Laser radar large scene repositioning method based on multi-line adaptive strategy

Abstract

This invention discloses a large-scene relocalization method for LiDAR based on a multi-line-count adaptive strategy, comprising: acquiring raw point cloud data collected by LiDAR and performing dynamic target removal to obtain a static point cloud; constructing a local sub-graph as a scene point cloud representation using the static point cloud based on the LiDAR's scan line count, projecting it as a multi-channel structured image, and extracting scene descriptors from it using a pre-trained deep neural network; performing nearest neighbor retrieval and local feature matching in a map database based on the scene descriptors to solve the current global pose and complete the relocalization. Through the scan line count adaptive strategy, a single frame of static point cloud is directly used for high-line-count LiDAR, while multiple frames of static point cloud are accumulated to construct a local sub-graph for low-line-count LiDAR, balancing the information integrity of low-line-count LiDAR with the computational efficiency of high-line-count LiDAR; and removing dynamic targets through range graph connected component segmentation and multi-feature scoring to reduce the interference of the dynamic environment on matching accuracy.

Description

Technical Field

[0001] This invention relates to the field of lidar positioning technology, and in particular to a lidar large-scene relocalization method based on a multi-line number adaptive strategy. Background Technology

[0002] 3D lidar directly acquires environmental geometric information, is unaffected by lighting, and is suitable for both day and night and complex environments. LiDAR-based relocation technology has become an important research direction for high-precision positioning in large urban scenes.

[0003] Early LiDAR relocalization methods relied on manually designed geometric features or global descriptors to encode the shape distribution, spatial structure, or projection information of point clouds for scene matching and recognition. While these methods are simple to implement and require no training, their matching performance is unstable in large-scale urban environments due to changes in viewpoint, occlusion, and repetitive structures. To address this issue, deep learning-based methods learn global descriptors with stronger discriminative capabilities through neural networks, improving relocalization accuracy and robustness in complex scenes. Given the massive volume of 3D point cloud data and the high computational complexity of direct processing, researchers have further converted point clouds into structured 2D representations such as distance maps or bird's-eye views, combining this with convolutional neural networks for feature extraction. This approach significantly improves processing speed while preserving spatial structure information, becoming the current mainstream technical approach.

[0004] The methods described above typically assume that point clouds have relatively stable density and distribution characteristics. In actual deployments, point cloud densities vary significantly among different line count lidar systems, making it difficult to adapt to a unified projection and encoding method. Low-line-count lidar systems exhibit sparse point clouds in single frames, resulting in incomplete representation of the environmental structure; high-line-count lidar systems suffer from data redundancy and high computational overhead. Existing methods are mostly designed for specific types of lidar, lacking stability under varying equipment conditions and failing to employ adaptive processing mechanisms.

[0005] In urban road environments, dynamic targets such as vehicles and pedestrians are ubiquitous. These dynamic objects form unstable structures in point clouds, and their direct participation in feature extraction and matching reduces the discriminative power of descriptors, affecting relocation accuracy. Semantic segmentation methods can identify dynamic targets, but their high computational complexity makes them difficult to meet real-time requirements. Geometric filtering methods have low computational cost, but they struggle to accurately distinguish between dynamic and static objects, resulting in significant issues of incorrect rejection or retention.

[0006] To address the issue of insufficient single-frame point cloud information in low-line-count LiDAR, existing multi-frame fusion methods rely on odometry or IMU for point cloud registration, leading to accumulated errors that amplify over time and impact repositioning accuracy. High-line-count LiDAR single-frame data already possesses strong expressive power; employing multi-frame fusion strategies would introduce unnecessary computational overhead. Existing technologies do not differentiate the applicability of single-frame processing versus multi-frame accumulation based on the LiDAR line count, making it difficult to simultaneously achieve both accuracy and efficiency across different devices.

[0007] The generation process of structured representation methods depends on the distribution characteristics of point clouds. Under different line count radars or different sampling densities, the structured representation after projection varies, and the feature representation of the same scene is inconsistent under different devices, affecting matching accuracy and system generalization ability.

[0008] Over long-term operation, the map size continues to grow. The fixed-interval keyframe strategy struggles to balance representational integrity and storage efficiency, and the lack of an efficient retrieval structure results in high matching computational overhead in large-scale maps, making it difficult to meet real-time positioning requirements. Summary of the Invention

[0009] Therefore, the technical problem to be solved by the present invention is to overcome the problems in the prior art, such as the lack of an adaptive processing mechanism for LiDAR with different line counts, the decrease in repositioning accuracy caused by dynamic target interference, the lack of local optimization in multi-frame accumulation, and the insufficient map management efficiency in large-scale scenarios.

[0010] To address the aforementioned technical problems, this invention provides a large-scene relocalization method for LiDAR based on a multi-line number adaptive strategy, comprising: S1: Acquire the raw point cloud data collected by the lidar, and perform dynamic target removal on the raw point cloud data to obtain a static point cloud; S2: Based on the number of scan lines of the LiDAR, a local sub-image is constructed using the static point cloud as a representation of the scene point cloud; let the number of LiDAR scan lines be... The preset line count threshold is The scene point cloud representation for: , in, This is the static point cloud of the current frame. For the first estimated by the odometer Frame to the Relative pose transformation of frames, This represents the cumulative frame count. When the number of scan lines is greater than a preset threshold, the static point cloud of the current frame is directly used as the scene point cloud representation; when the number of scan lines is less than or equal to the preset threshold, continuous scan lines are used. The static point cloud of a frame is accumulated into a local sub-image through pose transformation, which is used as the scene point cloud representation. S3: Project the scene point cloud representation into a structured representation, and extract scene descriptors from the structured representation; S4: Based on the scene descriptor, perform a search and matching in the map database to solve the current global pose and complete the relocalization.

[0011] In one embodiment of the present invention, the method for projecting the scene point cloud representation into a structured representation in step S3 is as follows: The scene point cloud representation is projected onto the ground plane and discretized into a two-dimensional regular grid. The point density distribution characteristics and height distribution characteristics of the point cloud within each grid are statistically analyzed. Based on the point density distribution characteristics and height distribution characteristics, a multi-channel structured image containing at least a point density channel and a height statistical channel is generated. In the multi-channel structured image, for sparse grid regions where the point density is lower than a preset value, adaptive smoothing processing is performed, in which the smoothing parameter adapts to the local point density.

[0012] In one embodiment of the present invention, the method for extracting scene descriptors from the structured representation in step S3 is as follows: The structured representation is processed using a pre-trained deep neural network to extract local feature maps. and global descriptor : , , in, This refers to the structured representation. For a convolutional encoder containing rotationally equivariant units, This is a global aggregation module. The number of local feature channels, This is a global description sub-dimension.

[0013] In one embodiment of the present invention, the method for solving the current global pose by searching and matching in the map database based on the scene descriptor in step S4 is as follows: Let the current scene descriptor be The historical description subset in the map database is ,in The prior pose of a historical keyframe or sub-image. For auxiliary information of local features, The number of historical descriptors in the map database; In the map database, the scene descriptor for the current frame Perform nearest neighbor search to obtain candidate matching frame indexes. The candidate matching frames are then obtained, wherein, This refers to Euclidean or cosine distance; Extract a set of key points from the structured representation corresponding to the candidate matching frame, and perform local feature descriptor matching with the set of key points extracted from the structured representation of the current frame; Based on the results of the local feature descriptor matching, the planar rigid body transformation between the current frame and the candidate matching frames is solved; Based on the planar rigid body transformation and the prior pose of the candidate matching frame Restore the global pose at the current moment. The formula is: , in, For pose vector, , For planar position coordinates, Yaw angle , , The relative pose components are obtained by solving the planar rigid body transformation. For pose composite operators.

[0014] In one embodiment of the present invention, after obtaining the raw point cloud data collected by the lidar in step S1, the method further includes: Based on the extrinsic parameters between the lidar and the vehicle body equipped with it, the original point cloud data is transformed from the lidar coordinate system to the vehicle coordinate system. The transformation formula is as follows: , in, For a moment The first collection The position of each point in the radar coordinate system The rotation matrix of the laser radar to the vehicle body, Let the laser radar be the translation vector to the vehicle body. These are the coordinates of the points after transformation to the vehicle coordinate system.

[0015] In one embodiment of the present invention, the method for dynamically removing targets from the original point cloud data in step S1 is as follows: projecting the original point cloud data onto a distance map, performing connected component segmentation on the distance map, and obtaining candidate target clusters; Each candidate target cluster is scored using multiple features. Based on the scoring results, dynamic target clusters are identified and eliminated to obtain the static point cloud.

[0016] In one embodiment of the present invention, the method for projecting the original point cloud data onto a distance map and performing connected component segmentation on the distance map to obtain candidate target clusters is as follows: Construct a distance graph ,in For scan line numbering, For azimuth index, To measure distance, Distance from the location on the map The corresponding original 3D points; Calculate adjacent pixels and Distance gradient between With height jump The formula is: , , in, For position z-axis coordinate; when and When two points belong to the same connected structure, they are determined to be disconnected on the distance graph; otherwise, they are disconnected. This is the distance gradient threshold. The threshold for height jump; After completing the connected component segmentation, the candidate target cluster is obtained.

[0017] In one embodiment of the present invention, the method for performing multi-feature scoring on each candidate target cluster, identifying and eliminating dynamic target clusters based on the scoring results, and obtaining the static point cloud is as follows: extracting size features, height features, and temporal consistency features for each candidate target cluster, and performing weighted scoring based on the size features, height features, and temporal consistency features; when the score of a candidate target cluster is higher than a preset scoring threshold, the candidate target cluster is determined to be a dynamic target and eliminated.

[0018] Based on the same inventive concept, this invention also provides a large-scene relocalization system for lidar based on a multi-line number adaptive strategy, comprising: The data acquisition and dynamic target removal module is used to acquire the raw point cloud data collected by the lidar, and to perform dynamic target removal on the raw point cloud data to obtain a static point cloud. The sub-graph construction module is used to construct a local sub-graph based on the number of scan lines of the LiDAR and the static point cloud, as a representation of the scene point cloud; let the number of scan lines of the LiDAR be... The preset line count threshold is The scene point cloud representation for: , in, This is the static point cloud of the current frame. For the first estimated by the odometer Frame to the Relative pose transformation of frames, This represents the cumulative frame count. When the number of scan lines is greater than a preset threshold, the static point cloud of the current frame is directly used as the scene point cloud representation; when the number of scan lines is less than or equal to the preset threshold, continuous scan lines are used. The static point cloud of a frame is accumulated into a local sub-image through pose transformation, which is used as the scene point cloud representation. The feature extraction module is used to project the scene point cloud representation into a structured representation and extract scene descriptors from the structured representation; The relocalization solution module is used to perform retrieval and matching in the map database based on the scene descriptor, solve the current global pose, and complete the relocalization.

[0019] The technical solution of the present invention has the following advantages compared with the prior art: This invention obtains a static point cloud by acquiring the original point cloud and performing dynamic target removal. A local sub-graph is adaptively constructed based on the number of LiDAR scan lines to represent the scene point cloud. This sub-graph is then projected into a multi-channel structured image, and scene descriptors are extracted. Global pose is then calculated by searching and matching against a map database. This method employs an adaptive strategy of single-frame processing or multi-frame accumulation for LiDARs with different line counts, balancing the information integrity of low-line-count LiDARs with the computational efficiency of high-line-count LiDARs. Dynamic targets are removed through range graph connectivity segmentation and multi-feature scoring, reducing the interference of the dynamic environment on matching accuracy from the source. The unordered point cloud is transformed into a regular structured representation, and rotation-robust descriptors are extracted using a deep network, reducing computational complexity while maintaining matching accuracy. Attached Figure Description

[0020] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0021] Figure 1 This is a flowchart illustrating the large-scene relocalization method for LiDAR based on a multi-line adaptive strategy provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the technical route of the local and remote management system for lidar in an embodiment of the present invention; Figure 3 This is a schematic diagram of the mobile robot platform structure in an embodiment of the present invention; Figure 4 This is a schematic diagram showing typical scenes and scale annotations in an embodiment of the present invention; Figure 5 This is a curve showing the accuracy-recall rate of loop closure detection in a KITTI urban scenario in this embodiment of the invention. Figure 6 This is a curve showing the accuracy-recall rate of loop closure detection in a park setting according to an embodiment of the present invention. Figure 7 This is an example diagram of city and park scene matching in an embodiment of the present invention; Figure 8 This is a global map display diagram in an embodiment of the present invention, where (a) and (b) are urban road scenes, and (c) and (d) are park scenes. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0023] Example 1: like Figure 1 As shown, this invention provides a large-scene relocalization method for LiDAR based on a multi-line number adaptive strategy, including: S1: Acquire the raw point cloud data collected by the lidar, and perform dynamic target removal on the raw point cloud data to obtain a static point cloud; S2: Based on the number of scan lines of the lidar, a local sub-image is constructed using the static point cloud as a representation of the scene point cloud; S3: Project the scene point cloud representation into a structured representation, and extract scene descriptors from the structured representation; S4: Based on the scene descriptor, perform a search and matching in the map database to solve the current global pose and complete the relocalization.

[0024] This invention provides a large-scene relocalization method for LiDAR based on a multi-line-count adaptive strategy. The method acquires raw point cloud data collected by LiDAR and performs dynamic target removal to obtain a static point cloud. Based on the LiDAR scan line count, a local sub-graph is adaptively constructed from the static point cloud as a representation of the scene point cloud. This sub-graph is projected into a structured representation, from which scene descriptors are extracted. Relocation is then completed by searching and matching in a map database to solve the global pose. This method, through a scan line count adaptive strategy, directly uses a single frame of static point cloud for high-line-count LiDAR and accumulates multiple frames of static point cloud to construct a local sub-graph for low-line-count LiDAR, balancing the information integrity of low-line-count LiDAR with the computational efficiency of high-line-count LiDAR. By performing dynamic target removal during the point cloud preprocessing stage, the interference of the dynamic environment on matching accuracy is reduced from the source. By projecting the point cloud into a structured representation and extracting scene descriptors for retrieval and matching, high-precision real-time relocalization is achieved, making it suitable for large-scale urban applications such as autonomous driving, unmanned inspection, and intelligent logistics.

[0025] In this embodiment of the invention, a three-dimensional lidar installed on a mobile vehicle body is used as the main sensing unit, and an industrial control computer is used as the local real-time computing carrier.

[0026] During the raw point cloud acquisition stage, a time frame is set. The collected raw point cloud data is denoted as ,in Indicates the first The position of each point in the radar coordinate system.

[0027] After receiving the point cloud data, the industrial control computer transforms the original point cloud data from the radar coordinate system to the vehicle coordinate system based on the extrinsic parameters between the LiDAR and the vehicle body equipped with the LiDAR. The transformation formula is as follows: ,in The rotation matrix of the laser radar to the vehicle body, Let the laser radar be the translation vector to the vehicle body. These are the coordinates of the points after transformation to the vehicle coordinate system. Coordinate unification ensures that all subsequent processing is performed in the vehicle coordinate system, eliminating the impact of differences in radar installation location on the algorithm.

[0028] Considering that the main deployment is on ground-based mobile robot platforms, and the moving subject satisfies approximately planar motion constraints, the global pose can be simplified to: ,in and For planar position coordinates, Using the yaw angle as the basis, the three-degree-of-freedom motion assumption can significantly reduce the computational complexity of pose recovery and improve the real-time performance of engineering applications.

[0029] In urban road environments, vehicles, pedestrians, and other moving objects introduce strong time-varying structures into the point cloud data acquired by LiDAR. If these dynamic targets directly participate in subsequent map construction and descriptor generation, it will lead to inconsistencies in the observation results of the same location at different times, thereby reducing the discriminative power of the descriptors and affecting the relocation accuracy. To address this issue, a fast dynamic target suppression mechanism is introduced, which effectively removes dynamic targets through two main steps: distance map segmentation and multi-feature scoring.

[0030] In the connected component segmentation stage, the original point cloud data is subjected to radius filtering and coarse ground separation to obtain a candidate static point set. With candidate dynamic point set This process initially separates obvious ground points from outliers, facilitating fine segmentation.

[0031] Furthermore, the original point cloud data is projected onto the distance map to construct the distance map. ,in The scan line number corresponds to the vertical channel index of the LiDAR. LiDARs with different line numbers have different numbers of scan lines. value, This is the azimuth index, corresponding to the scan point number of the lidar in the horizontal direction. For measuring distance, used to calculate points Radial distance to the radar center Distance from the location on the map The corresponding original 3D points.

[0032] After the distance map is constructed, a fast segmentation criterion is defined based on the distance gradient and height jump between adjacent pixels.

[0033] Specifically, calculate adjacent pixels and Distance gradient between The distance gradient reflects the degree of difference in radial depth between two adjacent points; simultaneously, the height jump is calculated. The height jump reflects the vertical elevation change between two adjacent points, where For position The elevation coordinates of the corresponding point.

[0034] By analyzing the changes in these two dimensions, it can be determined whether two adjacent points belong to the same object surface. When and When the changes in both radial depth and vertical height of adjacent pixels are within a preset threshold range, it can be determined that these two points belong to the same connected structure, i.e., the same object. Conversely, if the change in either dimension exceeds the corresponding threshold, it is determined that there is an object boundary between these two points, and they are separated on the distance map. This is the distance gradient threshold. These two thresholds, representing height abrupt changes, directly affect the precision of the segmentation. By using a connected component segmentation method based on distance gradient and height abrupt changes, it is possible to achieve higher precision. It completes approximate clustering of the entire frame point cloud within a time complexity of order of magnitude, efficiently obtaining a series of candidate target clusters.

[0035] Furthermore, for each candidate target cluster obtained from the segmentation... The size, height, and temporal consistency features of the target cluster are extracted, and a weighted scoring method is used to comprehensively determine whether the target cluster is a dynamic target.

[0036] Specifically, calculate the comprehensive score. ,in Size scoring measures whether the geometric dimensions of a cluster of targets fall within the size range of typical dynamic targets such as vehicles and pedestrians. The altitude score is used to measure whether the ground elevation of a target cluster matches the characteristics of a dynamic target. , , These are the weighting coefficients for the three scoring items mentioned above, which can be adjusted according to the actual application scenario.

[0037] In the above scoring, the time series consistency score It is a key indicator. Its calculation method is as follows: , in, This represents the BEV projection operator in a bird's-eye view, used to project a 3D point set onto a 2D ground plane. This represents the local static map constructed in the previous moment. To prevent a very small constant with a denominator of zero.

[0038] The essence of this formula is to calculate the complement of the overlap rate between the current candidate target cluster and the historical local static map on the BEV projection plane. Static objects, such as buildings, curbs, and traffic signs, maintain a stable position in observations at different times. Therefore, their current observation and the historical local static map have a high overlap rate on the BEV projection plane, resulting in a lower temporal consistency score. Dynamic objects, such as moving vehicles and pedestrians, have a lower overlap rate with the historical local static map due to their changing positions over time, resulting in a higher score. This is beneficial for effectively distinguishing between static structures and dynamic targets in a scene.

[0039] When the comprehensive score of the candidate target cluster Higher than the preset scoring threshold When a cluster is identified as a dynamic target, it is removed from the point set used for mapping and descriptor generation. This culling mechanism effectively removes interference from dynamic targets such as vehicles and pedestrians on scene representation while maintaining high real-time performance, ensuring that the final static point cloud primarily originates from stable geometric structures within the scene.

[0040] Furthermore, in step S2, a scene point cloud representation is constructed. Based on the number of scan lines of the LiDAR, a local sub-image is constructed using the static point cloud as the scene point cloud representation.

[0041] In this embodiment of the invention, differentiated scene point cloud representation construction strategies are adopted for LiDARs with different scan line counts. In actual engineering deployments, different types of LiDARs exhibit significant differences in scan line count, field of view, and point cloud density. When the radar line count is low, the vertical sampling of a single frame of static point cloud is relatively sparse. If it is directly projected to generate subsequent structured representations, mesh holes and structural gaps often occur, resulting in incomplete scene information representation and affecting the discriminative ability of subsequent scene descriptor extraction. Conversely, when the radar line count is high, a single frame of static point cloud already possesses sufficient information density and scene structure representation capability. If a multi-frame accumulation method is still used at this time, it will not only introduce additional computational overhead and data processing latency, but also reduce the quality of scene point cloud representation due to the continuous amplification of odometer accumulation error.

[0042] Therefore, by setting a threshold for the number of scan lines, the system can adaptively select whether to use a single frame of static point cloud directly as the scene point cloud representation, or to accumulate multiple consecutive frames of static point cloud into a local sub-image through pose transformation and then use it as the scene point cloud representation, based on the actual number of scan lines of the lidar.

[0043] Specifically, let the number of laser radar scan lines be... The preset line count threshold is Then the scene point cloud representation Determined according to the following adaptive construction strategy. When When the current LiDAR scan line count exceeds a preset threshold, falling into the category of high-line-count LiDAR, a single frame of static point cloud is sufficient to effectively represent the geometric structure information of the scene. Therefore, the current frame of static point cloud is directly used as the scene point cloud representation. .when When the current scan line count of the LiDAR is lower than or equal to a preset threshold, falling into the category of low-line-count LiDAR, the information content of a single frame of static point cloud is insufficient to fully represent the scene structure, and the system will continuously... The static point cloud of a frame is accumulated into a local sub-image through pose transformation as a representation of the scene point cloud, that is... In the above formula, Indicates the current frame, i.e., the time. The corresponding static point cloud, The first number estimated by the odometer Frame to the Relative pose transformation of frames, This indicates the number of frames in the continuous static point cloud that participated in the accumulation. Line count threshold. The preferred resolution is 16 lines. When the LiDAR has 16 lines or less, a multi-frame static point cloud accumulation method is used to construct a local sub-map. When the LiDAR has more than 16 lines, a single frame static point cloud is directly used as the scene point cloud representation. Through this adaptive judgment and differentiated processing based on the number of scan lines, the same relocalization framework can simultaneously meet the requirements of low-line-count LiDAR for scene information integrity and high-line-count LiDAR for real-time processing efficiency, achieving effective adaptation to the hardware characteristics of different types of LiDAR.

[0044] To further suppress error propagation introduced by low-line-count LiDAR in constructing local sub-maps using multi-frame static point cloud accumulation, a time-decay weight is introduced into the relative pose of each frame within the accumulation window. The relative pose within the local sub-map is then locally optimized and corrected by minimizing the weighted residual. During continuous multi-frame accumulation, the further away a frame is from the current time, the greater the uncertainty in its pose estimation, and its contribution to the construction of the scene point cloud representation should be correspondingly reduced.

[0045] Based on this pattern, the first participating in the accumulation... Frames are assigned decay weights ,in This is the decay coefficient, used to control the decay rate of the weight as the time interval increases. For the time index of the current frame, This represents the time index corresponding to each frame participating in the accumulation. The exponentially decaying weight function assigns higher weights to frames closer to the current time, while the weights of more distant frames decrease exponentially. This aligns with the general rule that recent observations are more reliable in motion estimation.

[0046] After obtaining the attenuation weights for each frame, the relative poses of each frame within the local sub-image are locally corrected by minimizing the weighted residuals. The objective function is: , In this optimization objective, This represents the relative pose consistency residual function, used to measure the degree of deviation in relative pose transformation. This represents the pose update operator, used to update the pose correction amount. Apply to the original relative pose transformation superior, The first one to be solved The pose correction amount corresponding to the frame.

[0047] Through the above-mentioned weighted residual minimization optimization process, the integrity of scene point cloud representation information can be improved while effectively controlling error propagation during multi-frame accumulation, thereby improving the pose consistency between frames within a local sub-image.

[0048] Furthermore, in step S3, the scene point cloud representation is projected into a structured representation, and then scene descriptors are extracted from the structured representation to provide a feature representation with strong discriminative ability for subsequent map database retrieval and matching.

[0049] The scene point cloud representation is projected onto the ground plane and discretized into a two-dimensional regular grid, connecting the disordered point cloud with the regularized convolutional network.

[0050] Specifically, let the BEV grid resolution be... The coverage area is , No. Each grid corresponds to a planar region. For scene point cloud representation Each three-dimensional point in Through the planar projection operator This is then mapped to the corresponding grid cells. To improve the ability to distinguish repetitive structures and complex urban boundaries, instead of simply retaining single point density statistics, a multi-channel structured image is constructed, namely the BEV tensor in the multi-channel bird's-eye view. ,in This is the point density channel, reflecting the density of the point cloud within each grid region; This is a height statistics channel, reflecting the elevation distribution characteristics of point clouds within each grid area; This is the local geometric variance channel, reflecting the degree of dispersion of the point cloud geometry within each grid region; This is the reflection intensity channel, which reflects the statistical characteristics of the reflection intensity of the point cloud within each grid region.

[0051] Dot density channels The definition of is: , in It is a normalization constant used to map density values ​​to a reasonable numerical range.

[0052] Altitude Statistics Channel The definition of is: , in To fall into the grid The point set, For point The z-axis coordinate is the elevation value. To prevent extremely small constants with a denominator of zero.

[0053] Local geometric variance channels The definition of is: , in Let be the covariance matrix of the point cloud within the grid. The three-dimensional mean of the point cloud within the grid. This represents the trace of the matrix. This channel encodes the discreteness of local geometry by summing the eigenvalues ​​of the covariance matrix. It responds well to structures with regular geometric features, such as building facades and road boundaries, but responds differently to uniformly distributed ground points or chaotic dynamic objects, which helps to improve the ability to distinguish scenes with complex urban boundaries and repetitive structures.

[0054] To address the sparsity issue that may arise in scene point cloud representation under low-line-count LiDAR conditions, this embodiment of the invention performs adaptive smoothing processing on the point density channel. When the LiDAR line count is low or there are distant regions in the scene, the point density statistics within some grids may fluctuate significantly due to the insufficient number of points, affecting the stability and discriminative ability of subsequent scene descriptor extraction.

[0055] Therefore, for point density channels Perform a Gaussian smoothing filter with smoothing parameters that adaptively change with local point density. The smoothed point density channels are: , in For The neighborhood centered on, The smooth bandwidth parameter is adaptively varied with local point density. This is a normalization factor to ensure that the sum of the smoothing weights is 1. When the local point density is high, Smaller values ​​result in weaker smoothing effects while preserving detailed structural information; when the local point density is low... Larger values ​​enhance the smoothing effect and effectively fill in the information gaps in sparse regions.

[0056] This density-adaptive smoothing strategy avoids the problem of reduced discriminative power caused by oversmoothing or undersmoothing in low-line-count LiDAR or long-range areas, resulting in better consistency of the structured representation generated under different device conditions.

[0057] Furthermore, the structured representation is processed using a pre-trained deep neural network, and a two-level feature representation is used to extract rotation-robust scene descriptors.

[0058] Through local feature encoder From structured representation Extracting local feature maps ,in and The spatial dimensions of the feature map. This represents the number of local feature channels. The local feature encoder is designed as a convolutional encoder containing rotationally equivariant units, making the encoder approximately equivariant in response to rotations of the BEV image. ,in For the planar rotation operator of the BEV image, This represents the corresponding rotational mapping in the feature space. This rotationally equivariant design ensures that when a mobile robot collects data at the same location with different orientations, the extracted local feature maps maintain a consistent spatial correspondence.

[0059] After extracting local feature maps, the global aggregation module is used. Aggregating local feature maps yields a global descriptor. ,in This refers to the global descriptor dimension. The global aggregation module eliminates the influence of rotation phase, ensuring that the final global descriptor has approximately rotation-invariant properties. This further enhances the matching stability of the scene descriptor under varying yaw angle conditions.

[0060] During the model training phase, the deep neural network is trained using a weakly supervised learning approach based on triplet loss, requiring no precise pose ground truth supervision. Let the descriptors for the query sample, positive sample, and negative sample be respectively... , , The triplet loss function is defined as follows: ,in The interval is a constant. This loss function minimizes the distance between the query sample and the descriptor of the positive sample (i.e., the same location), and maximizes the distance between the query sample and the descriptor of the negative sample (i.e., different locations). By minimizing this loss function, the network can learn scene descriptors with strong discriminative power. Weakly supervised training based on coarse-grained location labels makes the model easier to deploy in engineering projects, without relying on high-precision pose ground truth annotations.

[0061] Further, in step S4, relocalization is performed. The scene descriptor of the current frame is searched and matched in the map database to solve for the current global pose.

[0062] Let the scene descriptor of the current frame be... The historical description subset stored in the map database is ,in For the first A scene description of a historical keyframe or sub-graph. This refers to the prior pose corresponding to the historical keyframe or subgraph. For auxiliary information of local features, This represents the total number of historical descriptors in the map database. The map database can be deployed on a host computer management platform for global storage and indexing, or it can maintain a subset of commonly used scenarios in the local cache of an industrial control computer to balance the retrieval coverage of large-scale scenarios with the response speed of real-time queries.

[0063] Specifically, in the map database, the scene descriptor of the current frame Perform nearest neighbor search to obtain candidate matching frame indexes. ,in This is the distance metric function, which can be either Euclidean distance or cosine distance. When using Euclidean distance... When using cosine distance, Both distance metrics have their applicable scenarios. Euclidean distance is sensitive to the absolute magnitude of descriptors, while cosine distance focuses more on the directional consistency of descriptors. The choice can be made based on the data characteristics of the actual deployment environment. This nearest neighbor search step transforms the problem of matching high-dimensional scene descriptors into an efficient distance ranking problem. By adopting a vector index structure (such as KD-tree, LSH, or a quantized approximate nearest neighbor retrieval structure), sublinear retrieval complexity can be achieved. Even when the map database reaches a scale of hundreds of thousands, it can still maintain a millisecond-level retrieval speed, meeting the real-time requirements of online relocalization for mobile robots.

[0064] After obtaining candidate matching frames, keypoint sets are extracted from the structured representation corresponding to the candidate matching frames, and keypoint sets are also extracted from the structured representation of the current frame, based on the local feature maps extracted in step S3 by the pre-trained deep neural network. and Local feature descriptor matching is performed on the two sets of key points. Let the set of key points extracted from the query frame BEV be... The set of key points extracted from candidate frame BEV is By calculating the distance between the local feature descriptors corresponding to each key point, a correspondence between two sets of key points is established. Based on the point pair correspondence obtained by matching the local feature descriptors, the RANSAC algorithm is used to solve the planar rigid body transformation between the current frame and the candidate matching frame. This planar rigid body transformation is performed by a 2D rotation matrix. and a 2D translation vector Composed of, its matrix form is The RANSAC algorithm iteratively estimates the optimal value through a random sampling consensus strategy. The parameters enable robust recovery of the relative pose relationship between two frames even with a small number of mismatched point pairs.

[0065] Furthermore, based on the planar rigid body transformation and the prior pose of the candidate matching frame... Restore the global pose at the current moment. .in For pose vector, and These are the planar position coordinates of the mobile robot in the global coordinate system. Yaw angle , , The relative pose components from the candidate matching frame coordinate system to the current frame coordinate system are obtained from the planar rigid body transformation solution. For pose composite operators.

[0066] Because the entire pose recovery process employs a two-stage hierarchical framework combining global descriptor retrieval and local feature matching, the search range is quickly narrowed down to a single or a few candidate matching frames through nearest neighbor retrieval using global descriptors. Then, the relative pose is accurately recovered through local feature matching and geometric verification. Pose calculation only needs to be performed on a three-degree-of-freedom plane, resulting in a small number of parameters and low computational complexity, making it suitable for real-time execution on an industrial control computer embedded platform. After solving for the current global pose, accurate localization of the current frame within a large scene map is achieved.

[0067] like Figure 2As shown, to further illustrate the feasibility of the actual deployment of the method in this embodiment of the invention, the system hardware composition is described from the perspective of experimental platform construction. The system adopts a layered architecture of local computing by the mobile robot and remote management by the host computer. The whole system includes a mobile robot platform, a lidar sensor module, a multi-sensor auxiliary module, an industrial control computer local computing unit, a communication module, and a host computer management platform. Each part is connected and interacts with data through standard interfaces.

[0068] The mobile robot platform uses a differential drive or wheeled mobile robot as the system carrier, possessing stable motion control capabilities. LiDAR and other sensors are fixedly mounted on the top of the robot or at designated support positions to ensure unobstructed field of view and reduce vibration interference. The LiDAR sensor module uses a 3D LiDAR as the primary environmental perception device. Depending on application requirements, different line counts of LiDAR, such as 16-line, 32-line, or 64-line LiDARs, can be selected. The LiDAR connects to the industrial control computer via an Ethernet interface, outputting environmental point cloud data in real time. The multi-sensor auxiliary module includes an inertial measurement unit (IMU) and a wheel velocimeter. The IMU measures the platform's angular velocity and linear acceleration to assist in estimating short-term motion states; the wheel velocimeter acquires the linear velocity information of the mobile platform, providing initial pose estimation for point cloud stitching and sub-graph construction. These sensors connect to the industrial control computer via serial ports or CAN buses, and their data undergoes local time synchronization and fusion processing to support multi-frame accumulation for low-line-count LiDAR.

[0069] The industrial control computer's local computing unit, serving as the core computing unit of the system, employs an embedded computing platform with CPU and GPU acceleration capabilities. It is responsible for executing the entire relocation algorithm process described in this embodiment, including point cloud data reception and preprocessing, rapid dynamic target removal, adaptive point cloud construction, bird's-eye view generation, depth feature extraction and descriptor calculation, and descriptor matching and pose estimation. To ensure real-time performance, all core algorithms are executed locally on the industrial control computer, with only key results such as keyframes, sub-maps, and their descriptors uploaded to the host computer, thereby reducing communication burden. The communication module facilitates data interaction between the industrial control computer and the host computer via wireless LAN, 4G / 5G cellular networks, or industrial Ethernet. It primarily uploads locally generated keyframes or sub-map data to the host computer and receives map indexes or matching results returned by the host computer. The host computer management platform, acting as the global management center, is responsible for large-scale map data storage, keyframe and sub-map index construction, descriptor database maintenance, and global retrieval and matching support. The host computer receives keyframes or sub-map data uploaded from the industrial control computer and organizes them into a structured map database. By establishing efficient indexes such as vector retrieval or hash structures, it can quickly query and match large-scale scene data. At the same time, it can also optimize and update historical data to achieve map maintenance and expansion under long-term operating conditions.

[0070] In actual operation, the modules work together: the mobile robot moves in the environment, and the LiDAR collects point cloud data; the point cloud and auxiliary sensor data are transmitted to the industrial control computer; the industrial control computer executes the relocation algorithm of this embodiment to obtain the current pose; key frames or sub-maps are generated according to the strategy and uploaded to the host computer; the host computer stores and manages the data and supports subsequent relocation queries.

[0071] Example 2: Based on the same inventive concept as Embodiment 1, the present invention also provides a large-scene relocalization system for lidar based on a multi-line number adaptive strategy, comprising: The data acquisition and dynamic target removal module is used to acquire the raw point cloud data collected by the lidar, and to perform dynamic target removal on the raw point cloud data to obtain a static point cloud. The sub-graph construction module is used to construct a local sub-graph based on the number of scan lines of the lidar using the static point cloud, as a representation of the scene point cloud; The feature extraction module is used to project the scene point cloud representation into a structured representation and extract scene descriptors from the structured representation; The relocalization solution module is used to perform retrieval and matching in the map database based on the scene descriptor, solve the current global pose, and complete the relocalization.

[0072] The data acquisition and dynamic target removal module acquires raw point cloud data collected by the LiDAR and performs dynamic target removal on the raw point cloud data to obtain a static point cloud. This module first transforms the raw point cloud data from the LiDAR coordinate system to the vehicle coordinate system based on the extrinsic parameters between the LiDAR and the vehicle body mounting it, eliminating the influence of differences in radar installation position on subsequent processing. After coordinate transformation, the module projects the raw point cloud data onto a distance map and performs connected component segmentation on the distance map based on the distance gradient and height jump of adjacent pixels to obtain candidate target clusters. Subsequently, it extracts size features, height features, and temporal consistency features from each candidate target cluster and performs a weighted score. When the score of a candidate target cluster exceeds a preset score threshold, the corresponding candidate target cluster is determined to be a dynamic target cluster and removed. The final retained static point cloud mainly comes from stable geometric structures in the scene. The static point cloud output by this module will serve as input data for the subgraph construction module.

[0073] The sub-image construction module is used to construct local sub-images based on the number of scan lines of the LiDAR and the static point cloud output by the data acquisition and dynamic target culling module, serving as the scene point cloud representation. This module is the core adaptive unit of this system. By setting a scan line number threshold, it adaptively selects the construction method of the scene point cloud representation according to the actual scan line number of the LiDAR. When the scan line number is higher than the preset line number threshold, the static point cloud of the current frame is directly used as the scene point cloud representation; when the scan line number is lower than or equal to the preset line number threshold, multiple consecutive frames of static point clouds are accumulated into a local sub-image through relative pose transformation estimated by odometry, serving as the scene point cloud representation. To suppress error propagation during the multi-frame accumulation process, this module also introduces time decay weights to the relative pose of each frame within the accumulation window, and performs local optimization and correction of the relative pose within the local sub-image by minimizing the weighted residual, thereby improving the information integrity and pose consistency of the scene point cloud representation.

[0074] The feature extraction module projects the scene point cloud representation output by the subgraph construction module into a structured representation and extracts scene descriptors from the structured representation. This module first projects the scene point cloud representation onto a ground plane, discretizes it into a two-dimensional regular grid, and statistically analyzes the point density and height distribution features within each grid to generate a multi-channel structured image containing at least a point density channel and a height statistical channel. For sparse grid regions where the point density is below a preset value, adaptive smoothing processing is performed, where the smoothing parameter adapts to the local point density. After generating the structured representation, this module uses a pre-trained deep neural network for processing. It extracts local feature maps using a convolutional encoder containing rotationally equivariant units, and then aggregates these local feature maps using a global aggregation module to obtain a rotationally robust global descriptor as the scene descriptor.

[0075] The relocalization solution module is used to search and match the scene descriptor output by the feature extraction module in the map database to solve the current global pose and complete the relocalization. This module first performs a nearest neighbor search on the scene descriptor of the current frame in the map database to obtain candidate matching frame indices and determine the candidate matching frames. Then, it extracts the keypoint set from the structured representation corresponding to the candidate matching frames and performs local feature descriptor matching with the keypoint set extracted from the structured representation of the current frame. Based on the matching result, it solves the planar rigid body transformation between the current frame and the candidate matching frames. Finally, based on this planar rigid body transformation and the prior pose of the candidate matching frames, it recovers the current global pose through pose composition operations.

[0076] Example 3: To verify the effectiveness of the LiDAR large-scene relocalization method based on multi-line number adaptive strategy proposed in this invention, a complete experimental system was constructed and tested in a real environment.

[0077] This invention presents an experimental system built on a wheeled mobile robot platform. Its hardware components include a 3D LiDAR, an inertial measurement unit (IMU), a wheel velocities meter (WVTM), an industrial control computer (ICC), a host computer server, and a communication module. A 16-line and a 64-line IMU are used for comparative experiments. The 16-line IMU is used to verify the multi-frame accumulation effect under low line count conditions, while the 64-line IMU is used to verify the real-time performance and accuracy of single-frame processing. The IMU acquires the platform's angular velocity and acceleration information, and the WVTM acquires the platform's linear velocity information, providing initial pose estimation for point cloud stitching and sub-graph construction. The IMU, configured with an Intel i7 processor and an NVIDIA embedded GPU, serves as the local computing unit and is used to run the algorithm of this invention in real time. The host computer server, configured with a multi-core CPU and a high-performance GPU, is used for map storage and management. The communication module uses a wireless LAN to achieve data transmission between the IMU and the host computer. The LiDAR is mounted on the top of the mobile robot platform to ensure an unobstructed field of view. The IMU and WVTM are installed inside the chassis, and the IMU is fixed inside the robot, connecting to the sensors via Ethernet or serial ports. like Figure 3 The schematic diagram of the mobile robot platform shown is as follows: Figure 4 The image shows a typical experimental scenario and scale annotations.

[0078] In terms of software and algorithm implementation, the system software is developed based on the Ubuntu operating system and implemented using a hybrid programming approach of C++ and Python. Point cloud processing and coordinate transformation are completed based on the ROS framework, and the deep learning model is implemented based on PyTorch.

[0079] In this embodiment of the invention, the main algorithm parameters are set as follows: the BEV grid resolution is 0.2 meters, the BEV coverage area is [-50m, 50m] × [-50m, 50m]; the distance gradient threshold for dynamic object removal is 0.5 meters, and the height difference threshold is 0.3 meters; the cumulative frame count of the low-line-count radar sub-image is 5 to 10 frames, the keyframe selection interval is 10 frames, and the descriptor dimension is 256. The deep learning model adopts a convolutional network structure based on bird's-eye view, with multi-channel BEV images as input and global descriptors as output. After pre-training on public datasets such as KITTI, the model is deployed to an industrial control computer for online inference.

[0080] In actual operation, the system executes the following steps: The mobile robot moves in the environment, and the LiDAR collects point cloud data in real time; the industrial control computer receives the point cloud data and performs preprocessing, including coordinate transformation, filtering, and dynamic object removal; the point cloud construction method is selected according to the number of LiDAR lines. For a 16-line LiDAR, odometry is used to register and accumulate multiple consecutive frames of point cloud data to construct a local sub-map; for a 64-line LiDAR, a single frame of point cloud is directly processed; the processed point cloud is projected as a BEV image and multi-channel input data is generated; the BEV image is input into a deep learning model to extract global descriptors; descriptor matching is performed in the local or host computer database to obtain candidate positions; the relative pose is estimated through local feature matching and the RANSAC method to obtain the current global localization result; keyframes or sub-maps are generated according to the strategy and uploaded to the host computer via the network; the host computer stores and indexes the received data for subsequent relocalization queries.

[0081] The embodiments of this invention were tested in typical scenarios such as urban roads and industrial parks. Figure 4 As shown, the urban road scene covers an area of ​​approximately 3 kilometers and is characterized by regular road structures, dense intersections, and abundant dynamic targets; the park scene covers an area of ​​approximately 800 meters and is characterized by irregular structures, strong local repetition, and numerous weakly structured areas. The two types of scenes differ significantly in spatial structural complexity and point cloud distribution characteristics, placing higher demands on the generalization ability of the relocalization algorithm.

[0082] Table 1 shows the comparison results of matching performance in multiple KITTI city scenarios (sequences 00, 02, 05, and 08). The method of this invention achieves significantly higher average accuracy than the comparative methods in all sequences, reaching 0.887 in sequence 00 and 0.960 in sequence 08. Compared to the traditional BoW3D method, the matching accuracy is improved by more than 3 times; compared to the STD method based on structure description, the accuracy is improved by more than 2 times. In terms of computational efficiency, the average matching time of the method of this invention remains between 3 and 14 milliseconds, significantly reducing computational overhead compared to approximately 80 to 90 milliseconds for the STD method, thus maintaining high accuracy while exhibiting good real-time performance.

[0083] Table 1:

[0084] like Figure 5 The curves shown are the corresponding precision-recall curves. The curves of the method of this invention are generally higher than those of other methods. It can still maintain high precision in the high recall region (Recall greater than 0.6) and quickly reach precision close to 1 in the low recall region, indicating that the algorithm has good robustness and the descriptors have stronger discriminative ability.

[0085] Table 2 shows the experimental results on the park scene dataset (sequences 00, 01, 02, and 03). The method of this invention achieves an average accuracy close to 1 in all sequences, reaching a maximum of 0.999, significantly outperforming the BoW3D and STD methods. It maintains extremely high matching accuracy even in complex structures and weakly textured regions. In terms of time performance, the average matching time of this method is approximately 3 to 10 milliseconds, significantly better than the approximately 130 to 150 milliseconds of the STD method, comparable to the BoW3D method but with a substantial improvement in accuracy.

[0086] Table 2:

[0087] like Figure 6 As shown in the figure, the PR curve further demonstrates that the method of the present invention maintains high accuracy throughout the entire recall range, the curve is close to the ideal state, the false match rate is extremely low, and it still has good stability in complex environments.

[0088] like Figure 7 As shown, this paper presents a matching example in urban road and park scenarios. The matching relationship between the source BEV and the target BEV is shown by connecting the lines in the figure. In the urban road scenario, the matching results show good geometric consistency along the road structure. In the complex structure scenario of the park, the matching points can still be correctly aligned with the key structural areas. The matching distribution is uniform and there is no obvious drift, which shows that the descriptor has good spatial consistency and verifies the effectiveness of the two-stage strategy of global retrieval and local registration.

[0089] like Figure 8 As shown, global point cloud maps constructed under different scenarios are illustrated, where (a) and (b) are global point cloud maps for urban road scenarios, and (c) and (d) are global point cloud maps for park scenarios. The map structure is continuous and without obvious breaks, indicating that the relocalization results are stable. No significant drift occurred under long-distance operation at the kilometer level, and the complex areas of the park still maintain structural consistency, verifying that the hierarchical map management mechanism proposed in this invention can effectively support long-term operation in large-scale scenarios.

[0090] Experimental results show that under low-line-count LiDAR conditions, constructing a sub-map through multi-frame accumulation significantly improves point cloud density, resulting in a marked increase in relocation success rate. Under high-line-count LiDAR conditions, single-frame processing significantly reduces computation time while maintaining accuracy, enabling real-time relocation. Introducing dynamic object culling significantly improves descriptor matching stability and reduces mismatch rate. Through the host computer map management mechanism, the system maintains stable relocation performance and effectively controls data scale under long-term operation. As can be seen from the above embodiments, the relocation method proposed in this invention can operate stably under different types of LiDAR conditions and maintain high positioning accuracy in complex dynamic environments. Furthermore, the system architecture of local computation and host computer collaborative management effectively improves the system's scalability and engineering practicality.

[0091] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0092] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0093] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0094] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0095] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A large-scene relocalization method for LiDAR based on a multi-line number adaptive strategy, characterized in that, include: S1: Acquire the raw point cloud data collected by the lidar, and perform dynamic target removal on the raw point cloud data to obtain a static point cloud; S2: Based on the number of scan lines of the LiDAR, a local sub-image is constructed using the static point cloud as a representation of the scene point cloud; let the number of LiDAR scan lines be... The preset line count threshold is The scene point cloud representation for: ， in, This is the static point cloud of the current frame. For the first estimated by the odometer Frame to the Relative pose transformation of frames, This represents the cumulative frame count. When the number of scan lines is greater than a preset threshold, the static point cloud of the current frame is directly used as the scene point cloud representation; when the number of scan lines is less than or equal to the preset threshold, continuous scan lines are used. The static point cloud of a frame is accumulated into a local sub-image through pose transformation, which is used as the scene point cloud representation. S3: Project the scene point cloud representation into a structured representation, and extract scene descriptors from the structured representation; S4: Based on the scene descriptor, perform a search and matching in the map database to solve the current global pose and complete the relocalization.

2. The large-scene relocalization method for LiDAR based on a multi-line adaptive strategy according to claim 1, characterized in that: In step S3, the method for projecting the scene point cloud representation into a structured representation is as follows: The scene point cloud representation is projected onto the ground plane and discretized into a two-dimensional regular grid. The point density distribution characteristics and height distribution characteristics of the point cloud within each grid are statistically analyzed. Based on the point density distribution characteristics and height distribution characteristics, a multi-channel structured image containing at least a point density channel and a height statistical channel is generated. In the multi-channel structured image, for sparse grid regions where the point density is lower than a preset value, adaptive smoothing processing is performed, in which the smoothing parameter adapts to the local point density.

3. The large-scene relocalization method for LiDAR based on a multi-line adaptive strategy according to claim 1, characterized in that: In step S3, the method for extracting scene descriptors from the structured representation is as follows: The structured representation is processed using a pre-trained deep neural network to extract local feature maps. and global descriptor : ， ， in, This refers to the structured representation. For a convolutional encoder containing rotationally equivariant units, For global aggregation module, The number of local feature channels, This is a global description sub-dimension.

4. The large-scene relocalization method for LiDAR based on a multi-line adaptive strategy according to claim 1, characterized in that: In step S4, the method for retrieving and matching the current global pose based on the scene descriptor in the map database is as follows: Let the current scene descriptor be The historical description subset in the map database is ,in The prior pose of a historical keyframe or sub-image. For auxiliary information of local features, The number of historical descriptors in the map database; In the map database, the scene descriptor for the current frame Perform nearest neighbor search to obtain candidate matching frame indexes. The candidate matching frames are then obtained, wherein, This refers to Euclidean or cosine distance; Extract a set of key points from the structured representation corresponding to the candidate matching frame, and perform local feature descriptor matching with the set of key points extracted from the structured representation of the current frame; Based on the results of the local feature descriptor matching, the planar rigid body transformation between the current frame and the candidate matching frames is solved; Based on the planar rigid body transformation and the prior pose of the candidate matching frame Restore the global pose at the current moment. The formula is: ， in, For pose vector, , For planar position coordinates, Yaw angle , , The relative pose components are obtained by solving the planar rigid body transformation. For pose composite operators.

5. The large-scene relocalization method for LiDAR based on a multi-line adaptive strategy according to claim 1, characterized in that: In step S1, after acquiring the raw point cloud data collected by the lidar, the following steps are also included: Based on the extrinsic parameters between the lidar and the vehicle body equipped with it, the original point cloud data is transformed from the lidar coordinate system to the vehicle coordinate system. The transformation formula is as follows: ， in, For a moment The first collection The position of each point in the radar coordinate system The rotation matrix of the laser radar to the vehicle body, Let the laser radar be the translation vector to the vehicle body. These are the coordinates of the points after transformation to the vehicle coordinate system.

6. The large-scene relocalization method for LiDAR based on a multi-line adaptive strategy according to claim 1, characterized in that: In step S1, the method for dynamically removing targets from the original point cloud data is as follows: project the original point cloud data onto a distance map, perform connected component segmentation on the distance map, and obtain candidate target clusters; Each candidate target cluster is scored using multiple features. Based on the scoring results, dynamic target clusters are identified and eliminated to obtain the static point cloud.

7. The large-scene relocalization method for LiDAR based on a multi-line number adaptive strategy according to claim 6, characterized in that: The method for projecting the original point cloud data onto a distance map and performing connected component segmentation on the distance map to obtain candidate target clusters is as follows: Construct a distance graph ,in For scan line numbering, For azimuth index, To measure distance, Distance from the location on the map The corresponding original 3D points; Calculate adjacent pixels and Distance gradient between With height jump The formula is: ， ， in, For position z-axis coordinate; when and When two points belong to the same connected structure, they are determined to be disconnected on the distance graph; otherwise, they are disconnected. This is the distance gradient threshold. The threshold for height jump; After completing the connected component segmentation, the candidate target cluster is obtained.

8. The large-scene relocalization method for LiDAR based on a multi-line adaptive strategy according to claim 6, characterized in that: The method for obtaining the static point cloud by performing multi-feature scoring on each candidate target cluster and identifying and eliminating dynamic target clusters based on the scoring results is as follows: extract size features, height features, and temporal consistency features for each candidate target cluster, and perform weighted scoring based on the size features, height features, and temporal consistency features; when the score of a candidate target cluster is higher than a preset scoring threshold, the candidate target cluster is determined to be a dynamic target and eliminated.

9. A large-scene relocalization system for lidar based on a multi-line number adaptive strategy, characterized in that, include: The data acquisition and dynamic target removal module is used to acquire the raw point cloud data collected by the lidar, and to perform dynamic target removal on the raw point cloud data to obtain a static point cloud. The sub-graph construction module is used to construct a local sub-graph based on the number of scan lines of the LiDAR and the static point cloud, as a representation of the scene point cloud; let the number of scan lines of the LiDAR be... The preset line count threshold is The scene point cloud representation for: ， in, This is the static point cloud of the current frame. For the first estimated by the odometer Frame to the Relative pose transformation of frames, This represents the cumulative frame count. When the number of scan lines is greater than a preset threshold, the static point cloud of the current frame is directly used as the scene point cloud representation; when the number of scan lines is less than or equal to the preset threshold, continuous scan lines are used. The static point cloud of a frame is accumulated into a local sub-image through pose transformation, which is used as the scene point cloud representation. The feature extraction module is used to project the scene point cloud representation into a structured representation and extract scene descriptors from the structured representation; The relocalization solution module is used to perform retrieval and matching in the map database based on the scene descriptor, solve the current global pose, and complete the relocalization.

Images