Automobile wind noise point cloud data preprocessing method, system, device and storage medium

Abstract

This application belongs to the field of automotive wind noise prediction technology. It provides a method, system, device, and storage medium for preprocessing automotive wind noise point cloud data, including: identifying the coordinates, normal vectors, and label fields of the point cloud file; performing attitude enhancement on the point coordinates and normal vectors using the same rotation matrix; dividing the target surface and reference surface according to the normal vector direction; performing unilateral translation along the normal vector of the target surface at multiple distance levels and merging it with the reference surface; performing structure and label completeness checks, label mapping, and point order randomization on the enhanced point cloud; and generating training, validation, and testing indexes by grouping according to relative directories. This application can maintain the consistency of the geometric and vector attributes of the enhanced data, achieve interpretable local geometric variant generation, and improve the traceability of data preprocessing and the quality of training data.

Description

Technical Field

[0001] This application belongs to the field of automotive wind noise prediction technology, and in particular relates to a method, system, device and storage medium for preprocessing automotive wind noise point cloud data. Background Technology

[0002] Developing automotive wind noise mitigation systems typically requires extensive styling iterations for the front grille, rearview mirrors, A-pillars, windshield wipers, and other body components. Traditional methods rely primarily on wind tunnel testing or computational fluid dynamics simulations to evaluate different solutions, with each solution requiring a lengthy process from geometric preparation and mesh generation to solution development. Employing point cloud deep learning models to establish the mapping between component geometry and wind noise response can shorten the solution selection time; however, model performance depends on the quantity of training data, physical consistency, and mesh quality.

[0003] Point cloud files used for wind noise prediction do not only contain three-dimensional coordinates. Depending on the data source, the files also include surface normal vectors and component labels, for example, using a seven-column structure of "X, Y, Z, nx, ny, nz, label" per row, or a four-column structure of "X, Y, Z, label". Existing preprocessing methods often treat point clouds as ordinary numerical tables, which has at least the following problems: (1) The number of samples is limited by simulation cost. Directly copying the point cloud cannot increase geometric changes, and simply changing the arrangement of points cannot form a new design sample; if only the coordinates are rotated without rotating the normal vector, the surface orientation will be inconsistent with the geometric position, resulting in data distortion.

[0004] (2) Wind noise sensitive components are often designed by varying the spacing, thickness, or surface position. General random disturbances can change multiple areas simultaneously, making it difficult to correspond to specific engineering design variables and also hindering the interpretation of model output. Manually selecting a surface point by point and then moving it involves a large amount of work, and different personnel may obtain inconsistent surface ranges.

[0005] (3) The original files and enhancement files are usually scattered in multi-level directories, and there are often problems such as incorrect column count, missing labels, inconsistent label numbers, non-standard line breaks at the end of the file and fixed point order. If they are not checked uniformly before entering the training phase, the erroneous files will only be exposed during the training phase; if only a global random sampling is performed on all files, small-scale subdirectories may not be sampled into a certain data subset, resulting in uneven data distribution.

[0006] (4) Currently, enhancement, verification, point order processing, and dataset list generation are often completed by multiple independent scripts, and the output files lack a unified source identifier and processing record. When anomalies occur, it is difficult to determine which original file, which set of rotation parameters, or which translation distance generated the sample. Summary of the Invention

[0007] In view of this, this application aims to provide a method, system, device and storage medium for preprocessing automotive wind noise point cloud data to solve at least one of the above problems.

[0008] To achieve the above objectives, the technical solution of this application is implemented as follows: Firstly, this application provides a method for preprocessing automotive wind noise point cloud data, including: Obtain point cloud files of automotive external components, identify the field patterns of each point record, and convert point coordinates, surface normal vectors, and component labels into a unified data structure; A rotation matrix is ​​constructed based on the set of rotation parameters in the traversed point cloud file. The point coordinates are rotated based on the rotation matrix. When the point record contains a surface normal vector, the surface normal vector is transformed synchronously using the same rotation matrix while keeping the part label unchanged, thus obtaining the attitude-enhanced point cloud. For a point cloud containing surface normal vectors, the target surface and reference surface are determined according to the direction of the normal vectors. The points on the target surface are moved along the direction of the normal vector of the target surface according to the translation distance set, while keeping the points on the reference surface unchanged. The moved target surface and reference surface are merged to obtain a locally geometrically enhanced point cloud. Based on the field pattern, the point cloud containing component labels and used for model training is subjected to field structure checks, label mapping, and label set completeness checks. The point cloud that passes the checks is randomized in point order, and the dataset is divided into groups according to relative directories to generate point cloud files and dataset indexes for wind noise prediction model training.

[0009] Secondly, based on the same inventive concept, this application also provides a vehicle wind noise point cloud data preprocessing system, including: The directory scanning and format parsing module is used to acquire point cloud files, identify field patterns, and form a unified data structure containing point coordinates, surface normals, and component labels. The coordinate-normal vector co-rotation module is used to construct a rotation matrix based on the rotation axis and rotation angle, and to perform homogeneous rotation on the point coordinates and the surface normal vectors contained therein; The normal vector facet and orientation translation module is used to determine the target surface and the reference surface according to the normal vector direction, and move the target surface along the normal vector direction of the target surface by multiple translation distances; The structure and semantic quality check module is used to check the number of fields, numerical parsability, normal vector quantity constraints, and label set completeness, and outputs the reasons for file-level anomalies. The label mapping and point order processing module is used to uniformly number the labels according to the mapping table and perform random permutations on the point records; The directory hierarchy sampling and indexing module is used to partition the dataset according to relative directory groups and generate a dataset index for training the wind noise prediction model. The results recording module is used to save enhancement files, anomaly lists, processing parameters, and dataset lists, so that each output sample can be traced back to the original file and enhancement parameters.

[0010] Thirdly, based on the same inventive concept, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in the first aspect.

[0011] Fourthly, based on the same inventive concept, this application also provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method as described in the first aspect.

[0012] Compared with existing technologies, the automotive wind noise point cloud data preprocessing method, system, device, and storage medium described in this application have the following advantages: This application achieves automated preprocessing of point cloud training data by performing adaptive field parsing, coordinate and normal vector co-enhancement, local surface orientation transformation based on normal vector constraints, and data quality detection and index generation on point cloud data for wind noise prediction. While ensuring the consistency of geometric structure and surface semantics, it effectively improves the diversity, reliability and traceability of training samples for wind noise prediction models, reduces manual processing costs, and improves the efficiency of dataset construction and the quality of model training data. Attached Figure Description

[0013] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a method for preprocessing automotive wind noise point cloud data according to an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a vehicle wind noise point cloud data preprocessing system according to an embodiment of this application; Figure 3 This is a schematic diagram of the hardware structure of the electronic device described in an embodiment of this application. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0015] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0016] This application accepts the original wind noise point cloud catalog and preprocessing parameters. The preprocessing parameters include at least the allowed data field patterns, rotation axis and rotation angle sets, target surface selection rules, translation distance sets, expected label sets, label mapping relationships, random seeds, and dataset partitioning ratios.

[0017] First, the directory is scanned and the point cloud files are parsed, converting different delimiters and field patterns into a unified in-memory data structure. Then, rotation enhancement (overall pose enhancement) and / or local surface enhancement are performed according to the task configuration. Overall pose enhancement uses the same rotation matrix for point coordinates and normal vectors, while local surface enhancement separates the target surface from the reference surface based on the normal vector distribution, moving only the target surface and preserving the reference surface. The enhancement results enter the quality processing chain, undergoing structural and label set checks, followed by label unification, point order randomization, and directory hierarchy index generation. The final output includes a qualified point cloud, an anomaly list, and training, validation, and test set indices.

[0018] Rotational enhancement and local surface enhancement can be applied to the original sample separately, or sequentially after the field format is standardized. To avoid confusion regarding the source of the enhanced samples, the original file identifier, enhancement type, and enhancement parameters are recorded in the output file name or accompanying metadata.

[0019] The embodiments of this application are described in detail below with reference to the accompanying drawings.

[0020] Please see Figure 1 As shown in the figure, this embodiment provides a method for preprocessing automotive wind noise point cloud data, which specifically includes the following steps: Step S10: Obtain the point cloud file of the vehicle's external components, identify the field pattern of each point record, and convert the point coordinates, surface normal vectors, and component labels into a unified data structure.

[0021] Specifically, in this embodiment, the root directory of the original point cloud is obtained, and text files in each subdirectory are recursively scanned. For each file, the root directory name, relative directory, original file name, and extension are recorded. Backslashes in relative paths are uniformly converted to forward slashes, and the point cloud file extensions are removed from the index to adapt to training programs on different operating systems. If a point cloud file does not exist in the directory, "No file to process" is recorded; if a single file cannot be opened, the file path and read exception are recorded to prevent the exception from interrupting the batch processing of other files.

[0022] Read non-empty lines from the file, split the data into fields based on whitespace characters or configured delimiters, and determine the number of fields. Four columns of data are parsed as "coordinates + label", six columns as "coordinates + normal vector", and seven columns as "coordinates + normal vector + label". Lines not belonging to the allowed field patterns are marked as illegal lines, and the filename and line number are recorded.

[0023] All coordinate and normal vector fields are converted to floating-point numbers, and the label field is converted to a numeric category identifier. If different records in the file use different numbers of fields, or if there are numerical values ​​that cannot be converted, then that file will not proceed to the subsequent enhancement steps.

[0024] For the first point cloud Let there be a point, and define its three-dimensional coordinates as follows: ; In the formula, , and They represent the first The coordinate components of a point in a three-dimensional coordinate system. Its surface normal vector is defined as: ; In the formula, , and They represent the first The normal vector components of each point; For component labels. A unified data record can be represented as: ; When the original file has a four-column structure, the records do not contain It can perform coordinate rotation and label processing, but does not enter the surface translation step that depends on the normal vector; when the original file has a six-column structure, the records contain coordinates and normal vectors but no labels, and local surface enhancement can be performed, and then labels are added according to the data source or output as unlabeled geometric samples; when the original file has a seven-column structure, all processing steps can be performed.

[0025] Step S20: Construct a rotation matrix based on the set of rotation parameters in the traversed point cloud file, perform rotation on the point coordinates based on the rotation matrix, and synchronously transform the surface normal vector using the same rotation matrix when the point record contains the surface normal vector, keeping the part label unchanged, to obtain the attitude-enhanced point cloud.

[0026] Specifically, in this embodiment, for files that pass the initial check, the system iterates through the set of rotation parameters: ; In the formula, For the first The rotation angle corresponding to the group of rotation parameters. For the axis of rotation, This represents the number of rotation parameter sets. The system converts angles to radians and constructs a rotation matrix based on the rotation axis, around... axis, shaft and The rotation matrices of the axes are as follows: ; ; ; Perform the following transformation on each point: ; If the file contains surface normal vectors, use the same rotation matrix for the surface normal vectors as for the point coordinates: ; Component labels do not change with spatial orientation: ; The above matrix satisfies Therefore, rigid body rotation does not change the distance between points; the coordinates and normal vectors use the same matrix, which also preserves the angular relationship between them. In the program implementation, when points are stored as row vectors, the same matrix is ​​used... Equivalent form of multiplication.

[0027] Each set of rotation parameters generates an augmented file, with the filename containing the original file's backbone, rotation axis, and rotation angle. If the original sample size is... The number of rotation parameters is Then this step generates One rotated sample; whether to retain the original sample is determined by the output configuration.

[0028] Step S30: For a point cloud containing surface normal vectors, determine the target surface and reference surface according to the direction of the normal vectors. Move the points on the target surface along the direction of the normal vector of the target surface according to the translation distance set, while keeping the points on the reference surface unchanged. Merge the moved target surface and the reference surface to obtain a locally geometrically enhanced point cloud.

[0029] Specifically, in this embodiment, this step automatically distinguishes the target surface and the reference surface using the two principal normal vector directions in the point cloud, and moves the target surface along the normal vector direction of the target surface at multiple distance levels. After moving, the original surface is replaced with a new surface, and then merged with the unchanged reference surface.

[0030] Furthermore, for files requiring local surface enhancement, the normal vector matrix is ​​extracted and the directions of different normal vectors are statistically analyzed. A preferred implementation requires the point cloud to consist of two principal surfaces, and the set of normal vectors to satisfy: ; Based on the preset target direction, label rules, or normal vector sorting results, select one of the two principal directions as the target normal vector. Points satisfying this direction constitute the target surface. The remaining points constitute the reference surface. : ; ; When the normal vectors in the input data are discrete and standardized fixed values, the normal vector components can be directly compared; when the normal vectors are derived from floating-point calculations, clustering can be performed based on the angle between their directions. The similarity of the angle between the normal vector and the target normal vector can be expressed as: ; when If the value is not less than a set threshold, the point is assigned to the target surface. If two main surfaces that meet the requirements cannot be formed, "abnormal normal vector type" is recorded and the translation enhancement of the file is skipped. For curved point clouds, cosine similarity clustering, K-means clustering, or region growing algorithms can also be used to identify the main surfaces; the target surface can also be determined based on normal vectors, component labels, spatial bounding boxes, or a combination of the three.

[0031] After determining the target surface, iterate through the set of translation distances: ; In the formula, This is the translation distance for the first gear. This refers to the number of translation distance increments.

[0032] For points on the target surface, translate along the direction of the target normal vector; for points on the reference surface, keep the coordinates unchanged, and the coordinate update relationship is as follows: ; Before performing the translation, normalize the target normal vector: ; When the input normal vector is already a unit vector The distance can be calculated directly using the vector multiplication method in the code. After translation, the coordinates are retained to the configured decimal places, preferably four. The original coordinates of the target surface are replaced by the translated coordinates, and then merged with the unmoved reference surface. The number of output points is the same as the number of input points.

[0033] Each distance setting generates a separate file, with the filename containing "translated" and the distance value. The local enhancement is not an unconstrained random perturbation, but rather uses the normal vector as both the surface identity and the direction of design change to simulate parameter variations in surface spacing, blade position, or local thickness. The translation distance can be generated using a fixed list, or at equal intervals, randomly sampled, or in a Latin hypercube manner within given upper and lower limits.

[0034] Step S40: Based on the field pattern, perform field structure checks, label mapping, and label set completeness checks on the point clouds containing component labels and used for model training. Randomize the point order of the point clouds that pass the checks, and divide the dataset into groups according to relative directories to generate point cloud files and dataset indexes for wind noise prediction model training.

[0035] Specifically, in this embodiment, the labels are unified according to a configurable mapping table. For example, when conflicting numbers are used in different data sources, the original labels 0 and 1 can be mapped to 50 and 51; the specific numbers mentioned above are only examples, and the mapping relationship is configured according to the actual data source.

[0036] For the seven-column point cloud file used for model training, the system checks the number of fields row by row and extracts the actual set of labels from the file: ; Suppose that the set of labels required for this type of point cloud is . Then the set of missing labels is: ; The conditions for a document to pass quality gate control are: ; In the formula, This is an indicator function that, when the number of fields in each row meets the requirements and the set of missing labels is empty, A score of 1 indicates the file passes quality gating. Files that fail are added to an exception list, recording illegal lines or missing tags; files that pass proceed to point-order processing. Tag mapping can be performed either before completeness checks or during the check by transforming the expected tag set according to the same mapping relationship; both should use the same version of the mapping table.

[0037] For documents that pass quality gates, let's assume they have a total of For each point, the system generates an index set. random permutations The output order is as follows: ; This step only changes the arrangement of the point records, without altering the coordinates, normal vector, or label of any single point. Necessary line ending characters are added during reading, extra blank lines are avoided during output, and a random seed can be recorded for reproduction.

[0038] Create file mappings using relative directories as grouping keys: ; In the formula, For relative directory grouping sets, This represents the number of groups.

[0039] For the In each subdirectory The files are first shuffled under a fixed random seed, then sorted according to a configurable training set ratio. Validation set ratio and test set ratio Divide, and satisfy: ; The number of files in each subdirectory can be determined by the following formula: ; ; ; The partitioning results for each subdirectory are aggregated into training, validation, and testing lists, respectively. Compared to a single global sampling, this method ensures that all source subdirectories participate in the partitioning according to the same rules, reducing the possibility of small directories being completely absent from a certain subset of data. The output path adopts the format "root directory name / relative directory / file name", uniformly using forward slashes and removing file extensions.

[0040] When multiple augmented samples are generated from the same original file and strict prevention of augmentation lineages from crossing sets is required, lineage groups can be established first based on the original backbone in the filename, and then sampling can be performed using these lineage groups as indivisible units. Dataset partitioning can also employ K-fold cross-validation based on training requirements.

[0041] The system ultimately outputs a rotation-enhanced point cloud file, a local surface translation-enhanced point cloud file, a standard point cloud file that has passed quality gating and completed point order processing, an anomaly list containing filenames, line numbers, anomaly types, and missing labels, a JSON index for the training / validation / test sets, and a processing record containing the original file, enhancement method, rotation parameters, translation distance, label mapping version, and random seed. The processing record can be saved as a separate JSON file or written to a database or data management platform.

[0042] like Group rotation parameters and The translation distance of the gear is respectively applied to If there are 1 original sample, then the number of augmented samples that can be added is 1. .

[0043] When rotational enhancement and local surface translation enhancement are combined sequentially, the actual augmentation amount is determined based on the enhancement arrangement. Point cloud files can be in TXT, CSV, NPY, HDF5, or other data formats capable of storing coordinates, surface normals, and part labels. Quality checks can also be expanded to include label point ratio, coordinate range, normal vector magnitude, and duplicate point checks, depending on task requirements.

[0044] This application achieves automated preprocessing of point cloud training data by performing adaptive field parsing, coordinate and normal vector co-enhancement, local surface orientation transformation based on normal vector constraints, and data quality detection and index generation on point cloud data for wind noise prediction. While ensuring the consistency of geometric structure and surface semantics, it effectively improves the diversity, reliability and traceability of training samples for wind noise prediction models, reduces manual processing costs, and improves the efficiency of dataset construction and the quality of model training data.

[0045] It should be noted that the above description describes some embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0046] Based on the same inventive concept, and corresponding to the methods of any of the above embodiments, the embodiments of this application also provide an automotive wind noise point cloud data preprocessing system.

[0047] like Figure 2 As shown, the vehicle wind noise point cloud data preprocessing system includes: The directory scanning and format parsing module 11 is used to acquire point cloud files, identify field patterns, and form a unified data structure containing point coordinates, surface normal vectors, and component labels. The coordinate-normal vector cooperative rotation module 12 is used to construct a rotation matrix based on the rotation axis and rotation angle, and to perform homogeneous rotation on the point coordinates and the surface normal vectors contained therein; The normal vector facet and orientation translation module 13 is used to determine the target surface and the reference surface according to the normal vector direction, and move the target surface along the normal vector direction of the target surface by multiple translation distances; The structure and semantic quality inspection module 14 is used to check the number of fields, numerical parsability, normal vector quantity constraints, and label set completeness, and outputs the reasons for file-level anomalies. The label mapping and point order processing module 15 is used to uniformly number the labels according to the mapping table and perform random permutations on the point records; Directory-level sampling and indexing module 16 is used to partition the dataset according to relative directory groups and generate a dataset index for training the wind noise prediction model. The results recording module 17 is used to save enhancement files, anomaly lists, processing parameters, and dataset lists so that each output sample can be traced back to the original file and enhancement parameters.

[0048] For ease of description, the above apparatus is described in terms of its functions, divided into various modules. Of course, in implementing the embodiments of this application, the functions of each module can be implemented in one or more software and / or hardware.

[0049] The apparatus of the above embodiments is used to implement the corresponding method in any of the foregoing embodiments and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0050] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, embodiments of this application also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the methods described in any of the above embodiments.

[0051] Figure 3 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0052] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0053] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0054] The input / output interface 1030 is used to connect input / output modules to realize information input and output. The input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0055] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0056] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0057] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0058] The electronic devices described above are used to implement the corresponding methods in any of the foregoing embodiments and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0059] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium that stores computer instructions for causing the computer to perform the methods described in any of the above embodiments.

[0060] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0061] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to perform the methods described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0062] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0063] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the claims of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. A method for preprocessing automobile wind noise point cloud data, characterized in that, include: Obtain point cloud files of automotive external components, identify the field patterns of each point record, and convert point coordinates, surface normal vectors, and component labels into a unified data structure; A rotation matrix is ​​constructed based on the set of rotation parameters in the traversed point cloud file. The point coordinates are rotated based on the rotation matrix. When the point record contains a surface normal vector, the surface normal vector is transformed synchronously using the same rotation matrix while keeping the part label unchanged, thus obtaining the attitude-enhanced point cloud. For a point cloud containing surface normal vectors, the target surface and reference surface are determined according to the direction of the normal vectors. The points on the target surface are moved along the direction of the normal vector of the target surface according to the translation distance set, while keeping the points on the reference surface unchanged. The moved target surface and reference surface are merged to obtain a locally geometrically enhanced point cloud. Based on the field pattern, the point cloud containing component labels and used for model training is subjected to field structure checks, label mapping, and label set completeness checks. The point cloud that passes the checks is randomized in point order, and the dataset is divided into groups according to relative directories to generate point cloud files and dataset indexes for wind noise prediction model training.

2. The method according to claim 1, characterized in that: The field patterns include four-column, six-column, and seven-column field patterns. The four-column field pattern includes three-dimensional coordinates and part labels, the six-column field pattern includes three-dimensional coordinates and surface normals, and the seven-column field pattern includes three-dimensional coordinates, surface normals, and part labels. When reading point cloud files, records that do not belong to the allowed field patterns are marked as illegal rows, and the corresponding file name and row number are recorded.

3. The method according to claim 1, characterized in that: The set of rotation parameters includes multiple sets of rotation parameters consisting of rotation axes and rotation angles. A corresponding orthogonal rotation matrix is ​​constructed according to each set of rotation parameters. The coordinates of each point and the surface normal vector contained therein are transformed using the same orthogonal rotation matrix, while keeping the component label unchanged. The product of the transpose of the orthogonal rotation matrix and the original matrix is ​​the identity matrix. Each set of rotation parameters generates an enhancement file, and the original file identifier, rotation axis, and rotation angle are written into the output file name or the accompanying processing record.

4. The method according to claim 1, characterized in that, The process of determining the target surface and the reference surface based on the normal vector direction includes: The direction distribution of the statistical normal vector matrix is ​​analyzed. When the set of normal vectors forms two main directions, one of the main directions is selected as the target normal vector based on the preset target direction, component label, or normal vector sorting result. The points whose normal vectors are the same as the target normal vectors are defined as the target surface, and the remaining points are defined as the reference surface. When the normal vector is obtained by floating-point calculation, clustering is performed based on the similarity of the angle between the normal vector and the target normal vector, and points with an angle similarity of not less than a set threshold are assigned to the target surface.

5. The method according to claim 1, characterized in that, The step of moving points on the target surface along the normal vector direction of the target surface according to the translation distance set includes: Traverse the set of translation distances containing multiple translation distances and normalize the target normal vector to a unit target normal vector; For points on the target surface, update their coordinates to the sum of the original coordinates, the unit target normal vector, and the current translation distance; for points on the reference surface, keep their coordinates unchanged. The updated target surface replaces the original target surface and is merged with the reference surface, so that the number of output points is the same as the number of input points, and a local geometrically enhanced point cloud is generated for each translation distance.

6. The method according to claim 1, characterized in that, The field structure check and tag set completeness check include: The number of fields is checked line by line and the actual label set of the point cloud file is extracted. The difference between the expected label set required to be included in the corresponding point cloud file and the actual label set is determined as the missing label set. When the number of fields in each row meets the requirements and the set of missing tags is empty, the file is determined to pass the quality gate; otherwise, the file, illegal lines, or missing tags are written to the exception list. Tag mapping is performed before the completeness check, or the expected tag set is transformed using the same tag mapping relationship during the completeness check.

7. The method according to claim 1, characterized in that, The point order randomization includes: For point cloud files that have passed quality gating, a random permutation is generated based on the point record index of the point cloud file according to the random seed, and the point records are rearranged according to the random permutation, wherein only the arrangement order of the point records is changed without changing the coordinates, surface normal vectors and component labels of any point. The processing record should at least save the original file, enhancement method, rotation parameters, translation distance, label mapping version, and random seed.

8. A preprocessing system for automotive wind noise point cloud data, characterized in that, include: The directory scanning and format parsing module is used to acquire point cloud files, identify field patterns, and form a unified data structure containing point coordinates, surface normals, and component labels. The coordinate-normal vector co-rotation module is used to construct a rotation matrix based on the rotation axis and rotation angle, and to perform homogeneous rotation on the point coordinates and the surface normal vectors contained therein; The normal vector facet and orientation translation module is used to determine the target surface and the reference surface according to the normal vector direction, and move the target surface along the normal vector direction of the target surface by multiple translation distances; The structure and semantic quality check module is used to check the number of fields, numerical parsability, normal vector quantity constraints, and label set completeness, and outputs the reasons for file-level anomalies. The label mapping and point order processing module is used to uniformly number the labels according to the mapping table and perform random permutations on the point records; The directory hierarchy sampling and indexing module is used to partition the dataset according to relative directory groups and generate a dataset index for training the wind noise prediction model. The results recording module is used to save enhancement files, anomaly lists, processing parameters, and dataset lists, so that each output sample can be traced back to the original file and enhancement parameters.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium, characterized in that, in, The non-transitory computer-readable storage medium stores computer instructions for causing a computer to perform the method described in any one of claims 1-7.

Images