Map element association matching method and system based on centroid vector

Abstract

The application belongs to the technical field of high-precision positioning, and particularly relates to a map element correlation matching method and system based on a centroid vector, which comprises the following steps: for two map element sets to be matched, a centroid vector matrix is constructed to record the relative position relationship between different types of elements in the set in the form of a vector; for a target element in the first set, a candidate element of the same type is searched for in the second set, a local candidate matching pair is generated by comparing the neighbor distribution similarity of the target element and each candidate element, and voting statistics are performed on all local candidate matching pairs to determine the final matching corresponding relationship according to the number of votes. The application can effectively resist spatial transformation, avoid cross-type mis-matching, and does not depend on a large amount of labeled data, and the matching process is clear and interpretable.

Description

Technical Field

[0001] This invention belongs to the field of high-precision positioning technology, specifically relating to a method and system for map feature association and matching based on centroid vectors. Background Technology

[0002] With the development of technologies such as autonomous driving, intelligent assisted driving, and high-precision positioning, map elements play a crucial role in applications such as vehicle environment modeling, positioning correction, and map updates. Feature-level association and matching are needed between map data or perception results from different sources to achieve information fusion and consistency maintenance. However, due to differences in data acquisition time, sensor type, coordinate system, etc., map elements from different data sources generally suffer from spatial transformation, positional deviation, and missing elements. Currently, the main matching techniques include methods based on global features, methods based on point cloud registration, methods based on deep learning, and methods based on direct distance matching.

[0003] The aforementioned solutions each have their limitations when dealing with the complex and ever-changing differences in map data in practical applications. The main problem lies in the lack of a unified matching framework that can effectively resist spatial transformations, take into account the semantic and spatial contextual relationships of elements, and does not rely on a large amount of labeled data. This makes it difficult for existing methods to achieve a good balance between accuracy and reliability. Summary of the Invention

[0004] The technical problem to be solved by this invention is to overcome the fact that existing map matching methods cannot simultaneously and effectively resist spatial transformation, take into account the semantics of elements and local spatial relationships, and do not rely on a large amount of labeled data.

[0005] To address the aforementioned technical problems, a first aspect of this invention discloses a map feature association and matching method based on centroid vectors, the method comprising:

[0006] For the first and second map feature sets to be matched, centroid vector matrices are constructed respectively. The centroid vector matrices are used to record the relative positional relationships between different types of features within the set in vector form.

[0007] For a target element in the first map element set, similar elements are searched for as candidates in the second map element set. Local candidate matching pairs are generated by comparing the distribution similarity between the target element and the neighbor elements of each candidate element.

[0008] A voting statistics are performed on all the generated local candidate matching pairs, and the final matching correspondence in the second map element set is determined for the elements in the first map element set based on the voting results.

[0009] As an optional implementation, in the first aspect of the present invention, constructing centroid vector matrices for the first set of map features and the second set of map features to be matched respectively includes:

[0010] For the first set of elements Second element set , respectively construct the corresponding centroid vector matrix for the two sets;

[0011] Wherein, for each of the centroid vector matrices, its i-th Line number Column storage from features Point to feature The centroid coordinate difference vector, and only if the element With elements When the types are different, the centroid coordinate difference vector is stored as the effective vector.

[0012] As an optional implementation, in a first aspect of the invention, generating local candidate matching pairs by comparing the distribution similarity of the target feature with the neighboring features of each candidate feature includes:

[0013] For the first set of elements Each of the first elements In the second set of elements Search for all elements that match the first element. Second elements of the same type constitute a candidate set;

[0014] For each candidate second element in the candidate set Perform a neighbor matching operation, which includes: extracting the first element. The effective neighbor vector set and its neighbor types, and the candidate second element. The set of valid neighbor vectors and their neighbor types;

[0015] For the first element Each neighbor vector in the candidate second element In the set of neighbor vectors, find candidate vectors that are of the same type and are not occupied, and calculate the comprehensive similarity score between the neighbor vector and each candidate vector that meets the conditions.

[0016] Based on the comprehensive similarity score, the first element is... Each neighboring element determines its position in the candidate second element. The best match or non-match among the neighbors is used to generate a batch of candidates for the second element. Relevant local candidate matching pairs.

[0017] As an optional implementation, in the first aspect of the present invention, the calculation of the comprehensive similarity score includes:

[0018] Calculate the directional similarity between the neighbor vector and the candidate vector. and length similarity ;

[0019] The directional similarity and length similarity Calculated using the following formula:

[0020] ;

[0021] ;

[0022] in, Let be the angle between the neighbor vector and the candidate vector. The difference in length between the neighbor vector and the candidate vector is... For directional tolerance parameters, This refers to the distance tolerance parameter.

[0023] The comprehensive similarity score Calculated by weighted sum:

[0024] ,in This represents the directional similarity weight.

[0025] As an optional implementation, in a first aspect of the invention, the neighbor matching operation further includes, while traversing the first element When performing a match on neighboring elements, the order of their neighbor indexes is randomized.

[0026] As an optional implementation, in the first aspect of the invention, the directional tolerance parameter The distance tolerance parameter The directional similarity weight and the minimum score threshold used to determine the validity of a match. It can adaptively adjust according to the element density and road topology characteristics of the application scenario.

[0027] As an optional implementation, in the first aspect of the present invention, the step of performing a voting statistics on all generated local candidate matching pairs and determining the final matching correspondence in the second map feature set for the elements in the first map feature set based on the voting results includes:

[0028] Statistical analysis of each set of first elements Elements With elements in the second element set B The total number of times the candidate matching pairs appear is taken as the number of votes for that candidate matching pair;

[0029] For the first set of elements For each element in the process, the second element with the highest number of votes among all its candidate matching pairs that have received votes is determined as the final matching result.

[0030] If the first element set If an element in a dataset fails to receive a valid vote in any of the local candidate matching pairs, it is marked as having no match.

[0031] As an optional implementation, in the first aspect of the invention, the centroid vector matrix is ​​stored in a sparse matrix format.

[0032] A second aspect of this invention discloses a centroid vector-based map feature association and matching system for implementing the centroid vector-based map feature association and matching method described in any of the above embodiments. The system includes:

[0033] The matrix construction module is used to construct centroid vector matrices for the first and second map feature sets to be matched, respectively. The centroid vector matrices are used to record the relative positional relationships between different types of features within the set in vector form.

[0034] The candidate matching generation module is used to find similar elements as candidates in the second map element set for the target element in the first map element set, and generate local candidate matching pairs by comparing the distribution similarity of the target element and the neighbor elements of each candidate element.

[0035] The voting decision module is used to perform voting statistics on all generated local candidate matching pairs, and determine the final matching correspondence of the elements in the second map element set based on the voting results.

[0036] A third aspect of this invention discloses another map feature association and matching system based on centroid vectors, the system comprising:

[0037] Memory containing executable program code;

[0038] A processor coupled to the memory;

[0039] The processor calls the executable program code stored in the memory to execute the map feature association and matching method based on centroid vector disclosed in the first aspect of the present invention.

[0040] The fourth aspect of the present invention discloses a computer-readable storage medium storing computer instructions, which, when invoked by a processor, are used to execute the centroid vector-based map feature association and matching method disclosed in the first aspect of the present invention.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] By using relative position vectors to record relationships between features, rather than relying on absolute coordinates, the method possesses inherent invariance to translations and rotations of the overall map data. This effectively overcomes spatial transformation problems caused by differences in coordinate systems, positioning deviations, or different data sources, providing a stable foundation for subsequent matching. Candidate pairs are generated by comparing the similarity of neighbor distributions between target features and candidate features, rather than relying solely on the absolute position or a single feature of the feature itself. This matching strategy based on local spatial context relationships can more accurately identify truly corresponding features, especially in scenarios where features appear similar or where location noise exists, improving matching accuracy and discriminative power. By voting on all locally generated candidate matching pairs and determining the final matching relationship based on the number of votes, this mechanism achieves global consistency verification of the matching results, effectively filtering out accidental and inconsistent mismatches, thereby improving the overall reliability of the final matching results. Attached Figure Description

[0043] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, wherein:

[0044] Figure 1 This is a flowchart illustrating the map feature association and matching method based on centroid vectors disclosed in an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram of the structure of the map feature association and matching system based on centroid vectors disclosed in an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of another map feature association and matching system based on centroid vectors disclosed in an embodiment of the present invention. Detailed Implementation

[0047] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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.

[0048] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0049] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0050] This invention discloses a map feature association matching method and system based on centroid vectors. By using relative position vectors to record the relationships between features, rather than relying on absolute coordinates, the method has inherent invariance to the translation and rotation of the overall map data. This effectively overcomes spatial transformation problems caused by differences in coordinate systems, positioning deviations, or different data sources, providing a stable foundation for subsequent matching. Candidate pairs are generated by comparing the similarity of the neighbor distribution between the target feature and candidate features, rather than relying solely on the absolute position or a single feature of the feature itself. This matching strategy based on local spatial context can more accurately identify the truly corresponding features, especially in scenarios where features have similar appearances or exist in location noise, thus improving the accuracy and discriminative power of matching.

[0051] Example 1

[0052] Please see Figure 1 , Figure 1 This is a flowchart illustrating the map feature association and matching method based on centroid vectors disclosed in an embodiment of the present invention. Figure 1 The described centroid-based map feature association and matching method is applied to a data processing chip, processing terminal, or processing server, which can be a local server or a cloud server; this embodiment of the invention does not impose any limitations. Figure 1 As shown, this centroid vector-based map feature association and matching method can include the following operations:

[0053] 101. For the first set of map features and the second set of map features to be matched, construct centroid vector matrices respectively. The centroid vector matrices are used to record the relative positional relationships between different types of features in the set in vector form.

[0054] Specifically, this step makes the matching method no longer rely on the absolute position of each element in the global coordinate system when processing data.

[0055] As can be seen, this design makes the matching process naturally invariant to the overall translation and rotation that may exist between the two sets of maps to be matched, thus establishing a unified benchmark for subsequent similarity comparison and decision-making that does not change with the coordinate system.

[0056] 102. For a target element in the first map element set, search for similar elements in the second map element set as candidates, and generate local candidate matching pairs by comparing the distribution similarity of the target element and each candidate element's neighboring elements.

[0057] Specifically, this step matches target elements by comparing the distribution patterns of their local neighbors with those of their candidate elements, and in doing so, it enforces semantic type consistency constraints on the elements. This mechanism makes the matching decision not only dependent on spatial proximity, but also on the structural relationships and category information between elements.

[0058] As can be seen, the present invention can effectively identify corresponding elements with similar local contexts without relying on the global shape or integrity of the map. At the same time, it fundamentally avoids the incorrect association of elements with different semantic types (such as stop lines and pedestrian crossings). Its matching judgment is based on calculable geometric similarity, and the process is clear and interpretable.

[0059] 103. Perform voting statistics on all generated local candidate matching pairs, and determine the final matching correspondence in the second map element set for the elements in the first map element set based on the voting results.

[0060] Specifically, the frequency of occurrence of all local candidate matching pairs is counted, and the final matching relationship is determined for each element based on the majority principle.

[0061] It is evident that this voting mechanism can effectively identify and adopt consistent matching relationships that have been identified in multiple independent local comparisons, while filtering out accidental mismatches that only occasionally occur. Thus, without relying on complex models or prior knowledge, it improves the confidence and overall consistency of the final matching results, and the process has clear statistical interpretability.

[0062] As an optional embodiment, the step described above, which involves constructing centroid vector matrices for the first and second map feature sets to be matched, respectively, includes:

[0063] For the first set of elements Second element set , respectively construct the corresponding centroid vector matrix for the two sets;

[0064] Wherein, for each of the centroid vector matrices, its i-th Line number Column storage from features Point to feature The centroid coordinate difference vector, and only if the element With elements When the types are different, the centroid coordinate difference vector is stored as the effective vector.

[0065] In this embodiment of the invention, a specific data structure, namely a centroid vector matrix, is constructed for each element set to encode the spatial relationships between elements within the set. The core design of this matrix is ​​that its first... Line number The elements of a column are defined from the elements. Point to feature The vector formed by the difference in centroid coordinates, and through "only if the element With elements The rule of "only storing vectors with different types" enables selective recording of relationships between elements of different semantic types at the data structure level, while vectors between elements of the same type are marked as invalid or ignored.

[0066] It is evident that this data organization, based on relative vectors and incorporating a semantic type isolation mechanism, establishes a spatial relationship expression that does not rely on an absolute coordinate system. This allows subsequent similarity comparisons to be based entirely on the relative directions and distances between elements, thus ensuring invariance to translation and rotation transformations of the overall data. Furthermore, since direct vector associations between elements of the same type are excluded at the basic data layer, this not only optimizes the storage structure but, more importantly, avoids meaningless comparisons based on the relative positions of similar elements from the outset. This ensures that subsequent matching calculations always focus on the more discriminative local spatial patterns formed by neighbors of different types, providing an underlying guarantee for obtaining semantically consistent matching results.

[0067] As an optional embodiment, the step described above, generating local candidate matching pairs by comparing the distribution similarity of the target feature with the neighboring features of each candidate feature, includes:

[0068] For the first set of elements Each of the first elements In the second set of elements Search for all elements that match the first element. Second elements of the same type constitute a candidate set;

[0069] For each candidate second element in the candidate set Perform a neighbor matching operation, which includes: extracting the first element. The effective neighbor vector set and its neighbor types, and the candidate second element. The set of valid neighbor vectors and their neighbor types;

[0070] For the first element Each neighbor vector in the candidate second element In the set of neighbor vectors, find candidate vectors that are of the same type and are not occupied, and calculate the comprehensive similarity score between the neighbor vector and each candidate vector that meets the conditions.

[0071] Based on the comprehensive similarity score, the first element is... Each neighboring element determines its position in the candidate second element. The best match or non-match among the neighbors is used to generate a batch of candidates for the second element. Relevant local candidate matching pairs.

[0072] In this embodiment of the invention, this step constructs a set of element similarity calculation and matching processes based on local spatial context and semantic consistency. The core of this process is that, for each element to be matched, the algorithm does not directly compare its own coordinates, but treats it and each candidate element of the same type as a local reference center, and systematically compares the vector sets formed by each of them and their surrounding neighbors of different types. The matching operation finds the corresponding items of the same type and the most similar vectors among the neighbors of the candidate elements for each neighbor of the reference element, and calculates a quantitative comprehensive similarity score, thereby establishing a set of fine-grained correspondence hypotheses with confidence assessment in the local scope.

[0073] As can be seen, this process reduces the reliance on global data integrity by decomposing the global matching problem into a large number of parallel pairwise comparisons based on local neighbor patterns. It utilizes relative vectors for similarity calculation, ensuring that the comparison remains invariant to overall coordinate translation and rotation. Mandatory type consistency checks are implemented throughout both the candidate search and vector matching stages, fundamentally preventing erroneous associations across semantic categories. This invention generates a similarity assessment for each possible pair of elements based on explicit geometric calculations (direction and length differences). Its output is a series of traceable and verifiable local hypotheses, providing reliable and rich input for subsequent global decisions based on statistical consensus, rather than an uninterpretable black-box judgment.

[0074] As an optional embodiment, the calculation of the comprehensive similarity score in the above steps includes:

[0075] Calculate the directional similarity between the neighbor vector and the candidate vector. and length similarity ;

[0076] The directional similarity and length similarity Calculated using the following formula:

[0077] ;

[0078] ;

[0079] in, Let be the angle between the neighbor vector and the candidate vector. The difference in length between the neighbor vector and the candidate vector is... For directional tolerance parameters, This refers to the distance tolerance parameter.

[0080] The comprehensive similarity score Calculated by weighted sum:

[0081] ,in This represents the directional similarity weight.

[0082] In this embodiment of the invention, this step defines a quantitative, multi-metric geometric similarity evaluation function to accurately measure the consistency between two spatial vectors (i.e., a pair of neighbor relationship vectors of the same type from different sets). Instead of using a single Euclidean distance, this function decouples the differences between vectors into two independent geometric dimensions: directional difference (angle between the vectors). ) and length difference ( The similarity scores were then converted into values ​​between 0 and 1 using a Gaussian kernel-based mapping. and Ultimately, this is achieved through a configurable weight. The two scores are linearly weighted to obtain a comprehensive similarity score. .

[0083] As can be seen, this similarity calculation model transforms the qualitative "whether similar" judgment into a continuous, differentiable numerical score, providing fine-grained confidence criteria for subsequent matching decisions. It decouples the evaluation of direction and length, allowing the method to differentiate and tolerate rotation errors (primarily affecting direction) and translation / scaling errors (primarily affecting length) to varying degrees. This is achieved through parameters... and It can flexibly adapt to the error characteristics of different scenarios, and the weighted sum mechanism introduces a clear trade-off dimension, enabling the method to adjust the relative importance of directional consistency and distance consistency in the final decision according to actual needs. The entire calculation process is based entirely on vector geometry and does not contain any data-driven black-box operations, thus ensuring the determinism of the scoring process and making the source of each matching score clearly traceable and verifiable.

[0084] As an optional embodiment, the neighbor matching operation in the above steps further includes, while traversing the first element When performing a match on neighboring elements, the order of their neighbor indexes is randomized.

[0085] In this embodiment of the invention, when traversing the neighbor list of the first element to perform one-to-one vector matching, a preprocessing operation is introduced, namely, the index order of all its neighbor elements is randomly rearranged, and then the best match is found for each neighbor in this random order.

[0086] As can be seen, this randomization process breaks the potential pattern of traversing the neighbor list in a fixed order (such as by index size or storage order). During the matching process, when a neighbor successfully matches and occupies a corresponding neighbor of a candidate element, this occupancy status affects the available choices of subsequent neighbors. A fixed traversal order may cause higher-ranked neighbors to always prioritize occupying the best or second-best matching objects, while lower-ranked neighbors may be forced to accept lower-quality matches or even fail to match due to limited choices. This order dependency can introduce unnecessary systematic biases into the matching results. By randomizing the order, each neighbor has an equal chance to be prioritized in each matching attempt, thus completely returning the dominant factor in the matching decision to the geometric similarity calculation between vectors themselves, rather than the accidental order of their storage or retrieval. This enhances adaptability to different data arrangements, making the generated set of local candidate matching pairs more reflective of the true geometric correspondence between elements and improving the stability of the matching process.

[0087] As an optional embodiment, in the above steps, the orientation tolerance parameter The distance tolerance parameter The directional similarity weight and the minimum score threshold used to determine the validity of a match. It can adaptively adjust according to the element density and road topology characteristics of the application scenario.

[0088] In this embodiment of the invention, this step explicitly states that the key parameters used to measure vector similarity in the method, including orientation tolerance, distance tolerance, orientation weight, and matching judgment threshold, are not fixed hyperparameters, but are designed as variables that can be dynamically adjusted according to the specific conditions of the actual application scenario.

[0089] As can be seen, this adjustable parameter design endows the method with the adaptability to cope with the inherent error characteristics and matching challenges of different scenarios. For example, in urban intersections with dense features and complex structures, the distance tolerance can be reduced to enhance the discriminative power of the matching; in highway scenarios with sparse features and large spatial spans, the distance tolerance can be appropriately increased to expand the effective search range. By adaptively configuring these parameters based on prior or real-time information such as feature density and road topology, the matching strategy of the method can be matched with the geometric noise level and structural complexity of the current scenario, thereby maintaining a balance between matching accuracy and reliability under a wider range of application conditions. At the same time, this also provides algorithm engineers with a clear parameter adjustment interface that is related to physical meaning, enabling them to optimize and fine-tune the behavior of the method in a targeted manner based on domain knowledge, thus enhancing the engineering practicality of the method.

[0090] As an optional embodiment, the step described above, namely, performing a vote count on all generated local candidate matching pairs and determining the final matching correspondence in the second map feature set based on the voting results for the features in the first map feature set, includes:

[0091] Statistical analysis of each set of first elements Elements With elements in the second element set B The total number of times the candidate matching pairs appear is taken as the number of votes for that candidate matching pair;

[0092] For the first set of elements For each element in the process, the second element with the highest number of votes among all its candidate matching pairs that have received votes is determined as the final matching result.

[0093] If the first element set If an element in a dataset fails to receive a valid vote in any of the local candidate matching pairs, it is marked as having no match.

[0094] In this embodiment of the invention, the core of this step is to establish a voting record table to accumulate the frequency of identical element correspondences identified by different reference elements in all previous independent local neighbor matching processes. Then, for each element in the first set, the algorithm selects the element in the second set with the highest number of votes from all its vote-receiving correspondences as the final match; if an element has never received any votes, it is explicitly output as having no match.

[0095] As can be seen, this voting and decision-making mechanism summarizes and integrates numerous local assumptions that may contain noise and conflict, generated in previous steps. A true match will receive a high number of votes because it maintains geometric consistency in multiple rounds of local comparisons with different factors as references; while accidental mismatches usually only occur in a few local scenarios and receive fewer votes. By adopting the match with the highest number of votes, the algorithm essentially selects the most globally consistent correspondence supported by the broadest evidence, thereby effectively suppressing ambiguities and errors that may arise in the local matching stage. At the same time, allowing the output of "no match" results enables the method to objectively reflect the possible addition or absence of elements in the data, enhancing its practicality. The entire decision-making process is based entirely on statistically significant and traceable frequency data, with clear and verifiable logic.

[0096] As an optional embodiment, in the above steps, the centroid vector matrix is ​​stored in a sparse matrix format.

[0097] In this embodiment of the invention, it is clarified that a sparse matrix format is used instead of a traditional two-dimensional dense array when storing the centroid vector matrix in the computer. The reason for this is that, since a valid vector is only stored when two feature types are different, the values ​​at most positions in the matrix are invalid or zero, thus its data structure has high sparsity. Using a sparse storage format, such as a compressed sparse row format or an adjacency list, means that only non-zero valid vectors and their corresponding row and column index information are actually stored in memory.

[0098] As can be seen, this storage strategy reduces memory consumption from being proportional to the square of the total number of features to being proportional to the product of the total number of features and the average number of their neighbors. For large-scale map data containing tens of thousands of features, this can result in a reduction of memory usage by several orders of magnitude. This not only enables the algorithm to process map data with wider areas and denser features without sacrificing computational accuracy, improving the scalability of the method, but also lowers the hardware resource threshold, enhancing its deployment feasibility on embedded systems or real-time computing platforms.

[0099] Furthermore, the specific steps of this invention are as follows:

[0100] 1. Build a "neighbor map" for each feature, recording its spatial relationships with other surrounding features. Given two sets of map features to be matched: the first set A contains m features. The second set of elements, B, contains n elements. Each element has a unique identifier. Type identifier (e.g., 1 = stop line, 2 = pedestrian crossing, 3 = speed bump, etc.) and centroid coordinates (3D coordinates) );

[0101] Construct centroid vector matrices for sets A and B respectively. and Taking set A as an example, iterating through any two elements in the set... and Fill the matrix according to the following rules: if it is the same feature ( ) or of the same type ( Marked as invalid. Otherwise, calculate from point to vector ;

[0102] This representation method has the following characteristics and advantages:

[0103] (1) Spatial invariance: Only the relative position vectors between different types of features are recorded, rather than the absolute coordinates. Therefore, it has natural robustness to map translation and rotation. The vectors retain their original lengths without normalization, because the vector magnitude represents the distance between features. This information is used to determine the consistency of spatial distribution in subsequent similarity calculations.

[0104] (2) Matrix sparsity and storage optimization: Since each map feature is usually only associated with a finite number of neighboring features of different types (features of the same type are excluded), the centroid vector matrix and It exhibits highly sparse characteristics. In the algorithm description of this invention, a two-dimensional array is used for ease of understanding and indexing. Logical representation is performed; in engineering implementation, sparse matrix storage formats (such as Compressed Sparse Row or adjacency lists) can be used for memory optimization. Taking an adjacency list as an example, each element... This corresponds to a list that stores the indices of all its valid neighbors. and the corresponding vector .by Taking one element as an example for memory comparison: if traditional dense matrix storage is used, it requires... Bytes (approximately 23MB, with each vector occupying 3 double-type data points, totaling 24 bytes); while using sparse storage, assuming each feature has an average of 5 neighbors of different types, only requires The size is reduced to approximately 140KB (including 4 bytes for the index and 24 bytes for the vector), thus reducing memory usage and improving the algorithm's practicality and scalability.

[0105] 2. This step finds possible correspondences by comparing the "neighbor distribution patterns" of elements. The basic idea is: if the stopping line in set A... The stop line in set B Then their surrounding neighbors should be similarly distributed—for example There is a pedestrian crossing 5 meters to the left and a speed bump 3 meters to the right. There should be similar distributions in the surrounding area.

[0106] Iterate through each element in set A ( Then, perform the following steps sequentially. First, extract all valid vectors from the i-th row of the centroid vector matrix to obtain... (from (The set of vectors pointing to each different type of neighbor) and (The corresponding set of neighboring feature indexes), for example It may contain three vectors, pointing to a pedestrian crossing, a speed bump, and a road marking, respectively. Then, search for all vectors in set B that correspond to... Similar elements (satisfying) ) constitute the candidate set The reason for only searching for the same type here is that stop lines can only match stop lines and not pedestrian crossings.

[0107] Next, we will examine the candidate set. Each element in Perform neighbor matching. Extract... Neighbor information (Neighbor vector set) (neighbor index set) and (Set of neighbor types). Assumption There are 3 neighbors There are 5 neighbors, and we need to determine the relationships between these neighbors. Initialize the tag array. (Length 5, initial values ​​all 0) to record Which neighbors have been matched? The neighbor index order is shuffled (to avoid fixed order deviation), and then each neighbor is processed sequentially. .

[0108] For the neighbors Extract query vector ,exist Find the most similar vector among all its neighbors. Iterate through... Each vector in Filter the data, with the following criteria: not matched. ), type consistent ( The vector length is greater than the minimum value. Calculate the similarity of the candidate vectors that pass the screening: first obtain the length of the query vector. Candidate vector length and included angle Then, the directional similarity is calculated using a Gaussian function. and length similarity (in Both are in the range [0, 1]. The overall similarity score is... Find the candidate with the highest score and record it as If the highest score is lower than If no match is found, then the candidate pair is recorded. Otherwise, record the candidate pair. (in ) and set .

[0109] The parameters in the similarity calculation described above are used to characterize the geometric consistency scale and structural stability of road features under different observation conditions. Among them, distance tolerance... The maximum acceptable spatial offset of local features under multi-source observation and multi-trip mapping conditions can be physically understood as the combined effect scale of positioning error, perception extraction error, and map building error. In urban road scenarios, this combined error is usually in the meter range, therefore this paper takes... m is the default tolerance threshold. Orientation tolerance. To characterize the direction invariance assumption of semantic road elements, considering that elements such as stop lines and pedestrian crossings have strong structural constraints on the main road direction under correct matching conditions, their cross-observation direction deviation is usually small, this paper sets... rad (approximately) This is to suppress false matches caused by inconsistency in orientation while ensuring matching recall. Orientation similarity weight. To balance the two complementary constraints of spatial distance consistency and directional structural consistency, this paper adopts... To avoid a single geometric factor dominating the matching results; minimum score threshold. This is used to filter low-confidence candidate matches and improve overall matching reliability. The parameters mentioned above are not fixed hyperparameters, but rather scale parameters reflecting scene scale and structural complexity. They can be adaptively adjusted based on local feature density and road topology characteristics: in scenarios with sparse features and large spatial spans, such as highways, the parameters can be appropriately increased. (e.g., up to 5.0 m) to expand the search area; in densely populated areas such as urban intersections, the search area can be reduced. (e.g., up to 1.0 m) to enhance discrimination, thereby maintaining matching stability and discrimination ability in different scenarios.

[0110] Here is the pseudocode for the core similarity calculation algorithm:

[0111] The function `FindNearestNeighbor(query vector v_query, candidate vector set V_cand, query type type_query, candidate type set T_cand, used flag)` initializes the following: `bestIdx = -1, bestScore = -∞`. It calculates the query vector length: `L_query = ||v_query||`. If `L_query < 1e-6`, it returns -1. For each candidate vector `v_cand[k]` (k = 0, 1, ..., |V_cand|-1): / / Filtering conditions. If `used[k] = 1`, skip. / / Already matched. If `T_cand[k] ≠ type_query`, skip. / / Type mismatch. It calculates the candidate vector length: `L_cand = ||v_cand[k]||`. If `L_cand < 1e-6`, skip. It calculates the similarity angle: `cos_θ = (v_query · v_cand[k]) / `. (L_query × L_cand) Limit range: cos_θ = clamp(cos_θ, -1, 1) Angle: θ = arccos(cos_θ) Length difference: ΔL = |L_query - L_cand| / / Gaussian similarity: Orientation similarity: s_dir = exp(-θ² / (2σ_angle²)) Length similarity: s_len = exp(-ΔL² / (2σ_len²)) / / Overall score: score = α × s_dir + (1-α) × s_len If score > bestScore: bestScore = score bestIdx = k / / Threshold judgment If bestScore < τ_score: Return -1 used[bestIdx] = 1 Return bestIdx

[0112] For example: Suppose set A has a stopping line. There are around (Pedestrian crossing, 5 meters to the left) and (Speed ​​bump, 3 meters to the right). There is a stop line at assembly point B. There are around (Pedestrian crossing, 5.1 meters to the left) (Speed ​​bump, 2.9 meters on the right) and (Road markings, 10 meters ahead). The matching process will attempt to match. and (Vectors with similar directions and lengths score highly and are successful) Matching and (Also successful), and because There are no corresponding road marker neighbors, so it won't be matched. The fact that all neighbors can be matched indicates... and It is very likely that this corresponds to the generation of candidate pairs. and .

[0113] Note that for each element in set A, the above-mentioned neighbor matching must be performed one by one with all candidates of the same type in set B. Each matching generates a batch of candidate pairs. The same element may generate different candidate pairs in different matching processes, which requires a subsequent voting mechanism to filter them.

[0114] 3. A voting fusion mechanism is employed to identify reliable matches from a large number of candidate pairs. True matches are identified and receive a high number of votes in multiple local comparisons based on different reference factors, while accidental mismatches only appear in a few local comparisons and receive fewer votes. By statistically analyzing the frequency of occurrence and selecting the pair with the highest number of votes, true matches can be effectively identified and mismatches can be suppressed.

[0115] Specifically, for each candidate pair of correspondences, the frequency of its occurrence in all local alignments is counted. A voting record table is then established. Record candidate pairs The number of times a candidate pair is identified. Iterate through all the candidate pairs generated above: if a candidate pair points to a valid match (i.e., a corresponding element exists), then add one vote to that correspondence; if a candidate pair indicates that no match was found, then it is not counted in the statistics.

[0116] Let's illustrate with a practical example: Suppose we want to match pedestrian crossings. In multiple rounds of local comparison using different stop lines, speed bumps, and other reference elements, five instances of identification were detected. correspond It was identified once. correspond There were also two instances where the virus could not be found in certain local comparisons. The correspondence might be that the reference elements themselves didn't match. Therefore, in the voting records... and The corresponding candidate received 5 votes. and The corresponding result receives 1 vote, while the two "not found" results do not participate in the vote.

[0117] After the voting statistics are completed, for each element in set A, the candidate corresponding to the element that received the most votes is selected as the final match. If an element does not receive any valid votes in all local comparisons, it is marked as no match. Continuing with the example above, Among the two candidate correspondences With 5 votes, far exceeding With 1 vote, the final result was determined. match This "majority rule" mechanism effectively filters out accidental false matches, significantly improving matching reliability.

[0118] Final output matching mapping table Record the matching results for each feature in set A: if the feature If there is a clear correspondence, it is recorded as: If no corresponding entry is found, record it as: .

[0119] Furthermore, in the centroid vector matrix construction stage, relative position vectors are used instead of absolute coordinates to represent the spatial relationships between elements, making the method translation-invariant and rotation-invariant. Even if there is a large translation or rotation between two sets of elements, accurate matching is still possible. In the local candidate matching relationship generation stage, the semantic type information of elements is fully utilized. Only the relationships between different types of elements are recorded during matrix construction, only elements of the same type are matched during candidate search, and the consistency of neighbor types is enforced during similarity calculation, avoiding semantic errors such as mismatching stop lines as pedestrian crossings.

[0120] Example 2

[0121] Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of the map feature association and matching system based on centroid vectors disclosed in an embodiment of the present invention. Figure 2 The described centroid-based map feature association and matching system can be applied to data processing chips, processing terminals, or processing servers, and the processing server can be a local server or a cloud server; this embodiment of the invention does not limit the application. Figure 2 As shown, this centroid-based map feature association and matching system can include the following operations:

[0122] The matrix construction module 201 is used to construct centroid vector matrices for the first set of map features and the second set of map features to be matched, respectively. The centroid vector matrices are used to record the relative positional relationships between different types of features in the set in vector form.

[0123] Specifically, this step makes the matching method no longer rely on the absolute position of each element in the global coordinate system when processing data.

[0124] As can be seen, this design makes the matching process naturally invariant to the overall translation and rotation that may exist between the two sets of maps to be matched, thus establishing a unified benchmark for subsequent similarity comparison and decision-making that does not change with the coordinate system.

[0125] The candidate matching generation module 202 is used to find similar elements as candidates in the second map element set for the target element in the first map element set, and generate local candidate matching pairs by comparing the distribution similarity of the target element and the neighbor elements of each candidate element.

[0126] Specifically, this step matches target elements by comparing the distribution patterns of their local neighbors with those of their candidate elements, and in doing so, it enforces semantic type consistency constraints on the elements. This mechanism makes the matching decision not only dependent on spatial proximity, but also on the structural relationships and category information between elements.

[0127] As can be seen, the present invention can effectively identify corresponding elements with similar local contexts without relying on the global shape or integrity of the map. At the same time, it fundamentally avoids the incorrect association of elements with different semantic types (such as stop lines and pedestrian crossings). Its matching judgment is based on calculable geometric similarity, and the process is clear and interpretable.

[0128] The voting decision module 203 is used to perform voting statistics on all generated local candidate matching pairs, and determine the final matching correspondence in the second map element set for the elements in the first map element set based on the voting results.

[0129] Specifically, the frequency of occurrence of all local candidate matching pairs is counted, and the final matching relationship is determined for each element based on the majority principle.

[0130] It is evident that this voting mechanism can effectively identify and adopt consistent matching relationships that have been identified in multiple independent local comparisons, while filtering out accidental mismatches that only occasionally occur. Thus, without relying on complex models or prior knowledge, it improves the confidence and overall consistency of the final matching results, and the process has clear statistical interpretability.

[0131] Example 3

[0132] Please see Figure 3 , Figure 3 This is a schematic diagram of another map feature association and matching system based on centroid vectors disclosed in an embodiment of the present invention. Figure 3 As shown, the device may include:

[0133] Memory 301 storing executable program code;

[0134] Processor 302 coupled to memory 301;

[0135] The processor 302 calls the executable program code stored in the memory 301 to execute some or all of the steps in the centroid vector-based map feature association and matching method disclosed in Embodiment 1 of the present invention.

[0136] Example 4

[0137] This invention discloses a computer storage medium storing computer instructions. When these computer instructions are invoked, they are used to execute some or all of the steps in the centroid vector-based map feature association and matching method disclosed in Embodiment 1 of this invention.

[0138] Example 5

[0139] This invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps of the centroid vector-based map feature association and matching method described in Embodiment 1.

[0140] The system embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0141] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0142] Finally, it should be noted that the map feature association and matching method and system based on centroid vector disclosed in the embodiments of the present invention are only preferred embodiments of the present invention, and are only used to illustrate the technical solutions of the present invention, and not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A map feature association and matching method based on centroid vectors, characterized in that, The method includes: For the first and second map feature sets to be matched, centroid vector matrices are constructed respectively. The centroid vector matrices are used to record the relative positional relationships between different types of features within the set in vector form. For a target element in the first map element set, similar elements are searched for as candidates in the second map element set. Local candidate matching pairs are generated by comparing the distribution similarity between the target element and the neighbor elements of each candidate element. A voting statistics are performed on all the generated local candidate matching pairs, and the final matching correspondence in the second map element set is determined for the elements in the first map element set based on the voting results.

2. The map feature association and matching method based on centroid vectors according to claim 1, characterized in that, The construction of centroid vector matrices for the first and second map feature sets to be matched includes: For the first set of elements Second element set , respectively construct the corresponding centroid vector matrix for the two sets; Wherein, for each of the centroid vector matrices, its i-th Line number Column storage from features Point to feature The centroid coordinate difference vector, and only if the element With elements When the types are different, the centroid coordinate difference vector is stored as the effective vector.

3. The map feature association and matching method based on centroid vectors according to claim 2, characterized in that, The step of generating local candidate matching pairs by comparing the distribution similarity between the target feature and the neighboring features of each candidate feature includes: For the first set of elements Each of the first elements In the second set of elements Search for all elements that match the first element. Second elements of the same type constitute a candidate set; For each candidate second element in the candidate set Perform a neighbor matching operation, which includes: extracting the first element. The effective neighbor vector set and its neighbor types, and the candidate second element. The set of valid neighbor vectors and their neighbor types; For the first element Each neighbor vector in the candidate second element In the set of neighbor vectors, find candidate vectors that are of the same type and are not occupied, and calculate the comprehensive similarity score between the neighbor vector and each candidate vector that meets the conditions. Based on the comprehensive similarity score, the first element is... Each neighboring element determines its position in the candidate second element. The best match or non-match among the neighbors is used to generate a batch of candidates for the second element. Relevant local candidate matching pairs.

4. The map feature association and matching method based on centroid vectors according to claim 3, characterized in that, The calculation of the comprehensive similarity score includes: Calculate the directional similarity between the neighbor vector and the candidate vector. and length similarity ; The directional similarity and length similarity Calculated using the following formula: ； ； in, Let be the angle between the neighbor vector and the candidate vector. The difference in length between the neighbor vector and the candidate vector is... For directional tolerance parameters, This refers to the distance tolerance parameter. The comprehensive similarity score Calculated by weighted sum: ,in This represents the directional similarity weight.

5. The map feature association and matching method based on centroid vectors according to claim 3, characterized in that, The neighbor matching operation further includes traversing the first element. When performing a match on neighboring elements, the order of their neighbor indexes is randomized.

6. The map feature association and matching method based on centroid vectors according to claim 4, characterized in that, The directional tolerance parameter The distance tolerance parameter The directional similarity weight and the minimum score threshold used to determine the validity of a match. It can adaptively adjust according to the element density and road topology characteristics of the application scenario.

7. The map feature association and matching method based on centroid vectors according to claim 1, characterized in that, The step of voting and counting all generated local candidate matching pairs, and determining the final matching correspondence in the second map feature set based on the voting results for the elements in the first map feature set includes: Statistical analysis of each set of first elements Elements With elements in the second element set B The total number of times the candidate matching pairs appear is taken as the number of votes for that candidate matching pair; For the first set of elements For each element in the process, the second element with the highest number of votes among all its candidate matching pairs that have received votes is determined as the final matching result. If the first element set If an element in a dataset fails to receive a valid vote in any of the local candidate matching pairs, it is marked as having no match.

8. The map feature association and matching method based on centroid vectors according to claim 2, characterized in that, The centroid vector matrix is ​​stored in a sparse matrix format.

9. A centroid-based map feature association and matching system, used to implement the centroid-based map feature association and matching method according to any one of claims 1-8, characterized in that, The system includes: The matrix construction module is used to construct centroid vector matrices for the first and second map feature sets to be matched, respectively. The centroid vector matrices are used to record the relative positional relationships between different types of features within the set in vector form. The candidate matching generation module is used to find similar elements as candidates in the second map element set for the target element in the first map element set, and generate local candidate matching pairs by comparing the distribution similarity of the target element and the neighbor elements of each candidate element. The voting decision module is used to perform voting statistics on all generated local candidate matching pairs, and determine the final matching correspondence of the elements in the second map element set based on the voting results.

10. A map feature association and matching system based on centroid vectors, characterized in that, The system includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the map feature association and matching method based on centroid vector as described in any one of claims 1-8.

Images