Automatic driving position re-identification method considering urban road structure distribution

Abstract

The invention discloses an automatic driving position re-identification method considering urban road structure distribution, and relates to the technical field of automatic driving positioning. The method comprises the following steps: constructing a city contour descriptor according to city road distribution and point cloud data measured by a laser radar; constructing a terrain descriptor considering measurement characteristics and traffic participant interference based on the point cloud data; and on the basis of the city contour descriptor and the terrain descriptor, firstly carrying out descending sorting on four subspaces of a local area, then removing area space descriptions of a first-layer radial concentric circle r1 and a second-layer radial concentric circle r2, and finally carrying out retrieval identification by utilizing KD-Tree to obtain a final identification position. The method can solve the problems that a traditional point cloud descriptor is single in information and insufficient in environment representation capability, and measurement errors caused by a low-resolution laser radar are not considered.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving positioning technology, and in particular to an autonomous driving location re-identification method that takes into account the distribution of urban road structure. Background Technology

[0002] With the deep integration of intelligent, information-based, and digital technologies in the transportation sector, intelligent vehicles are being widely applied in typical scenarios such as urban mobility, logistics transportation, and automated parking, placing higher demands on environmental perception and high-precision positioning capabilities. LiDAR, as a core sensor, can acquire rich 3D environmental information, supporting precise positioning and scene understanding. However, limited by its low resolution, non-instantaneous sampling characteristics, and frequent changes in illumination, viewing angle differences, and occlusion interference in complex urban environments, the acquired point cloud data often contains significant errors, making it difficult for single-frame information to reliably support accurate positioning. Simultaneously, efficiently identifying similar frames in historical records in the face of surging point cloud data incurs significant computational overhead, severely restricting the system's real-time performance and stability. To improve the robustness of positioning in urban environments, high-precision maps are widely used to enhance global consistency; however, due to the cumulative drift caused by sensor measurement errors and state estimation errors, map construction quality remains difficult to guarantee. Therefore, LiDAR, as an important means of location re-identification, plays a crucial role in positioning in complex urban environments.

[0003] Existing loop closure detection methods are mostly based on the measurement characteristics of lidar, constructing a compact representation structure from point cloud observations and achieving location re-identification by comparing the similarity between the current scan and the point cloud structure in the historical database. However, due to the inherent measurement noise and non-instantaneous sampling errors of lidar, coupled with the highly similar structural features in many areas of urban environments, it is difficult to clearly distinguish the structure of point clouds at different locations. This limitation is particularly prominent in large-scale scenarios, where the expressive power of traditional point cloud structures is insufficient, easily leading to mismatches or missed detections, seriously affecting the stability and accuracy of loop closure detection, and thus weakening the application effect of lidar in high-precision positioning tasks. Summary of the Invention

[0004] The purpose of this invention is to provide an autonomous driving location re-identification method that takes into account the distribution of urban road structure, aiming to solve or improve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, the present invention provides the following solution: An autonomous driving location re-identification method considering the distribution of urban road structure includes: Based on the distribution of urban roads and point cloud data measured by lidar, an urban outline descriptor is constructed. A terrain descriptor considering measurement characteristics and traffic participant interference is constructed based on the point cloud data; Based on the city outline descriptor and the terrain descriptor, the four subspaces of the local area are first sorted in descending order, then the regional spatial descriptions of the first layer of radial concentric circles r1 and the second layer of radial concentric circles r2 are removed, and finally the KD-Tree is used for retrieval and identification to obtain the final identification location.

[0006] Optionally, the step of constructing an urban outline descriptor based on the urban road distribution and point cloud data measured by lidar specifically includes: First, a polar coordinate system with the lidar as the measurement center is constructed according to the ring measurement method of the laser point cloud. ;in, r Polar radius, Polar angle, For right-handed coordinate system Axis data, n This represents the number of point clouds; Based on the polar coordinates, the point cloud space is divided into radial and circumferential directions respectively. and Each part forms Each subspace is defined so that any laser point cloud will fall within a specific sub-interval, and the corresponding circumferential index is calculated for each sub-interval. and radial initial index ; A nonlinear correction mechanism is introduced in the radial direction, and a centrosymmetric nonlinear adjustment strategy is adopted to adjust the boundary of each radial interval, thus obtaining the point cloud subspace corresponding to any laser point. ; For each of the aforementioned point cloud subspaces A local multi-valued statistical strategy is adopted. :reserve The largest Each point is used as the average height value, which is then used as the encoding value for the corresponding subspace. By arranging and combining the encoded values ​​of all point cloud subspaces, a size of [size missing] is obtained. City outline descriptor D .

[0007] Optionally, the expression for adjusting the boundary of each radial interval is: in, This is a nonlinear correction term. The average resolution angle in the vertical direction of the lidar. For the first The polar radius corresponding to each laser point This represents the maximum effective observation range of the lidar in the current frame. This represents the boundary of the radial interval.

[0008] Optionally, the expression for the local multi-valued statistical strategy is: in, This is a local multi-valued statistical strategy. and This refers to the horizontal resolution angle and bus beam of the current lidar.

[0009] Optionally, the step of constructing a terrain descriptor based on the point cloud data, taking into account measurement characteristics and interference from traffic participants, specifically includes: Urban feature segmentation based on urban road and building distribution characteristics: First, a set of concentric circles centered at the origin is used to radially layer the space to obtain preliminary segmentation results; Based on the preliminary segmentation results, a directional judgment is made, and two parabolas opening to the left and right respectively are designed in the lidar coordinate system to define the boundaries of the "front and back" regions: Among them, y r Let y be a parabola opening to the right. l For a parabola opening to the left, the parameters are... This controls the opening degree of the parabola; a larger value means more space is reserved on both sides. Indicates the current lane width; When the laser point is located between these two parabolas, it is considered to be in the "front" or "rear" area of ​​the vehicle; otherwise, it is determined based on the point... The coordinates, positive or negative, are used to divide the space into "left" or "right" regions; ultimately, the entire planar space is divided into... Each sector-shaped sub-region corresponds to a combination of each radial level and four directions; the radial level is composed of... A series of concentric circles are formed; The outer points that are not effectively covered are assigned to an additional spatial region. And construct a set of spatial regions: ;in, Indicates the first Each radial layer corresponds to a spatial region in the four directions of "front, back, left, and right"; Construct a vector based on the results of the city feature segmentation. To represent the encoding result: Where γ is the normalization factor, H The installation height of the lidar relative to the ground is used to eliminate installation deviations between different platforms. The terrain descriptor of each area is obtained by accumulating the normalized height of all points in the area.

[0010] Optionally, the process of using KD-Tree for retrieval and identification specifically includes: First, the revisit location and candidate sample space are determined from the historical access data based on the terrain descriptor. Then, the terrain descriptor of the current scanned area is compared with the historical descriptor corresponding to the revisit location and the candidate sample space. The difference in significant features is calculated. By accumulating the difference in the most significant features of each subspace, the terrain similarity score between the two revisits is obtained. Invalid candidates are then filtered out according to the set threshold. For the remaining candidate set after screening, a robust structure-aware weighted similarity measure is used for final retrieval and identification.

[0011] Optionally, the final retrieval and identification of the remaining revisit candidate set after the screening process using a robust structure-aware weighted similarity measurement method specifically includes: Based on the revisit candidate set, an explicit background ignoring mechanism is introduced in the similarity calculation. At the same time, an amplitude normalization strategy is adopted to reduce the dominant influence of feature value scale differences, and a cyclic translation of column vectors is used to simulate rotation alignment. Finally, all possible rotation offsets are matched, and the minimum error is taken as the best score to determine the final retrieval and recognition position.

[0012] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention discloses an autonomous driving location re-identification method considering the distribution of urban road structures. The method includes constructing an urban contour descriptor based on the urban road distribution and point cloud data measured by LiDAR; constructing a terrain descriptor based on the point cloud data, considering measurement characteristics and interference from traffic participants; based on the urban contour descriptor and the terrain descriptor, firstly sorting the four subspaces of a local area in descending order, then removing the regional spatial descriptions of the first layer of radial concentric circles r1 and the second layer of radial concentric circles r2; finally, using a KD-Tree for retrieval and identification to obtain the final identified location. This invention can solve the problems of traditional point cloud descriptors having limited information, insufficient environmental representation capabilities, and not considering measurement errors caused by low-resolution LiDAR. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating the autonomous driving location re-identification method of the present invention that takes into account the distribution of urban road structure. Figure 2 This is a schematic diagram illustrating the construction of terrain descriptors and city outline descriptors in this embodiment; Figure 3 This is a schematic diagram of the location re-identification results under similar measurement scenarios in this embodiment. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] The purpose of this invention is to provide an autonomous driving location re-identification method that takes into account the distribution of urban road structure, aiming to solve or improve at least one of the above-mentioned technical problems.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, this invention provides an autonomous driving location re-identification method considering the distribution of urban road structures. Using the scan data from an RS-Ruby 128 LiDAR scanner as the initial input, a compact LiDAR point cloud description structure with rich information and reasonable logic is constructed. The point cloud output by RS-Ruby 128 satisfies a right-handed coordinate system. The point cloud obtained from the LiDAR scan... The specific process includes: S1. Based on the urban road distribution and point cloud data obtained from LiDAR, a city contour descriptor is constructed. Traditional point cloud descriptors mostly start from point cloud measurement results, extracting the environmental features revealed by the measurement results to construct the point cloud descriptor. This does not consider the sensor's observation process, and is prone to errors when faced with insufficient sensor expressive capabilities. This embodiment, however, takes into account the urban road distribution and the measurement process and characteristics of LiDAR, and designs a city contour location descriptor, such as... Figure 2As shown, the left side represents the spatial segmentation method, and the right side represents the descriptor construction result obtained from a single laser scan.

[0019] S11. Reconstruct the spatial division of the point cloud in the bird's-eye view based on the vertical resolution angle of the lidar: First, a polar coordinate representation centered on the lidar is constructed based on the ring-shaped measurement method of the laser point cloud. .

[0020] The polar radius represents the measurement distance of the planar lidar. The polar angle represents the angle between the current laser beam and the starting direction of a laser frame. For ease of subsequent modeling and processing, this embodiment divides the point cloud space radially and circumferentially based on the aforementioned polar coordinate system. and Each part forms Each subspace, denoted as ,like Figure 2 As shown. Any laser point cloud will fall within a defined sub-interval. The corresponding circumferential index can be calculated using the following formula. and radial initial index .

[0021] in and They represent the first The polar angle and polar radius corresponding to each laser point This indicates the maximum effective observation range of the lidar in the current frame. To further enhance the physical rationality and perceptual consistency of point cloud partitioning, this embodiment introduces a nonlinear correction mechanism in the radial direction to adapt to the changing characteristics of the lidar's vertical angular resolution at different observation distances. Traditional methods often use uniform radius partitioning, as described above... This can be used as a radial index value under a uniform partitioning method. However, the point cloud density of lidar is higher at close range and decreases significantly at long range. To more reasonably reflect this spatial distribution characteristic, this embodiment designs a centrally symmetric nonlinear adjustment strategy to fine-tune the boundary of each radial interval.

[0022] in This is a nonlinear correction term, which is related to the vertical angular resolution and horizontal measurement distance of the lidar. This represents the average resolution angle in the vertical direction of the lidar. The final subspace index for any lidar point was obtained through correction. .

[0023] S12. Consider the Top-K feature encoding of low-resolution lidar: Through the above process, we obtain several segmented point cloud subspaces. The above formula can be used to quickly obtain the point cloud space to which any laser point belongs. Traditional encoding methods directly use the maximum height or intensity in each space to represent the characteristics of the current space. Although this can enhance the difference between each subspace, considering that lidar is a low-resolution sensor, it is difficult to guarantee that the position is absolutely the same when visiting the same location multiple times; there will always be different degrees of rotation and translation. This may lead to differences in measurements at the same location. In view of this, this embodiment designs a novel encoding method. Considering that the intensity feature is easily affected by reflectivity, incident angle, and distance, resulting in instability, we still select the height feature as the main characterization information to enhance the robust expression of urban outline features. Specifically, for any point cloud subspace... Instead of simply selecting the height of the highest point as a representative value, a local multi-value statistical strategy is adopted. :reserve The largest Each point is used as the encoding value for that subspace, and its average height value is used as the encoding value for that subspace.

[0024] in It is obtained by calculating the theoretical number of laser points in the current subspace region. and The horizontal resolution angle and bus beam of the current lidar are given. This mean-based encoding method preserves significant high-point features in the spatial structure (such as buildings, columns, and walls) while avoiding noise effects from single-point anomalies. Furthermore, under repeated observation conditions, even with slight attitude perturbations or occlusion, averaging multiple high values ​​maintains the relative stability of the features, thus enhancing the robustness of the descriptor in large-scale scene relocalization and loop closure detection. Finally, we can arrange and combine the encoding results of all subspaces to construct a matrix of size [size missing]. For the Urban Contour Descriptor (UCD), we use a matrix. This indicates that the descriptor stores... The calculation results fully preserve the elevation distribution pattern in the urban structure.

[0025] S2. Construct a terrain descriptor based on the point cloud data, taking into account measurement characteristics and interference from traffic participants. Although the urban contour descriptor in S1 has already provided a compact expression of the measurement results, it still mainly focuses on the measurement process and data results. If directly applied to scene recognition, it may be affected by terrain fluctuations or dynamic obstacles. Therefore, this embodiment starts from the distribution characteristics of urban roads and constructs a more robust urban terrain descriptor, aiming to characterize stable terrain structure features and possessing strong adaptability to traffic participants.

[0026] S21. Segmenting urban features based on the distribution of urban roads and buildings: Given that urban roads are mostly linear in shape and buildings are typically densely distributed in the left and right directions, this embodiment uses the lidar coordinate system as the center and constructs a perception area division method based on polar coordinates. First, a series of concentric circles centered at the origin are used to radially layer the space.

[0027] in, Indicates the first The radii of the concentric circles are first terms of the sum of the road width and the safety distance, and then increase exponentially with respect to the vertical angular resolution, forming a geometric sequence. Parameters As a regulating factor, This is a small offset term. This indicates the average vertical angular resolution of the lidar; The number of concentric circles must satisfy the following conditions: ,in To effectively sense distance, and further refine the spatial structure, directional judgment was introduced, dividing the plane into four regions: "front, back, left, and right." In the lidar coordinate system, The axis indicates the left and right directions. The axis represents the front-back direction. Therefore, two parabolas, opening to the left and right respectively, were designed to define the boundaries of the "front-back" region: Among them, parameters The value controls the opening degree of the parabola; the larger the value, the more space is reserved on the left and right sides. This indicates the current lane width. When the laser point is located between these two parabolas, it is considered to be in the "front" or "rear" area of ​​the vehicle. Otherwise, it depends on the point... The coordinates, being positive or negative, are used to divide the space into "left" or "right" regions. Ultimately, the entire planar space is divided into... Each sector-shaped sub-region corresponds to a radial level (by... (Formed by concentric circles) and combinations of four directions (front, back, left, and right). Furthermore, laser points outside the designated area will be grouped into an additional spatial region. This is used to represent peripheral points that are not effectively covered. To facilitate subsequent encoding and feature extraction, this embodiment constructs a set of spatial regions: in Indicates the first Each radial layer corresponds to a spatial region in the four directions of "front, back, left, and right". This structure provides a stable and directionally-aware spatial division for describing urban terrain.

[0028] S22. Perform hierarchical feature coding according to the characteristics of urban regions: Following the city feature segmentation method described above, we obtained a set of spatial regions Ω that conform to city characteristics. To highlight the feature differences of each region and enhance feature stability, we performed height accumulation and normalization on the LiDAR height of each subspace. For ease of representation, we constructed a vector T to represent the encoding result: Where γ is the normalization factor, H The installation height of the lidar relative to the ground is used to eliminate installation deviations between different platforms. Characteristic values ​​for each area are obtained by accumulating the normalized heights of all points in the area.

[0029] S3. Based on the city outline descriptor and the terrain descriptor, the four subspaces of the local area are first sorted in descending order, then the regional space descriptions of the first layer of radial concentric circles r1 and the second layer of radial concentric circles r2 are removed, and finally the KD-Tree is used for retrieval and identification to obtain the final identification location.

[0030] Specifically, one of the key tasks in location re-identification is to quickly find the revisited location from a large amount of historical access data. The core objective of the retrieval process is to rapidly narrow down the search space. This embodiment constructs an efficient retrieval strategy by combining the urban terrain descriptor T with a KD-Tree. Specifically, we first sort the four subspaces of the local area in descending order to ensure that the terrain descriptor has rotation invariance. Next, the spatial descriptions of regions r1 and r2 are removed to reduce the interference of dynamic objects on the identification process. Finally, using the fast search function of the KD-Tree, the C candidate results most similar to the current scan are retrieved from the historical access data.

[0031] S31. Rapid filtering mechanism for significant features: Although the retrieval significantly narrows the candidate sample space, directly performing the final similarity judgment still incurs high time costs. Furthermore, for results without revisiting, the filtering results also return potential candidates, and directly comparing similarity would waste significant computational resources. Therefore, to further accelerate the identification process, this embodiment designs a filtering mechanism based on the most salient features. Specifically, for the currently scanned terrain descriptor... and candidate history descriptors The similarity is assessed by calculating the differences between the most salient features of each region subspace. A similarity score is defined. as follows: Among them, the significant feature difference function Defined as: By summing the differences in the most salient features of each subspace region, a terrain similarity score can be obtained between two revisits. If the score exceeds the set threshold... If a candidate is not found to be valid, then the current candidate is considered invalid. This filtering mechanism can effectively reduce unnecessary computation and improve the computational efficiency of the overall recognition process.

[0032] S32. Robust similarity measurement method: Although we have obtained the most likely revisit candidate set through efficient retrieval and saliency filtering, some uncertainty remains because the first two steps mainly rely on compressed terrain information. Traditional methods use cosine distance, which is highly susceptible to differences in feature distribution, such as... Figure 3 As shown, the left side displays two LiDAR scans at different locations in the KITTI 08 dataset, while the right side shows descriptors and recognition results for various handcrafted feature methods. The values ​​in parentheses represent the recognition thresholds corresponding to achieving the highest F1 score (red indicates incorrect recognition at the current threshold, and green indicates correct recognition). To further improve recognition accuracy, we designed a robust structure-aware weighted similarity metric. This method is used to accurately evaluate each candidate against the current measurement. It introduces an explicit background ignoring mechanism in the similarity calculation, effectively eliminating interfering information that does not contribute to the recognition results. Simultaneously, it employs an amplitude normalization strategy to weaken the dominant influence of feature value scale differences, thereby improving the overall accuracy and robustness of the recognition. Specifically, for the current and possible candidate city contour descriptors... and We define its similarity score as: Among them, only and Only cases with at least one non-zero value are included in the calculation to ensure that the background region does not affect the score. However, in practical applications, the vehicle's heading angle may change during revisits, causing misalignment of the angles (i.e., columns) of the descriptors between the two visits. To address this issue, we introduce a cyclic translation of the column vector to simulate rotational alignment: that is, for each candidate descriptor... Translation in angular dimension The column is rotated and aligned to the current frame, and the rotation and translation operators are defined as follows: Finally, all possible rotational offsets are matched, and the smallest error is taken as the best score. This strategy significantly enhances the robustness and adaptability to rotation of descriptors while maintaining efficiency, effectively improving recognition accuracy in urban environments.

[0033] In summary, compared with traditional technologies, the present invention has the following characteristics: (1) Based on the measurement characteristics of low-resolution lidar, the original point cloud is non-uniformly divided according to the vertical resolution angle, which improves the reliability and integrity of information in the local subspace.

[0034] (2) Combining the structural distribution characteristics of urban roads and buildings, a spatial segmentation and coding strategy oriented towards terrain is proposed, which effectively eliminates dynamic traffic interference and improves the robustness of spatial expression in urban environments.

[0035] (3) Design a weighted similarity measurement method for structure perception, introduce local normalization and structural weights, enhance the adaptability to different observation perspectives and environmental structural changes, and improve matching accuracy and generalization ability.

[0036] (4) The overall method can be mapped to a compact complex structure representation, which balances expressive power and computational efficiency, and is suitable for efficient location re-identification tasks in real urban environments.

[0037] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0038] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for re-identifying the location of an autonomous driving system that considers the distribution of urban road structure, characterized in that, include: Based on the distribution of urban roads and point cloud data measured by lidar, an urban outline descriptor is constructed. A terrain descriptor considering measurement characteristics and traffic participant interference is constructed based on the point cloud data; Based on the city outline descriptor and the terrain descriptor, the four subspaces of the local area are first sorted in descending order, then the regional spatial descriptions of the first layer of radial concentric circles r1 and the second layer of radial concentric circles r2 are removed, and finally the KD-Tree is used for retrieval and identification to obtain the final identification location.

2. The autonomous driving location re-identification method considering urban road structure distribution according to claim 1, characterized in that, The process of constructing an urban outline descriptor based on urban road distribution and point cloud data measured by lidar specifically includes: First, a polar coordinate system with the lidar as the measurement center is constructed according to the ring measurement method of the laser point cloud. ;in, r Polar radius, Polar angle, For right-handed coordinate system Axis data, n This represents the number of point clouds; Based on the polar coordinates, the point cloud space is divided into radial and circumferential directions respectively. and Each part forms Each subspace is defined so that any laser point cloud will fall within a specific sub-interval, and the corresponding circumferential index is calculated for each sub-interval. and radial initial index ; A nonlinear correction mechanism is introduced in the radial direction, and a centrosymmetric nonlinear adjustment strategy is adopted to adjust the boundary of each radial interval, thus obtaining the point cloud subspace corresponding to any laser point. ; For each of the aforementioned point cloud subspaces A local multi-valued statistical strategy is adopted. :reserve The largest Each point is used as the average height value, which is then used as the encoding value for the corresponding subspace. By arranging and combining the encoded values ​​of all point cloud subspaces, a size of [size missing] is obtained. City outline descriptor D .

3. The autonomous driving location re-identification method considering urban road structure distribution according to claim 2, characterized in that, The expression for adjusting the boundary of each radial interval is: in, This is a nonlinear correction term. The average resolution angle in the vertical direction of the lidar. For the first The polar radius corresponding to each laser point This represents the maximum effective observation range of the lidar in the current frame. This represents the boundary of the radial interval.

4. The autonomous driving location re-identification method considering urban road structure distribution according to claim 2, characterized in that, The expression for the local multi-valued statistical strategy is: in, This is a local multi-valued statistical strategy. and This refers to the horizontal resolution angle and bus beam of the current lidar.

5. The autonomous driving location re-identification method considering urban road structure distribution according to claim 1, characterized in that, The construction of a terrain descriptor based on the point cloud data, considering measurement characteristics and interference from traffic participants, specifically includes: Urban feature segmentation based on urban road and building distribution characteristics: First, a set of concentric circles centered at the origin is used to radially layer the space to obtain preliminary segmentation results; Based on the preliminary segmentation results, a directional judgment is made, and two parabolas opening to the left and right respectively are designed in the lidar coordinate system to define the boundaries of the "front and back" regions: Among them, y r Let y be a parabola opening to the right. l For a parabola opening to the left, the parameters are... This controls the opening degree of the parabola; a larger value means more space is reserved on both sides. Indicates the current lane width; When the laser point is located between these two parabolas, it is considered to be in the "front" or "rear" area of ​​the vehicle; otherwise, it is determined based on the point... The coordinates, positive or negative, are used to divide the space into "left" or "right" regions; ultimately, the entire planar space is divided into... Each sector-shaped sub-region corresponds to a combination of each radial level and four directions; the radial level is composed of... A series of concentric circles are formed; The outer points that are not effectively covered are assigned to an additional spatial region. And construct a set of spatial regions: ;in, Indicates the first Each radial layer corresponds to a spatial region in the four directions of "front, back, left, and right"; Construct a vector based on the results of the city feature segmentation. To represent the encoding result: Where γ is the normalization factor, H The installation height of the lidar relative to the ground is used to eliminate installation deviations between different platforms. The terrain descriptor of each area is obtained by accumulating the normalized height of all points in the area.

6. The autonomous driving location re-identification method considering urban road structure distribution according to claim 1, characterized in that, The process of using KD-Tree for retrieval and identification specifically includes: First, the revisit location and candidate sample space are determined from the historical access data based on the terrain descriptor. Then, the terrain descriptor of the current scanned area is compared with the historical descriptor corresponding to the revisit location and the candidate sample space. The difference in significant features is calculated. By accumulating the difference in the most significant features of each subspace, the terrain similarity score between the two revisits is obtained. Invalid candidates are then filtered out according to the set threshold. For the remaining candidate set after screening, a robust structure-aware weighted similarity measure is used for final retrieval and identification.

7. The autonomous driving location re-identification method considering urban road structure distribution according to claim 6, characterized in that, The remaining candidate set after the screening process is then used for final retrieval and identification using a robust structure-aware weighted similarity metric, specifically including: Based on the revisit candidate set, an explicit background ignoring mechanism is introduced in the similarity calculation. At the same time, an amplitude normalization strategy is adopted to reduce the dominant influence of feature value scale differences, and a cyclic translation of column vectors is used to simulate rotation alignment. Finally, all possible rotation offsets are matched, and the minimum error is taken as the best score to determine the final retrieval and recognition position.

Images