and the minimum number of neighborhood points of the core point Execute local DBSCAN clustering function Output the set of second-level sub-clusters ; where, definition And based on the M value, perform the following feedback decision in reverse: If Then discard the first-level coarse cluster. Store the M secondary sub-clusters into the final target point cloud cluster set F; if M=1, discard the secondary clustering results. Revert and fully inherit the first-level coarse cluster Store the final target point cloud cluster set F; if M=0, then determine the first-level coarse cluster. Invalid; it is rejected from being included in the final target point cloud cluster set F.

[0025] The fitting output module is used to: use principal component analysis to solve for the geometric center of each point cloud cluster in the final target point cloud cluster set F, perform multi-angle scanning to calculate the optimal heading angle in the preset yaw angle search space, accurately calculate the length, width and height of the directed enclosing sum, and publish it after performing coordinate system mirror mapping transformation.

[0026] Furthermore, in the acquisition and filtering module, the vertical filtering interval is... The horizontal filtering interval satisfies The height axis filtering range is: removing items that meet the following conditions. ground point or High altitude point.

[0027] 10. A bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation device according to claim 8, characterized in that, in the output construction module, the global first-level Euclidean coarse clustering function... By utilizing the large radius connectivity property, all sparse edge points extending around the object are completely enclosed in each independent first-order coarse cluster. middle.

[0028] This invention offers at least the following advantages: The method of this technical solution effectively overcomes the edge feature loss problem caused by pure density clustering through first-level Euclidean coarse clustering and second-level local DBSCAN fine clustering, significantly improving the dimensional accuracy of obstacle perception. Furthermore, this invention performs local DBSCAN clustering only on first-level coarse clusters that meet the quantity threshold condition in the second sub-step of step 3, and the clustering space is strictly limited to the bounded range within a single first-level cluster. This mechanism shrinks the computationally dense DBSCAN operation from the global space to a local subspace, significantly reducing the data scale of neighborhood search and greatly decreasing the overall computational complexity. This allows for stable operation on cost-constrained embedded computing platforms for low-speed autonomous vehicles, meeting the requirements of high-frequency real-time perception. Through the positive skip constraint mechanism in the first sub-step, such low-density small targets are identified early in the clustering process and directly saved to the final result set, ensuring the effective detection of small obstacles by the perception system and improving the perception integrity of low-speed autonomous vehicles in complex environments. On the other hand, this invention also provides a corresponding system, the advantages of which are similar to the method and will not be repeated here. This invention is mainly applicable to the field of vehicle technology. Attached Figure Description

[0029] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0030] Figure 1 This is a flowchart of the steps of the bidirectional dynamic constraint Euclidean-DBSCAN point cloud clustering and segmentation method;

[0031] Figure 2 This is a flowchart of the Euclidean-DBSCAN point cloud clustering and segmentation device with bidirectional dynamic constraints.

[0032] Figure 3 This is a schematic diagram of the system structure of the bidirectional dynamic constraint Euclidean-DBSCAN point cloud clustering and segmentation device;

[0033] Figure 4 This is a framework diagram of a point cloud clustering and segmentation device according to another embodiment. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0035] It should be noted that although functional modules are divided in the system diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the system or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0036] The following is an explanation of the necessary terms:

[0037] Euclidean Clustering: A clustering algorithm based on spatial Euclidean distance, which divides spatial points into the same cluster when the distance between them is less than a set distance tolerance.

[0038] DBSCAN (Density-Based Spatial Clustering of Applications with Noise): A density-based spatial clustering algorithm that identifies high-density point clusters by relying on neighborhood radius (ϵ) and minimum number of points (MinPts) and can isolate and filter low-density noise points.

[0039] Bidirectional Dynamic Constraint: This refers to a feedback mechanism established between two-level clustering algorithms. The positive constraint uses the size index of the preceding cluster to dynamically constrain the starting conditions of the subsequent algorithm; the negative constraint uses the state of the subsequent density cluster to dynamically determine whether the boundary set of the preceding cluster is backed up and inherited.

[0040] KdTree (K-Dimensional Tree): A tree-like data structure for partitioning points in a multidimensional space, often used for efficient nearest neighbor search of point cloud data.

[0041] Oriented Bounding Box (OBB): A 3D bounding box with a 3D orientation angle (Yaw) used to accurately represent the physical size and spatial orientation of obstacles.

[0042] ROI (Region of Interest): In 3D point cloud processing, to reduce the amount of computation of invalid data, a specific spatial range containing the target to be detected is pre-defined or extracted (such as a field of view truncated by setting thresholds for length, width and height).

[0043] Voxel grid filter: A commonly used point cloud downsampling method. By dividing the 3D space into several tiny cubic grids (voxels) and replacing all points in the grid with the center point of each grid, the number of points in the point cloud is significantly reduced while preserving the macroscopic geometric features of the space.

[0044] Principal Component Analysis (PCA) is a statistical feature extraction algorithm based on orthogonal transformation. In 3D object detection, PCA is often used to extract the principal orientation of the spatial distribution of point cloud clusters, which is then used to accurately solve for and fit the physical orientation angle of the object.

[0045] ROS (Robot Operating System): An open-source framework that provides underlying hardware abstraction and control for robot software development. Its nodes typically use specific topics for continuous data publishing and subscription.

[0046] Research reveals that in existing technologies for low-speed autonomous driving, non-deep learning point cloud clustering techniques for in-vehicle environmental perception primarily rely on single-stage global retrieval schemes. These schemes are typically pure global Euclidean clustering or pure global DBSCAN clustering. The pure global Euclidean clustering scheme relies on a single large distance tolerance to construct a KdTree for partitioning within the global region. While this scheme ensures no loss of macroscopic contours, it easily leads to blurred target boundaries and spatial overlap in close-range multi-target groups. The pure global DBSCAN clustering scheme relies on a global density threshold to filter low-density points. Although it has noise removal capabilities, its removal of sparse edge points results in target detection sizes smaller than the actual physical contours, making it difficult to meet the close-range obstacle avoidance requirements of low-speed vehicles; furthermore, the global O(NlogN) computational overhead cannot achieve high-frequency real-time operation on low-performance chips in low-speed vehicles. Existing point cloud clustering and segmentation methods generally employ unidirectional serial logic. This approach not only fails to solve the edge feature loss, encirclement, and shrinkage problems caused by density clustering but also incurs significant unnecessary computational overhead. Therefore, how to solve the problems of edge feature loss, encirclement and shrinkage caused by density clustering and improve computing power utilization efficiency are technical issues that urgently need to be studied in the industry.

[0047] refer to Figure 1 and Figure 2 , Figure 1This is a flowchart of the steps of the bidirectional dynamic constraint Euclidean-DBSCAN point cloud clustering and segmentation method. Figure 2 This is a flowchart illustrating the process of the Euclidean-DBSCAN point cloud clustering and segmentation device with bidirectional dynamic constraints.

[0048] To address the problems existing in the prior art, this application discloses a bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation method, which can be executed by a smart device. When executed by the smart device, the steps include:

[0049] Step 1: Obtain the original 3D point cloud around the low-speed unmanned vehicle. Adaptive resolution parameters using a voxel grid filter. Uniform downsampling is performed, and spatial pass-through filtering is applied based on preset vertical, horizontal, and height axis filtering intervals to extract the effective detection point set within the near-field region of interest. .

[0050] In step 1, after the low-speed autonomous vehicle starts, the onboard LiDAR sensor begins to collect environmental point cloud data around the vehicle in real time, acquiring the original 3D point cloud at a predetermined sampling frequency. Subsequently, the system calls the VoxelGridFilter from the Point Cloud Library (PCL) to adapt the resolution parameters. For the original 3D point cloud The data is uniformly downsampled. In a specific implementation, for example, when the autonomous vehicle is traveling at a low speed (e.g., 5 km / h) in a structured road scenario, the system adaptively downsamples the data. The setting is 0.1m to preserve sufficient environmental detail; when the autonomous vehicle is driving in an open area, the system will... The depth was adjusted to 0.2m to reduce computational load. After downsampling, the system performs spatial pass-through filtering on the 3D point cloud based on preset vertical, horizontal, and height axis filtering intervals to extract the effective detection point cloud set within the near-field region of interest. .

[0051] In some further specific embodiments, in step 1, the vertical filtering interval is: The horizontal filtering interval satisfies The height axis filtering range is: removing items that meet the following conditions. ground point or High-altitude points. Among them, Represented as the vertical nearest boundary point of the near-field region of interest. This is represented by the furthest vertical boundary point of the region of interest upon entry. This is represented as the left horizontal boundary point of the near-field region of interest. Represented as the right-hand boundary point of the near-field region of interest; This indicates the boundary point with the lowest elevation. This indicates the highest boundary point.

[0052] In this system, the x-axis (longitudinal direction) typically refers to the vehicle's forward and backward direction, with the front of the vehicle being positive and the rear being negative. The y-axis (lateral direction) typically refers to the vehicle's left and right direction, with the left side being positive and the right side being negative. The z-axis (vertical direction) typically refers to the vehicle's vertical direction, with the top being positive and the bottom being negative.

[0053] In a specific example, the system establishes a three-dimensional filtering interval based on a pre-calibrated near-field sensing range: the longitudinal filtering interval is... The horizontal filtering interval satisfies The height axis filtering range is: removing items that meet the following conditions. ground point or High-altitude points. Through the above filtering operations, the effective detection points cluster in the near-field space surrounding the low-speed unmanned vehicle are finally obtained. .

[0054] Step 2: For the valid detection point set Construct a global first search space and apply large distance tolerance. Execute the global first-level Euclidean coarse clustering function Produces a set of first-level coarse clusters. And initialize a final target point cloud cluster set to store the final clustering results; The initial state is an empty set.

[0055] In some further specific embodiments, in step 2, the global first-order Euclidean coarse clustering function By utilizing the large radius connectivity property, all sparse edge points extending around the object are completely enclosed in each independent first-order coarse cluster. middle.

[0056] For the effective detection point set A global KdTree index structure is constructed, establishing the global first search space. Subsequently, the system calls the first-level Euclidean clustering function. Set large distance tolerance =0.8, perform global coarse clustering. This clustering process uses a radius search mechanism based on KdTree, traversing all point cloud data and clustering points with a spatial distance less than 0.8. Points are grouped into the same cluster. Due to the large distance tolerance, the system can completely encompass the sparse edge points extending outwards from the object (such as low-reflection areas like vehicle rearview mirrors and pedestrian arms) into the corresponding first-level coarse clusters. Finally, the system outputs a set of first-level coarse clusters. The final target point cloud cluster set F is initialized as an empty set to store the effective clustering results after subsequent bidirectional dynamic constraint decision-making.

[0057] Taking the clustering of a standard sedan as an example: the main body of the sedan has a high point cloud density, while the side mirrors, due to their small size and low surface reflectivity, only produce a small number of sparse edge points, with their spatial location approximately 0.3m to 0.5m from the main body of the sedan. Because... =0.8 is greater than the distance from the sparse points of the rearview mirror to the main body of the vehicle. When performing coarse clustering, it can connect the sparse edge points of the rearview mirror to the point cloud of the main body of the vehicle into the same cluster. Similarly, the point clouds of pedestrians' outstretched parts such as arms and backpacks are also completely encapsulated within the corresponding first-level coarse clusters of pedestrians.

[0058] Of course, in a preferred embodiment... The value is also adaptively adjusted based on the actual scanning beam and point cloud density of the lidar. For a 16-line lidar (where the point cloud is relatively sparse). Set to a relatively large 1.0m; for a 64-line lidar (where the point cloud is relatively dense). It is set to a smaller value of 0.6m. However, it meets the requirements regardless of the configuration. It is significantly larger than the spatial distance between the main body of the object and the extended edge, thus ensuring the integrity of the envelope in the coarse clustering stage.

[0059] Step 3: Serially traverse each first-level coarse cluster in the first-level coarse cluster set. The two-way dynamic constraint decision-making is performed, which includes the first sub-step and the second sub-step.

[0060] The first sub-step includes: extracting the current first-level coarse clusters. Point cloud quantity ;

[0061] If the number of point clouds < This triggers a positive skip constraint, causing the first-level coarse clusters to be... Save directly to the final target point cloud cluster set F; if ≥ Then proceed to the second sub-step, where, The preset adaptive traffic splitting threshold;

[0062] The second sub-step includes: extracting the current first-level coarse clusters. The internal points construct locally bounded Utilizing a narrow sensing radius and the minimum number of neighborhood points of the core point Execute local DBSCAN clustering function Output the set of second-level sub-clusters ; where, definition And based on the M value, perform the following feedback decision in reverse: If Then discard the first-level coarse cluster. Store the M secondary sub-clusters into the final target point cloud cluster set F; if M=1, discard the secondary clustering results. Revert and fully inherit the first-level coarse cluster Store the final target point cloud cluster set F; if M=0, then determine the first-level coarse cluster. Invalid; refuse to incorporate it into the final target point cloud cluster set F.

[0063] In step 3, each cluster in the first-level coarse clustering cluster set is traversed sequentially according to the cluster number order. A two-way dynamic constraint decision is performed on it. The two-way constraint decision includes a first sub-step and a second sub-step.

[0064] The first sub-step mainly involves a positive skip constraint based on the number of point clouds. The second-level clustering is initiated using this cardinality dynamic constraint, with the following mapping rules:

[0065]

[0066] in This is a preset adaptive splitting threshold. If the value is less than this threshold, it is determined to be a low-density, small target, and forcibly performing density partitioning will cause it to be filtered out as a whole. The system triggers a positive skip constraint here, exempting it from secondary retrieval and directly pushing it into the final target point cloud cluster set F. Conversely, if the condition is met, this macroscopic boundary will serve as a secondary external bounded constraint, proceeding to the second sub-step.

[0067] The first sub-step is explained in more detail below. In the first sub-step, the current first-level coarse cluster is counted. Number of three-dimensional points included In one embodiment, an adaptive shunt threshold is set. If the number of points in the current cluster is less than 30, the system determines that the cluster is a low-density, small target (such as a small obstacle in the distance or sparse vegetation). In this case, a positive skip constraint is triggered—the system skips the secondary DBSCAN clustering retrieval for that cluster and directly saves the cluster to the final target point cloud cluster set F. This mechanism prevents low-density, small targets from being filtered out and discarded entirely during density clustering because they do not meet the core point condition. The system then proceeds to the second sub-step.

[0068] The second sub-step involves inverse boundary correction constraints based on connected states. Specifically, for the first-level coarse clusters entering this stage, the system removes them from the global full-scale point cloud retrieval space and only extracts... The internal points construct locally bounded Based on this, a narrow sensing radius is utilized. and the minimum number of neighborhood points of the core point Execute local DBSCAN clustering function Output the set of second-level sub-clusters :

[0069]

[0070] in This represents the number of secondary sub-clusters formed. Based on the M value, the following three different modalities of feedback decision-making are executed in reverse. To determine the final elements to be included in the final target point cloud cluster set F:

[0071]

[0072] Modality A (precise segmentation, when M>1): Determines that there is spatial adhesion of multiple targets within the current first-level coarse cluster. In this case, the second-level clustering result takes effect, discarding the first-level coarse cluster and replacing it with a cluster formed by the second-level clustering. Each sub-cluster is replaced and stored in the final target point cloud cluster set. .

[0073] Modal B (marginal regression, when M=1): Determines that the current first-level coarse cluster is internally continuous and has no multi-objective adhesion. To avoid the second-level density clustering removing sparse points on the periphery due to the core point threshold, the system activates a reverse constraint mechanism: completely discarding the current second-level clustering results. Re-adopt, revert, and fully inherit the first-level coarse cluster. To ensure that its original 3D physical dimensions are not distorted, it is then added to the final target point cloud cluster set. .

[0074] Modality C (noise removal, when M=0): Determines that no effective core points can be formed within the current first-level coarse cluster, classifying it as spatial random noise or a false alarm. The system directly determines the first-level coarse cluster to be invalid and refuses to include it in the final target point cloud cluster set. (The output is an empty set) ).

[0075] The second sub-step is described in detail as follows: In the second sub-step, the system first extracts the current first-level coarse cluster from the global point cloud space. All points contained within the system are used to construct a locally bounded KdTree. Subsequently, the system sets a narrow sensing radius. Minimum number of neighborhood points of the core point It performs local DBSCAN clustering. This process only performs neighborhood searches within the local space of the current cluster, significantly reducing computational complexity.

[0076] After execution, the system obtains a set of secondary sub-clusters. Where M is the number of sub-clusters. The system performs three types of feedback decisions based on the value of M:

[0077] When M > 1 (Modal A), it is determined that there is multi-objective spatial adhesion within the current first-level coarse cluster, and the second-level clustering successfully achieves accurate segmentation between different objects. At this point, the original first-level coarse cluster is discarded. All M secondary sub-clusters are stored in the final target point cloud cluster set F.

[0078] When M=1 (Modal B), the current first-level coarse cluster is determined to be continuous and without multiple targets adhering to each other. To avoid target size distortion caused by DBSCAN excluding sparse outer points due to the core point threshold, the reverse boundary correction constraint is activated, and the current second-level clustering result is discarded. Revert and fully inherit the first-level coarse cluster The entire point cloud is stored in the final target point cloud cluster set F. This mechanism ensures the integrity of the object's original three-dimensional physical dimensions.

[0079] When M=0 (modality C), it is determined that no valid core point can be formed within the current first-level coarse cluster, and it belongs to spatial random noise or false alarm targets. The cluster is directly determined to be invalid and is rejected from being included in the final target point cloud cluster set F.

[0080] In some further specific embodiments, in the first sub-step, during the positive skip constraint, when < If the current primary coarse cluster is identified as a low-density, small target, its secondary search is waived to prevent the low-density, small target from being filtered out as a whole. In one specific embodiment, the system presets an adaptive traffic splitting threshold. When the system traverses to a certain first-level coarse cluster... At that time, the number of point clouds was counted. Since 12 is less than 25, the cluster is determined to be a low-density, small target. In real-world scenarios, this cluster might correspond to obstacles such as: pedestrians at a distance, small traffic cones, small obstacles scattered on the road, or locally sparse point clouds of bicycles / motorcycle riders in front of a vehicle. In this case, after triggering the positive skip constraint, the local DBSCAN clustering in the second sub-step is not performed on this cluster; instead, the cluster is directly... The complete point cloud is stored in the final target point cloud cluster set F. This decision logic is based on the following physical understanding: low-density small targets have a limited number of points and do not have the conditions for multiple targets to stick together inside, so there is no need to subdivide them through density clustering; on the contrary, if DBSCAN processing is forcibly performed on them, their sparse point cloud distribution cannot meet the neighborhood conditions of the core points, and they are very likely to be judged as noise and filtered out, resulting in the target being missed.

[0081] By employing a positive skip constraint mechanism, such targets are identified and directly retained early in the clustering process, fundamentally eliminating the risk of missed detection and improving the recall rate of low-speed autonomous vehicles for perceptual minor obstacles. Furthermore, for point cloud data smaller than [a certain number], [this method is effective]. Clusters can be processed without constructing a local KdTree or performing neighborhood search and density determination using local DBSCAN, directly skipping secondary retrieval. This not only saves computational resources but also reduces unnecessary algorithm branch decisions, further improving the overall system efficiency. Furthermore, the forward skip constraint avoids false filtering by DBSCAN at the source level, while the edge regression mechanism prevents DBSCAN from removing edge points when dealing with a large number of clusters. These two mechanisms work together at different scales to ensure the integrity of the target point cloud.

[0082] In some further specific embodiments, in the second sub-step, when M=1, it is determined that the current first-level coarse cluster is continuous and has no multi-target adhesion, the reverse constraint mechanism is activated, and the current second-level clustering result is discarded. Completely inherits the first-level coarse cluster. This is to ensure that its original three-dimensional physical dimensions are not distorted.

[0083] In one specific embodiment, the system performs coarse clustering at a certain level. After performing local DBSCAN clustering, a set of secondary sub-clusters is obtained. According to statistics, This means that the local DBSCAN only divided the cluster into a sub-cluster within that cluster. At this point, the system determines... In terms of space, it is a single, continuous target, and there is no situation where multiple targets are stuck together.

[0084] However, it should be noted that the DBSCAN algorithm has a core threshold. In fact, it only includes The core points that satisfy the density condition and their density-reachable points, while Sparse outreach points at the edges were marked as boundary points or noise points and discarded during the DBSCAN process. Therefore, The physical dimensions represented are compared to the actual dimensions. It's a shrunken version.

[0085] In this case, the system activates the reverse boundary correction constraint: completely discarding it. The clustering results are reverted to the state before the local DBSCAN was executed, and the original first-level coarse clusters are reverted to their original state. The complete target point cloud cluster set F is stored. Taking a specific numerical example: the first-level coarse cluster of a car target. It contains 1200 points, with a directed bounding area and dimensions of 4.8m * 1.9m * 1.5m (including rearview mirrors); after local DBSCAN... It contains only 980 points, with a reduced enclosure and size of 4.8m*1.8m*1.5m (rearview mirror points are removed). Through an inverse constraint mechanism, the system's final output is... The full dimensions ensure the accuracy of obstacle size perception.

[0086] This specific embodiment ensures that the final output target size is strictly equal to the complete size encompassed by the Euclidean coarse cluster by activating the reverse boundary correction constraint when M=1. This allows the downstream path planning and obstacle avoidance modules to make safety decisions based on the actual physical size of obstacles, effectively reducing collision risks. It fundamentally solves the industry pain point of underestimating the target physical size caused by DBSCAN density clustering.

[0087] Step 4: Use principal component analysis to solve for the geometric center of each point cloud cluster in the final target point cloud cluster set F. Perform multi-angle scanning within the preset yaw angle search space to calculate the optimal heading angle, accurately calculate the length, width, and height of the directed enclosing sum, and then perform coordinate system mirror mapping transformation before publishing it externally.

[0088] In step 4, the system traverses each point cloud cluster in the final target point cloud cluster set F, calls the Principal Component Analysis (PCA) algorithm to calculate the covariance matrix of each cluster, and solves for eigenvalues ​​and eigenvectors to determine the geometric center and principal direction of the point cloud cluster. Subsequently, the system performs multi-angle scanning within a preset yaw angle search space with a fixed step size, calculates the bounding sum and area at each angle, selects the angle with the smallest area as the optimal heading angle, and accurately calculates the length, width, and height of the directed bounding sum. Finally, the system performs coordinate system mirroring transformation, and publishes the three-dimensional directed bounding sum and state information of the obstacle through ROS topics in a custom message format for use by the downstream path planning module and obstacle avoidance control module.

[0089] This invention effectively overcomes the edge feature loss problem caused by pure density clustering by using first-level Euclidean coarse clustering and second-level local DBSCAN fine clustering, significantly improving the dimensional accuracy of obstacle perception. Furthermore, this invention performs local DBSCAN clustering only on first-level coarse clusters that meet the quantity threshold condition in the second sub-step of step 3, and the clustering space is strictly limited to the bounded range within a single first-level cluster. This mechanism shrinks the computationally dense DBSCAN operation from the global space to a local subspace, significantly reducing the data scale of neighborhood search and greatly decreasing the overall computational complexity. This allows for stable operation on cost-constrained embedded computing platforms for low-speed autonomous vehicles, meeting the requirements of high-frequency real-time perception. Through the positive skip constraint mechanism in the first sub-step, such low-density small targets are identified early in the clustering process and directly saved to the final result set, ensuring the effective detection of small obstacles by the perception system and improving the perception integrity of low-speed autonomous vehicles in complex environments.

[0090] In some further specific embodiments, the adaptive resolution parameters of the voxel grid filter The narrow sensing radius of the local DBSCAN And the large distance tolerance of the global first-level Euclidean coarse clustering Satisfying the spatial scale hierarchical progression inequality: < < .

[0091] In one specific embodiment, the system configures three key spatial scale parameters according to the following values: voxel grid resolution. The narrow sensing radius of the local DBSCAN is 0.08m. With a distance tolerance of 0.15m, the global first-level Euclidean coarse clustering exhibits a large range tolerance. The value is 0.8m. The above parameters are clearly satisfied. < < The physical meaning of this parameter constraint is as follows: the voxel downsampling has the lowest resolution, ensuring that the point cloud is uniformly sparsified without destroying the basic geometric structure of the object; the local DBSCAN has a perception radius at an intermediate scale, which can perform fine density connectivity determination within a single target while avoiding cross-target adhesion; the first-order Euclidean coarse clustering has the highest distance tolerance, which can fully connect the main body and sparse edge parts of the same object to form a complete macroscopic envelope.

[0092] By imposing such constraints on the parameters, the hierarchical structure and functional synergy of the multi-scale clustering architecture are mathematically guaranteed. The three spatial scales are arranged sequentially from coarse to fine and from small to large, each undertaking a different functional role: Responsible for uniform sparsity at the data level. Responsible for determining the fine density within the target. It is responsible for the complete envelope connectivity at the macro level. The three elements do not overlap and are progressively layered, ensuring the mathematical consistency and effectiveness of the entire clustering process.

[0093] refer to Figure 3 , Figure 3 This is a schematic diagram of the system structure of a bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation device. Furthermore, this invention also provides a bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation device, comprising: an acquisition and filtering module, a output construction module, a decision execution module, and a fitting output module.

[0094] The acquisition and filtering module is used to: acquire the original 3D point cloud surrounding the low-speed unmanned vehicle. Adaptive resolution parameters using a voxel grid filter. Uniform downsampling is performed, and spatial pass-through filtering is applied based on preset vertical, horizontal, and height axis filtering intervals to extract the effective detection point set within the near-field region of interest. .

[0095] After the low-speed autonomous vehicle starts, the acquisition and filtering module requests the onboard LiDAR sensor to begin real-time acquisition of environmental point cloud data around the vehicle, obtaining the original 3D point cloud at a predetermined sampling frequency. Subsequently, the filtering module calls the VoxelGridFilter from the Point Cloud Library (PCL) to adapt the resolution parameters. For the original 3D point cloud The data is uniformly downsampled. In a specific implementation, for example, when the autonomous vehicle is traveling at a low speed (e.g., 5 km / h) in a structured road scenario, the system adaptively downsamples the data. The setting is 0.1m to preserve sufficient environmental detail; when the autonomous vehicle is driving in an open area, the system will... The value was adjusted to 0.2m to reduce computational load. After downsampling, the acquisition and filtering module performed spatial pass-through filtering on the 3D point cloud based on preset vertical, horizontal, and height axis filtering intervals to extract the effective detection point cloud set within the near-field region of interest. .

[0096] In some further specific embodiments, in the acquisition filtering module, the vertical filtering interval is: The horizontal filtering interval satisfies The height axis filtering range is: removing items that meet the following conditions. ground point or High-altitude points. Among them, Represented as the vertical nearest boundary point of the near-field region of interest. This is represented by the furthest vertical boundary point of the region of interest upon entry. This is represented as the left horizontal boundary point of the near-field region of interest. Represented as the right-hand boundary point of the near-field region of interest; This indicates the boundary point with the lowest elevation. This indicates the highest boundary point.

[0097] In this system, the x-axis (longitudinal direction) typically refers to the vehicle's forward and backward direction, with the front of the vehicle being positive and the rear being negative. The y-axis (lateral direction) typically refers to the vehicle's left and right direction, with the left side being positive and the right side being negative. The z-axis (vertical direction) typically refers to the vehicle's vertical direction, with the top being positive and the bottom being negative.

[0098] In a specific example, the acquisition filtering module establishes a three-dimensional filtering interval based on a pre-calibrated near-field sensing range: the vertical filtering interval is... The horizontal filtering interval satisfies The height axis filtering range is: removing items that meet the following conditions. ground point or High-altitude points. Through the above filtering operations, the effective detection points cluster in the near-field space surrounding the low-speed unmanned vehicle are finally obtained. .

[0099] The output module is used for: targeting the effective detection point set Construct a global first search space and apply large distance tolerance. Execute the global first-level Euclidean coarse clustering function Produces a set of first-level coarse clusters. And initialize a final target point cloud cluster set to store the final clustering results; The initial state is an empty set.

[0100] In some further specific embodiments, in the output construction module, the global first-level Euclidean coarse clustering function By utilizing the large radius connectivity property, all sparse edge points extending around the object are completely enclosed in each independent first-order coarse cluster. middle.

[0101] For the effective detection point set A global KdTree index structure is constructed to establish the global first retrieval space. Subsequently, the output module calls the first-level Euclidean clustering function. Set large distance tolerance =0.8, perform global coarse clustering. This clustering process uses a radius search mechanism based on KdTree, traversing all point cloud data and clustering points with a spatial distance less than 0.8. Points are grouped into the same cluster. Due to the large distance tolerance, the output module can completely encompass the sparse edge points extending outwards from the object (such as low-reflection areas like vehicle rearview mirrors and pedestrian arms) into the corresponding first-level coarse clusters. Finally, the system outputs a set of first-level coarse clusters. The final target point cloud cluster set F is initialized as an empty set to store the effective clustering results after subsequent bidirectional dynamic constraint decision-making.

[0102] Taking the clustering of a standard sedan as an example: the main body of the sedan has a high point cloud density, while the side mirrors, due to their small size and low surface reflectivity, only produce a small number of sparse edge points, with their spatial location approximately 0.3m to 0.5m from the main body of the sedan. Because... =0.8 is greater than the distance from the sparse points of the rearview mirror to the main body of the vehicle. When performing coarse clustering, it can connect the sparse edge points of the rearview mirror to the point cloud of the main body of the vehicle into the same cluster. Similarly, the point clouds of pedestrians' outstretched parts such as arms and backpacks are also completely encapsulated within the corresponding first-level coarse clusters of pedestrians.

[0103] Of course, in a preferred embodiment... The value is also adaptively adjusted based on the actual scanning beam and point cloud density of the lidar. For a 16-line lidar (where the point cloud is relatively sparse). Set to a relatively large 1.0m; for a 64-line lidar (where the point cloud is relatively dense). It is set to a smaller value of 0.6m. However, it meets the requirements regardless of the configuration. It is significantly larger than the spatial distance between the main body of the object and the extended edge, thus ensuring the integrity of the envelope in the coarse clustering stage.

[0104] The decision execution module is used to: serially traverse each first-level coarse cluster in the first-level coarse cluster set. The two-way dynamic constraint decision-making is performed, which includes the first sub-step and the second sub-step.

[0105] The first sub-step includes: extracting the current first-level coarse clusters. Point cloud quantity .

[0106] If the number of point clouds < This triggers a positive skip constraint, causing the first-level coarse clusters to be... Save directly to the final target point cloud cluster set F; if ≥ Then proceed to the second sub-step, where, The preset adaptive shunting threshold;

[0107] The second sub-step includes: extracting the current first-level coarse clusters. The internal points construct locally bounded Utilizing a narrow sensing radius and the minimum number of neighborhood points of the core point Execute local DBSCAN clustering function Output the set of second-level sub-clusters ; where, definition And based on the M value, perform the following feedback decision in reverse: If Then discard the first-level coarse cluster. Store the M secondary sub-clusters into the final target point cloud cluster set F; if M=1, discard the secondary clustering results. Revert and fully inherit the first-level coarse cluster Store the final target point cloud cluster set F; if M=0, then determine the first-level coarse cluster. Invalid; refuse to incorporate it into the final target point cloud cluster set F.

[0108] In the decision execution module, each cluster in the first-level coarse clustering cluster set is traversed sequentially according to the cluster number order. A two-way dynamic constraint decision is performed on it. The two-way constraint decision includes a first sub-step and a second sub-step.

[0109] The first sub-step mainly involves a positive skip constraint based on the number of point clouds. The second-level clustering is initiated using this cardinality dynamic constraint, with the following mapping rules:

[0110]

[0111] in This is a preset adaptive diversion threshold. If the value is less than this threshold, it is determined to be a low-density, small target, and forcibly performing density division will cause it to be filtered out as a whole. The decision execution module triggers a positive skip constraint here, exempting it from secondary retrieval and directly pushing it into the final target point cloud cluster set F. Conversely, if the condition is met, this macroscopic boundary will serve as a secondary external bounded constraint, proceeding to the second sub-step.

[0112] The first sub-step is explained in more detail below. In the first sub-step, the current first-level coarse cluster is counted. Number of three-dimensional points included In one embodiment, an adaptive shunt threshold is set. If the number of points in the current cluster is less than 30, the decision execution module determines that the cluster is a low-density, small target (e.g., a small obstacle in the distance or sparse vegetation). In this case, a positive skip constraint is triggered—the system skips the secondary DBSCAN clustering retrieval for that cluster and directly saves the cluster to the final target point cloud cluster set F. This mechanism prevents low-density, small targets from being filtered out and discarded entirely during density clustering because they do not meet the core point condition. The system then proceeds to the second sub-step.

[0113] The second sub-step involves inverse boundary correction constraints based on connected states. Specifically, for the first-level coarse clusters entering this stage, the system removes them from the global full-scale point cloud retrieval space and only extracts... The internal points construct locally bounded Based on this, a narrow sensing radius is utilized. and the minimum number of neighborhood points of the core point Execute local DBSCAN clustering function Output the set of second-level sub-clusters :

[0114]

[0115] in This represents the number of secondary sub-clusters formed. Based on the M value, the following three different modalities of feedback decision-making are executed in reverse. To determine the final elements to be included in the final target point cloud cluster set F:

[0116]

[0117] Modality A (precise segmentation, when M>1): Determines that there is spatial adhesion of multiple targets within the current first-level coarse cluster. In this case, the second-level clustering result takes effect, discarding the first-level coarse cluster and replacing it with a cluster formed by the second-level clustering. Each sub-cluster is replaced and stored in the final target point cloud cluster set. .

[0118] Modal B (marginal regression, when M=1): Determines that the current first-level coarse cluster is internally continuous and has no multi-objective adhesion. To avoid the second-level density clustering removing sparse points on the periphery due to the core point threshold, the system activates a reverse constraint mechanism: completely discarding the current second-level clustering results. Re-adopt, revert, and fully inherit the first-level coarse cluster. To ensure that its original 3D physical dimensions are not distorted, it is then added to the final target point cloud cluster set. .

[0119] Modality C (noise removal, when M=0): Determines that no effective core points can be formed within the current first-level coarse cluster, classifying it as spatial random noise or false alarms. The system directly determines the first-level coarse cluster to be invalid and refuses to include it in the final target point cloud cluster set. (The output is an empty set) ).

[0120] The second sub-step is described in detail as follows: In the second sub-step, the system first extracts the current first-level coarse cluster from the global point cloud space. All points contained within the system are used to construct a locally bounded KdTree. Subsequently, the system sets a narrow sensing radius. Minimum number of neighborhood points of the core point It performs local DBSCAN clustering. This process only performs neighborhood searches within the local space of the current cluster, significantly reducing computational complexity.

[0121] After execution, the system obtains a set of secondary sub-clusters. Where M is the number of sub-clusters. The system performs three types of feedback decisions based on the value of M:

[0122] When M > 1 (Modal A), it is determined that there is multi-objective spatial adhesion within the current first-level coarse cluster, and the second-level clustering successfully achieves accurate segmentation between different objects. At this point, the original first-level coarse cluster is discarded. All M secondary sub-clusters are stored in the final target point cloud cluster set F.

[0123] When M=1 (Modal B), the current first-level coarse cluster is determined to be continuous and without multiple targets adhering to each other. To avoid target size distortion caused by DBSCAN excluding sparse outer points due to the core point threshold, the reverse boundary correction constraint is activated, and the current second-level clustering result is discarded. Revert and fully inherit the first-level coarse cluster The entire point cloud is stored in the final target point cloud cluster set F. This mechanism ensures the integrity of the object's original three-dimensional physical dimensions.

[0124] When M=0 (modality C), it is determined that no valid core point can be formed within the current first-level coarse cluster, and it belongs to spatial random noise or false alarm targets. The cluster is directly determined to be invalid and is rejected from being included in the final target point cloud cluster set F.

[0125] In some further specific embodiments, in the first sub-step, during the positive skip constraint, when < If the current primary coarse cluster is identified as a low-density, small target, its secondary search is waived to prevent the low-density, small target from being filtered out as a whole. In one specific embodiment, the decision execution module presets an adaptive diversion threshold. When the decision execution module traverses to a certain first-level coarse cluster... At that time, the number of point clouds was counted. Since 12 is less than 25, the cluster is determined to be a low-density, small target. In real-world scenarios, this cluster might correspond to obstacles such as: pedestrians at a distance, small traffic cones, small obstacles scattered on the road, or locally sparse point clouds of bicycles / motorcycle riders in front of a vehicle. In this case, after triggering the positive skip constraint, the local DBSCAN clustering in the second sub-step is not performed on this cluster; instead, the cluster is directly... The complete point cloud is stored in the final target point cloud cluster set F. This decision logic is based on the following physical understanding: low-density small targets have a limited number of points and do not have the conditions for multiple targets to stick together inside, so there is no need to subdivide them through density clustering; on the contrary, if DBSCAN processing is forcibly performed on them, their sparse point cloud distribution cannot meet the neighborhood conditions of the core points, and they are very likely to be judged as noise and filtered out, resulting in the target being missed.

[0126] By employing a positive skip constraint mechanism, such targets are identified and directly retained early in the clustering process, fundamentally eliminating the risk of missed detection and improving the recall rate of low-speed autonomous vehicles for perceptual minor obstacles. Furthermore, for point cloud data smaller than [a certain number], [this method is effective]. Clusters can be processed without constructing a local KdTree or performing neighborhood search and density determination using local DBSCAN, directly skipping secondary retrieval. This not only saves computational resources but also reduces unnecessary algorithm branch decisions, further improving the overall system efficiency. Furthermore, the forward skip constraint avoids false filtering by DBSCAN at the source level, while the edge regression mechanism prevents DBSCAN from removing edge points when dealing with a large number of clusters. These two mechanisms work together at different scales to ensure the integrity of the target point cloud.

[0127] In some further specific embodiments, in the second sub-step, when M=1, it is determined that the current first-level coarse cluster is continuous and has no multi-target adhesion, the reverse constraint mechanism is activated, and the current second-level clustering result is discarded. Completely inherits the first-level coarse cluster. This is to ensure that its original three-dimensional physical dimensions are not distorted.

[0128] In one specific embodiment, the decision execution module performs a coarse clustering analysis at a certain level. After performing local DBSCAN clustering, a set of secondary sub-clusters is obtained. According to statistics, This means that the local DBSCAN only divides the cluster into a sub-cluster within that cluster. At this point, the decision execution module determines... In terms of space, it is a single, continuous target, and there is no situation where multiple targets are stuck together.

[0129] However, it should be noted that the DBSCAN algorithm has a core threshold. In fact, it only includes The core points that satisfy the density condition and their density-reachable points, while Sparse outreach points at the edges were marked as boundary points or noise points and discarded during the DBSCAN process. Therefore, The physical dimensions represented are compared to the actual dimensions. It's a shrunken version.

[0130] In this case, the decision execution module activates the reverse boundary correction constraint: completely discard. The clustering results are reverted to the state before the local DBSCAN was executed, and the original first-level coarse clusters are reverted to their original state. The complete target point cloud cluster set F is stored. Taking a specific numerical example: the first-level coarse cluster of a car target. It contains 1200 points, with a directed bounding area and dimensions of 4.8m * 1.9m * 1.5m (including rearview mirrors); after local DBSCAN... It contains only 980 points, with a reduced enclosure and size of 4.8m*1.8m*1.5m (rearview mirror points are removed). Through an inverse constraint mechanism, the system's final output is... The full dimensions ensure the accuracy of obstacle size perception.

[0131] This specific embodiment ensures that the final output target size is strictly equal to the complete size encompassed by the Euclidean coarse cluster by activating the reverse boundary correction constraint when M=1. This allows the downstream path planning and obstacle avoidance modules to make safety decisions based on the actual physical size of obstacles, effectively reducing collision risks. It fundamentally solves the industry pain point of underestimating the target physical size caused by DBSCAN density clustering.

[0132] The fitting output module is used to: use principal component analysis to solve for the geometric center of each point cloud cluster in the final target point cloud cluster set F, perform multi-angle scanning to calculate the optimal heading angle in the preset yaw angle search space, accurately calculate the length, width and height of the directed enclosing sum, and publish it after performing coordinate system mirror mapping transformation.

[0133] The fitting output module iterates through each point cloud cluster in the final target point cloud cluster set F, calling the Principal Component Analysis (PCA) algorithm to calculate the covariance matrix of each cluster, and solving for eigenvalues ​​and eigenvectors to determine the geometric center and principal direction of the point cloud cluster. Subsequently, the system performs multi-angle scanning within a preset yaw angle search space with a fixed step size, calculating the bounding sum and area at each angle, selecting the angle with the smallest area as the optimal heading angle, and accurately calculating the length, width, and height of the directed bounding sum. Finally, the system performs coordinate system mirroring transformation, publishing the 3D directed bounding sum and state information of the obstacle through ROS topics in a custom message format for use by the downstream path planning and obstacle avoidance control modules.

[0134] refer to Figure 4 , Figure 4 This is a framework diagram of a point cloud clustering and segmentation device according to another embodiment. The bidirectional dynamic constraint Euclidean-DBSCAN point cloud clustering and segmentation method of this invention can be run within this framework. This point cloud clustering and segmentation device includes: a point cloud preprocessing module, a first-level global coarse clustering module, a bidirectional dynamic constraint decision module, and a state calculation and output module. The point cloud preprocessing module receives the raw point cloud and extracts the effective point set of the near-field ROI; the first-level global coarse clustering module performs global Euclidean clustering to establish the macroscopic envelope boundary of the target; the bidirectional dynamic constraint decision module is the core processing unit, which further includes a forward quantity splitting submodule, a second-level local density clustering submodule, and a reverse boundary feedback submodule, which cooperate to achieve local computational dimensionality reduction and adaptive backtracking inheritance of edge features; the state calculation and output module calculates the geometric center and orientation angle of the processed point cloud clusters, fits a three-dimensional oriented bounding sum, and completes data communication and publishing between systems.

[0135] On the other hand, a computer-readable storage medium is provided, wherein a processor-executable program is stored, which, when executed by a processor, is used to implement the bidirectional dynamic constraint Euclidean-DBSCAN point cloud clustering and segmentation method as described in any of the above specific embodiments.

[0136] This application also discloses a computer program product, including a computer program or computer instructions, which are stored in a computer-readable storage medium. The processor of the computer device reads the computer program or computer instructions from the computer-readable storage medium and executes the computer program or computer instructions, causing the computer device to perform the bidirectional dynamic constraint Euclidean-DBSCAN point cloud clustering and segmentation method as described in any of the preceding embodiments.

[0137] 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, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, 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 apparatuses.

[0138] 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.

[0139] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, 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 indirect coupling or communication connection through some interfaces, apparatuses, or units, and may be electrical, mechanical, or other forms.

[0140] 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.

[0141] 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.

[0142] 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 several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0143] Although the description of this application has been quite detailed and particularly focused on several of the described embodiments, it is not intended to limit itself to any of these details or embodiments or any particular embodiment. Rather, it should be considered as effectively covering the intended scope of this application by referring to the appended claims and taking into account the prior art, which provides for a broad possible interpretation of these claims. Furthermore, the foregoing description of this application with respect to embodiments foreseeable by the inventors is intended to provide a useful description, and non-substantial modifications to this application that have not yet been foreseen may still represent equivalent modifications.

[0144] 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.

Claims

1. A bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation method, characterized in that, include: Step 1: Obtain the original 3D point cloud around the low-speed unmanned vehicle. Adaptive resolution parameters using a voxel grid filter. Uniform downsampling is performed, and spatial pass-through filtering is applied based on preset vertical, horizontal, and height axis filtering intervals to extract the effective detection point set within the near-field region of interest. ; Step 2: For the valid detection point set Construct a global first search space and apply large distance tolerance. Execute the global first-level Euclidean coarse clustering function Produces a set of first-level coarse clusters. ; And initialize a final target point cloud cluster set to store the final clustering results. The initial state is an empty set; Step 3: Serially traverse each first-level coarse cluster in the first-level coarse cluster set. The two-way dynamic constraint decision-making is performed, which includes: the first sub-step and the second sub-step. The first sub-step includes: extracting the current first-level coarse clusters. Point cloud quantity ; If the number of point clouds < This triggers a positive skip constraint, causing the first-level coarse clusters to be... Save directly to the final target point cloud cluster set F; if ≥ Then proceed to the second sub-step, where, The preset adaptive traffic splitting threshold; The second sub-step includes: extracting the current first-level coarse clusters. The internal points construct locally bounded Utilizing a narrow sensing radius and the minimum number of neighborhood points of the core point Execute local DBSCAN clustering function Output the set of second-level sub-clusters ; where, definition And based on the M value, perform the following feedback decision in reverse: If Then discard the first-level coarse cluster. Store the M secondary sub-clusters into the final target point cloud cluster set F; if M=1, discard the secondary clustering results. Revert and fully inherit the first-level coarse cluster Store the final target point cloud cluster set F; if M=0, then determine the first-level coarse cluster. Invalid; it is rejected from being included in the final target point cloud cluster set F. Step 4: Use principal component analysis to solve for the geometric center of each point cloud cluster in the final target point cloud cluster set F. Perform multi-angle scanning within the preset yaw angle search space to calculate the optimal heading angle, accurately calculate the length, width, and height of the directed enclosing sum, and then perform coordinate system mirror mapping transformation before publishing it externally.

2. The bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation method according to claim 1, characterized in that, In step 1, the longitudinal filtering interval is The horizontal filtering interval satisfies The height axis filtering range is: removing items that meet the following conditions. ground point or High altitude point.

3. The bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation method according to claim 1, characterized in that, In step 2, the global first-order Euclidean coarse clustering function By utilizing the large radius connectivity property, all sparse edge points extending around the object are completely enclosed in each independent first-order coarse cluster. middle.

4. The bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation method according to claim 1, characterized in that, In the first sub-step, during the positive skip constraint, when < If the current primary coarse cluster is determined to be a low-density small target, its secondary search is waived to avoid the low-density small target being filtered out as a whole.

5. The bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation method according to claim 1, characterized in that, In the second sub-step, when M=1, it is determined that the current first-level coarse cluster is continuous and has no multi-objective adhesion, the reverse constraint mechanism is activated, and the current second-level clustering result is discarded. Completely inherits the first-level coarse cluster. This is to ensure that its original three-dimensional physical dimensions are not distorted.

6. The bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation method according to claim 1, characterized in that, The adaptive resolution parameters of the voxel grid filter The narrow sensing radius of the local DBSCAN And the large distance tolerance of the global first-level Euclidean coarse clustering Satisfying the spatial scale hierarchical progression inequality: < < .

7. The bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation method according to claim 1, characterized in that, In step 4, the method of external publication is to package and publish the three-dimensional oriented enclosure and state of the obstacle through a ROS topic.

8. A bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation device, characterized in that, include: The module includes a filter module, an output construction module, a decision execution module, and a fitting output module. The acquisition and filtering module is used to: acquire the original 3D point cloud surrounding the low-speed unmanned vehicle. Adaptive resolution parameters using a voxel grid filter. Uniform downsampling is performed, and spatial pass-through filtering is applied based on preset vertical, horizontal, and height axis filtering intervals to extract the effective detection point set within the near-field region of interest. ; The output module is used for: targeting the effective detection point set Construct a global first search space and apply large distance tolerance. Execute the global first-level Euclidean coarse clustering function Produces a set of first-level coarse clusters. And initialize a final target point cloud cluster set to store the final clustering results; The initial state is an empty set; The decision execution module is used to: serially traverse each first-level coarse cluster in the first-level coarse cluster set. The two-way dynamic constraint decision-making is performed, which includes: the first sub-step and the second sub-step. The first sub-step includes: extracting the current first-level coarse clusters. Point cloud quantity ; If the number of point clouds < This triggers a positive skip constraint, causing the first-level coarse clusters to be... Save directly to the final target point cloud cluster set F; if ≥ Then proceed to the second sub-step, where, The preset adaptive traffic splitting threshold; The second sub-step includes: extracting the current first-level coarse clusters. The internal points construct locally bounded Utilizing a narrow sensing radius and the minimum number of neighborhood points of the core point Execute local DBSCAN clustering function Output the set of second-level sub-clusters ; where, definition And based on the M value, perform the following feedback decision in reverse: If Then discard the first-level coarse cluster. Store the M secondary sub-clusters into the final target point cloud cluster set F; if M=1, discard the secondary clustering results. Revert and fully inherit the first-level coarse cluster Store the final target point cloud cluster set F; if M=0, then determine the first-level coarse cluster. Invalid; it is rejected from being included in the final target point cloud cluster set F. The fitting output module is used to: use principal component analysis to solve for the geometric center of each point cloud cluster in the final target point cloud cluster set F, perform multi-angle scanning to calculate the optimal heading angle in the preset yaw angle search space, accurately calculate the length, width and height of the directed enclosing sum, and publish it after performing coordinate system mirror mapping transformation.

9. The bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation device according to claim 8, characterized in that, In the acquisition and filtering module, the vertical filtering interval is: The horizontal filtering interval satisfies The height axis filtering range is: removing items that meet the following conditions. ground point or High altitude point.

10. The bidirectional dynamically constrained Euclidean-DBSCAN point cloud clustering and segmentation device according to claim 8, characterized in that, In the output module, the global first-order Euclidean coarse clustering function By utilizing the large radius connectivity property, all sparse edge points extending around the object are completely enclosed in each independent first-order coarse cluster. middle.

Images