A global positioning method and device based on lidar data, medium and program product

Abstract

The application discloses a global positioning method and device based on laser radar data, a medium and a program product, and is applied to the technical field of global positioning, so as to avoid that high-precision global positioning cannot be realized in a complex scene. The horizontal plane of point cloud data is discretely processed based on a preset resolution, a BEV color image corresponding to each key frame is generated, scene map data is generated based on the pose data, the point cloud data and the BEV color image of the key frame, the coverage area of all key frames is calculated according to the pose data, the position index corresponding to each key frame is generated based on the coverage area, real-time laser point cloud data is acquired, the real-time laser point cloud data is converted into a target BEV image, the target BEV image is input into a pre-trained target classification network, a prediction result is obtained, a local map is constructed based on the prediction result, and the real-time laser point cloud data and the pre-constructed local map are registered to obtain a global positioning result.

Description

Technical Field

[0001] This invention relates to the field of global positioning technology, and in particular to a global positioning method, device, medium, and program product based on lidar data. Background Technology

[0002] Currently, global positioning typically refers to determining the absolute pose of a platform in a global coordinate system, i.e., its three-dimensional position and attitude. Unlike relative positioning or local tracking, the core task of global positioning is to match current observations with known maps, semantic landmarks, or topological structures through external sensing, thereby recovering the target's state within a unified reference frame. It is primarily used to provide the correct initial pose for SLAM (Simultaneous Localization and Mapping) or navigation systems to initiate or reset them; to achieve coordinate alignment during multi-machine collaboration and map sharing, enabling the fusion of local maps and observations from different platforms within the same global framework; and to project online-generated local maps onto a geographic reference frame to support long-term operation, asset management, and geographic information system integration. Therefore, global positioning is not only a fundamental function for achieving reliable autonomous navigation but also a key capability for building scalable, multi-platform, and long-term autonomous systems.

[0003] Existing global localization methods, such as the common point cloud to bird's-eye view (BEV) representation process, only use the three-dimensional coordinates (x, y, z) of points or project the point cloud into a single-channel grayscale image, ignoring information such as intensity, echo number, point density distribution, and height layering. This results in insufficient recognition in scenes with similar structures but different semantics or surface characteristics, thus increasing the false matching rate. Secondly, the traditional feature extraction, feature aggregation, and feature matching process, which is generally based on single-frame images, is extremely sensitive to changes in observation viewpoint, sensor pose, local occlusion, and dynamic objects. This manifests as a significant decrease in localization accuracy with viewpoint shifts or local scene changes, making it impossible to achieve high-precision global localization in complex scenes. Summary of the Invention

[0004] This invention provides a global positioning method, device, medium, and program product based on lidar data to improve the accuracy of global positioning in complex scenarios.

[0005] To address the aforementioned technical problems, this invention provides a global positioning method based on lidar data, comprising: Based on a preset laser SLAM algorithm, several key frames of historical laser point cloud data are extracted, and pose data and point cloud data of each key frame are obtained. The horizontal plane of the point cloud data is discretized based on a preset resolution to generate BEV color images corresponding to each key frame, and scene map data is generated based on the pose data of the key frames, point cloud data and BEV color images. The coverage area of ​​all keyframes is calculated based on the pose data, and the location index corresponding to each keyframe is generated based on the coverage area. Training data is generated based on the scene map data and the location index. The preset image classification neural network is trained and optimized based on the training data to obtain the target classification network; Acquire real-time laser point cloud data, convert the real-time laser point cloud data into a target BEV image; input the target BEV image into the target classification network to obtain prediction results; The prediction results are verified. If the verification is successful, a local map is constructed based on the prediction results, and the real-time laser point cloud data and the pre-constructed local map are registered to obtain the global positioning result.

[0006] This invention extracts several keyframes from historical laser point cloud data based on a preset laser SLAM algorithm and simultaneously acquires their pose and point cloud data. This ensures that scene map construction relies on the highly consistent pose results optimized over a long period of time by SLAM, guaranteeing spatial consistency between subsequent map data and localization results from the source. Secondly, by discretizing the horizontal plane of the keyframe point cloud data according to a preset resolution and generating BEV color images, the original 3D point cloud is transformed into a 2D multi-channel representation adapted to convolutional neural networks. This not only reduces data dimensionality and computational complexity but also fully preserves discriminative spatial features such as point density and reflectivity, thereby improving the ability to distinguish between different but structurally similar scenes. Thirdly, by calculating the coverage area based on the pose data of the keyframes and generating a location index, continuous spatial locations are discretized into finite location units, and training data is constructed accordingly. This transforms the global localization problem from a regression or feature matching problem into a stable classification problem. The problem is that the training and inference difficulty of the model are significantly reduced. Furthermore, by using training data to train and optimize the image classification neural network, the network can automatically learn the high-level semantic and spatial structure features in BEV images, avoiding reliance on manually designed features, thereby improving the generalization ability of location recognition and its adaptability to complex environments. Subsequently, the real-time laser point cloud is converted into the target BEV image and input into the trained target classification network for prediction, which can achieve fast end-to-end coarse localization of the current position, significantly shortening the global search time and improving real-time performance. Finally, by verifying the prediction results and constructing a local map when the verification is successful, and registering the real-time point cloud with the local map, not only can the localization risk caused by misclassification be effectively suppressed, but also the pose error can be further corrected through fine geometric registration, achieving high-precision and reliable global localization results. This balances localization speed and accuracy, and improves the overall stability and practicality of the system in complex scenarios.

[0007] Furthermore, the step of discretizing the horizontal plane of the point cloud data based on a preset resolution to generate BEV color images corresponding to each keyframe, and generating scene map data based on the pose data of the keyframes, the point cloud data, and the BEV color images, includes: The point cloud data of each key frame is downsampled based on a preset voxel grid, and a two-dimensional grid is constructed on the horizontal plane of the point cloud data according to a preset resolution. Calculate the point cloud features of the two-dimensional grid, and generate BEV color images corresponding to each keyframe based on the point cloud features; Scene map data is generated based on keyframe pose data, point cloud data, and BEV color images.

[0008] This invention achieves effective compression and structured representation of the original dense point cloud by first performing voxel downsampling on the keyframe point cloud and constructing a two-dimensional grid on a horizontal plane according to a preset resolution, then calculating the gridded point cloud features and generating BEV color images. On the one hand, voxel downsampling reduces computational and storage overhead and homogenizes the point distribution, which is beneficial to the stability of subsequent statistical features; on the other hand, generating BEV color images corresponding to each keyframe based on the point cloud features provides a compatible and more discriminative input representation for the convolutional neural network, improving the classifier's ability to distinguish similar structural scenes and the efficiency of utilizing intensity information, thereby improving the robustness of coarse localization and the real-time performance and scalability of the overall algorithm.

[0009] Furthermore, the step of calculating the coverage area of ​​all keyframes based on the pose data, generating a position index corresponding to each keyframe based on the coverage area, and generating training data based on the scene map data and the position index includes: Calculate the minimum rectangular bounding region of all keyframes based on the position and pose data of the keyframes, and determine the boundary and size of the minimum rectangular bounding region; The grid cell matrix is ​​divided into a horizontal plane within the area enclosed by the smallest rectangle based on the preset grid cell side length. A unique index is assigned to each cell in the grid cell matrix to generate a grid cell index mapping table; The location index corresponding to each keyframe is generated based on the grid cell index mapping table, and training data is generated based on the scene map data and the location index.

[0010] This invention achieves spatial discretization and standardized management of training samples by calculating the minimum rectangular bounding region based on keyframe pose, dividing the grid cell matrix with preset grid cell side lengths, assigning a unique index to each grid cell, and establishing a mapping table. This transforms the continuous spatial localization problem into a discrete classification problem, simplifying the model output dimension and inference complexity. Furthermore, the adjustable grid cell side lengths balance accuracy and class size, allowing for flexible configuration between localization resolution and training burden according to task requirements, facilitating deployment under different scene densities and resource constraints.

[0011] Furthermore, the step of generating location indices corresponding to each keyframe based on the grid cell index mapping table, and generating training data based on the scene map data and location indices, includes: Obtain the projection coordinates of each keyframe on the horizontal plane, and map the keyframes onto the grid cells based on the projection coordinates and the grid cell index mapping table to generate the position index corresponding to each keyframe; Location label categories are generated based on the location index corresponding to the keyframe, and training data is generated based on the scene map data and the location label categories.

[0012] This invention maps the projected coordinates of each keyframe on the horizontal plane to grid cells and uses the cell index as a location label to form location label categories. Training data is then generated based on the scene map and this label set, achieving consistency and repeatability in training set construction. Multiple BEV images within the same location cell are labeled as the same category, which can naturally integrate observations under different viewpoints, times, and occlusion conditions, improving the diversity and robustness of category samples. This results in the trained classification network having better generalization ability when facing viewpoint shifts, local occlusion, and short-term environmental changes. At the same time, the storage of the mapping table and metadata facilitates subsequent data balancing, sample selection, and online incremental expansion.

[0013] Furthermore, the prediction result is verified. If the verification is successful, a local map is constructed based on the prediction result, and the real-time laser point cloud data and the pre-constructed local map are registered to obtain a global positioning result, including: Obtain the prediction results, which include score vectors for each location category; If any score vector is greater than the preset threshold, the verification is successful, and the position category with the highest score is determined as the candidate position. A local map is constructed based on the candidate locations, and the real-time laser point cloud data and the local map are registered using a preset ICP algorithm to obtain global positioning results.

[0014] This invention utilizes a score threshold to effectively suppress erroneous localization caused by classifier misidentification, thereby reducing the false alarm rate. Furthermore, performing local map registration on candidates that pass the threshold not only verifies the classification results using geometric information but also significantly improves localization accuracy and confidence through fine point cloud registration, enhancing the reliability and security of the system in the presence of similar backgrounds or dynamic objects.

[0015] Furthermore, the step of constructing a local map based on the candidate locations and registering the real-time laser point cloud data and the local map using a preset ICP algorithm to obtain a global positioning result includes: Obtain the set of target keyframes corresponding to the candidate locations, and construct a local map centered on the set of target keyframes; Using the candidate location as an initial estimate, the real-time laser point cloud data and the local map are registered using a preset ICP algorithm to obtain a global positioning result.

[0016] This invention constructs a local map centered on a set of keyframes corresponding to candidate locations, and uses these candidate locations as initial estimates combined with the Iterative Closest Point (ICP) algorithm for registration, resulting in significant improvements in registration convergence and efficiency. The local map constructed with the keyframe set provides a rich reference geometry with location semantics, making ICP more likely to converge quickly and less prone to getting trapped in local minima under good initial conditions. The initial estimate reduces the registration search space and the number of iterations, improving the registration success rate and real-time performance. Simultaneously, using the keyframe set as a reference increases robustness to occlusion and local dynamic changes, comprehensively improving the stability and accuracy of the final global localization.

[0017] In a second aspect, the present invention provides a communication device including a module for performing the method.

[0018] Thirdly, the present invention provides a communication device, including a processor and an interface circuit, wherein the interface circuit is used to receive signals from other communication devices and transmit them to the processor or to send signals from the processor to other communication devices, and the processor is used to implement the method through logic circuits or execution code instructions.

[0019] Fourthly, the present invention provides a computer-readable storage medium storing a computer program or instructions that, when executed by a communication device, implement the method described therein.

[0020] Fifthly, the present invention provides a computer program product, including a computer program or instructions, which, when executed by a communication device, implement the method described therein. Attached Figure Description

[0021] Figure 1 A flowchart illustrating a global positioning method based on lidar data provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the training process of a target classification network provided in an embodiment of the present invention.

[0022] Figure 3 This is another flowchart illustrating a global positioning method based on lidar data provided in an embodiment of the present invention. Detailed Implementation

[0023] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0024] The terms "first" and "second," etc., in the specification, claims, and drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0025] 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. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0026] Example 1 See Figure 1 , Figure 1 This is a schematic flowchart illustrating a global positioning method based on lidar data, provided by an embodiment of the present invention. The embodiment of the present invention provides a global positioning method based on lidar data, including steps 101 to 106, as detailed below: Step 101: Extract several key frames from historical laser point cloud data based on the preset laser SLAM algorithm, and obtain the pose data and point cloud data of each key frame; In this embodiment, a laser inertial SLAM system, taking LIO-SAM as an example, is used to synchronize, distort, and match the original sensor stream with the front-end scan, thereby obtaining real-time inter-frame pose estimation.

[0027] In this embodiment, based on a preset keyframe selection strategy, such as: translational displacement exceeding 0.3–1.0 meters, rotation angle exceeding 2°–10°, or time interval exceeding 0.5–2 seconds since the last keyframe, or observation of a significant new structure, scans that meet the conditions are marked as keyframes.

[0028] In one optional embodiment, Lidar keyframes of historical laser point cloud data are obtained using a laser SLAM mapping algorithm.

[0029] In this embodiment, when a keyframe is selected, the point cloud data of that frame, after filtering and optional voxel downsampling, is encapsulated together with the real-time pose data and written into the local map database as a keyframe entry. Subsequently, the SLAM backend constructs a pose graph using keyframes as nodes and periodically performs graph optimization and loop closure detection, writing the globally consistent pose obtained by the backend optimization back to the keyframe table to ensure spatial consistency.

[0030] In this embodiment, each keyframe entry includes, but is not limited to: a keyframe identifier, a timestamp, a six-DOF pose in the global coordinate system, raw or downsampled point cloud data, sensor model and extrinsic parameters, parameters used for point cloud processing, and associated metadata, such as frame aggregation range and visibility mask. This provides accurate and searchable foundational data for subsequent map construction, training data generation, and online localization.

[0031] Step 102: Discretize the horizontal plane of the point cloud data based on a preset resolution to generate BEV color images corresponding to each key frame, and generate scene map data based on the pose data of the key frames, point cloud data and BEV color images. In this embodiment, the step of discretizing the horizontal plane of the point cloud data based on a preset resolution to generate BEV color images corresponding to each keyframe, and generating scene map data based on the pose data of the keyframes, the point cloud data, and the BEV color images, includes: The point cloud data of each key frame is downsampled based on a preset voxel grid, and a two-dimensional grid is constructed on the horizontal plane of the point cloud data according to a preset resolution. Calculate the point cloud features of the two-dimensional grid, and generate BEV color images corresponding to each keyframe based on the point cloud features; Scene map data is generated based on keyframe pose data, point cloud data, and BEV color images.

[0032] In this embodiment, the point cloud of each key frame is transformed to a local reference frame centered on the key frame according to the pose of the key frame, and the point cloud is cropped with a pre-set overlay window to define the region of interest; wherein, the overlay window can be a rectangular region centered on the key frame with a preset side length.

[0033] In this embodiment, the cropped point cloud is downsampled using a voxel grid based on a preset voxel size to homogenize the point distribution and reduce the computational burden. Then, a two-dimensional grid is divided on a horizontal plane according to the same voxel size or a preset resolution. The downsampled point cloud is mapped point by point to each grid cell, and the number of points and the reflectance set of each point are counted within each cell. For each cell, the number of cells is mapped to R-channel pixel values ​​based on a pre-determined normalized reference value, and the average reflectance of the cells is mapped to G-channel pixel values. The B-channel can be set to a constant value to form a three-channel input.

[0034] In this embodiment, taking point cloud P as an example, point cloud P is first downsampled using a voxel grid with a unit size of g meters to uniformly distribute the points; then, the xy plane of the point cloud is discretized into a grid with a resolution of g meters. Considering a cubic window centered at the origin with a side length of [-D meters, D meters], the BEV image I(u,v) can be regarded as having a size of The matrix contains RGB three channels, each channel being a pixel value normalized to the range [0,1].

[0035] In this embodiment, the R channel I of the BEV pixel value R (u,v) is calculated based on the normalized point density: (1) Where, N g This represents the number of points within the grid at position (u,v). This is the normalization factor, set to the maximum value of the point cloud density.

[0036] In this embodiment, the G channel I of the BEV pixel value G (u,v) is calculated using normalized reflectance: (2) in, (k=0,...,N g ) represents the reflectance of each point within the grid at location (u,v). max This is a fixed maximum reflectivity value for the lidar.

[0037] In this embodiment, the B channel I of the BEV pixel value R (u,v) are both set to a fixed value of 0.

[0038] In this embodiment, the generated BEV image, along with the pose data of the corresponding keyframe, the original or downsampled point cloud, the timestamp, and the metadata such as the normalization parameters and voxel grid resolution used, are written into the scene map database to form a searchable and traceable scene map entry.

[0039] This invention achieves effective compression and structured representation of the original dense point cloud by first performing voxel downsampling on the keyframe point cloud and constructing a two-dimensional grid on a horizontal plane according to a preset resolution, then calculating the gridded point cloud features and generating BEV color images. On the one hand, voxel downsampling reduces computational and storage overhead and homogenizes the point distribution, which is beneficial to the stability of subsequent statistical features; on the other hand, generating BEV color images corresponding to each keyframe based on the point cloud features provides a compatible and more discriminative input representation for the convolutional neural network, improving the classifier's ability to distinguish similar structural scenes and the efficiency of utilizing intensity information, thereby improving the robustness of coarse localization and the real-time performance and scalability of the overall algorithm.

[0040] Step 103: Calculate the coverage area of ​​all keyframes based on the pose data, generate the position index corresponding to each keyframe based on the coverage area, and generate training data based on the scene map data and the position index. In this embodiment, the step of calculating the coverage area of ​​all keyframes based on the pose data, generating a location index corresponding to each keyframe based on the coverage area, and generating training data based on the scene map data and the location index includes: Calculate the minimum rectangular bounding region of all keyframes based on the position and pose data of the keyframes, and determine the boundary and size of the minimum rectangular bounding region; The grid cell matrix is ​​divided into horizontal planes within the area enclosed by the smallest rectangle based on the preset grid cell side lengths. A unique index is assigned to each cell in the grid cell matrix to generate a grid cell index mapping table; The location index corresponding to each keyframe is generated based on the grid cell index mapping table, and training data is generated based on the scene map data and the location index.

[0041] In this embodiment, based on the pose projection of keyframes on the plane, the smallest rectangular enclosing region covering these keyframes is calculated, and the boundary position and size of the enclosing region are recorded. Then, according to the pre-set grid cell side length, the region is regularly divided into several grid cell matrices of equal size on the x–y plane (horizontal plane) of the enclosing region, and a unique index is assigned to each grid cell in the matrix to establish a grid cell index mapping table.

[0042] In this embodiment, generating location indices corresponding to each keyframe based on the grid cell index mapping table, and generating training data based on the scene map data and location indices, includes: Obtain the projection coordinates of each keyframe on the horizontal plane, and map the keyframes onto the grid cells based on the projection coordinates and the grid cell index mapping table to generate the position index corresponding to each keyframe; Location label categories are generated based on the location index corresponding to the keyframe, and training data is generated based on the scene map data and the location label categories.

[0043] In this embodiment, for each keyframe, its projected coordinates on the horizontal plane are read and the grid cell to which those coordinates belong is found, thereby determining the position index corresponding to the keyframe. When processing points located at cell boundaries, a predetermined boundary determination rule is adopted, such as preferentially assigning them to the side with the smaller index or taking the nearest cell, to ensure mapping consistency. Through the above mapping relationship, the BEV color images of all keyframes belonging to the same grid cell are uniformly labeled as the category label of that cell, and training sample pairs required for supervised learning are generated accordingly. The training sample pairs include BEV images and position indices.

[0044] This invention maps the projected coordinates of each keyframe on the horizontal plane to grid cells and uses the cell index as a location label to form location label categories. Training data is then generated based on the scene map and this label set, achieving consistency and repeatability in training set construction. Multiple BEV images within the same location cell are labeled as the same category, which can naturally integrate observations under different viewpoints, times, and occlusion conditions, improving the diversity and robustness of category samples. This results in the trained classification network having better generalization ability when facing viewpoint shifts, local occlusion, and short-term environmental changes. At the same time, the storage of the mapping table and metadata facilitates subsequent data balancing, sample selection, and online incremental expansion.

[0045] This invention achieves spatial discretization and standardized management of training samples by calculating the minimum rectangular bounding region based on keyframe pose, dividing the grid cell matrix with preset grid cell side lengths, assigning a unique index to each grid cell, and establishing a mapping table. This transforms the continuous spatial localization problem into a discrete classification problem, simplifying the model output dimension and inference complexity. Furthermore, the adjustable grid cell side lengths balance accuracy and class size, allowing for flexible configuration between localization resolution and training burden according to task requirements, facilitating deployment under different scene densities and resource constraints.

[0046] Step 104: Train and optimize the preset image classification neural network based on the training data to obtain the target classification network; Please refer to Figure 2 , Figure 2 This is a schematic diagram of the training process of a target classification network provided in an embodiment of the present invention.

[0047] In this embodiment, the BEV color images and their corresponding location indices are first divided into training set, validation set and test set according to a ratio.

[0048] In this embodiment, the image classification neural network preferably adopts a deep residual network, such as ResNet-50, as the backbone. The number of output units in the fully connected layers at the ends is consistent with the total number of position indices, and softmax is used to output the class probability vector. During training, cross-entropy loss or its weighted variant is selected, combined with Adam or SGD optimizers, learning rate preheating and cosine annealing and step decay learning rate scheduling strategies, and hyperparameters such as batch size, weight decay and momentum are tuned through cross-validation or grid search.

[0049] In this embodiment, to prevent overfitting, the loss and accuracy on the validation set should be monitored during training, and an early stopping strategy should be introduced. At the same time, the weights of the best-performing model on the validation set should be saved. After training, the classification accuracy, confusion matrix, Top-k accuracy, precision, recall and F1 score of the model should be evaluated using the test set. Based on the validation set results, the online inference threshold S_min and confidence calibration, such as temperature scaling, should be determined.

[0050] Step 105: Acquire real-time laser point cloud data, convert the real-time laser point cloud data into a target BEV image; input the target BEV image into the target classification network to obtain the prediction result; In this embodiment, real-time laser point cloud data is acquired, and time synchronization and distortion correction are performed on the real-time laser point cloud data to ensure the geometric consistency of the point cloud under a unified time reference frame. Then, the distortion-corrected point cloud is cropped according to a preset window range centered on a keyframe and transformed to the local reference frame of the keyframe. Based on the voxel size g used during training, voxel grid downsampling is performed on the cropped point cloud to homogenize the point distribution and reduce computational load. Next, a two-dimensional grid is constructed on the horizontal plane at resolution g. The downsampled point cloud is mapped to each grid cell, and the number of points and average reflectivity within each cell are counted. According to the normalization parameters determined during the training phase, the cell statistics are mapped to three-channel pixels. Cells without points are filled with default pixel values ​​agreed upon during the training phase. Finally, a BEV color image consistent with the one used during training is formed according to the row and column sequence, and the same pixel normalization and data type conversion are performed.

[0051] In this embodiment, the target BEV image is fed into the deployed target classification network in batch=1 for forward inference, and the network returns the score vector of each location category.

[0052] Step 106: Verify the prediction result. If the verification is successful, construct a local map based on the prediction result and register the real-time laser point cloud data and the pre-constructed local map to obtain the global positioning result.

[0053] In this embodiment, the step of verifying the prediction result, and if the verification is successful, constructing a local map based on the prediction result, and registering the real-time laser point cloud data and the pre-constructed local map to obtain a global positioning result, includes: Obtain the prediction results, which include score vectors for each location category; If any score vector is greater than the preset threshold, the verification is successful, and the position category with the highest score is determined as the candidate position. A local map is constructed based on the candidate locations, and the real-time laser point cloud data and the local map are registered using a preset ICP algorithm to obtain global positioning results.

[0054] In this embodiment, for the prediction output of the classification network, the prediction score vector is first read and sorted from largest to smallest to obtain several candidate positions; then, the candidate scores are judged by a preset confidence threshold S_min: if the score of any candidate is higher than S_min, the verification is passed and the candidate position with the highest score is selected as the primary candidate position; otherwise, it is judged as an unknown location and the failure handling strategy is triggered.

[0055] In this embodiment, after verification, a set of keyframes associated with the selected candidate location is retrieved, and the downsampled point clouds of these keyframes are stitched together according to their respective global poses to form a local reference point cloud map. Subsequently, the pose corresponding to the candidate location is used as the initial transformation estimate from the real-time point cloud to the local map. The real-time distortion-free point cloud is transformed to the local map coordinate system, and the ICP registration process is initiated. This invention utilizes a score threshold to effectively suppress erroneous localization caused by classifier misidentification, thereby reducing the false alarm rate. Furthermore, performing local map registration at candidate locations that pass the threshold not only verifies the classification results using geometric information but also significantly improves positioning accuracy and confidence through refined point cloud registration, enhancing the reliability and security of the system in the presence of similar backgrounds or dynamic objects.

[0056] In this embodiment, if the verification fails, it can be considered that the laser point cloud data is a new location not covered by the training database, and global positioning is rejected.

[0057] In this embodiment, the step of constructing a local map based on the candidate locations and registering the real-time laser point cloud data and the local map using a preset ICP algorithm to obtain a global positioning result includes: Obtain the set of target keyframes corresponding to the candidate locations, and construct a local map centered on the set of target keyframes; Using the candidate location as an initial estimate, the real-time laser point cloud data and the local map are registered using a preset ICP algorithm to obtain a global positioning result.

[0058] In this embodiment, a point-to-surface registration strategy with normal vectors is preferably adopted for registration to accelerate convergence and improve accuracy. During the registration process, KD-tree is applied to accelerate nearest point search, an iteration upper limit and a convergence threshold are set, and the registration quality index is evaluated after each iteration. When the registration reaches convergence and the quality index meets the preset acceptance conditions, such as the residual being lower than the threshold and the overlap rate being higher than the threshold, the registration result is used as the final global pose output.

[0059] This invention constructs a local map centered on a set of keyframes corresponding to candidate locations, and uses these candidate locations as initial estimates in conjunction with the ICP algorithm for registration, resulting in significant improvements in registration convergence and efficiency. The local map constructed using the keyframe set provides a rich reference geometry with location semantics, making ICP more likely to converge quickly and less prone to getting trapped in local minima under good initial conditions. The initial estimate reduces the registration search space and the number of iterations, improving the registration success rate and real-time performance. Simultaneously, using the keyframe set as a reference increases robustness to occlusion and local dynamic changes, comprehensively improving the stability and accuracy of the final global localization.

[0060] Please refer to Figure 3 , Figure 3 This is another flowchart illustrating a global positioning method based on lidar data provided in an embodiment of the present invention.

[0061] In this embodiment, firstly, in the offline stage, keyframe point clouds extracted by laser SLAM are used as raw data. The point cloud of each keyframe is projected and discretized into a color BEV image using a preset voxel and grid resolution. Simultaneously, this batch of BEV images is used to train a convolutional neural network, and the corresponding model weights are saved. After training, a classifier capable of mapping BEV images to discrete location categories is obtained. During online localization, the real-time acquired point cloud is also converted into a target BEV image and input into the trained network to obtain a location index. Based on this location index, the corresponding set of keyframe point clouds or a local map is retrieved from the scene map. Then, using the pose of this candidate location as an initial value, geometric registration algorithms such as ICP are used to register the real-time point cloud with the pre-constructed local point cloud map. Finally, a refined point cloud pose is output as the global localization result.

[0062] In this embodiment, several keyframes are extracted from historical laser point cloud data based on a preset laser SLAM algorithm, and their pose data and point cloud data are acquired simultaneously. This ensures that scene map construction relies on the highly consistent pose results after long-term optimization of SLAM, guaranteeing the spatial consistency between subsequent map data and positioning results from the source. Secondly, by discretizing the horizontal plane of the keyframe point cloud data according to a preset resolution and generating BEV color images, the original 3D point cloud is transformed into a 2D multi-channel representation adapted to convolutional neural networks. This not only reduces data dimensionality and computational complexity but also fully preserves discriminative spatial features such as point density and reflectivity, thereby improving the ability to distinguish between different but structurally similar scenes. Thirdly, by calculating the coverage area based on the pose data of the keyframes and generating a location index, continuous spatial locations are discretized into finite location units, and training data is constructed accordingly. This transforms the global localization problem from a regression or feature matching problem into a stable one. The classification problem significantly reduces the difficulty of model training and inference. Furthermore, by training and optimizing the image classification neural network using training data, the network can automatically learn high-level semantic and spatial structural features in BEV images, avoiding reliance on manually designed features, thereby improving the generalization ability of location recognition and its adaptability to complex environments. Subsequently, the real-time laser point cloud is converted into a target BEV image and input into the trained target classification network for prediction, enabling fast end-to-end coarse localization of the current location, significantly shortening the global search time and improving real-time performance. Finally, by verifying the prediction results and constructing a local map when verification is successful, and registering the real-time point cloud with the local map, not only can the localization risk caused by misclassification be effectively suppressed, but also pose errors can be further corrected through fine geometric registration, achieving high-precision and reliable global localization results. This balances localization speed and accuracy, and improves the overall stability and practicality of the system in complex scenarios.

[0063] The present invention provides a communication device, including a module for performing the method.

[0064] The present invention provides a communication device, including a processor and an interface circuit. The interface circuit is used to receive signals from other communication devices and transmit them to the processor, or to send signals from the processor to other communication devices. The processor is used to implement the method through logic circuits or execution code instructions.

[0065] The present invention provides a computer program product, including a computer program or instructions, which, when executed by a communication device, implement the method described therein.

[0066] In this embodiment of the invention, a terminal device is also provided, including 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 above-described global positioning method based on lidar data.

[0067] In this embodiment of the invention, a computer-readable storage medium is also provided, which includes a stored computer program, wherein the computer program controls the device where the computer-readable storage medium is located to execute the above-described global positioning method based on lidar data when it is running.

[0068] For example, a computer program can be divided into one or more modules, one or more of which are stored in memory and executed by a processor to perform the present invention. The one or more modules can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in a terminal device.

[0069] 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, memory, and display. Those skilled in the art will understand that the above components are merely examples of terminal devices and do not constitute a limitation on the terminal device. It may include more or fewer components, or combinations of certain components, or different components. For example, the terminal device may also include input / output devices, network access devices, buses, etc.

[0070] 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 through various interfaces and lines.

[0071] Memory can be used to store computer programs and / or modules. The processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, application programs required for at least one function (such as sound playback, text conversion, etc.), etc.; the data storage area can store data created based on the use of the mobile phone (such as audio data, text message data, etc.). In addition, memory can include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital cards (SD cards), flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0072] In this invention, the module for global positioning based on lidar data, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments 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 computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. Those skilled in the art can understand and implement this invention without any inventive effort.

[0073] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A global positioning method based on lidar data, characterized in that, include: Based on a preset laser SLAM algorithm, several key frames of historical laser point cloud data are extracted, and pose data and point cloud data of each key frame are obtained. The horizontal plane of the point cloud data is discretized based on a preset resolution to generate BEV color images corresponding to each key frame, and scene map data is generated based on the pose data of the key frames, point cloud data and BEV color images. The coverage area of ​​all keyframes is calculated based on the pose data, and the location index corresponding to each keyframe is generated based on the coverage area. Training data is generated based on the scene map data and the location index. The preset image classification neural network is trained and optimized based on the training data to obtain the target classification network; Acquire real-time laser point cloud data, convert the real-time laser point cloud data into a target BEV image; input the target BEV image into the target classification network to obtain prediction results; The prediction results are verified. If the verification is successful, a local map is constructed based on the prediction results, and the real-time laser point cloud data and the pre-constructed local map are registered to obtain the global positioning result.

2. The global positioning method based on lidar data as described in claim 1, characterized in that, The step involves discretizing the horizontal plane of the point cloud data based on a preset resolution to generate BEV color images corresponding to each keyframe, and generating scene map data based on the pose data of the keyframes, the point cloud data, and the BEV color images, including: The point cloud data of each key frame is downsampled based on a preset voxel grid, and a two-dimensional grid is constructed on the horizontal plane of the point cloud data according to a preset resolution. Calculate the point cloud features of the two-dimensional grid, and generate BEV color images corresponding to each keyframe based on the point cloud features; Scene map data is generated based on keyframe pose data, point cloud data, and BEV color images.

3. The global positioning method based on lidar data as described in claim 2, characterized in that, The step of calculating the coverage area of ​​all keyframes based on the pose data, generating a position index corresponding to each keyframe based on the coverage area, and generating training data based on the scene map data and the position index includes: Calculate the minimum rectangular bounding region of all keyframes based on the position and pose data of the keyframes, and determine the boundary and size of the minimum rectangular bounding region; The grid cell matrix is ​​divided into horizontal planes within the area enclosed by the smallest rectangle based on the preset grid cell side lengths. A unique index is assigned to each cell in the grid cell matrix to generate a grid cell index mapping table; The location index corresponding to each keyframe is generated based on the grid cell index mapping table, and training data is generated based on the scene map data and the location index.

4. The global positioning method based on lidar data as described in claim 3, characterized in that, The step of generating location indices for each keyframe based on the grid cell index mapping table, and generating training data based on the scene map data and location indices, includes: Obtain the projection coordinates of each keyframe on the horizontal plane, and map the keyframes onto the grid cells based on the projection coordinates and the grid cell index mapping table to generate the position index corresponding to each keyframe; Location label categories are generated based on the location index corresponding to the keyframe, and training data is generated based on the scene map data and the location label categories.

5. The global positioning method based on lidar data as described in claim 4, characterized in that, The step involves verifying the prediction result. If the verification is successful, a local map is constructed based on the prediction result, and the real-time laser point cloud data and the pre-constructed local map are registered to obtain a global positioning result, including: Obtain the prediction results, which include score vectors for each location category; If any score vector is greater than the preset threshold, the verification is successful, and the position category with the highest score is determined as the candidate position. A local map is constructed based on the candidate locations, and the real-time laser point cloud data and the local map are registered using a preset ICP algorithm to obtain global positioning results.

6. The global positioning method based on lidar data as described in claim 5, characterized in that, The process of constructing a local map based on the candidate locations and registering the real-time laser point cloud data and the local map using a preset ICP algorithm to obtain a global positioning result includes: Obtain the set of target keyframes corresponding to the candidate locations, and construct a local map centered on the set of target keyframes; Using the candidate location as an initial estimate, the real-time laser point cloud data and the local map are registered using a preset ICP algorithm to obtain a global positioning result.

7. A communication device, characterized in that, Includes a module for performing the method as described in any one of claims 1 to 6.

8. A communication device, characterized in that, The device includes a processor and an interface circuit, wherein the interface circuit is used to receive signals from other communication devices and transmit them to the processor or to send signals from the processor to other communication devices, and the processor is used to implement the method as described in any one of claims 1 to 6 through logic circuits or execution code instructions.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program or instructions, which, when executed by a communication device, implement the method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the communication device, the method as described in any one of claims 1 to 6 is implemented.

Images