An omnidirectional intelligent perception method and system for a double-track tunnel car washer

Abstract

The present application relates to the field of three-dimensional image processing, and particularly relates to an omnidirectional intelligent sensing method and system for a double-track tunnel car washer, the method comprising: collecting three-dimensional point cloud data of a vehicle to obtain a three-dimensional point cloud sequence; calculating a comprehensive quality score of each frame of point cloud from the dimensions of direction consistency and spatial symmetry; screening candidate templates of each frame of point cloud in a three-dimensional vehicle template library, and calculating a matching score between each frame of point cloud and the candidate templates; calculating a matching total score of each candidate template based on the matching score and the comprehensive quality score, selecting a candidate template with the highest matching total score as a template matched with the vehicle, and controlling the car washer to clean the vehicle according to the car washing control parameters bound with the template. The present application automatically gives a higher weight to reliable frames with high definition and high symmetry, while suppressing the influence of low-quality frames that are severely disturbed, thereby improving the accuracy and robustness of the identification process.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional image processing. In particular, it relates to an omnidirectional intelligent sensing method and system for a dual-track tunnel car wash machine. Background Technology

[0002] To achieve adaptive and precise vehicle cleaning, tunnel car wash machines need to quickly and accurately identify the vehicle type after it enters the tunnel but before the cleaning process begins. The reliability of the identification results directly determines the targeted nature and effectiveness of subsequent cleaning operations.

[0003] Currently, to achieve vehicle model recognition, existing technologies mainly rely on sensors such as optical cameras and LiDAR to acquire 3D point cloud data of vehicles, and then match it with pre-stored 3D vehicle model templates to complete the recognition. However, in the specific application scenario of tunnel car wash machines, the large amount of water mist, foam generated during the pre-cleaning stage and in the tunnel environment, as well as water droplets remaining on the vehicle surface, can cause scattering and refraction interference to the optical sensors and adhere to the vehicle surface, altering its local geometric features. Existing solutions treat the continuous multiple frames of point cloud collected during the vehicle's passage as uniform data and process them, such as through simple averaging or direct multi-frame matching. This fails to distinguish low-quality frames caused by instantaneous water mist obstruction, strong reflection, or sensor noise, resulting in these unreliable data being indiscriminately introduced into the recognition process, causing problems such as mismatches and fluctuations in recognition results.

[0004] Therefore, there is a need in this field for an omnidirectional intelligent sensing method and system for dual-track tunnel car wash machines to solve the technical problems of mismatch and fluctuations in recognition results caused by the existing solutions. Summary of the Invention

[0005] To address the technical problems of mismatch between vehicle and model templates and fluctuations in recognition results caused by the existing solutions, the present invention provides solutions in the following aspects.

[0006] In the first aspect, an omnidirectional intelligent sensing method for a dual-track tunnel car wash machine includes: Three-dimensional point cloud data of vehicles are collected at a fixed frequency and then subjected to coordinate unification and data preprocessing to obtain a three-dimensional point cloud sequence. Based on the three-dimensional point cloud sequence, a template matching the vehicle is selected from a pre-built three-dimensional vehicle template library, and the car wash machine is controlled to clean the vehicle according to the car wash control parameters bound to the template. The process of selecting templates to match the vehicle includes: taking any frame of the 3D point cloud sequence as the target frame point cloud; calculating the directional entropy of each point based on the unit normal vector of the points in the target frame point cloud; calculating the directional consistency score based on the directional entropy of all points; taking the vertical plane where the center line of the dual rails is located as the reference, taking the points in the target frame point cloud as the source points, calculating the symmetric points of each source point, obtaining the echo intensity vector of the source point and the echo intensity vector of the symmetric point, and calculating the spatial symmetry score; multiplying the directional consistency score and the spatial symmetry score to obtain the comprehensive quality score of the target frame point cloud; obtaining candidate templates for the target frame point cloud based on the 3D dimensions of the axis-aligned bounding box of the target frame point cloud and the 3D dimensions of the axis-aligned bounding box of each template, and calculating the matching score between the target frame point cloud and each candidate template; calculating the total matching score of each candidate template based on the matching score between each frame point cloud and its candidate templates and the comprehensive quality score of each frame point cloud, and selecting the candidate template with the highest total matching score as the template to match the vehicle.

[0007] Preferably, the method for obtaining the unit normal vector of a point in the target frame point cloud includes: taking any point in the target frame point cloud as the target point, constructing a local point set formed by the target point and its nearest neighbor, performing principal component analysis on the local point set to obtain three eigenvectors and the eigenvalues ​​corresponding to the three eigenvectors respectively, selecting the eigenvector corresponding to the smallest eigenvalue, and normalizing the eigenvector to obtain the unit normal vector of the target point.

[0008] Preferably, the calculation of the orientation entropy of each point based on the unit normal vector of the point in the target frame point cloud includes: dividing the range of cosine values ​​into a preset number of discrete intervals; calculating the cosine value of the angle between the unit normal vector of the target point and the unit normal vector of each nearest neighbor point, and counting the total number of cosine values ​​and the number falling into each discrete interval; for each discrete interval, calculating the ratio of the number of cosine values ​​falling into the discrete interval to the total number of cosine values, and obtaining the discrete probability distribution of the discrete interval; and calculating the orientation entropy of the target point using the Shannon entropy formula based on the discrete probability distribution of each discrete interval.

[0009] Preferably, the calculation of the orientation consistency score based on the orientation entropy of all points includes: performing negative correlation normalization on the orientation entropy of each point to obtain the normalized value of the orientation entropy; calculating the average value of the normalized values ​​of the orientation entropy of all points in the target frame point cloud to obtain the orientation consistency score of the target frame point cloud.

[0010] Preferably, obtaining the source point echo intensity vector and the symmetric point echo intensity vector includes: each source point and its symmetric point form a pair; after removing overlapping pairings, a full set of point pairs is obtained; the laser echo intensity value of each source point in the full set of point pairs is obtained to form the source point echo intensity vector; for each symmetric point in the full set of point pairs, if there is a point in the target frame point cloud that matches its coordinates, the laser echo intensity value of the symmetric point is set to the laser echo intensity value of the matching point; if there is no point in the target frame point cloud that matches its coordinates, the laser echo intensity value of the symmetric point is set to 0, thereby obtaining the symmetric point echo intensity vector.

[0011] Preferably, the method for calculating the spatial symmetry score includes: calculating the cosine similarity value between the source point echo intensity vector and the symmetry point echo intensity vector to obtain the spatial symmetry score of the target frame point cloud.

[0012] Preferably, the step of obtaining candidate templates for the target frame point cloud includes: calculating the three-dimensional dimensions of the axis-aligned bounding box of the target frame point cloud, retrieving the three-dimensional dimensions of the axis-aligned bounding box of each template in the three-dimensional vehicle model template library; for each dimension, calculating the difference rate between the dimension of the target frame point cloud and the dimension of each template; if the difference rate of the dimensions of the three dimensions does not exceed a preset threshold, then the template is used as a candidate template for the target frame point cloud.

[0013] Preferably, the step of calculating the total matching score of each candidate template based on the matching score between each frame point cloud and its candidate template and the comprehensive quality score of each frame point cloud includes: extracting the intersection of the candidate templates of all frames of point clouds in the 3D point cloud sequence to obtain a global candidate template set; for each candidate template in the global candidate template set, using the comprehensive quality score of each frame point cloud as a weight, weighted summing the matching scores of the candidate template and each frame point cloud to obtain the total matching score of the candidate template.

[0014] Preferably, the method for constructing the three-dimensional vehicle model template library includes: pre-collecting complete three-dimensional point cloud models of multiple typical vehicles in clean and dry states; performing standardization processing operations such as coordinate decentering, scale normalization, and main direction alignment on each collected original point cloud model in sequence to obtain a three-dimensional vehicle model template library composed of standardized point cloud models, with each template associated with a vehicle model tag and each template bound to a set of car wash control parameters.

[0015] Secondly, an omnidirectional intelligent sensing system for a dual-track tunnel car wash machine includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned omnidirectional intelligent sensing method for a dual-track tunnel car wash machine is implemented.

[0016] The present invention has the following effects: 1. This invention quantifies the degree of surface geometric feature disorder caused by water mist and noise by calculating the directional entropy of each point. By analyzing the echo intensity of symmetrical point pairs on the left and right sides of the vehicle body, it effectively captures asymmetrical interference such as unilateral occlusion and foam adhesion, fundamentally solving the problem that existing technologies cannot distinguish data quality.

[0017] 2. This invention uses the comprehensive quality score of each frame's point cloud as a weight to fuse the matching score between the frame's point cloud and the candidate template, enabling the system to automatically assign higher weights to reliable frames with high definition and high symmetry, while suppressing the influence of low-quality frames that are severely interfered with, thus greatly improving the accuracy and robustness of the recognition process. Attached Figure Description

[0018] Figure 1 This is a flowchart of steps S1-S2 in an omnidirectional intelligent sensing method for a dual-track tunnel car wash machine according to an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of steps S20-S24 in an omnidirectional intelligent sensing method for a dual-track tunnel car wash machine according to an embodiment of the present invention. Detailed Implementation

[0020] 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 some embodiments of the present invention, but not all embodiments.

[0021] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] Reference Figure 1 An omnidirectional intelligent sensing method for a dual-track tunnel car wash machine includes steps S1-S2, as follows: S1: Collect 3D point cloud data of the vehicle at a fixed frequency, perform coordinate unification and data preprocessing to obtain a 3D point cloud sequence.

[0023] To achieve high-quality acquisition of omnidirectional 3D point clouds of vehicles, this invention deploys three area-array LiDARs (or high-density line laser profilometers in other embodiments) on a fixed gantry at the tunnel entrance. The specific deployment is as follows: two area-array LiDARs are located on the left and right sides of the vehicle's direction of travel, respectively, and are strictly symmetrical about the vertical plane containing the center line of the dual tracks; the third area-array LiDAR is located at the top center of the fixed gantry, i.e., on the vertical plane containing the center line of the dual tracks. This deployment allows the area-array LiDARs to effectively scan the front, top, and left and right sides of the vehicle as it passes, forming a complete omnidirectional point cloud coverage.

[0024] Based on the above deployment method, a standardized cuboid detection space is defined in the tunnel's fixed coordinate system to delineate the core scanning area of ​​the vehicle, accommodating mainstream vehicle sizes. This detection space is specifically defined as follows: Axle: The lateral direction of the vehicle, ranging from left to right. In this embodiment, The center of symmetry is ; Axis: The longitudinal direction of vehicle travel, ranging from... In this embodiment, The starting point is the reference surface at the tunnel entrance; Axis: Vertical height direction, range is In this embodiment, The starting point is the track plane.

[0025] The area array lidar uses its built-in synchronous triggering unit to synchronously acquire 3D point cloud data of the vehicle surface at a sampling frequency of 100Hz (100 frames per second). After all the 3D point cloud data acquired by the area array lidar are aligned according to the sampling timestamp, they are uniformly transformed into the aforementioned tunnel fixed coordinate system using the pre-calibrated extrinsic parameter matrices of each area array lidar, forming a multi-frame omnidirectional 3D point cloud dataset.

[0026] To improve the accuracy and efficiency of subsequent identification and evaluation, the following preprocessing operations were performed on the raw 3D point cloud data collected above: Statistical Outlier Removal (SOR) was used to remove obvious outliers caused by sensor noise or environmental interference; voxel grid downsampling (grid size of 0.01m) was used to reduce the amount of data without losing key geometric features; and based on the detection space range defined above, pass-through filtering was used to extract the point cloud of the vehicle passage area.

[0027] After the above preprocessing, a high-quality 3D point cloud sequence is obtained that can be used for subsequent quality assessment and matching: ,in Total number of frames The sampling time is specified. Each frame of the point cloud in the 3D point cloud sequence includes the set of all points of the vehicle at the corresponding sampling time.

[0028] To provide an accurate benchmark for vehicle model identification and comparison, this embodiment pre-constructs a standardized 3D vehicle model template library. The specific construction steps are as follows: Complete 3D point cloud models of multiple typical vehicles in clean and dry conditions are pre-collected, covering common vehicle types, including but not limited to sedans, SUVs, MPVs, and small trucks; for each collected original point cloud model, standardized processing operations such as coordinate decentering, scale normalization, and principal direction alignment are sequentially performed (existing techniques will not be elaborated upon) to ensure template consistency and comparability; finally, a 3D vehicle model template library composed of standardized point cloud models is obtained. ,in This represents the total number of templates in the template library. The template number represents the number of each template. Each template is uniquely associated with a vehicle type tag, such as Sedan-Type A, SUV-Type B, and each template... A set of car wash control parameters is bound to the system. These parameters are preset according to the characteristics of the car model, including but not limited to: spray bar angle, water pressure level, washing time, and brush speed.

[0029] This template library not only serves as a comparison benchmark for vehicle model recognition, but also enables a direct mapping between recognition results and cleaning processes, supporting the system to automatically trigger a refined cleaning operation that matches the vehicle model after recognition is completed.

[0030] S2: Based on the 3D point cloud sequence, select a template that matches the vehicle from the pre-built 3D vehicle template library, and control the car wash machine to clean the vehicle according to the car wash control parameters bound to the template.

[0031] Reference Figure 2 Step S2 includes steps S20-S24, as detailed below: S20: Take any frame point cloud in the 3D point cloud sequence as the target frame point cloud, calculate the orientation entropy of each point based on the unit normal vector of the points in the target frame point cloud, and calculate the orientation consistency score based on the orientation entropy of all points.

[0032] Any frame of point cloud in the 3D point cloud sequence is taken as the target frame point cloud, and any point in the target frame point cloud is taken as the target point. A local point set is constructed by the target point and its nearest neighbor (in this embodiment, the total number of points in the local point set is 16, which can be adjusted according to the point cloud density). The unit normal vector of the target point is obtained based on the local point set, thereby obtaining the unit normal vector of each point. The orientation entropy of the target point is calculated based on the unit normal vector of the target point and its nearest neighbor. The orientation consistency score of the target frame point cloud is calculated based on the orientation entropy of all points in the target frame point cloud.

[0033] The process of obtaining the unit normal vector of the target point based on the local point set includes: performing principal component analysis (PCA) on the local point set to obtain three eigenvectors and the eigenvalues ​​corresponding to the three eigenvectors; selecting the eigenvector corresponding to the smallest eigenvalue; and normalizing the eigenvector to obtain the unit normal vector of the target point.

[0034] Principal component analysis (PCA) of a local point set can identify the three principal directions that best describe the distribution of these points, as well as the degree of dispersion along these three principal directions. The direction corresponding to the largest eigenvalue is the direction in which the data is most dispersed, which can be understood as the direction of the plane (such as the direction of the extension of a car door surface); the direction corresponding to the smallest eigenvalue is the direction in which the data is least dispersed and thinnest, that is, the direction perpendicular to the plane, which is also the direction of the normal vector.

[0035] Calculating the orientation entropy of a target point based on the unit normal vectors of the target point and its nearest neighbor includes: dividing the range of cosine values ​​into a predetermined number (e.g., 10) of discrete intervals; calculating the cosine of the angle between the unit normal vector of the target point and the unit normal vector of each nearest neighbor point, and counting the total number of cosine values ​​and the number falling into each discrete interval; for each discrete interval, calculating the ratio of the number of cosine values ​​falling into that discrete interval to the total number of cosine values, thus obtaining the discrete probability distribution of that discrete interval; and calculating the orientation entropy of the target point using the Shannon entropy formula based on the discrete probability distribution of each discrete interval.

[0036] The larger the directional entropy, the more chaotic the distribution of the normal vector direction in the local region where the local point set is located, and the more serious the geometric noise or surface interference (such as water mist); the smaller the directional entropy, the more concentrated and consistent the distribution of the normal vector direction in the local region where the local point set is located, and the clearer and more reliable the surface geometric features.

[0037] Calculating the orientation consistency score of the target frame point cloud based on the orientation entropy of all points in the target frame point cloud includes: performing negative correlation normalization on the orientation entropy of each point to obtain a normalized value of the orientation entropy; and calculating the average of the normalized values ​​of the orientation entropy of all points in the target frame point cloud to obtain the orientation consistency score of the target frame point cloud. The specific formula is as follows: In the formula, Indicates the sampling time The corresponding orientation consistency score of the whole frame point cloud; Indicates the sampling time The total number of points in the corresponding full-frame point cloud; Indicates the sampling time The corresponding full-frame point cloud The directional entropy of each point; Represented by natural constant It is a negative exponential function with base 0.

[0038] The larger the value, the higher the geometric orientation consistency of the entire point cloud surface, the lower the overall geometric noise level, and the higher the data reliability.

[0039] S21: Using the vertical plane where the center line of the double track is located as the reference, take the points in the target frame point cloud as the source points, calculate the symmetrical points of each source point, obtain the source point echo intensity vector and the symmetrical point echo intensity vector, and calculate the spatial symmetry score.

[0040] Using the vertical plane containing the center line of the double rails as the reference (i.e., based on) Using the left and right mirror symmetry center planes as a reference, points in the target frame point cloud are taken as source points. Symmetric points are obtained for each source point, and each source point and its symmetric point form a point pair, thus obtaining the initial set of point pairs. If two point pairs are mirror pairs of each other, they are considered as one set of mirror pairs. Point pair satisfy , If we keep only one set of point pairs and remove the overlapping point pairs from the initial set of point pairs, we get the full set of point pairs.

[0041] The laser echo intensity value of each source point in the full set of point pairs is obtained to form a source point echo intensity vector. For each symmetrical point in the full set of point pairs, if a point with matching coordinates exists in the target frame point cloud, the laser echo intensity value of the symmetrical point is set to the laser echo intensity value of the matching point; otherwise, the laser echo intensity value of the symmetrical point is set to 0, thus obtaining the symmetrical point echo intensity vector. The cosine similarity value between the source point echo intensity vector and the symmetrical point echo intensity vector is calculated to obtain the spatial symmetry score of the target frame point cloud. The specific formula is as follows: In the formula, Indicates the sampling time The corresponding spatial symmetry score of the whole frame point cloud; This represents the total number of point pairs in the complete set of point pairs; Represents the set of all point pairs. The laser echo intensity value of the source point of the point pair; Represents the set of all point pairs. The laser echo intensity value of the symmetrical point of the pair of points.

[0042] The closer it is to 1, the more consistent the echo intensity distribution of the left and right symmetrical point pairs in the entire frame point cloud is, with no unilateral interference or occlusion, optimal symmetry, and reliable data quality. The closer the value is to 0, the more likely it is to indicate significant unilateral complete occlusion, severe partial occlusion, or sensor noise, resulting in significant differences in echo intensity between symmetrical point pairs, poor point cloud symmetry, and low data quality.

[0043] S22: Multiply the orientation consistency score and the spatial symmetry score to obtain the overall quality score of the target frame point cloud.

[0044] The overall quality score of the target frame point cloud is obtained by multiplying its directional consistency score and spatial symmetry score. A higher overall quality score indicates that the point cloud in the frame has both high geometric directional consistency and good vehicle body structural symmetry, and the overall data quality is more reliable; conversely, a lower overall quality score indicates that the point cloud in the frame has obvious geometric noise or structural asymmetry, and the data quality is lower.

[0045] Similarly, the overall quality score of each frame of the point cloud in the 3D point cloud sequence can be obtained.

[0046] S23: Based on the 3D dimensions of the axis-aligned bounding box of the target frame point cloud and the 3D dimensions of the axis-aligned bounding box of each template, obtain the candidate templates of the target frame point cloud and calculate the matching score between the target frame point cloud and each candidate template.

[0047] Calculate the 3D dimensions of the axis-aligned bounding box of the target frame point cloud, including length, width, and height; retrieve the 3D dimensions of the axis-aligned bounding box of each template in the 3D vehicle template library; for each dimension, calculate the difference rate between the dimension of the target frame point cloud in that dimension and the dimension of each template in that dimension. If the difference rate of the dimensions in all three dimensions does not exceed a preset threshold (e.g., 30%, the threshold can be set and adjusted based on experience), then the template is used as a candidate template for the target frame point cloud; use the standard iterative nearest-point algorithm to calculate the matching score between the target frame point cloud and each of its candidate templates. The process of calculating the matching score is existing technology and will not be described in detail here.

[0048] Similarly, we can obtain the candidate templates for each frame of the point cloud in the 3D point cloud sequence and the matching score between each frame of the point cloud and its candidate template.

[0049] S24: Calculate the total matching score of each candidate template based on the matching score between each frame point cloud and its candidate template and the comprehensive quality score of each frame point cloud, and select the candidate template with the highest total matching score as the template to be matched with the vehicle.

[0050] Extract the intersection of candidate templates of all frame point clouds in the 3D point cloud sequence to obtain the global candidate template set and the matching score set of each frame point cloud with the template in the global candidate template set.

[0051] For each candidate template in the global candidate template set, the overall quality score of each frame's point cloud is used as a weight to weight the matching scores between the candidate template and each frame's point cloud, resulting in a weighted sum of the matching scores. The specific formula is as follows: In the formula, Indicates the first The total score of matching candidate templates with the 3D point cloud sequence; This represents the total number of point cloud frames in a 3D point cloud sequence. Indicates the sampling time The overall quality score of the corresponding whole-frame point cloud; Indicates the sampling time The corresponding whole frame point cloud and the first Matching scores between candidate templates. The larger the value, the higher the value. The higher the overall match between the candidate template and the vehicle.

[0052] The candidate template with the highest total matching score is selected as the final identification template for the vehicle. At the same time, the unique vehicle model tag and car wash control parameters associated with the template are extracted. The dual-track tunnel car wash machine is then controlled to clean the vehicle based on the extracted car wash control parameters.

[0053] This invention innovatively uses the comprehensive quality score of a single-frame point cloud as an adaptive weight to drive the weighted fusion decision of multi-frame matching results, enabling the system to have autonomous judgment capabilities, dynamically trust high-quality data frames and suppress interference from low-quality frames. This significantly improves the accuracy and robustness of vehicle model recognition results at the data source level, ensuring stable and reliable output in complex environments.

[0054] This application also discloses an omnidirectional intelligent sensing system for a dual-track tunnel car wash machine. The system includes a processor and a memory. The memory stores computer program instructions. When the computer program instructions are executed by the processor, the omnidirectional intelligent sensing method for a dual-track tunnel car wash machine according to the above embodiments of the present invention is implemented.

[0055] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0056] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. An omnidirectional intelligent sensing method for dual-track tunnel car wash machines, characterized in that, include: Three-dimensional point cloud data of vehicles are collected at a fixed frequency and then subjected to coordinate unification and data preprocessing to obtain a three-dimensional point cloud sequence. Based on the three-dimensional point cloud sequence, a template matching the vehicle is selected from a pre-built three-dimensional vehicle template library, and the car wash machine is controlled to clean the vehicle according to the car wash control parameters bound to the template. The process of selecting templates to match the vehicle includes: taking any frame of the 3D point cloud sequence as the target frame point cloud; calculating the directional entropy of each point based on the unit normal vector of the points in the target frame point cloud; calculating the directional consistency score based on the directional entropy of all points; taking the vertical plane where the center line of the dual rails is located as the reference, taking the points in the target frame point cloud as the source points, calculating the symmetric points of each source point, obtaining the echo intensity vector of the source point and the echo intensity vector of the symmetric point, and calculating the spatial symmetry score; multiplying the directional consistency score and the spatial symmetry score to obtain the comprehensive quality score of the target frame point cloud; obtaining candidate templates for the target frame point cloud based on the 3D dimensions of the axis-aligned bounding box of the target frame point cloud and the 3D dimensions of the axis-aligned bounding box of each template, and calculating the matching score between the target frame point cloud and each candidate template; calculating the total matching score of each candidate template based on the matching score between each frame point cloud and its candidate templates and the comprehensive quality score of each frame point cloud, and selecting the candidate template with the highest total matching score as the template to match the vehicle.

2. The omnidirectional intelligent sensing method for a dual-track tunnel car wash machine according to claim 1, characterized in that, The method for obtaining the unit normal vector of a point in the target frame point cloud includes: taking any point in the target frame point cloud as the target point, constructing a local point set formed by the target point and its nearest neighbor, performing principal component analysis on the local point set to obtain three eigenvectors and the eigenvalues ​​corresponding to the three eigenvectors respectively, selecting the eigenvector corresponding to the smallest eigenvalue, and normalizing the eigenvector to obtain the unit normal vector of the target point.

3. The omnidirectional intelligent sensing method for a dual-track tunnel car wash machine according to claim 2, characterized in that, The calculation of the orientation entropy of each point based on the unit normal vector of the point in the target frame point cloud includes: dividing the range of cosine values ​​into a preset number of discrete intervals; calculating the cosine value of the angle between the unit normal vector of the target point and the unit normal vector of each nearest neighbor point, and counting the total number of cosine values ​​and the number falling into each discrete interval; for each discrete interval, calculating the ratio of the number of cosine values ​​falling into the discrete interval to the total number of cosine values, and obtaining the discrete probability distribution of the discrete interval; and calculating the orientation entropy of the target point using the Shannon entropy formula based on the discrete probability distribution of each discrete interval.

4. The omnidirectional intelligent sensing method for a dual-track tunnel car wash machine according to claim 1, characterized in that, The calculation of the directional consistency score based on the directional entropy of all points includes: performing negative correlation normalization on the directional entropy of each point to obtain the normalized value of the directional entropy; calculating the average value of the normalized values ​​of the directional entropy of all points in the target frame point cloud to obtain the directional consistency score of the target frame point cloud.

5. The omnidirectional intelligent sensing method for a dual-track tunnel car wash machine according to claim 1, characterized in that, The process of obtaining the source point echo intensity vector and the symmetric point echo intensity vector includes: each source point and its symmetric point form a pair; after removing overlapping pairs, a full set of point pairs is obtained; the laser echo intensity value of each source point in the full set of point pairs is obtained to form the source point echo intensity vector; for each symmetric point in the full set of point pairs, if there is a point in the target frame point cloud that matches its coordinates, the laser echo intensity value of the symmetric point is set to the laser echo intensity value of the matching point; if there is no point in the target frame point cloud that matches its coordinates, the laser echo intensity value of the symmetric point is set to 0, thereby obtaining the symmetric point echo intensity vector.

6. The omnidirectional intelligent sensing method for a dual-track tunnel car wash machine according to claim 1, characterized in that, The method for calculating the spatial symmetry score includes: calculating the cosine similarity value between the source point echo intensity vector and the symmetry point echo intensity vector to obtain the spatial symmetry score of the target frame point cloud.

7. The omnidirectional intelligent sensing method for a dual-track tunnel car wash machine according to claim 1, characterized in that, The process of obtaining candidate templates for the target frame point cloud includes: calculating the three-dimensional dimensions of the axis-aligned bounding box of the target frame point cloud, and retrieving the three-dimensional dimensions of the axis-aligned bounding box of each template in the three-dimensional vehicle template library; for each dimension, calculating the difference rate between the dimension of the target frame point cloud and the dimension of each template; if the difference rate of the dimensions of the three dimensions does not exceed a preset threshold, then the template is used as a candidate template for the target frame point cloud.

8. The omnidirectional intelligent sensing method for a dual-track tunnel car wash machine according to claim 1, characterized in that, The calculation of the total matching score of each candidate template based on the matching score between each frame point cloud and its candidate template and the comprehensive quality score of each frame point cloud includes: extracting the intersection of the candidate templates of all frames of point clouds in the 3D point cloud sequence to obtain a global candidate template set; for each candidate template in the global candidate template set, using the comprehensive quality score of each frame point cloud as a weight, the matching scores of the candidate template and each frame point cloud are weighted and summed to obtain the total matching score of the candidate template.

9. The omnidirectional intelligent sensing method for a dual-track tunnel car wash machine according to claim 1, characterized in that, The method for constructing the 3D vehicle template library includes: pre-collecting complete 3D point cloud models of multiple typical vehicles in clean and dry states; performing standardization processing operations such as coordinate decentering, scale normalization, and principal direction alignment on each collected original point cloud model to obtain a 3D vehicle template library composed of standardized point cloud models, with each template associated with a vehicle model tag and each template bound to a set of car wash control parameters.

10. An omnidirectional intelligent sensing system for a dual-track tunnel car wash machine, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the omnidirectional intelligent sensing method for a dual-track tunnel car wash machine according to any one of claims 1-9.

Images