A positioning method, system, device and medium
By semantically segmenting and constructing multi-scale feature trees from LiDAR point cloud data, and combining a coarse-fine matching strategy, the problems of low positioning accuracy and efficiency of LiDAR were solved, achieving a highly efficient and accurate positioning effect.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGDONG POWER GRID CO LTD
- Filing Date
- 2026-02-25
- Publication Date
- 2026-06-02
AI Technical Summary
Existing lidar positioning methods have low positioning accuracy and efficiency in complex and ever-changing environments.
By acquiring target point cloud data from target sensors, the data is divided based on preset semantic categories, and a multi-scale hierarchical feature tree is constructed. A coarse-fine two-level semantic filtering and hierarchical feature matching strategy is adopted. Candidate regions are quickly locked using large-scale features, and precise alignment is performed by combining small-scale fine features, thereby reducing positioning errors and computational load.
It improves the accuracy and efficiency of lidar positioning, and enables high-precision positioning in complex environments.
Smart Images
Figure CN122134805A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of autonomous navigation and robotics, and in particular to a positioning method, system, device, and medium. Background Technology
[0002] LiDAR (Light Detection and Ranging) technology is widely used in autonomous driving, environmental modeling, and robot navigation. As an important sensor, LiDAR provides high-precision 3D environmental perception information. Global localization (GMR) is a crucial step in these applications, aiming to determine the precise pose (position and orientation) of the LiDAR within a global map, providing a foundation for subsequent tasks such as navigation and path planning. Existing localization methods mainly rely on single-scale or simple feature extraction strategies. However, these methods suffer from low accuracy and efficiency in complex and variable environments. Summary of the Invention
[0003] This invention provides a positioning method, system, device, and medium that can solve the problem of low positioning accuracy in existing technologies.
[0004] This invention provides a positioning method, comprising: Acquire target point cloud data from the target sensor; Based on preset semantic categories and the semantic levels corresponding to each semantic category, the target point cloud data is divided to obtain at least one first semantic layer point cloud and at least one second semantic layer point cloud. For each point cloud of the first semantic layer and each point cloud of the second semantic layer, a multi-scale hierarchical feature tree is constructed to obtain the first feature tree and the second feature tree. Based on the features at a first scale in the first feature tree, a first match is performed with preset map data to obtain a candidate region, and based on the features at a second scale in the second feature tree, a second match is performed with the preset map data of the candidate region to obtain the positioning result of the target sensor, wherein the first scale is greater than the second scale.
[0005] This invention upgrades geometric point clouds to semantic-geometric joint point clouds, ensuring all subsequent operations are performed only within the same semantic category, thus improving matching accuracy. Semantic categories are further mapped to semantic levels, enabling coarse-to-fine semantic filtering and laying the data foundation for subsequent coarse-to-fine two-stage matching. The original point cloud is compressed into a tree-node-feature structure, reducing memory usage, and the tree structure facilitates rapid multi-scale retrieval, improving matching speed. Leveraging the insensitivity of large-scale features to viewpoint, occlusion, and partial missing features, several possible regions are quickly identified, reducing the full-image search space. Within the narrowed candidate regions, fine-grained small-scale features are used for precise alignment, significantly reducing localization errors and computational load, improving both localization accuracy and efficiency.
[0006] Further, the step of obtaining candidate regions by performing a first match between the features at a first scale in the first feature tree and preset map data includes: In the preset map data, first map data corresponding to the semantic category of each first feature tree is determined, and first map features corresponding to the first scale in the first map data are determined. Calculate the similarity between the features at the first scale and each of the first map features; Map regions with a similarity greater than a first preset threshold are identified as candidate regions.
[0007] This approach, based on coarse matching, further refines the semantic-scale dual filtering strategy, ensuring both the recall rate of candidate regions and reducing the computational cost of subsequent fine matching.
[0008] Further, the step of performing a second match between the features at the second scale in the second feature tree and the preset map data of the candidate region to obtain the localization result of the target sensor specifically involves: In each candidate region, second map data corresponding to the semantic category of each second feature tree is determined, and second map features corresponding to the second scale in the second map data are determined; The matching error between the features at the second scale and the features on the second map is calculated using a preset feature point pair matching algorithm, and the positioning result is determined based on the matching error.
[0009] This allows for high-precision matching within the candidate region provided by coarse matching, achieving decoupling between accuracy and speed. In other words, coarse matching improves speed, while fine matching improves accuracy.
[0010] Furthermore, the construction of the multi-scale hierarchical feature tree specifically involves: Based on the multiple preset scale levels corresponding to each semantic layer point cloud, determine the neighborhood radius corresponding to each scale level; Based on each of the neighborhood radii, features corresponding to each of the scale levels are extracted respectively; Using the semantic category corresponding to the semantic layer point cloud as the root node and each scale level as a child node, the features corresponding to each scale level are attached to the corresponding child node to obtain the multi-scale hierarchical feature tree.
[0011] Using radius instead of grid to define scale allows for compatibility with non-uniform and unstructured point clouds, avoiding quantization errors caused by voxelization. By extracting features at corresponding scale levels based on the radius of each neighborhood, multi-scale features from local to global can be generated at once for the same set of point clouds. Subsequent readings only require reading by scale, eliminating the need for repeated calculations. The tree structure supports top-down or bottom-up traversal strategies, facilitating quick switching between accuracy-first and speed-first modes in different applications, thus improving positioning efficiency.
[0012] Furthermore, the extraction of features corresponding to each scale level based on each of the neighborhood radii specifically involves: When the neighborhood radius is greater than or equal to a preset radius threshold, for each point in the semantic layer point cloud, the point cloud density and normal vector distribution entropy are calculated based on all neighboring points within the current neighborhood radius. The point cloud density and the normal vector distribution entropy are then concatenated to obtain the features at the first scale. When the neighborhood radius is less than a preset radius threshold, the target bounding box of the semantic layer point cloud is determined, and the size ratio of the target bounding box, the centroid coordinates of the point cloud, and the shape features based on the Fourier descriptor are calculated. The size ratio, the centroid coordinates of the point cloud, and the shape features are then combined to obtain the features of the second scale.
[0013] This combination of density and entropy is robust to first-scale features with significant local shape but drifting absolute coordinates, making it suitable for coarse matching. Meanwhile, the bounding box + Fourier descriptor can characterize both the overall contour and detailed shape, exhibiting invariance to rotation, scaling, and mirroring, making it suitable for fine matching. Automatic feature type switching via radius-threshold ensures that the same framework can capture both macroscopic contours and microscopic shapes, guaranteeing a balance between coarse matching recall and fine matching resolution.
[0014] Furthermore, the step of dividing the target point cloud data based on preset semantic categories and the semantic levels corresponding to each semantic category to obtain at least one first semantic layer point cloud and at least one second semantic layer point cloud specifically involves: The target point cloud data is processed by a semantic classification model to obtain the semantic category label corresponding to each target point in the target point cloud data; The target points are divided according to the semantic category labels to obtain multiple sets of point cloud data; Based on the semantic level corresponding to each group of point cloud data, determine each first semantic layer point cloud and each second semantic layer point cloud.
[0015] This approach uses deep learning to complete segmentation and classification in one go, significantly improving semantic accuracy. Furthermore, only one forward inference is needed in the online phase, and all subsequent modules are processed separately according to semantic labels, greatly improving parallelism.
[0016] Furthermore, the acquisition of target point cloud data from the target sensor specifically includes: Acquire the raw point cloud data collected by the target sensor; Calculate the number of neighboring points of each point in the original point cloud data within a preset first neighborhood, and determine points whose number of neighboring points is less than a second preset threshold as noise points. Remove all noise points to obtain the denoised original point cloud data. The denoised original point cloud data is divided into multiple voxels, and the target point corresponding to each voxel is determined to obtain the target point cloud data.
[0017] This approach employs a statistical filtering algorithm, which identifies and removes noise points by setting a neighborhood radius and threshold. Then, a voxel grid downsampling algorithm is used to divide the voxel grid with a set side length and retain representative points. Statistical filtering can effectively remove isolated noise points and improve data quality. Voxel grid downsampling, on the other hand, significantly reduces the amount of data while retaining key information, thereby reducing the computational burden of subsequent processing, improving overall processing efficiency, and providing a higher quality and more efficient data foundation for subsequent accurate positioning.
[0018] Another embodiment of the present invention provides a positioning system, including: an acquisition module, a semantic segmentation module, a construction module, and a positioning module; The acquisition module is used to acquire target point cloud data from the target sensor; The semantic segmentation module is used to segment the target point cloud data based on preset semantic categories and the semantic levels corresponding to each semantic category, so as to obtain at least one first semantic layer point cloud and at least one second semantic layer point cloud. The construction module is used to construct multi-scale hierarchical feature trees for each first semantic layer point cloud and each second semantic layer point cloud, respectively, to obtain a first feature tree and a second feature tree; The positioning module is used to perform a first match between the features at a first scale in the first feature tree and preset map data to obtain a candidate region, and to perform a second match between the features at a second scale in the second feature tree and the preset map data of the candidate region to obtain the positioning result of the target sensor, wherein the first scale is greater than the second scale.
[0019] Another embodiment of the present invention provides a terminal device, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements the steps of the positioning method of the present invention.
[0020] Another embodiment of the present invention provides a computer-readable storage medium item, including: a stored computer program, which, when the computer program is running, controls the device where the computer-readable storage medium is located to perform the steps of the positioning method of the present invention. Attached Figure Description
[0021] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0022] Figure 1 This is a flowchart illustrating a positioning method provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating a hierarchical feature matching strategy provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating a multi-scale feature tree construction step provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a positioning system provided in an embodiment of the present invention. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art relating to this application; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims and foregoing description of the drawings of this application are intended to cover non-exclusive inclusion.
[0025] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.
[0026] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will understand, explicitly and implicitly, that the embodiments described herein can be combined with other embodiments.
[0027] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.
[0028] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).
[0029] In the description of the embodiments of this application, unless otherwise expressly specified and limited, the technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.
[0030] See Figure 1 To address the problem of low positioning accuracy in existing technologies, an embodiment of the present invention provides a positioning method comprising: Step S101: Obtain target point cloud data collected by the target sensor.
[0031] In the above steps, taking LiDAR as an example, the target sensor acquires raw point cloud data collected by LiDAR. LiDAR is a laser detection and ranging system that emits a laser beam towards the target and measures the return time of the reflected light to obtain information such as the target sensor's distance, azimuth, and height, thereby generating point cloud data. Considering that the raw point cloud data directly acquired by LiDAR is an unprocessed collection of data containing a large number of discrete points in space, each point typically contains three-dimensional coordinate information and may also contain information such as reflection intensity, the target point cloud data can be obtained by preprocessing the raw point cloud data through denoising and downsampling. Denoising removes noise points generated by environmental interference, equipment errors, and other factors from the point cloud data. These noise points interfere with subsequent positioning and feature extraction operations. Common denoising methods include statistical methods and distance-based methods. Downsampling reduces the number of points in the point cloud while maintaining the overall characteristics of the point cloud data, thereby reducing the data volume and improving the efficiency of subsequent processing.
[0032] Step S102: Based on preset semantic categories and the semantic levels corresponding to each semantic category, the target point cloud data is divided to obtain at least one first semantic layer point cloud and at least one second semantic layer point cloud.
[0033] In the above steps, based on preset spatial semantic categories (such as roads, buildings, and obstacles), the preprocessed point cloud data is semantically layered to obtain semantic layer point clouds corresponding to different semantic categories. The preset spatial semantic categories are predefined categories that can describe the attributes of different objects or regions in space. Semantic layering is to classify the preprocessed point cloud data according to different semantic categories based on the preset spatial semantic categories, and to group points belonging to the same semantic category into the same layer to form different semantic layer point clouds. The semantic layer point cloud is a set of point cloud data corresponding to different semantic categories obtained after semantic layering, and the points in each set have the same semantic attributes. Based on the preset semantic hierarchy, at least one set of point cloud data with a higher semantic hierarchy is identified as the first semantic layer point cloud, and at least one set of point cloud data with a lower semantic hierarchy is identified as the second semantic layer point cloud. The higher semantic layer is the first semantic layer with a higher degree of semantic abstraction and a wider coverage in the semantic layering; the lower semantic layer is the second semantic layer with a lower degree of semantic abstraction and a greater focus on details. For example, the building semantic layer is selected as the first (higher) semantic layer, and the obstacle semantic layer or the road semantic layer is selected as the second (lower) semantic layer.
[0034] Step S103: For each point cloud of the first semantic layer and each point cloud of the second semantic layer, construct a multi-scale hierarchical feature tree to obtain the first feature tree and the second feature tree.
[0035] Multi-scale refers to the analysis and processing of point cloud data at different spatial scales or resolutions. Large scales can capture overall global features, while small scales can obtain detailed local features. Therefore, local features correspond to features at the second scale, and global features correspond to features at the first scale. Local features are features that describe the local geometry and structure within a small area of point cloud data, such as the curvature of points, normal vectors, and local surface changes. These features reflect the detailed information of the point cloud in a local region. Global features are features that describe the overall shape, distribution, and other macroscopic characteristics of the entire point cloud data or a larger area. A multi-scale hierarchical feature tree is a data structure that includes extracted local and global features organized according to hierarchical relationships at different scales.
[0036] In the above steps, for each semantic layer point cloud, multiple feature extraction parameters at different scales are set; features of the semantic layer point cloud are extracted at each scale; the extracted features at each scale are organized into a tree data structure according to the scale hierarchy to form a multi-scale hierarchical feature tree corresponding to the semantic layer, wherein the first feature tree corresponds to the first semantic layer point cloud and the second feature tree corresponds to the second semantic layer point cloud.
[0037] Step S104: Based on the features at the first scale in the first feature tree, perform a first match with preset map data to obtain a candidate region, and based on the features at the second scale in the second feature tree, perform a second match with the preset map data of the candidate region to obtain the positioning result of the target sensor, wherein the first scale is greater than the second scale.
[0038] In the above steps, such as Figure 2 The diagram illustrates the flowchart of the hierarchical feature matching strategy. Based on the multi-scale hierarchical feature tree, a hierarchical feature matching strategy is used for global localization. First, coarse global feature matching is performed at the high-level semantic layer to filter candidate localization regions. Then, fine local feature matching is performed at the low-level semantic layer to determine the final localization result, thereby improving the accuracy and efficiency of the matching. Coarse feature matching uses global features and large-scale local features from the high-level semantic layer to perform a preliminary matching operation with preset map data such as the global map or other reference data. This step does not require highly precise matching; its main purpose is to quickly narrow down the search range and obtain candidate localization regions. Fine local feature matching uses local features from the low-level semantic layer to perform a precise matching operation within the filtered candidate localization regions to determine the precise location of the LiDAR in the global map. This step requires high matching accuracy. The obtained localization result is the location information of the LiDAR in the global map. The location information includes the precise location and attitude information of the LiDAR in the global map. The location is usually represented by three-dimensional coordinates (x, y, z), and the attitude is usually represented by rotation matrices or Euler angles.
[0039] As an example of an embodiment of the present invention, the step of obtaining a candidate region by performing a first match between the features at a first scale in the first feature tree and preset map data includes: determining, in the preset map data, first map data corresponding to the semantic category of each of the first feature trees, and determining, in the first map data, first map features corresponding to the first scale; calculating the similarity between the features at the first scale and each of the first map features; and determining the map regions with similarity greater than a first preset threshold as the candidate regions.
[0040] In this embodiment, the building semantic layer serves as a high-level semantic layer, extracting multi-scale global features from the current LiDAR building semantic layer. Similarity calculations are performed between the data and multi-scale global features of the building semantic layer in the global map (using cosine similarity or Euclidean distance); a similarity threshold is set. Retain similarity greater than The map area is used as a candidate location area. , This represents the number of candidate regions.
[0041] As an example of an embodiment of the present invention, the step of obtaining the positioning result of the target sensor by performing a second matching based on the features of the second scale in the second feature tree and the preset map data of the candidate region specifically involves: in each candidate region, determining the second map data corresponding to the semantic category of each second feature tree, and determining the second map features in the second map data corresponding to the second scale; calculating the matching error between the features of the second scale and the second map features using a preset feature point pair matching algorithm, and determining the positioning result based on the matching error.
[0042] In this embodiment, the obstacle semantic layer is selected as the low-level semantic layer, and for each candidate region... Extract the current lidar and Multi-scale local features of the semantic layer of internal obstacles; a feature point pair matching algorithm is used to calculate the error of local feature matching. , The sum of squared coordinate deviations of the matching point pairs; selection error Minimum candidate region To make the current lidar in The pose within the map is used as the preliminary localization result; the preliminary localization result is iteratively optimized to minimize the local feature matching error, resulting in the final localization result. The final localization result is then converted into the pose parameters of the LiDAR in the global map coordinate system. ,in For translation coordinates, The rotation angle; Furthermore, the validity of the pose parameters is verified. Specifically, the overlap between the current LiDAR point cloud and the corresponding area point cloud of the global map is calculated: ; in, The number of overlapping points. This represents the total number of points in the current lidar point cloud. like , If the preset overlap threshold is met, the pose is deemed valid and output; otherwise, the above steps are repeated for localization.
[0043] As an example of an embodiment of the present invention, the construction of the multi-scale hierarchical feature tree specifically involves: determining the neighborhood radius corresponding to each of the multiple preset scale levels corresponding to each semantic layer point cloud; extracting the features corresponding to each scale level based on each neighborhood radius; taking the semantic category corresponding to the semantic layer point cloud as the root node and each scale level as the child node, attaching the features corresponding to each scale level to the corresponding child node to obtain the multi-scale hierarchical feature tree.
[0044] In this embodiment, as Figure 3 The flowchart shown illustrates the steps of constructing a multi-scale feature tree. Within each semantic layer, features are extracted sequentially from large scale to small scale, and these features are organized in a tree structure. The structure of the multi-scale hierarchical feature tree is a three-level hierarchical structure, with the semantic layer as the root node and each scale level as a child node. Each scale node carries the local feature set and global features corresponding to that scale.
[0045] It should be noted that the foregoing embodiments have shown that local features correspond to features at the second scale, and global features correspond to features at the first scale.
[0046] As an example of an embodiment of the present invention, the step of extracting features corresponding to each scale level based on each of the neighborhood radii specifically involves: when the neighborhood radius is greater than or equal to a preset radius threshold, for each point in the semantic layer point cloud, calculating the point cloud density and normal vector distribution entropy based on all neighboring points within the current neighborhood radius, and concatenating the point cloud density and the normal vector distribution entropy to obtain the features of the first scale; when the neighborhood radius is less than the preset radius threshold, determining the target bounding box of the semantic layer point cloud, calculating the size ratio of the target bounding box, the centroid coordinates of the point cloud, and the shape features based on the Fourier descriptor, and concatenating the size ratio, the centroid coordinates of the point cloud, and the shape features to obtain the features of the second scale.
[0047] In this embodiment, local and global features at different scales are extracted within each semantic layer, and then organized into a feature tree structure according to the scale hierarchy, including: setting up a feature tree structure for each semantic layer point cloud. scale level Corresponding scale parameters , Indicates the first Feature extraction neighborhood radius at each scale; for each point in the semantic layer point cloud, extract local features at each scale: taking the first... Taking scale as an example, for points in the semantic layer point cloud Take its neighborhood radius Calculate the local point cloud density for all neighboring points within the area: ; in, Indicates the number of points in the neighborhood. Normal vector distribution entropy: ; in, The normal vector is represented in the th case. The probability distribution in each principal direction will and splicing as the first scale point Local features ; For each semantic layer point cloud, global features are extracted at various scales: taking the first... Taking scale as an example, for semantic layer point clouds , Calculate the aspect ratio of its minimum bounding box. Point cloud centroid coordinates and shape features based on Fourier descriptors ,Will , and splicing as the first Global features of this semantic layer at this scale .
[0048] Furthermore, the lowest semantic layer may not need to extract global features.
[0049] As an example of an embodiment of the present invention, the step of dividing the target point cloud data based on preset semantic categories and the semantic levels corresponding to each semantic category to obtain at least one first semantic layer point cloud and at least one second semantic layer point cloud specifically involves: processing the target point cloud data through a semantic classification model to obtain a semantic category label corresponding to each target point in the target point cloud data; dividing each target point according to each semantic category label to obtain multiple sets of point cloud data; and determining each first semantic layer point cloud and each second semantic layer point cloud according to the semantic level corresponding to each set of point cloud data.
[0050] In this embodiment, a semantic classification model is constructed: a point cloud-based deep learning classification network (such as PointNet or PointConv) is used as training samples, with point cloud data labeled with semantic categories as training samples, to train a model that can distinguish the semantic categories of roads, buildings, and obstacles; the preprocessed point cloud data is then used to construct the model. Input the semantic classification model to obtain each point semantic category tags ; by semantic category tag The points are divided to obtain the road semantic layer point cloud. Architectural semantic layer point cloud Obstacle semantic layer point cloud .
[0051] As an example of an embodiment of the present invention, the acquisition of target point cloud data from the target sensor specifically includes: acquiring the raw point cloud data collected by the target sensor; calculating the number of neighboring points of each point in the raw point cloud data within a preset first neighborhood, and determining points whose number of neighboring points is less than a second preset threshold as noise points, removing all noise points to obtain denoised raw point cloud data; dividing the denoised raw point cloud data into multiple voxels, and determining the target point corresponding to each voxel to obtain the target point cloud data.
[0052] In this embodiment, the raw three-dimensional point cloud data output by the lidar is acquired. ,in Indicates the first The three-dimensional coordinates of the points This represents the total number of points in the original point cloud; a statistical filtering algorithm is used for noise reduction, such as calculating the number of points... Within the preset neighborhood radius The number of neighboring points within the range; if the number of neighboring points is less than a preset threshold. If the point is identified as a noise point, it is removed, resulting in a denoised point cloud. A voxel grid downsampling algorithm is used for downsampling, such as setting the voxel grid side length. denoised point clouds Mapped onto a voxel mesh, with a representative point retained within each voxel, the preprocessed point cloud is obtained. .
[0053] like Figure 4 As shown, based on the above-described method embodiments, an embodiment of the present invention provides a positioning system 400, including: an acquisition module 401, a semantic segmentation module 402, a construction module 403, and a positioning module 404; The acquisition module 401 is used to acquire target point cloud data from the target sensor; The semantic segmentation module 402 is used to segment the target point cloud data based on preset semantic categories and the semantic levels corresponding to each semantic category, so as to obtain at least one first semantic layer point cloud and at least one second semantic layer point cloud. The construction module 403 is used to construct multi-scale hierarchical feature trees for each first semantic layer point cloud and each second semantic layer point cloud, respectively, to obtain a first feature tree and a second feature tree; The positioning module 404 is used to perform a first match between the features at a first scale in the first feature tree and preset map data to obtain a candidate region, and to perform a second match between the features at a second scale in the second feature tree and preset map data of the candidate region to obtain the positioning result of the target sensor, wherein the first scale is greater than the second scale.
[0054] It is understood that the above system item embodiments correspond to the method item embodiments of the present invention, and can implement the positioning method provided by any of the above method item embodiments of the present invention.
[0055] It should be noted that the system embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.
[0056] For ease of description and brevity, the embodiments of the system items of the present invention include all the implementation methods in the above positioning method embodiments, and will not be repeated here.
[0057] Based on the above-described embodiments of the positioning method, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the positioning method of any embodiment of the present invention.
[0058] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.
[0059] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.
[0060] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.
[0061] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the positioning method described in any of the above-described method embodiments of the present invention.
[0062] Based on the above-described method embodiments, this invention also provides a computer program / program product, which is stored in a storage medium and executed by at least one processor to implement the various processes of any of the above-described method embodiments, and can achieve the same technical effect. To avoid repetition, it will not be described again here.
[0063] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.
[0064] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. A positioning method, characterized in that, include: Acquire target point cloud data from the target sensor; Based on preset semantic categories and the semantic levels corresponding to each semantic category, the target point cloud data is divided to obtain at least one first semantic layer point cloud and at least one second semantic layer point cloud. For each point cloud of the first semantic layer and each point cloud of the second semantic layer, a multi-scale hierarchical feature tree is constructed to obtain the first feature tree and the second feature tree. Based on the features at a first scale in the first feature tree, a first match is performed with preset map data to obtain a candidate region, and based on the features at a second scale in the second feature tree, a second match is performed with the preset map data of the candidate region to obtain the positioning result of the target sensor, wherein the first scale is greater than the second scale.
2. The positioning method as described in claim 1, characterized in that, The step of obtaining candidate regions by performing a first match between the features at a first scale in the first feature tree and preset map data includes: In the preset map data, first map data corresponding to the semantic category of each first feature tree is determined, and first map features corresponding to the first scale in the first map data are determined. Calculate the similarity between the features at the first scale and each of the first map features; Map regions with a similarity greater than a first preset threshold are identified as candidate regions.
3. The positioning method as described in claim 2, characterized in that, The second matching of the features at the second scale in the second feature tree with the preset map data of the candidate region to obtain the localization result of the target sensor is specifically as follows: In each candidate region, second map data corresponding to the semantic category of each second feature tree is determined, and second map features corresponding to the second scale in the second map data are determined; The matching error between the features at the second scale and the features on the second map is calculated using a preset feature point pair matching algorithm, and the positioning result is determined based on the matching error.
4. The positioning method as described in claim 1, characterized in that, The construction of the multi-scale hierarchical feature tree is specifically as follows: Based on multiple preset scale levels corresponding to each semantic layer point cloud, determine the neighborhood radius corresponding to each scale level; Based on each of the neighborhood radii, features corresponding to each of the scale levels are extracted respectively; Using the semantic category corresponding to the semantic layer point cloud as the root node and each scale level as a child node, the features corresponding to each scale level are attached to the corresponding child node to obtain the multi-scale hierarchical feature tree.
5. The positioning method as described in claim 4, characterized in that, The feature extraction based on each of the neighborhood radii for each scale level is specifically as follows: When the neighborhood radius is greater than or equal to a preset radius threshold, for each point in the semantic layer point cloud, the point cloud density and normal vector distribution entropy are calculated based on all neighboring points within the current neighborhood radius. The point cloud density and the normal vector distribution entropy are then concatenated to obtain the features at the first scale. When the neighborhood radius is less than a preset radius threshold, the target bounding box of the semantic layer point cloud is determined, and the size ratio of the target bounding box, the centroid coordinates of the point cloud, and the shape features based on the Fourier descriptor are calculated. The size ratio, the centroid coordinates of the point cloud, and the shape features are then combined to obtain the features of the second scale.
6. The positioning method according to any one of claims 1-5, characterized in that, The target point cloud data is divided based on preset semantic categories and the semantic levels corresponding to each semantic category to obtain at least one first semantic layer point cloud and at least one second semantic layer point cloud, specifically as follows: The target point cloud data is processed by a semantic classification model to obtain the semantic category label corresponding to each target point in the target point cloud data; The target points are divided according to the semantic category labels to obtain multiple sets of point cloud data; Based on the semantic level corresponding to each group of point cloud data, determine each first semantic layer point cloud and each second semantic layer point cloud.
7. The positioning method according to any one of claims 1-5, characterized in that, The acquisition of target point cloud data from the target sensor specifically involves: Acquire the raw point cloud data collected by the target sensor; Calculate the number of neighboring points of each point in the original point cloud data within a preset first neighborhood, and determine points whose number of neighboring points is less than a second preset threshold as noise points. Remove all noise points to obtain the denoised original point cloud data. The denoised original point cloud data is divided into multiple voxels, and the target point corresponding to each voxel is determined to obtain the target point cloud data.
8. A positioning system, characterized in that, include: The module includes an acquisition module, a semantic segmentation module, a construction module, and a location module. The acquisition module is used to acquire target point cloud data from the target sensor; The semantic segmentation module is used to segment the target point cloud data based on preset semantic categories and the semantic levels corresponding to each semantic category, so as to obtain at least one first semantic layer point cloud and at least one second semantic layer point cloud. The construction module is used to construct multi-scale hierarchical feature trees for each first semantic layer point cloud and each second semantic layer point cloud, respectively, to obtain a first feature tree and a second feature tree; The positioning module is used to perform a first match between the features at a first scale in the first feature tree and preset map data to obtain a candidate region, and to perform a second match between the features at a second scale in the second feature tree and the preset map data of the candidate region to obtain the positioning result of the target sensor, wherein the first scale is greater than the second scale.
9. A terminal device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements the positioning method as described in claim 1.
10. A computer-readable storage medium, characterized in that, include: A stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the positioning method as described in claim 1.