Intelligent facing control method and system for paver

Abstract

The invention provides a paver intelligent welt control method and system, and particularly relates to the technical field of road construction machinery intelligentization, the method comprises the following steps: firstly, using industrial cameras installed on two sides of a paver to collect an operation area image in real time, then extracting image edge features, and calculating a definition score; according to a scoring result, selecting different paths to generate target boundary position coordinates: if the score is greater than a set threshold value, adopting a gradient operator and a traditional edge detection algorithm of Hough transformation; and otherwise, applying a target detection model based on YOLO series deep learning. Afterwards, bird's-eye view transformation is carried out based on the target boundary position coordinates to eliminate perspective distortion, and transverse offset is obtained. And finally, generating a screed stretching control instruction according to the transverse offset, transmitting the screed stretching control instruction to a paver controller through a CAN (Controller Area Network) bus, and driving a telescopic oil cylinder to execute screed welting action so as to realize high-precision self-adaptive welting.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent road construction machinery, and in particular to an intelligent edge control method and system for a paver. Background Art

[0002] As road construction demands ever-increasing quality, efficiency, and intelligent technology, the operating accuracy and degree of automation of pavers, as key equipment for pavement shaping, directly impact project quality and cost. During paving operations, it is crucial to ensure that the edges of the screed (such as curbstones, bridge edges, lime lines, etc.) adhere closely and maintain high precision (usually requiring ≤2cm). However, traditional methods rely heavily on manual observation and operation by left and right assistant operators to adjust the extension and retraction of the screed. This not only significantly increases labor costs and management difficulty, but also, due to the subjectivity of manual judgment, delayed response, and environmental interference (such as strong light, shadows, and complex linear shapes), it is difficult to consistently meet the edge accuracy standards, limiting construction efficiency and becoming a bottleneck restricting the intelligent upgrade of paving operations and reducing costs and increasing efficiency.

[0003] To address the drawbacks of manual operation, existing technologies attempt to incorporate methods such as laser ranging, ultrasonic sensors, or GPS positioning to assist paving operations. However, these technologies often face challenges such as reduced perception accuracy, insufficient reliability, or high costs in complex and variable real-world road construction environments (such as those with shadows, bridge obstructions, non-standard edge structures, or strong light interference). Furthermore, existing publicly available intelligent control technologies for pavers primarily focus on paving thickness control, machine path navigation, or autonomous driving capabilities, and do not specifically provide effective solutions for the dynamic recognition and automatic linkage control of screed plate edges and road boundaries (such as curbstones and lime lines). These solutions lack real-time, robust visual perception of boundary features, and fail to establish a visually-recognized automatic screed plate extension and retraction control mechanism. Therefore, they cannot truly replace manual labor for high-precision, adaptive edge alignment.

[0004] In summary, how to establish a high-precision, adaptive linkage adjustment mechanism between the paver screed and the road boundary by integrating real-time visual recognition and automatic control technology to stably achieve intelligent edge-adhering operations without human intervention is an urgent problem that needs to be solved. Summary of the Invention

[0005] The main purpose of the present invention is to provide a method and system for intelligent edge-adapting control of a paver, so as to at least solve the technical problem of how to establish a high-precision, adaptive linkage adjustment mechanism between the paver screed and the road boundary by integrating real-time visual recognition and automatic control technology, so as to stably realize intelligent edge-adapting operation without human intervention. Therefore, by integrating dual-path visual recognition and closed-loop control mechanism, high-precision adaptive edge-adapting of the paver screed to the road boundary can be stably realized without human intervention.

[0006] In order to achieve the above objectives, the present invention provides a method and system for intelligent edge control of a paver.

[0007] In a first aspect, the present invention provides a method for intelligent edge control of a paver, the control method comprising: Acquire images of the paver's operating area, where the images are collected in real time by industrial cameras installed on both sides of the paver; Extracting edge features of the image of the work area and calculating an image clarity score; When the image clarity score is greater than a set threshold, the target boundary position coordinates are generated by a traditional edge detection algorithm using a gradient operator and Hough transform; otherwise, the target boundary position coordinates are generated by a target detection model based on the YOLO series deep learning; Performing a bird's-eye view transformation based on the target boundary position coordinates to eliminate perspective distortion and generate a lateral offset relative to the paver screed; An ironing plate extension and retraction control instruction is generated according to the lateral offset, and the extension and retraction control instruction is transmitted to the paver controller through the CAN bus to drive the extension cylinder to perform the ironing plate edge-adhering action.

[0008] Specifically, extracting edge features of the work area image and calculating the image clarity score includes: Performing edge detection on the work area image using a Sobel operator to extract edge gradient information of the work area image; Calculating a variance value of the gradient amplitude in the working area image based on the edge gradient information; The variance of the gradient amplitude is normalized to a value in the range of 0-100 and used as the image clarity score.

[0009] Specifically, the traditional edge detection algorithm using the gradient operator and Hough transform to generate the target boundary position coordinates includes: Using the Sobel operator to perform edge detection on the image of the working area to generate an edge gradient map; performing a Hough transform on the edge gradient map to identify a linear target boundary in the work area image; The pixel coordinates of the endpoints of the linear target boundary are extracted as the target boundary position coordinates.

[0010] Specifically, the target boundary position coordinates are generated by the target detection model based on the YOLO series deep learning, including: Input the work area image into the pre-trained YOLOv5 model; Outputting the detection frame coordinates of the target boundary in the work area image through the YOLOv5 model; The endpoint coordinates of the boundary line segment are parsed from the detection frame coordinates as the target boundary position coordinates.

[0011] Specifically, performing a bird's-eye view transformation based on the target boundary position coordinates to eliminate perspective distortion and generate a lateral offset relative to the screed of the paver includes: Establish a perspective transformation matrix based on the installation height and pitch angle parameters of the paver industrial camera; Converting the target boundary position coordinates from the image coordinate system to the world coordinate system using the perspective transformation matrix; The horizontal distance between the target boundary and the center line of the paver screed is calculated in the world coordinate system as a lateral offset.

[0012] Specifically, generating a screed extension and retraction control instruction according to the lateral offset, and transmitting the extension and retraction control instruction to the paver controller via the CAN bus to drive the extension cylinder to perform the screed edge contact action includes: Calculating the piston displacement value of the telescopic oil cylinder according to the lateral offset; generating a screed extension and retraction control instruction including a displacement instruction based on the piston displacement value; The telescopic control instruction is transmitted to the paver controller via the CAN bus, driving the telescopic cylinder to perform the edge-adhering action corresponding to the piston displacement value.

[0013] Specifically, after transmitting the telescopic control instruction to the paver controller via the CAN bus, the method further includes: The paver controller performs message verification on the received screed extension and retraction control command, and if the verification fails or no feedback signal from the paver controller is received within a preset response time, it is determined that a communication abnormality occurs; When the communication abnormality occurs, it automatically switches to manual control mode and triggers the sound and light alarm device; The manual control mode interrupts the automatic edge-adhering action, and the operator manually controls the extension and retraction of the screed plate through the paver control panel.

[0014] In a second aspect, the present invention provides an intelligent edge control system for a paver, wherein the control system applies the control method described in the first aspect, and the control system comprises: Image acquisition modules are installed on both sides of the paver, and are used to obtain images of the paver's working area. The images of the working area are collected in real time by industrial cameras; An edge scoring module is connected to the image acquisition module, and is used to extract edge features of the image of the work area and calculate an image clarity score; A coordinate generation module is connected to the edge scoring module, and is used to determine that when the image clarity score is greater than a set threshold, the target boundary position coordinates are generated by a traditional edge detection algorithm using a gradient operator and Hough transform; otherwise, the target boundary position coordinates are generated by a target detection model based on the YOLO series deep learning; an offset calculation module connected to the coordinate generation module, the offset calculation module being configured to perform a bird's-eye view transformation based on the target boundary position coordinates to eliminate perspective distortion and generate a lateral offset relative to the screed plate of the paver; An execution control module is connected to the offset calculation module, and is used to generate an ironing board extension and contraction control instruction according to the lateral offset, and transmit the extension and contraction control instruction to the paver controller through the CAN bus to drive the extension cylinder to perform the ironing board edge-adhering action.

[0015] Specifically, the edge scoring module includes: an edge feature extraction unit, configured to perform edge detection on the work area image using a Sobel operator, and extract edge gradient information of the work area image; a variance calculation unit connected to the edge feature extraction unit, the variance calculation unit being configured to calculate a variance value of a gradient amplitude in the work area image based on the edge gradient information; A score normalization unit is connected to the variance calculation unit, and is used to normalize the variance value of the gradient amplitude to a value in the range of 0-100 as the image clarity score.

[0016] Specifically, the coordinate generation module includes: a dynamic decision-making unit, configured to determine whether the image clarity score is greater than a set threshold, and if so, activate a traditional detection unit; otherwise, activate a deep learning detection unit; A traditional detection unit is connected to the dynamic decision unit, and is used to generate target boundary position coordinates by using a traditional edge detection algorithm of a gradient operator and a Hough transform; A deep learning detection unit is connected to the dynamic decision unit, and the deep learning detection unit is used to generate target boundary position coordinates through a target detection model based on the YOLO series deep learning.

[0017] The present application provides a method and system for intelligent edge-adhesion control of a paver. The method utilizes industrial cameras installed on both sides of the paver to collect images of the working area in real time, extract image edge features, and calculate a clarity score. Based on the comparison result of the image clarity score and the set threshold, different methods are used to generate the target boundary position coordinates: if the score is greater than the threshold, the traditional edge detection algorithm using the gradient operator and Hough transform is used; if the score is not greater than the threshold, the target detection model based on the YOLO series deep learning is used. A bird's-eye view transformation is performed based on the generated target boundary position coordinates to eliminate perspective distortion, thereby obtaining a lateral offset relative to the paver's screed. Based on the lateral offset, an screed extension and retraction control instruction is generated and transmitted to the paver controller via the CAN bus to drive the telescopic cylinder to perform the screed's edge-adhesion action. This method integrates dual-path visual recognition and a closed-loop control mechanism to achieve high-precision adaptive edging of the paver's screed to the road boundary. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings:

[0019] Figure 1 A schematic diagram of the process flow of the intelligent edge control method for a paver provided in this application; Figure 2 This is a connection diagram of the intelligent edge control system for the paver provided in this application.

[0020] The above drawings illustrate specific embodiments of the present application, which will be described in more detail below. These drawings and the textual description are not intended to limit the scope of the present application in any way, but rather to illustrate the concepts of the present application to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION

[0021] To make the purpose, technical solutions, and advantages of this application more clear, the technical solutions in this application will be clearly and completely described below in conjunction with the drawings in this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0022] The terms "first," "second," "third," "fourth," and so forth (if any) in the present description and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, such that the embodiments of the present invention described herein can be practiced in orders other than those illustrated or described herein.

[0023] In the present invention, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described in this application as "exemplary" or "for example" should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0024] This application provides a method and system for intelligent edge-adhesion control of a paver. This control method uses industrial cameras mounted on both sides of the paver to capture real-time images of the work area, extract edge features, and calculate a clarity score. Based on the score, a traditional edge detection algorithm or a YOLO-based deep learning-based object detection model is selected to generate target boundary coordinates. A bird's-eye view transformation eliminates perspective distortion, deriving a lateral offset. This generates screed extension and extension control commands, which are transmitted via the CAN bus to a controller that drives the extension cylinder to complete the edge-adhesion operation.

[0025] The following specific embodiments describe in detail the technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0026] Figure 1 The flow chart of the intelligent edge control method for the paver provided in this application is as follows: Figure 1 As shown, the intelligent edge control method of the paver provided in this embodiment includes: S101: Obtain an image of the working area of the paver, which is collected in real time by industrial cameras installed on both sides of the paver; extract edge features of the working area image and calculate the image clarity score.

[0027] Specifically, extracting edge features of the work area image and calculating the image clarity score includes: Performing edge detection on the work area image using a Sobel operator to extract edge gradient information of the work area image; Calculating a variance value of the gradient amplitude in the working area image based on the edge gradient information; The variance of the gradient amplitude is normalized to a value in the range of 0-100 and used as the image clarity score.

[0028] During implementation, step S101 specifically includes: Step S101.1: Industrial camera image acquisition 1.1 Install industrial cameras on the front brackets of the left and right side plates of the paver, with the camera at a height of 1.2 meters from the ground and the pitch angle set to -15° (tilted downward).

[0029] 1.2 The camera collects RGB images of the working area with a resolution of 1280×720 at a rate of 30 frames per second. The image coverage range includes the road surface area from 3 meters to 10 meters in front of the paver.

[0030] Step S101.2: Sobel operator edge detection 2.1 Convert the collected RGB work area image into a grayscale image. The conversion formula is: Gray = 0.299 * R + 0.587 * G + 0.114 * B, Where R, G, and B are the red, green, and blue channel values of the pixel respectively.

[0031] 2.2 Apply a 3×3 Sobel horizontal convolution kernel to the grayscale image ( ) and the vertical convolution kernel ( ): .

[0032] 2.3 Calculate the horizontal gradient component of each pixel and the vertical gradient component : .

[0033] in, represents a two-dimensional discrete convolution operation, Represents a grayscale image matrix.

[0034] 2.4 Output edge gradient information: including gradient amplitude and direction .

[0035] Step S101.3: Calculate the gradient amplitude variance 3.1 Extract the gradient magnitude G of all pixels in the whole image to form an magnitude set (n is the total number of pixels).

[0036] 3.2 Calculate the variance of the amplitude set S : , in is the mean amplitude.

[0037] Step S101.4: Normalization of clarity score 4.1 Variance value Linear mapping to the range 0-100: , in =0 (no gradient change), =106 (calibrated based on 100,000 samples).

[0038] 4.2 When Score>100, it is forced to be truncated to 100, and when Score<0, it is forced to be zero.

[0039] In this step, industrial cameras on both sides are deployed to collect 1280×720 resolution working images in real time, and the Sobel operator is used for accurate edge detection (3×3 horizontal / vertical convolution kernels are used to calculate the gradient amplitude G). Image clarity is quantified and linearly normalized to generate a clarity score ranging from 0 to 100. This method provides an objective quantitative basis for subsequent dual-path decision-making (S102-S104). It maintains scoring accuracy in strong light environments (>100,000 lux), replacing traditional subjective judgment and ensuring that the paver can reliably trigger the optimal recognition path even in complex lighting conditions.

[0040] S102: Determine whether the image clarity score is greater than a set threshold, if so, execute step S103, if not, execute step S104.

[0041] S103: Generate the target boundary position coordinates through the traditional edge detection algorithm of gradient operator and Hough transform.

[0042] Specifically, the traditional edge detection algorithm using the gradient operator and Hough transform to generate the target boundary position coordinates includes: Using the Sobel operator to perform edge detection on the image of the working area to generate an edge gradient map; performing a Hough transform on the edge gradient map to identify a linear target boundary in the work area image; The pixel coordinates of the endpoints of the linear target boundary are extracted as the target boundary position coordinates.

[0043] S104: Generate target boundary position coordinates through the target detection model based on the YOLO series deep learning.

[0044] Specifically, the target boundary position coordinates are generated by the target detection model based on the YOLO series deep learning, including: Input the work area image into the pre-trained YOLOv5 model; Outputting the detection frame coordinates of the target boundary in the work area image through the YOLOv5 model; The endpoint coordinates of the boundary line segment are parsed from the detection frame coordinates as the target boundary position coordinates.

[0045] During implementation, steps S102-S104 specifically include: S102: Clarity Score Decision Read the image clarity score generated by S101 (denoted as S, ranging from 0 to 100) and compare it with the preset threshold T: If S>T: activate the traditional edge detection path (execute S103); If S≤T: activate the deep learning detection path (execute S104); Where T=40.

[0046] S103: Traditional edge detection path (Sobel+Hough transform) Step S103.1: Generate Sobel edge gradient map 1.1 Apply the 3×3 Sobel operator to the grayscale working area image (size 1280×720) processed by S101: Horizontal gradient: , .

[0047] Vertical Gradient: , .

[0048] Gradient Magnitude: .

[0049] 1.2 Binarization: Setting the Threshold =50, when ≥ Set to 1 when , otherwise set to 0, generate edge gradient map (binary matrix ).

[0050] Step S103.2: Hough transform to identify straight line boundaries 2.1 pairs Perform a Hough transform: Parameter space definition: ( is the distance from the origin to the straight line, and θ is the angle); Accumulator matrix A dimensions: (step size 1 pixel), θ∈[0,180) (step size 1); Traversal Pixels with a median value of 1 ( ), calculate all ( ) combination, for A( ) accumulates counts, where i represents the edge pixel index and j represents the parameter combination index.

[0051] 2.2 Extract peak value: Select the value of accumulator A greater than 200 ( ), as candidate line parameters.

[0052] Step S103.3: Endpoint coordinate extraction 3.1 For each candidate line : Compute its intersection with the image boundary: Left boundary x=0: ; Right boundary x=1279: .

[0053] Intercept the endpoints of the line segment within the image range ( )and( ).

[0054] 3.2 Output target boundary position coordinates: {( ),( )} (one pair of coordinates for each target boundary).

[0055] S104: Deep Learning Detection Path (YOLOv5) Before executing step S104, the YOLOv5 model needs to be trained. The YOLOv5 model training method is as follows: 1. Training dataset construction The model was trained using a total of 10,000 images of paving scenes. 5,000 of these images were captured from actual paver operation video frames (covering various operating conditions, including strong lighting, shadows, rain, and fog), and 5,000 were generated using image enhancement techniques (including data augmentation using affine transformation and brightness adjustment). The training samples were annotated with three types of objects: curbstones with their bottom edge lines (3,200 sets of annotations), lime lines with complete continuous segments (4,500 sets of annotations), and bridge guardrail vertical boundaries (2,300 sets of annotations).

[0056] 2. Model structure and training parameters The YOLOv5 network architecture was used, consisting of a CSPDarknet53 backbone network, a PANet feature pyramid fusion structure, and three detection heads of different scales (output resolutions of 80×80, 40×40, and 20×20, respectively). The training input resolution was fixed at 640×640 pixels, and stochastic gradient descent (SGD) was used for 300 epochs of training. The momentum parameter was set to 0.937, the weight decay coefficient was 0.0005, and the learning rate was gradually decayed from an initial value of 0.01 to 0.001 using a cosine annealing strategy. Each training batch consisted of 32 images.

[0057] 3. Loss function configuration The model training uses a composite loss function: bounding box regression uses the CIoU loss function to measure the position deviation between the predicted box and the annotated box, target classification uses the binary cross entropy loss function (BCEWithLogitsLoss), and confidence prediction uses the focal loss function (Focal Loss) to solve the problem of sample category imbalance.

[0058] 4. Training hardware environment Model training was completed on a computing platform consisting of four NVIDIA Tesla V100 graphics processors, and the training time was approximately 48 hours.

[0059] After the training is completed, perform the following specific steps: Step S104.1: YOLOv5 model input preprocessing Scale the RGB work area image captured by S101 to 640×640 resolution; Normalize pixel values: ( is the original pixel value).

[0060] Step S104.2: YOLOv5 model inference 2.1 Input data: preprocessed images (Size 640×640×3); 2.2 Model Structure Backbone: CSPDarknet53 (convolution kernel size 3×3, stride 2 downsampling); Neck: PANet (feature pyramid fusion); Head: 3 detection heads (output sizes 80×80, 40×40, and 20×20).

[0061] 2.3 Output: Detection box parameters ( ): ( ): Bounding box center coordinates (relative to grid offset); ( ): width and height of the bounding box (relative to the image size); : target confidence (0-1); : Class label (kerbstone = 0, lime line = 1, bridge edge = 2).

[0062] Step S104.3: Boundary coordinate analysis 3.1 Screening Confidence Detection frame ≥0.5; 3.2 Convert relative coordinates to absolute pixel coordinates: .

[0063] 3.3 Extract boundary segment endpoints: Sort the vertices of the rectangular detection box and output the target boundary position coordinates: {( ),( )}.

[0064] This step intelligently switches between traditional edge detection (Sobel+Hough) and deep learning (YOLOv5) through a dynamic decision-making mechanism (threshold T=40). When the clarity score S>40, the Sobel operator is used to generate an edge gradient map, which is then transformed using a Hough transform to extract the coordinates of the line boundary endpoints. When S≤40, the image is fed into a pre-trained YOLOv5 model (CSPDarknet53 backbone + PANet feature fusion) to analyze the detection bounding box coordinates and generate boundary endpoints. These two complementary paths cover targets such as lime lines, curbstones, and bridge edges, and can output sub-pixel-level target boundary coordinates in both bright and shadowy environments. This provides reliable input for subsequent bird's-eye view transformations, eliminating the risk of failure of a single algorithm in complex working conditions.

[0065] S105: Performing a bird's-eye view transformation based on the target boundary position coordinates to eliminate perspective distortion and generate a lateral offset relative to the paver screed.

[0066] Specifically, performing a bird's-eye view transformation based on the target boundary position coordinates to eliminate perspective distortion and generate a lateral offset relative to the screed of the paver includes: Establish a perspective transformation matrix based on the installation height and pitch angle parameters of the paver industrial camera; Converting the target boundary position coordinates from the image coordinate system to the world coordinate system using the perspective transformation matrix; The horizontal distance between the target boundary and the center line of the paver screed is calculated in the world coordinate system as a lateral offset.

[0067] During implementation, step S105 specifically includes: S105.1: Perspective Transformation Matrix Construction 1.1 Camera parameter acquisition (same as S101): Installation height h = 1.2m (vertical distance from the camera optical center to the ground); Pitch angle θ = (the tilt angle of the camera optical axis to the ground); focal length =4.5mm (physical focal length of the camera); Image sensor size =6.287mm×4.712mm.

[0068] 1.2 Internal parameter matrix calculation: , Among them, 640×360 is the center point coordinate of the 1280×720 image.

[0069] 1.3 Calculation of external parameter matrix: Rotation Matrix (Rotation θ around the X axis).

[0070] Translation vector .

[0071] 1.4 Perspective transformation matrix: , Specific parameter values:

[0072] S105.2: Image Coordinate System to World Coordinate System Conversion 2.1 Input: Target boundary position coordinates generated by S103 / S104 (Endpoint pixel coordinates (u,v), such as ( )).

[0073] 2.2 Conversion formula: .

[0074] 2.3 World coordinate calculation ground =0):

[0075] 2.4 Output: Boundary point coordinates in the world coordinate system ( , ).

[0076] S105.3: Calculation of Lateral Offset 3.1 Screed centerline reference: The paver's travel direction is the positive direction of the Y axis; The position of the screed centerline in the world coordinate system: =0 (longitudinal symmetry plane of the vehicle body).

[0077] 3.2 Calculate horizontal distance: For each boundary endpoint ( , ): .

[0078] The median of Δd of all endpoints is taken as the lateral offset.

[0079] 3.3 Output: lateral offset Δd (unit: meter).

[0080] This step converts the target boundary endpoint coordinates (from S103 / S104) in the image coordinate system into the ground position in the world coordinate system (( , ) and generates a lateral offset by calculating the absolute distance |Δd| from the boundary point to the longitudinal symmetry plane of the paver (X=0). This transformation eliminates perspective distortion caused by camera tilt, ensuring that a 1-pixel error at a distance of 10 meters corresponds to an actual error of <0.02 meters. This provides a real-world spatial position reference for screed extension and retraction control, addressing the edge position deviation caused by uncorrected perspective distortion in traditional methods.

[0081] S106: Generate an extension and retraction control instruction for the screed plate according to the lateral offset, and transmit the extension and retraction control instruction to the paver controller via the CAN bus to drive the extension cylinder to perform the screed plate edge-adhering action.

[0082] Specifically, generating a screed extension and retraction control instruction according to the lateral offset, and transmitting the extension and retraction control instruction to the paver controller via the CAN bus to drive the extension cylinder to perform the screed edge contact action includes: Calculating the piston displacement value of the telescopic oil cylinder according to the lateral offset; generating a screed extension and retraction control instruction including a displacement instruction based on the piston displacement value; The telescopic control instruction is transmitted to the paver controller via the CAN bus, driving the telescopic cylinder to perform the edge-adhering action corresponding to the piston displacement value.

[0083] Optionally, after transmitting the telescopic control instruction to the paver controller via the CAN bus, the method further includes: The paver controller performs message verification on the received screed extension and retraction control command, and if the verification fails or no feedback signal from the paver controller is received within a preset response time, it is determined that a communication abnormality occurs; When the communication abnormality occurs, it automatically switches to manual control mode and triggers the sound and light alarm device; The manual control mode interrupts the automatic edge-adhering action, and the operator manually controls the extension and retraction of the screed plate through the paver control panel.

[0084] During implementation, step S106 specifically includes: S106.1: Calculation of Piston Displacement 1.1 Input: lateral offset generated by S105 (denoted as Δd, unit: meter); 1.2 Conversion formula:

[0085] in: L: Displacement of the telescopic cylinder piston (unit: mm); k: displacement conversion coefficient, value is 1500 mm / m; Δd: lateral offset (unit: meter).

[0086] S106.2: Generation of scaling control instructions 2.1 Instruction data structure (compliant with ISO 11898 standard): CAN identifier: 0x18FF01A0 (screed control dedicated ID); Data field format (as shown in Table 1): Table 1:

[0087] 2.2 Instruction generation steps: 2.2.1 Split the displacement value L into two bytes: Low byte B0= mod 256; High byte B1= .

[0088] 2.2.2 Combined data field: [B0, B1, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00].

[0089] S106.3: CAN bus transmission and execution 3.1 Physical layer parameters: Protocol: CAN 2.0B; Baud rate: 250 kbps; Cable: twisted-pair shielded cable (characteristic impedance 120Ω).

[0090] 3.2 Transmission process: 3.2.1 The instruction sender (Raspberry Pi edge computing platform) loads the data field into the CAN controller; 3.2.2 Converted into differential signal through CAN transceiver; 3.2.3 The paver controller receives and parses data: Verify identifier 0x18FF01A0; Extract bytes 0-1 and reassemble the displacement value L.

[0091] 3.3 Cylinder drive execution: The controller outputs a 4-20mA analog signal to the proportional valve; Cylinder piston moving speed v=50mm / s (constant); After reaching the target displacement L, the position closed-loop control is maintained.

[0092] S106.4: Communication exception handling (optional) Step S106.4.1: Message Verification (1) Checksum algorithm: CRC-16-CCITT (polynomial 0x1021); Calculate the CRC value of the received data field (8 bytes); Compare with the last 2 bytes of the data frame.

[0093] (2) Timeout determination: The preset response time is 200ms.

[0094] Step S106.4.2: Exception Response (1) Trigger action: Sound and light alarm: The buzzer outputs 2000Hz@90dB sound waves; The red LED flashes at a frequency of 2Hz.

[0095] Mode switch: Send 0x00000000 reset command to the console; The hydraulic system switches to manual servo mode.

[0096] Step S106.4.3: Manual Control Operation method: The control console joystick voltage signal range is 0-5V; Cylinder displacement L 手动 =300×V (unit: mm) (V is the joystick voltage value).

[0097] This step accurately converts the lateral offset Δd into the piston displacement value L through the displacement conversion coefficient k=1500mm / m, generates a CAN bus control command that complies with the ISO 11898 standard (identifier 0x18FF01A0, data field 8 bytes), and drives the telescopic cylinder to accurately position at a speed of 50mm / s at a baud rate of 250kbps. When a CRC check failure or a 200ms response timeout is detected, an audible and visual alarm (2000Hz beep + 2Hz red light) is immediately triggered and the system switches to manual control mode (joystick voltage value - displacement conversion rate 300mm / V). While ensuring the paving edge error is ≤2cm, the system downtime rate can be significantly reduced, achieving the dual requirements of construction safety and precision.

[0098] This embodiment provides a method for intelligent edge-adhesion control of a paver. This method uses industrial cameras installed on both sides of the paver to capture images of the work area in real time, extract edge features from the images, and calculate a clarity score. Based on a comparison of the image clarity score with a set threshold, different strategies are used to generate the target boundary position coordinates: when the score is greater than the threshold, a traditional edge detection algorithm using a gradient operator and Hough transform is used; when the score is not greater than the threshold, a target detection model based on the YOLO series of deep learning is used. A bird's-eye view transformation is performed based on the generated target boundary position coordinates to eliminate perspective distortion and obtain a lateral offset relative to the paver's screed. Based on this lateral offset, a screed extension and extension control command is generated and transmitted to the paver controller via the CAN bus, driving the extension cylinder to execute the screed's edge-adhesion action. This method combines dual-path visual recognition with a closed-loop control mechanism to achieve high-precision, adaptive, and intelligent edge-adhesion of the paver's screed to the road boundary without the need for human intervention.

[0099] Figure 2 The connection diagram of the intelligent edge control system of the paver provided in this application is as follows: Figure 2 As shown in the figure, the paver intelligent edge control system provided by this embodiment is applied Figure 1 The intelligent edge control method for a paver described in the embodiment, the control system includes: Image acquisition modules are installed on both sides of the paver, and are used to obtain images of the paver's working area. The images of the working area are collected in real time by industrial cameras; An edge scoring module is connected to the image acquisition module, and is used to extract edge features of the image of the work area and calculate an image clarity score; A coordinate generation module is connected to the edge scoring module, and is used to determine that when the image clarity score is greater than a set threshold, the target boundary position coordinates are generated by a traditional edge detection algorithm using a gradient operator and Hough transform; otherwise, the target boundary position coordinates are generated by a target detection model based on the YOLO series deep learning; an offset calculation module connected to the coordinate generation module, the offset calculation module being configured to perform a bird's-eye view transformation based on the target boundary position coordinates to eliminate perspective distortion and generate a lateral offset relative to the screed plate of the paver; An execution control module is connected to the offset calculation module, and is used to generate an ironing board extension and contraction control instruction according to the lateral offset, and transmit the extension and contraction control instruction to the paver controller through the CAN bus to drive the extension cylinder to perform the ironing board edge-adhering action.

[0100] Specifically, the edge scoring module includes: an edge feature extraction unit, configured to perform edge detection on the work area image using a Sobel operator, and extract edge gradient information of the work area image; a variance calculation unit connected to the edge feature extraction unit, the variance calculation unit being configured to calculate a variance value of a gradient amplitude in the work area image based on the edge gradient information; A score normalization unit is connected to the variance calculation unit, and is used to normalize the variance value of the gradient amplitude to a value in the range of 0-100 as the image clarity score.

[0101] Specifically, the coordinate generation module includes: a dynamic decision-making unit, configured to determine whether the image clarity score is greater than a set threshold, and if so, activate a traditional detection unit; otherwise, activate a deep learning detection unit; A traditional detection unit is connected to the dynamic decision unit, and is used to generate target boundary position coordinates by using a traditional edge detection algorithm of a gradient operator and a Hough transform; A deep learning detection unit is connected to the dynamic decision unit, and the deep learning detection unit is used to generate target boundary position coordinates through a target detection model based on the YOLO series deep learning.

[0102] The intelligent edge control system for a paver provided in this embodiment specifically includes the following steps when implemented: 1. Image acquisition module implementation Two industrial cameras are rigidly mounted on the front brackets of the left and right side panels of the paver, with their lenses facing the ground in front of the machine. The cameras are physically connected to the image processing unit via standard shielded GigE Vision cables. The cameras are configured to capture 30 frames per second of 1280×720 resolution images of the work area, covering the paving surface from 3 to 10 meters in front of the paver. The cameras are installed at a height of 1.2 meters from the optical center of the lens to the ground, with the optical axis tilted 15 degrees to the ground.

[0103] 2. Edge scoring module implementation The edge feature extraction unit receives the work area image from the industrial camera via the PCIe bus. It converts the work area image into a grayscale image and applies the Sobel operator with a specific convolution kernel (horizontally [-1, 0, 1; -2, 0, 2; -1, 0, 1], and vertically [-1, -2, -1; 0, 0, 0; 1, 2, 1]) to calculate the gradient component.

[0104] The variance calculation unit receives the gradient magnitude matrix output by the edge feature extraction unit through the data bus. This unit calculates the arithmetic mean of the gradient magnitudes of all pixels, calculates the sum of the squared differences between each gradient magnitude and the mean, and divides the sum by the total number of pixels.

[0105] The score normalization unit linearly maps the output value of the variance calculation unit to the range of 0-100: the minimum variance value corresponds to 0 points, the maximum variance value corresponds to 100 points, and the intermediate values are converted proportionally.

[0106] 3. Implementation of coordinate generation module The dynamic decision unit receives the image clarity score output by the edge scoring module via the GPIO port. When the score exceeds the set threshold of 40, the traditional detection unit is activated; when the score is less than or equal to 40, the deep learning detection unit is activated. The dynamic decision unit is designed to address the robustness of edge detection under varying lighting conditions.

[0107] The traditional detection unit extracts edge gradient information from the work area image using a Sobel edge detector, binarizes the gradient amplitude (threshold 50), and then performs Hough line detection (with an angle step of 1 degree and a distance step of 1 pixel in parameter space). It extracts lines with cumulative values exceeding 200 as target boundaries and outputs the pixel coordinates of the line endpoints.

[0108] The deep learning detection unit uses a pre-trained YOLOv5 target detection model (including the CSPDarknet53 backbone network and the PANet feature fusion structure), scales the work area image to a 640×640 resolution input model, and extracts the detection box vertices with a confidence level exceeding 0.5 as the target boundary coordinates.

[0109] 4. Implementation of the offset calculation module This module receives the boundary endpoint coordinates output by the coordinate generation module via the LVDS interface. Using industrial camera calibration parameters (focal length 4.5mm, sensor size 6.287mm × 4.712mm, mounting height 1.2m, and pitch angle 15°), a perspective transformation model is established to convert the endpoint pixel coordinates into X-direction distances relative to the paver screed in the world coordinate system. The median of the multiple transformed lateral distances is calculated to output the final lateral offset. This step eliminates lens distortion and ensures an offset accuracy of 0.02 meters per 10 meters.

[0110] 5. Implementation of the execution control module The module receives lateral offset data from the offset calculation module via the SPI bus. The offset is converted into displacement commands based on a ratio of 1500 mm per meter of distance to a cylinder piston stroke, and then assembled into a CAN bus message: message identifier 0x18FF01A0. The low byte of the data field stores the rounded-off lower 8 bits of the displacement value, and the high byte stores the rounded-off upper 8 bits of the displacement value. The control mode byte is fixed at 0x01, indicating automatic control. This message is transmitted to the paver controller via a CAN transceiver at a baud rate of 250 kbps. If the CAN bus response times out by 200 ms or the CRC check fails, a 2000 Hz buzzer is triggered, the red LED flashes twice per second, and the paver controller automatically switches to manual control mode to receive the analog voltage signal from the joystick.

[0111] When the intelligent edge-adhesion control system for the paver provided in this embodiment is in operation, an industrial camera captures images of the paving surface in real time. When the ambient light is sufficient, the traditional Sobel-Hough algorithm generates sub-pixel boundary coordinates. When rain, fog, or shadows blur the image, the YOLOv5 model maintains detection stability. After all boundary coordinates are dedistorted using a bird's-eye view transformation, a lateral offset is generated to drive the cylinder to perform the edge-adhesion action. Test data shows that under conditions ranging from strong light of 100,000 lux to weak light of 50 lux, the system maintains a stable paving edge error within 2 cm, significantly reducing the interruption rate of paving operations.

[0112] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present invention that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only.

[0113] It will be understood that the present application is not limited to the exact construction that has been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof.

Claims

1. A method for intelligent edge control of a paver, characterized in that: The method comprises: Acquire images of the paver's operating area, where the images are collected in real time by industrial cameras installed on both sides of the paver; Extracting edge features of the image of the work area and calculating an image clarity score; When the image clarity score is greater than a set threshold, the target boundary position coordinates are generated by a traditional edge detection algorithm using a gradient operator and Hough transform; otherwise, the target boundary position coordinates are generated by a target detection model based on the YOLO series deep learning; Performing a bird's-eye view transformation based on the target boundary position coordinates to eliminate perspective distortion and generate a lateral offset relative to the paver screed; An ironing plate extension and retraction control instruction is generated according to the lateral offset, and the extension and retraction control instruction is transmitted to the paver controller through the CAN bus to drive the extension cylinder to perform the ironing plate edge-adhering action.

2. The intelligent edge control method for a paver according to claim 1, characterized in that: The step of extracting edge features of the work area image and calculating the image clarity score includes: Performing edge detection on the work area image using a Sobel operator to extract edge gradient information of the work area image; Calculating a variance value of the gradient amplitude in the working area image based on the edge gradient information; The variance of the gradient amplitude is normalized to a value in the range of 0-100 and used as the image clarity score.

3. The intelligent edge control method for a paver according to claim 1, characterized in that: The conventional edge detection algorithm using the gradient operator and Hough transform generates the target boundary position coordinates, including: Using the Sobel operator to perform edge detection on the image of the working area to generate an edge gradient map; performing a Hough transform on the edge gradient map to identify a linear target boundary in the work area image; The pixel coordinates of the endpoints of the linear target boundary are extracted as the target boundary position coordinates.

4. The intelligent edge control method for a paver according to claim 1, characterized in that: The target boundary position coordinates are generated by the target detection model based on the YOLO series deep learning, including: Input the work area image into the pre-trained YOLOv5 model; Outputting the detection frame coordinates of the target boundary in the work area image through the YOLOv5 model; The endpoint coordinates of the boundary line segment are parsed from the detection frame coordinates as the target boundary position coordinates.

5. The intelligent edge control method for a paver according to claim 1, characterized in that: The step of performing a bird's-eye view transformation based on the target boundary position coordinates to eliminate perspective distortion and generate a lateral offset relative to the screed plate of the paver includes: Establish a perspective transformation matrix based on the installation height and pitch angle parameters of the paver industrial camera; Converting the target boundary position coordinates from the image coordinate system to the world coordinate system using the perspective transformation matrix; The horizontal distance between the target boundary and the center line of the paver screed is calculated in the world coordinate system as a lateral offset.

6. The intelligent edge control method for a paver according to claim 1, characterized in that: Generating a screed extension and retraction control instruction according to the lateral offset, and transmitting the extension and retraction control instruction to a paver controller via a CAN bus to drive a retractable oil cylinder to perform an screed edge-adhering action, includes: Calculating the piston displacement value of the telescopic oil cylinder according to the lateral offset; generating a screed extension and retraction control instruction including a displacement instruction based on the piston displacement value; The telescopic control instruction is transmitted to the paver controller via the CAN bus, driving the telescopic cylinder to perform the edge-adhering action corresponding to the piston displacement value.

7. The intelligent edge control method for a paver according to claim 6, characterized in that: After transmitting the telescopic control instruction to the paver controller via the CAN bus, the method further includes: The paver controller performs message verification on the received screed extension and retraction control command, and if the verification fails or no feedback signal from the paver controller is received within a preset response time, it is determined that a communication abnormality occurs; When the communication abnormality occurs, it automatically switches to manual control mode and triggers the sound and light alarm device; The manual control mode interrupts the automatic edge-adhering action, and the operator manually controls the extension and retraction of the screed plate through the paver control panel.

8. An intelligent edge control system for a paver, characterized in that: The control system applies the control method according to any one of claims 1 to 7, and the control system includes: Image acquisition modules are installed on both sides of the paver, and are used to obtain images of the paver's working area. The images of the working area are collected in real time by industrial cameras; An edge scoring module is connected to the image acquisition module, and is used to extract edge features of the image of the work area and calculate an image clarity score; A coordinate generation module is connected to the edge scoring module, and is used to determine that when the image clarity score is greater than a set threshold, the target boundary position coordinates are generated by a traditional edge detection algorithm using a gradient operator and Hough transform; otherwise, the target boundary position coordinates are generated by a target detection model based on the YOLO series deep learning; an offset calculation module connected to the coordinate generation module, the offset calculation module being configured to perform a bird's-eye view transformation based on the target boundary position coordinates to eliminate perspective distortion and generate a lateral offset relative to the screed plate of the paver; An execution control module is connected to the offset calculation module, and is used to generate an ironing board extension and contraction control instruction according to the lateral offset, and transmit the extension and contraction control instruction to the paver controller through the CAN bus to drive the extension cylinder to perform the ironing board edge-adhering action.

9. The intelligent edge control system of a paver according to claim 8, characterized in that: The edge scoring module includes: an edge feature extraction unit, configured to perform edge detection on the work area image using a Sobel operator, and extract edge gradient information of the work area image; a variance calculation unit connected to the edge feature extraction unit, the variance calculation unit being configured to calculate a variance value of a gradient amplitude in the work area image based on the edge gradient information; A score normalization unit is connected to the variance calculation unit, and is used to normalize the variance value of the gradient amplitude to a value in the range of 0-100 as the image clarity score.

10. The intelligent edge control system of a paver according to claim 8, characterized in that: The coordinate generation module includes: a dynamic decision-making unit, configured to determine whether the image clarity score is greater than a set threshold, and if so, activate a traditional detection unit; otherwise, activate a deep learning detection unit; A traditional detection unit is connected to the dynamic decision unit, and is used to generate target boundary position coordinates by using a traditional edge detection algorithm of a gradient operator and a Hough transform; A deep learning detection unit is connected to the dynamic decision unit, and the deep learning detection unit is used to generate target boundary position coordinates through a target detection model based on the YOLO series deep learning.

Images