Highway guardrail height detection method based on line knot light

Abstract

The invention relates to the technical field of highway guardrail height detection, in particular to a highway guardrail height detection method based on line knot light, and the method comprises the steps: installing a line structured light scanner, a high-definition camera and real-time dynamic differential positioning equipment at the side of a detection vehicle to construct a vehicle-mounted system; the method comprises the following steps: synchronously acquiring guardrail three-dimensional point cloud, images and vehicle positioning data when a vehicle runs according to preset parameters; identifying the type of the guardrail by using an instance segmentation model pre-trained by an edge calculation unit, and outputting a code; performing point cloud dynamic compensation, road reference surface extraction and guardrail feature extraction on the acquired data based on the guardrail type; and finally calculating the actual height of the guardrail according to the elevation difference between the road reference surface and the guardrail top surface. According to the method, through multi-sensor fusion, intelligent identification and a targeted data processing algorithm, high-precision and real-time dynamic detection of different types of guardrails is realized, the problems of low efficiency, poor precision and weak adaptability of a traditional method are solved, and the method can be widely applied to highway guardrail safety detection scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of highway guardrail height detection, in particular to a highway guardrail height detection method based on line joint light. BACKGROUND

[0002] With the continuous growth of highway mileage in China, the height of guardrails as an important facility for road safety directly affects the protection performance and driving safety; under the current background, the traditional detection technology has many shortcomings and cannot meet the modern highway maintenance needs, which are as follows:

[0003] Low efficiency and high risk of manual detection: manual detection of guardrail height using tools such as tape measure and level requires personnel to be close to the lane, and in the environment of high-speed driving vehicles, the safety of the detection personnel is threatened; at the same time, manual detection is slow, and the efficiency of data collection and recording is low, making it difficult to achieve rapid detection of long distances and large areas, and unable to meet the needs of regular and efficient detection of highway guardrails;

[0004] Poor flexibility of static detection equipment: some use fixedly installed laser radar or visual detection equipment to detect guardrails, such equipment relies on pre-deployed fixed positions and is difficult to adapt to detection needs in different road sections and different scenarios, and lacks flexibility and mobility; moreover, once the detection section changes, the equipment needs to be reinstalled and debugged, which is costly and time-consuming;

[0005] Weak adaptability of existing dynamic detection technology: existing vehicle-mounted dynamic detection methods mostly use single sensors or fixed detection modes; for example, only relying on laser radar point cloud data for guardrail height measurement without considering the influence of guardrail type differences on detection results; since there are many types of highway guardrails such as double-wave, three-wave and concrete, their structures and geometric characteristics are different, a single detection mode is difficult to accurately extract the top elevation of different types of guardrails, resulting in large errors in detection results and failing to meet the high-precision detection requirements;

[0006] Lack of real-time and intelligence: although some existing detection technologies can achieve dynamic detection, data processing and analysis are mostly performed after detection is completed, which cannot provide real-time feedback of detection results and cannot timely find abnormal problems in guardrail height; at the same time, lacking intelligent recognition and self-adaptive adjustment mechanism, it cannot automatically optimize the detection strategy according to different guardrail types and environmental conditions, and the detection efficiency and reliability need to be improved.

[0007] Therefore, a highway guardrail height detection method based on line joint light is proposed to solve the above problems. SUMMARY

[0008] The present application aims to provide a highway guardrail height detection method based on line joint light to solve the problems raised in the background.

[0009] To achieve the above object, the present application provides the following technical solutions:

[0010] A highway guardrail height detection method based on line structure light, comprising the following steps:

[0011] S1, a line structure light scanner, a high-definition camera and a real-time dynamic differential positioning device are installed on the side of a detection vehicle to form a vehicle-mounted detection system;

[0012] S2, the vehicle drives at a preset speed on a lane at a preset distance from the guardrail, and synchronously collects three-dimensional point cloud data, image data and vehicle positioning data of the guardrail;

[0013] S3, the edge computing unit performs intelligent identification of the guardrail type and outputs a guardrail type code;

[0014] S4, based on the guardrail type, point cloud dynamic compensation, road reference surface extraction and guardrail feature extraction are performed:

[0015] Motion compensation and denoising processing are performed on the point cloud data;

[0016] The road reference surface elevation is extracted;

[0017] According to the guardrail type, a waveform mathematical model or a plane fitting method is selected to extract the guardrail top surface elevation;

[0018] S5, the actual height of the guardrail is calculated according to the difference between the road reference surface elevation and the guardrail top surface elevation.

[0019] As a preferred scheme, the installation height of the line structure light scanner in step S1 is 1.2m±0.1m, the depression angle is 15°±2°, the high-definition camera is coaxially fixed with the line structure light scanner; in step S2, the preset distance of the vehicle from the guardrail is 1.5m±0.2m, the preset speed is 80km / h, and the following conditions are met:

[0020] The line structure light scanner collects three-dimensional point cloud data at a frequency of 100Hz;

[0021] The high-definition camera collects guardrail images at a rate of 5 frames per second;

[0022] The real-time dynamic differential positioning device outputs centimeter-level vehicle position coordinates.

[0023] As a preferred scheme, step S3 comprises:

[0024] A pre-trained instance segmentation model is used to process a 640x640 pixel input image, and outputs a guardrail type code, which includes 01 double waveform / 02 triple waveform / 03 concrete;

[0025] The training of the instance segmentation model satisfies:

[0026] Using 707 labeled guardrail images, the training set and validation set ratio is 8:2;

[0027] Training parameters: learning rate 0.001, batch size 8, training rounds 100 rounds;

[0028] The last 10 rounds of training turn off the mosaic enhancement.

[0029] As a preferred solution, the motion compensation and denoising processing of the point cloud data in step S4 includes:

[0030] Statistical filter denoising: remove outliers within a neighborhood radius of 5 cm with a distance exceeding the mean ± 3 times the standard deviation;

[0031] Dynamic point cloud motion compensation:

[0032] According to the three-dimensional displacement amount (Δx, Δy, Δz) and the heading angle change amount (Δθ) of the vehicle measured by the real-time dynamic difference positioning device, a 4×4 motion transformation matrix is constructed;

[0033] The first row elements of the motion transformation matrix are composed of the cosine value of the heading angle change amount, the negative sine value, 0 and the X-axis displacement amount;

[0034] The second row elements are composed of the sine value of the heading angle change amount, the cosine value, 0 and the Y-axis displacement amount;

[0035] The third row elements are composed of 0, 0, 1 and the Z-axis displacement amount;

[0036] The fourth row elements are composed of 0, 0, 0, 1;

[0037] Multiply the original point cloud coordinates by the motion transformation matrix to get the compensated point cloud coordinates.

[0038] As a preferred solution, the normal vector constrained random sample consensus algorithm is used to extract the road reference surface elevation in step S4:

[0039] The normal vector of the constraint plane is perpendicular to the vertical direction, and the component is greater than 0.9;

[0040] Fit the road plane equation through 1000 iterations;

[0041] The road reference surface elevation is determined by the ratio of the constant term to the vertical direction coefficient of the plane equation.

[0042] As a preferred solution, the multi-scale sinusoidal model matching is used to extract the guardrail top surface elevation in step S4:

[0043] The double-wave guardrail adopts a sinusoidal function model defined by amplitude, wavelength and phase angle parameters;

[0044] The three-wave guardrail adopts a composite model of superposition of two sine waves with different amplitudes;

[0045] The matching score calculation process comprises:

[0046] A corresponding template is generated for each scale factor alpha, and alpha takes values of 0.8, 0.9, 1.0, 1.1, 1.2;

[0047] The Euclidean distance square of the actual coordinates of the point cloud and the theoretical coordinates of the template is calculated;

[0048] The negative exponent of the distance square divided by twice the square of the tolerance threshold is taken;

[0049] The arithmetic mean of the exponent results of all points is obtained to obtain the matching score.

[0050] As a preferred scheme, the multi-scale matching of the multi-scale sine model is performed in the order of standard scale → reduced scale → enlarged scale → minimum scale → maximum scale; when the matching score is greater than or equal to 0.85, it is determined that the matching is successful, and the top surface elevation of the guardrail is taken as the maximum vertical coordinate value in the matched point cloud; when the matching score is less than 0.85 for five consecutive scanning periods, the guardrail type corresponding to the standard template output in step S3 is forcibly loaded.

[0051] As a preferred scheme, in step S4, the detection of the concrete guardrail comprises:

[0052] The random sample consensus algorithm with a constraint of a horizontal component of the normal vector being greater than 0.9 is used to fit the guardrail plane;

[0053] The maximum vertical coordinate value of the point cloud is taken as the top surface elevation.

[0054] As a preferred scheme, the height calculation in step S5 is specifically:

[0055] The actual height of the guardrail is obtained by subtracting the road datum surface elevation from the top surface elevation of the guardrail.

[0056] As a preferred scheme, the edge calculation unit processing delay satisfies:

[0057] The dynamic compensation calculation is less than or equal to 10ms;

[0058] The multi-scale template matching is less than or equal to 50ms;

[0059] The whole process processing is less than or equal to 100ms;

[0060] The height detection result is ensured to be output in real time at a vehicle speed of 80km / h.

[0061] As can be seen from the above technical scheme provided by the present application, the highway guardrail height detection method based on line junction light provided by the present application has the beneficial effects that:

[0062] High-efficiency dynamic detection, improve work efficiency: adopt vehicle-mounted mobile detection mode, when the vehicle travels at a speed of 80km / h, the system's full-process processing time is ≤100ms, and the guardrail height detection result can be output in real time; compared with traditional manual detection or static detection mode, the detection time is greatly reduced, the work efficiency of highway guardrail detection is improved, the labor cost and time cost are reduced, and it is especially suitable for long-distance and large-range guardrail detection tasks;

[0063] Multi-sensor fusion, ensure detection accuracy: fuse line structured light scanner, high-definition camera and real-time dynamic differential positioning equipment, the line structured light scanner obtains high-precision three-dimensional point cloud data, the high-definition camera assists in identifying the type of guardrail, and the centimeter-level positioning equipment provides accurate pose information; through point cloud motion compensation, denoising and accurate extraction of road reference surface, the influence of vehicle motion, environmental interference and other factors is effectively eliminated, and the accuracy and reliability of the guardrail height detection result are ensured;

[0064] Intelligent identification and adaptation, enhance scene adaptability: use a pre-trained instance segmentation model to intelligently identify the type of guardrail, output dual waveform, three waveform, concrete and other type codes, and use differential feature extraction algorithms (such as multi-scale sine model matching, plane fitting) based on different types of guardrails; this adaptive detection strategy enables the system to cope with various types of guardrails and complex road scenes, avoids detection errors caused by differences in guardrail types, and improves the universality and robustness of the detection method;

[0065] Real-time collaborative optimization, strengthen system reliability: realize closed-loop optimization of the detection model through a federated learning framework, the edge node and the central server cooperatively update the model parameters, simulate extreme scenarios to verify the model robustness through digital twin technology, and continuously improve the detection accuracy; at the same time, the inter-vehicle communication and cooperative control mechanism supports multi-detection vehicle cooperative operation, optimizes detection route and task allocation, and ensures efficient and stable detection process;

[0066] Edge computing empowerment, meet real-time requirements: the edge computing unit integrates high-performance hardware and optimized algorithm library to realize real-time data processing and local model inference, reducing data transmission delay and cloud dependence; even in areas with poor network coverage, the detection task can still be executed normally, and through encrypted communication and secure data synchronization, the data privacy and system security are guaranteed. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 It is a kind of highway guardrail height detection method based on line joint light step flow chart. DETAILED DESCRIPTION

[0068] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0069] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in conjunction with the drawings and specific embodiments.

[0070] As shown in Figure 1 The embodiment of the present application provides a highway guardrail height detection method based on line joint light, which comprises the following steps:

[0071] S1, a line structure light scanner, a high-definition camera and a real-time dynamic differential positioning device are installed on the side of a detection vehicle to form a vehicle-mounted detection system;

[0072] S2, the vehicle drives at a preset speed on a lane at a preset distance from the guardrail, and synchronously collects guardrail three-dimensional point cloud data, image data and vehicle positioning data;

[0073] S3, a guardrail type intelligent identification is performed by an edge computing unit, and a guardrail type code is output;

[0074] S4, point cloud dynamic compensation, road reference surface extraction and guardrail feature extraction are performed based on the guardrail type:

[0075] Motion compensation and denoising processing are performed on the point cloud data;

[0076] The road reference surface elevation is extracted;

[0077] The road reference surface elevation is extracted;

[0078] S5, the actual height of the guardrail is calculated according to the difference between the road reference surface elevation and the guardrail top surface elevation.

[0079] In this embodiment, the function of step S1 is to build a vehicle-mounted detection system, and by reasonably installing sensors and devices, a foundation is laid for subsequent collection of guardrail data and execution of detection tasks; the following are the detailed steps:

[0080] Step S1-1: selection of sensors and devices:

[0081] Selection of line structure light scanner: a line structure light scanner with high precision and high resolution is selected, which needs to achieve millimeter-level scanning accuracy to ensure that clear and accurate guardrail three-dimensional point cloud data can be obtained; at the same time, the environmental adaptability of the device is considered, so that it can work stably under different weather conditions (such as sunny, cloudy, light rain) and different light intensity conditions;

[0082] High-definition camera selection: Select a high-definition camera with high frame rate and high pixel, pixel no less than 12 million, frame rate no less than 30 frames / second, to ensure that the collected guardrail image is clear and detailed, and convenient for subsequent guardrail type recognition; the camera should have auto focus and auto exposure functions to reduce manual intervention;

[0083] Real-time kinematic positioning equipment selection: Adopt real-time kinematic positioning equipment (RTK) with centimeter-level positioning accuracy, which can obtain accurate position coordinates and attitude information of the vehicle in real time, providing accurate pose reference for motion compensation of point cloud data; the equipment should support multiple satellite positioning systems (such as GPS, Beidou), to improve the reliability and stability of positioning;

[0084] Step S1-2: Sensor and equipment installation:

[0085] Installation position determination: Install the line structured light scanner on the side of the detection vehicle, set the installation height to 1.2m±0.1m, and the depression angle to 15°±2°, to ensure that the line structured light emitted by the scanner can completely cover the guardrail area; the high-definition camera is coaxially fixed with the line structured light scanner, to ensure that the data collected by the two is consistent in space, facilitating subsequent data fusion processing; the real-time kinematic positioning equipment is installed on the top of the vehicle in an open area, to ensure good satellite signal reception conditions;

[0086] Installation method design: Design a special installation bracket and fixing structure to ensure that the sensors and equipment are installed firmly and will not loosen or shift during vehicle driving; at the same time, the installation bracket is treated with shock absorption to reduce the impact of vehicle driving on the quality of data collected by the equipment;

[0087] Step S1-3: System connection and debugging:

[0088] Hardware connection: Use a special data transmission cable to connect the line structured light scanner, high-definition camera, real-time kinematic positioning equipment and edge computing unit on the vehicle, to ensure stable and fast data transmission; at the same time, provide stable power supply for each device, equipped with a power management module to prevent power fluctuations from damaging the equipment;

[0089] System debugging: After completing the hardware connection, debug the entire vehicle-mounted detection system; check whether each device is working normally, whether the data transmission is stable, and whether the time synchronization between devices is accurate; through the collection of sample data, verify the accuracy and consistency of the three-dimensional point cloud data obtained by the line structured light scanner, the image data collected by the high-definition camera and the positioning data output by the real-time kinematic positioning equipment; optimize and adjust the system parameters to ensure that the system is in the best working state.

[0090] In this embodiment, the role of step S2 is to ensure the acquisition of high-quality guardrail data by controlling the vehicle driving state and data acquisition parameters, providing a reliable basis for subsequent detection and analysis; the following are the detailed steps:

[0091] Step S2-1: Vehicle driving path and speed planning:

[0092] Driving path determination: According to the detection requirements of the highway guardrail, the vehicle is planned to drive in the lane 1.5m±0.2m away from the guardrail, ensuring that the line structure light scanner and the guardrail maintain a suitable scanning distance, avoiding data loss or precision decline due to too close or too far distance; at the same time, the driving route is planned to avoid complex road sections such as curves and slopes as much as possible, if it cannot be avoided, special data acquisition strategies for the road section are developed in advance;

[0093] Driving speed setting: the vehicle driving speed is preset to 80km / h, which can ensure the detection efficiency and meet the frequency requirements of sensor data acquisition; during driving, through the vehicle's cruise control system or automatic driving assistance system, the speed fluctuation range is strictly controlled to ensure that the speed is stable within ±5km / h;

[0094] Step S2-2: Data acquisition parameter configuration:

[0095] Line structure light scanner parameter configuration: set the acquisition frequency of the line structure light scanner to 100Hz to ensure that the three-dimensional point cloud data of the guardrail can be continuously and quickly acquired during vehicle driving; at the same time, according to the detection requirements of the guardrail, adjust the scanning range, resolution and other parameters of the scanner to ensure that the acquired point cloud data can completely and clearly reflect the geometric characteristics of the guardrail;

[0096] High-definition camera parameter configuration: set the high-definition camera to acquire guardrail images at a frequency of 5 frames / second to ensure that the acquired images can cover the key parts and detailed information of the guardrail; adjust the shooting angle, focal length, aperture and other parameters of the camera to ensure clear and distortion-free images, providing high-quality image data for guardrail type identification;

[0097] Real-time dynamic differential positioning device parameter configuration: ensure that the real-time dynamic differential positioning device outputs centimeter-level vehicle position coordinates and attitude information at a frequency of not less than 10Hz, including the vehicle's longitude, latitude, elevation, heading angle, pitch angle, roll angle, etc., to provide accurate reference for point cloud data motion compensation and positioning;

[0098] Step S2-3: Data synchronous acquisition and storage:

[0099] Time synchronization setting: In the vehicle detection system, high-precision clock synchronization technology (such as GPS clock synchronization) is used to ensure that the collection times of line structured light scanners, high-definition cameras, and real-time dynamic differential positioning devices remain consistent, with an error controlled within milliseconds; each set of collected data is marked with a timestamp to facilitate subsequent data fusion and analysis;

[0100] Data acquisition and storage: During the vehicle's travel along the preset path and speed, the line structured light scanner, high-definition camera, and real-time dynamic differential positioning device are started simultaneously to begin collecting guardrail three-dimensional point cloud data, image data, and vehicle positioning data; the collected data is transmitted in real time to the edge computing unit on the vehicle and stored locally through a large-capacity storage device (such as a solid-state hard drive); during storage, the data is classified and organized, indexed by collection time, data type, etc., to facilitate subsequent data queries and calls; at the same time, data compression techniques are used to reduce data storage space without losing data accuracy, improving storage efficiency.

[0101] In this embodiment, the role of step S3 is to execute intelligent recognition algorithms through the edge computing unit to accurately determine the guardrail type and output the code, providing a basis for subsequent targeted point cloud processing and height detection; the following are the detailed steps:

[0102] Step S3-1: Image preprocessing:

[0103] Image selection and format conversion: From the guardrail image data collected by the high-definition camera, select images that meet the detection requirements (such as clear, complete coverage of the guardrail), and uniformly convert the image format to a format suitable for instance segmentation model processing (such as JPEG or PNG format with RGB channels);

[0104] Image size adjustment: uniformly scale the selected images to 640x640 pixels to meet the input size requirements of the pre-trained instance segmentation model; during scaling, algorithms such as bilinear interpolation are used to minimize image distortion and preserve guardrail details;

[0105] Step S3-2: Instance segmentation model deployment and inference:

[0106] Model deployment: deploy the pre-trained instance segmentation model to the edge computing unit to ensure that the model can efficiently run in a vehicle environment; according to the hardware resources (such as CPU and GPU performance) of the edge computing unit, perform lightweight processing (such as pruning and quantization) on the model to reduce the model's computational complexity and memory usage without affecting the model's accuracy, improving inference speed;

[0107] Model inference: input the preprocessed 640x640 pixel image into the instance segmentation model, the model extracts features, detects targets and performs instance segmentation on the image through structures such as convolutional neural network (CNN), identifies the guardrail area in the image, and outputs the guardrail type code (01 represents double-wave guardrail, 02 represents triple-wave guardrail, and 03 represents concrete guardrail); during the inference process, the running state and inference results of the model are monitored in real time to ensure the accuracy and stability of the output code;

[0108] Step S3-3: model training and optimization:

[0109] Dataset construction: 707 labeled guardrail images are collected, which are divided into training set and validation set according to the ratio of 8:2; the labeled content includes the type, bounding box and instance mask of the guardrail, etc., to ensure the accuracy and consistency of the labeling and provide high-quality sample data for model training;

[0110] Training parameter setting: set the model training parameters, the learning rate is 0.001, the batch size is 8, and the training rounds are 100; in the training process, use optimization algorithms such as stochastic gradient descent (SGD) to update the model parameters, so that the model can continuously learn the features of the guardrail image and improve the recognition accuracy; in the last 10 rounds of training, the mosaic enhancement and other data enhancement strategies are turned off to avoid overfitting and ensure the generalization ability of the model in actual application;

[0111] Model evaluation and optimization: evaluate the trained model using the validation set, calculate the accuracy, recall rate, F1 value and other indicators of the model; according to the evaluation results, analyze the problems of the model (such as low recognition accuracy for some types of guardrails), adjust the model structure or training parameters, and retrain until the model performance reaches the optimal state; save and deploy the optimized model to the edge computing unit for actual guardrail type recognition tasks.

[0112] In this embodiment, the motion compensation and denoising processing of the point cloud data in step S4 includes:

[0113] Statistical filter denoising: remove outliers within a neighborhood radius of 5 cm with a distance exceeding the mean value ± 3 times the standard deviation;

[0114] Dynamic point cloud motion compensation:

[0115] According to the three-dimensional displacement (Δx, Δy, Δz) and the heading angle change (Δθ) measured by the real-time dynamic differential positioning device, a 4x4 motion transformation matrix is constructed;

[0116] The first row elements of the motion transformation matrix are composed of the cosine value of the heading angle change, the negative sine value, 0 and the X-axis displacement;

[0117] The second row of elements consists of a sine value, a cosine value, 0, and a Y-axis displacement amount of the heading angle change amount;

[0118] The third row of elements consists of 0, 0, 1, and a Z-axis displacement amount;

[0119] The fourth row of elements consists of 0, 0, 0, 1;

[0120] The original point cloud coordinates are multiplied by the motion transformation matrix to obtain compensated point cloud coordinates;

[0121] The extraction of the road reference surface elevation in step S4 adopts a random sample consensus algorithm with normal vector constraint:

[0122] The normal vector of the constraint plane is constrained in the vertical direction with a component > 0.9;

[0123] The road plane equation is fitted through 1000 iterations;

[0124] The road reference surface elevation is determined by the ratio of the constant term to the vertical direction coefficient of the plane equation;

[0125] The extraction of the guardrail top surface elevation in step S4 adopts a multi-scale sinusoidal model matching:

[0126] The double-wave guardrail adopts a sinusoidal function model defined by amplitude, wavelength, and phase angle parameters;

[0127] The triple-wave guardrail adopts a composite model of superimposing two sinusoidal waves with different amplitudes;

[0128] The matching score calculation process includes:

[0129] A corresponding template is generated for each scale factor α, and α takes values 0.8, 0.9, 1.0, 1.1, 1.2;

[0130] The Euclidean distance square of the actual point cloud coordinates and the theoretical coordinates of the template is calculated;

[0131] The distance square is divided by twice the square of the tolerance threshold, and the negative exponent is taken;

[0132] The arithmetic mean of the index results of all points is taken to obtain the matching score;

[0133] The detection of the concrete guardrail includes:

[0134] The random sample consensus algorithm with a normal vector horizontal component > 0.9 constraint is used to fit the guardrail plane;

[0135] The maximum vertical coordinate value of the point cloud is taken as the top surface elevation;

[0136] Further, the role of step S4 is to realize point cloud dynamic compensation, road reference surface extraction and guardrail feature extraction based on the guardrail type through a series of data processing algorithms, so as to provide key data for accurately calculating the height of the guardrail; the following are the detailed steps:

[0137] Step S4-1: Point cloud dynamic compensation and denoising based on guardrail type:

[0138] Motion compensation and denoising preprocessing: transmit the three-dimensional point cloud data collected in step S2 to the edge computing unit, and according to the guardrail type code output in step S3, first perform a statistical filter denoising operation for all types of guardrails; set the neighborhood radius to 5 cm, remove the outliers whose distance in the neighborhood exceeds the mean value ± 3 times the standard deviation, eliminate abnormal data points caused by environmental interference or sensor errors, and improve the quality of point cloud data;

[0139] Dynamic point cloud motion compensation: based on the three-dimensional displacement Δx, Δy, Δz (unit: m) of the vehicle measured by the real-time dynamic differential positioning device and the change amount Δθ (unit: radian) of the heading angle of the vehicle, a motion transformation matrix M t is constructed for adjacent scanning periods

[0140] (wherein, M t is the motion transformation matrix for adjacent scanning periods; Δx is the displacement of the vehicle in the x-axis direction measured by the real-time dynamic differential positioning device, with the unit of m; Δy is the displacement of the vehicle in the y-axis direction, with the unit of m; Δz is the displacement of the vehicle in the z-axis direction, with the unit of m; Δθ is the change amount of the heading angle of the vehicle, with the unit of radian);

[0141] Perform P raw transformation on the original point cloud coordinates P corr = M t · P raw , to obtain the point cloud coordinates P corr after motion compensation (wherein, P raw is the original point cloud coordinates; P corr is the point cloud coordinates after motion compensation), eliminate the point cloud offset generated during the driving of the vehicle, and make the point cloud data collected at different times accurately correspond to the actual position of the guardrail;

[0142] Step S4-2: Road reference surface extraction:

[0143] Normal vector constraint setting: for all types of guardrails, the random sample consensus algorithm (RANSAC) with normal vector constraint is used to extract the road reference surface; the constraint condition is set as the component of the plane normal vector in the vertical direction being greater than 0.9, to ensure that the fitted plane meets the road level feature and excludes interference such as inclined surfaces or guardrail surfaces;

[0144] Plane fitting and elevation calculation: The road plane equation ax+by+cz+d=0 is fitted through 1000 iterations (where a, b, c, and d are the coefficients of the plane equation). In each iteration, some points in the point cloud data are randomly selected as samples for plane fitting, and the distance from other points to the plane is calculated. The distance threshold is used to determine whether the point belongs to the plane. After multiple iterations, the plane containing the largest number of points is selected as the final road plane. According to the plane equation, the road base elevation is calculated. (Among them, D road (Elevation of the road base surface), providing a benchmark for subsequent guardrail height calculations;

[0145] Step S4-3: Feature extraction based on guardrail type:

[0146] Waveform guardrail feature extraction (double waveform / triple waveform):

[0147] Model matching: If step S3 identifies a double-waveform or triple-waveform guardrail, a multi-scale sinusoidal model is used for matching; for a double-waveform guardrail, a matching model is used. (where y(x) is the height of the guardrail model at position x; A is the amplitude; λ is the wavelength; x is the horizontal coordinate; φ is the phase; y0 is the offset); the three-wave guardrail uses (where y(x) is the height of the guardrail model at position x; A1 is the fundamental wave amplitude; A2 is the second harmonic amplitude; and λ is the fundamental wave wavelength).

[0148] Matching score calculation: According to the scale factor α (values ​​0.8, 0.9, 1.0, 1.1, 1.2), combined with the standard wavelength λ0 (dual wave 33cm / triple wave 50cm), the formula Calculate the matching score S α (Among them, S α is the matching score corresponding to the scale factor α; N is the number of points involved in the matching; i is the sequence number of the point; p i is the actual coordinate of the i-th point in the point cloud; q i is the theoretical coordinate of the i-th point in the template model; α is the scale factor; λ0 is the standard wavelength; δ is the tolerance threshold, the default is 0.02m); the matching is performed in the order of α=1.0→0.9→1.1→0.8→1.2. When S α When ≥0.85, the match is determined to be successful, and the guardrail top surface elevation Y is taken. max =max(y i )(where Y max is the elevation of the guardrail top surface; i is the vertical coordinate value of the i-th point on the model after successful matching); when 5 consecutive scanning cycles S α When <0.85, the guardrail type output in step S3 is forced to load and the top surface elevation is determined by the standard template;

[0149] Concrete guardrail feature extraction: If step S3 identifies a concrete guardrail, use the random sample consensus algorithm with a normal vector horizontal component > 0.9 constraint to fit the guardrail plane, exclude vertical and other interference surfaces; take the maximum value of the vertical coordinates of all points in the point cloud as the top surface elevation Y max = Z max (where Y max is the top surface elevation of the guardrail; Z max is the maximum value of the vertical coordinates of all points in the point cloud), accurately obtain the top surface height feature of the concrete guardrail.

[0150] In this embodiment, the height calculation of step S5 is as follows:

[0151] Subtract the road datum elevation from the guardrail top surface elevation to obtain the actual height of the guardrail;

[0152] Further, the role of step S5 is to continuously improve the accuracy and reliability of the highway guardrail height detection method through closed-loop optimization and collaborative control, and to ensure real-time and efficient output of the detection results; the following are the detailed steps:

[0153] Step S5-1: Detection result feedback and data collection:

[0154] Feedback mechanism building: Establish a dedicated detection result feedback channel in the vehicle-mounted detection system, and feed back the guardrail height detection results (including actual height value, detection time, early warning state during detection, etc.) to the core processing module of the edge computing unit in real time and accurately; adopt redundant transmission design, feedback through multiple data links to avoid data loss or delay due to single link failure;

[0155] Multi-source data collection: Collect all kinds of data related to the detection results, including:

[0156] Device operation data: scanning frequency, working temperature of line structured light scanner; shooting parameters, image quality indicators of high-definition camera; positioning accuracy, satellite signal strength, etc. of real-time dynamic differential positioning device running parameters;

[0157] Detection process data: point cloud data density, integrity, image data clarity, guardrail feature integrity, positioning data continuity, etc. directly related to the detection process;

[0158] Environmental data: weather conditions (sunny, rainy, foggy, etc.), light intensity, environmental temperature, humidity, etc. environmental factor data for analyzing the influence of environment on detection results;

[0159] Historical comparison data: Historical inspection data of the same road section and the same type of guardrail is used to compare and analyze the changing trends of the inspection results; the collected data is classified and coded, and stored in the vehicle's large-capacity storage device according to chronological order and data type to form a complete data archive;

[0160] Step S5-2: Optimize the detection model using the federated learning framework:

[0161] Edge node data preprocessing: Each edge computing unit (deployed on the detection vehicle) preprocesses the locally collected data. First, data cleaning is performed, using filtering algorithms to remove noise points in the point cloud data and blurry or invalid images in the image data. Key features are then extracted, such as the geometric features of the guardrail from the point cloud data and the texture and shape features of the guardrail from the image data. Finally, encryption algorithms (such as AES) are used to encrypt the data to ensure privacy and security during transmission and processing.

[0162] Local model training: Each edge node trains the local guardrail height detection model based on pre-processed data using online learning algorithms such as stochastic gradient descent. The model parameters are adjusted according to the characteristics of the local data, such as the threshold parameters in the point cloud processing algorithm and the weight parameters in the image recognition model, to update the local model parameters. (where k represents the kth edge computing node, is the local model parameter of the kth edge computing node), so that the local model can better adapt to the local data characteristics;

[0163] Global model parameter aggregation: Each edge node aggregates the updated local model parameters Upload to the central server; the central server will process the local data volume N of each edge node. k (i.e., the number of data samples that the kth edge node participates in training) and the total amount of local data of all edge nodes (K is the total number of edge nodes), according to the formula Calculate the global model parameters θ global (where θ global is the aggregated global model parameter), achieving comprehensive learning of different detection vehicle data features;

[0164] Global model distribution and update: The central server calculates the global model parameters θ global The global model parameters are sent to each edge node. After receiving them, each edge node uses them to update the local model, making the local model consistent with the global model, completing the optimization and update of the entire detection model, and improving the accuracy and adaptability of the model to guardrail height detection in different scenarios.

[0165] Step S5-3: Simulate complex detection scenarios through digital twin technology:

[0166] Digital twin model construction: Based on the actual physical parameters of the highway guardrail (material, size, shape), the sensor parameters of the detection vehicle (scanning range, accuracy, viewing angle), and the actual situation of the detection environment (road slope, curve curvature, surrounding facility distribution), a high-precision digital twin model is constructed using three-dimensional modeling technology. This model can real-time synchronize the actual running state of the detection vehicle, the working state of the sensor, and the collected data, and truly simulate the guardrail detection process;

[0167] Complex scenario definition and simulation: Based on historical detection data and complex situations encountered in actual detection, a series of complex detection scenarios are defined, including:

[0168] Environmental complex scenarios: Guardrail detection scenarios in adverse environments such as heavy rain leading to road reflection, heavy fog reducing visibility, and low light at night;

[0169] Road complex scenarios: Guardrail detection scenarios in special road sections such as sharp curves, long downhills, and road construction;

[0170] Device abnormal scenarios: Guardrail detection scenarios in device abnormal situations such as sensor partial failure and data transmission interruption; simulate these complex scenarios in the digital twin model, input different environmental parameters, road parameters, and device parameters, and observe the detection process and results of the detection model in complex scenarios;

[0171] Model performance verification and optimization: In-depth analysis of the simulation results, evaluation of the performance of the detection model in complex scenarios, including the accuracy of the detection results, detection efficiency, anti-interference ability, etc.; check whether the model will misjudge or miss in complex scenarios, whether the detection time meets the real-time requirements, etc.; if problems are found in the model, feedback the relevant information to the model training link in step S5-2, adjust the model parameters or improve the algorithm accordingly, enhance the detection ability and reliability of the model in complex scenarios;

[0172] Step S5-4: Multi-detection vehicle cooperative optimization:

[0173] Communication network construction: Install a dedicated inter-vehicle communication module on each detection vehicle, use low-latency, high-bandwidth, and anti-interference communication protocols such as 5G-V2X, and build a stable inter-vehicle communication network; through this network, detection vehicles can exchange detection data, vehicle position, driving state, and other information in real-time and reliably, ensuring the accuracy and timeliness of information transmission;

[0174] Information sharing and interaction: Each detection vehicle shares its own detection data (such as detected guardrail height, type, detection time), vehicle position coordinates, speed, direction, equipment running status, etc. through the inter-vehicle communication network in real time; at the same time, it receives information shared by other detection vehicles, integrates to form a panoramic view of the guardrail detection information in the region, and realizes mutual sharing of detection data and vehicle status;

[0175] Cooperative optimization strategy execution: Based on shared information, a cooperative optimization algorithm runs on each detection vehicle; the algorithm dynamically adjusts detection task allocation and driving route according to the detection progress, detection area coverage, vehicle position and driving trajectory of each detection vehicle in the region; for example, when multiple detection vehicles are detecting adjacent road sections, the algorithm reasonably allocates undetected areas according to the detection progress of each vehicle to avoid repeated detection and blind spots; when a detection vehicle encounters road congestion or other obstacles, the algorithm guides the vehicle to adjust the driving route, while coordinating the detection plan of surrounding vehicles to ensure efficient and orderly completion of the guardrail detection task in the entire region;

[0176] Step S5-5: Edge computing ensures real-time decision-making:

[0177] Real-time data processing: The edge computing unit uses its powerful computing power to process real-time data (three-dimensional point cloud data, image data, positioning data) collected by local sensors and other vehicle information received through the inter-vehicle communication network; using parallel computing, pipeline processing and other technologies, it quickly completes operations such as data filtering, feature extraction, coordinate conversion, etc. to provide timely and effective data support for subsequent decision-making;

[0178] Localized model inference: In the edge computing unit, a localized guardrail height detection model and a cooperative optimization algorithm are deployed to perform real-time model inference on the processed data; based on the inference results of the detection model, it quickly determines whether the currently detected guardrail height meets the standard and whether it needs to be warned; based on the calculation results of the cooperative optimization algorithm, it adjusts the detection task and driving state of the vehicle in a timely manner, such as changing the detection route, adjusting the detection speed, etc., to realize real-time decision-making;

[0179] Data interaction with the cloud platform: The edge computing unit regularly uploads local processed data (including detection results, data processing intermediate results, model running status, etc.) and model update information to the cloud platform for storage and backup; at the same time, it downloads the latest global model parameters, cooperative optimization strategy configuration information, etc. from the cloud platform; During data interaction, encryption protocols such as SSL / TLS are used to ensure the security of data transmission and prevent data leakage and tampering; by optimizing data transmission strategies, the timeliness of data interaction is ensured, so that the edge computing unit always makes real-time decisions based on the latest models and strategies, ensuring the efficient operation of the detection system.

[0180] In this embodiment, the edge computing unit processing delay meets:

[0181] Dynamic compensation calculation ≤10ms;

[0182] Multi-scale template matching ≤50ms;

[0183] Full-process processing ≤100ms;

[0184] Ensure real-time output of height detection results at a speed of 80km / h;

[0185] Further, the edge computing unit as the core processing module of the vehicle-mounted detection system of the application undertakes key tasks such as real-time data processing, model inference and collaborative control, and needs to meet the requirements of high performance, low delay and high reliability; the following is a detailed description:

[0186] I. Hardware architecture design:

[0187] Core computing module:

[0188] Adopting a heterogeneous multi-core processor (such as an ARM+GPU combination), the ARM processor is responsible for system scheduling, data management and control logic, and the GPU is used for parallel computing tasks such as point cloud data processing and image recognition; the processor frequency is not less than 2.5GHz, and the GPU computing power reaches more than 5TFLOPS, ensuring that millions of point cloud data and multiple frames of high-definition images can be processed per second;

[0189] Equipped with cache (L1 / L2 / L3) and large-capacity memory (16GB DDR4 or above), it supports fast data reading and caching, reduces I / O delay, and ensures smooth operation of complex algorithms (such as instance segmentation and dynamic compensation);

[0190] Data storage and transmission module:

[0191] Built-in large-capacity solid state drive (SSD, capacity ≥512GB), supporting high-speed reading and writing (sequential reading speed ≥3000MB / s), used for storing raw collected data, intermediate processing results and model parameters;

[0192] Integrated multi-type communication interface:

[0193] High-speed data interface: PCIe4.0 interface connection line structure light scanner and high-definition camera, realizing GB-level data transmission per second; USB3.2 interface for device debugging and data backup;

[0194] Network communication interface: Gigabit Ethernet interface for connecting inter-vehicle communication module, supporting 5G-V2X protocol to realize data interaction between vehicles; Wi-Fi6 interface for non-real-time data synchronization with cloud platform;

[0195] Power and heat dissipation module:

[0196] Adopt wide voltage input power module (support 9-36VDC input), adapt to vehicle power supply system, and equipped with power filter and voltage stabilizing circuit to prevent voltage fluctuation affecting equipment operation;

[0197] The heat dissipation design adopts "heat-conducting silicone + metal heat sink + intelligent temperature control fan" combination to ensure the processor temperature control within 70℃, ensuring the stable operation of the equipment in high temperature environment;

[0198] II. Software system architecture:

[0199] Real-time operating system (RTOS):

[0200] Carry real-time operating system such as Linux-RT or QNX, support nanosecond-level task scheduling precision, ensure strict timing requirements of data acquisition, processing and decision output; The system kernel is optimized by cutting, reducing resource occupation and improving running efficiency;

[0201] Data processing and algorithm library:

[0202] Point cloud processing library: integrate open source framework such as PCL (Point Cloud Library) to realize point cloud filtering (statistical filter, voxel filter), registration (ICP algorithm) and feature extraction (normal calculation, curvature estimation) functions;

[0203] Image processing library: based on OpenCV and TensorRT acceleration library, support image preprocessing (scaling, noise reduction), instance segmentation model inference and feature recognition, inference delay control within 50ms;

[0204] Positioning and motion compensation algorithm: built-in pose solving module based on real-time kinematic differential positioning (RTK), combined with inertial navigation data (IMU) for fusion, through self-defined motion compensation algorithm to realize real-time calibration of point cloud;

[0205] Model deployment and update module:

[0206] Deploy pre-trained instance segmentation model and guardrail height detection model in ONNX or TensorFlow Lite format, support model quantization and lightweight processing, reduce memory occupation and calculation overhead;

[0207] Integrate federal learning client to realize local model training, parameter uploading and global model downloading function, communicate with central server through encrypted channel (SSL / TLS);

[0208] III. Function realization:

[0209] Real-time data processing:

[0210] Pipeline processing of multi-source data: receive 3D point cloud data of line structured light scanner (100 Hz), high-definition camera images (5 frames / second), and positioning data (10 Hz), and perform denoising, feature extraction, and time synchronization in parallel, ensuring that the total data processing delay is ≤100 ms;

[0211] Support for data compression storage: use point cloud compression algorithms (such as Octree encoding) to compress the original point cloud data by a factor of 1:10, saving storage space;

[0212] Intelligent decision-making and collaborative control:

[0213] According to the guardrail type identification result (double waveform / three waveform / concrete), automatically call the corresponding detection algorithm (multi-scale sinusoidal matching, plane fitting), calculate the guardrail height, and output the detection report;

[0214] Integrate inter-vehicle collaborative control module, receive surrounding vehicle information, avoid collision between detection vehicles through collaborative obstacle avoidance algorithm (such as speed adjustment, path planning), and optimize detection route coverage efficiency;

[0215] System monitoring and maintenance:

[0216] Built-in health monitoring module, real-time monitoring of processor load, memory usage, sensor connection state, etc. Parameters, trigger sound and light alarm when abnormal and upload fault log to cloud;

[0217] Support for remote firmware upgrade (FOTA), update the operating system, algorithm library, and model parameters of the edge computing unit through OTA technology, reduce maintenance cost;

[0218] Four, performance indicators:

[0219] Computing performance: support real-time inference of 1 million point cloud data and 5 frames of 640x640 images per second;

[0220] Delay indicators: dynamic compensation calculation ≤10 ms, multi-scale template matching ≤50 ms, and full-process processing ≤100 ms;

[0221] Reliability: mean time between failures (MTBF) ≥10,000 hours, support for -20℃ to 60℃ wide temperature working environment;

[0222] Through the above design, the edge computing unit realizes the full-process closed loop from data acquisition, processing to decision-making, providing efficient and reliable computing support for highway guardrail height dynamic detection.

[0223] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A highway guardrail height detection method based on line-knotted light, characterized by: The following steps are involved: S1. Install a line structured light scanner, a high-definition camera, and a real-time dynamic differential positioning device on the side of the inspection vehicle to form an on-board inspection system; S2. The vehicle travels at a preset speed in a lane at a preset distance from the guardrail, and simultaneously collects guardrail 3D point cloud data, image data, and vehicle positioning data; S3. Perform intelligent identification of guardrail types through the edge computing unit and output the guardrail type code; S4. Perform point cloud dynamic compensation, road reference plane extraction, and guardrail feature extraction based on guardrail type: Perform motion compensation and denoising on point cloud data; Extract road datum elevation; Select waveform mathematical model or plane fitting method according to guardrail type to extract guardrail top surface elevation; S5. Calculate the actual height of the guardrail based on the difference between the road base elevation and the guardrail top elevation.

2. The method for detecting the height of a highway guardrail based on line-knotted light according to claim 1, characterized in that: In step S1, the installation height of the line structured light scanner is 1.2m±0.1m, the depression angle is 15°±2°, and the high-definition camera is coaxially fixed with the line structured light scanner. In step S2, the preset distance between the vehicle and the guardrail is 1.5m±0.2m, the preset speed is 80km / h, and the following conditions are satisfied: The line structured light scanner collects 3D point cloud data at a frequency of 100 Hz; The high-definition camera captures images of the guardrail at 5 frames per second; The real-time dynamic differential positioning device outputs centimeter-level vehicle position coordinates.

3. The method for detecting the height of a highway guardrail based on line-knotted light according to claim 1, characterized in that: The step S3 comprises: A pre-trained instance segmentation model is used to process a 640×640 pixel input image and output a guardrail type code, which includes 01 double waveform / 02 triple waveform / 03 concrete; The training of the instance segmentation model satisfies: 707 annotated guardrail images were used, with a training set to validation set ratio of 8:2; Training parameters: learning rate 0.001, batch size 8, training epochs 100; Mosaic enhancement was turned off in the last 10 rounds of training.

4. The method for detecting the height of a highway guardrail based on line-knotted light according to claim 1, characterized in that: The motion compensation and denoising processing of the point cloud data in step S4 includes: Statistical filter denoising: remove outliers within a neighborhood radius of 5 cm whose distance exceeds the mean ±3 times the standard deviation; Dynamic point cloud motion compensation: A 4×4 motion transformation matrix is ​​constructed based on the vehicle's three-dimensional displacement (Δx, Δy, Δz) and heading angle change (Δθ) measured by the real-time dynamic differential positioning device. The first row of elements of the motion transformation matrix consists of the cosine value, negative sine value, 0 and X-axis displacement of the heading angle change; The second row of elements consists of the sine value, cosine value, 0 and Y-axis displacement of the heading angle change; The third row of elements consists of 0, 0, 1 and the Z-axis displacement; The fourth row of elements consists of 0, 0, 0, 1; Multiply the original point cloud coordinates by the motion transformation matrix to obtain the compensated point cloud coordinates.

5. The method for detecting the height of a highway guardrail based on line-knotted light according to claim 1, characterized in that: The road reference surface elevation is extracted in step S4 using a normal vector constrained random sampling consensus algorithm: Constrain the vertical component of the plane normal vector to be greater than 0.9; Fit the road plane equation through 1000 iterations; The road base elevation is determined by the ratio of the plane equation constant term to the vertical direction coefficient.

6. The method for detecting the height of a highway guardrail based on line-knotted light according to claim 1, characterized in that: In step S4, the guardrail top elevation is extracted by multi-scale sinusoidal model matching: The double-wave guardrail adopts a sinusoidal function model defined by amplitude, wavelength and phase angle parameters; The three-wave guardrail adopts a composite model with two sine waves of different amplitudes superimposed; The matching score calculation process includes: Generate a corresponding template for each scale factor α, with α taking values ​​of 0.8, 0.9, 1.0, 1.1, and 1.2; Calculate the square of the Euclidean distance between the actual coordinates of the point cloud and the theoretical coordinates of the template; Divide the square of the distance by twice the square of the tolerance threshold and take the negative exponent; The matching score is obtained by taking the arithmetic mean of the index results of all points.

7. The method for detecting the height of a highway guardrail based on line-knotted light according to claim 6, characterized in that: The multi-scale matching of the multi-scale sinusoidal model is performed in the order of standard scale → reduced scale → enlarged scale → minimum scale → maximum scale; when the matching score is ≥0.85, the matching is determined to be successful, and the elevation of the guardrail top surface takes the maximum vertical coordinate value in the matching point cloud; when the matching score is <0.85 for five consecutive scanning cycles, the standard template corresponding to the guardrail type output in step S3 is forcibly loaded.

8. The method for detecting the height of a highway guardrail based on line-knotted light according to claim 1, characterized in that: In step S4, the detection of the concrete guardrail includes: The guardrail plane is fitted using a random sampling consensus algorithm with the constraint that the horizontal component of the normal vector is >0.9; The maximum vertical coordinate value of the point cloud is taken as the top surface elevation.

9. The method for detecting the height of a highway guardrail based on line-knotted light according to claim 1, characterized in that: The height calculation in step S5 is specifically as follows: Subtract the road base elevation from the guardrail top elevation to obtain the actual guardrail height.

10. A method for detecting the height of a highway guardrail based on line-knotted light according to any one of claims 4, 6, and 7, characterized in that: The edge computing unit processing delay satisfies: Dynamic compensation calculation ≤10ms; Multi-scale template matching ≤ 50ms; Full process processing ≤100ms; Ensure real-time output of height detection results at a vehicle speed of 80km / h.

Images