Non-contact road load identification method based on multiple regions of interest

Abstract

The present invention discloses a non-contact road load identification method for multiple regions of interest. The method comprises: locating vehicle tires in a camera-captured image and generating multiple regions of interest based thereon; extracting the tire-ground contact line, the tire-rim boundary line, and the tire model marking on the tire sidewall from the multiple regions of interest using a computer vision algorithm; obtaining inherent parameters of the tires through a table lookup based on the identified tire model markings; and calculating the axle weight of each axle on which the tire resides, combining the inherent parameters of the tires with the tire-ground contact line, and summing all axle weights to obtain the gross vehicle weight. The present invention has higher computational efficiency, thereby enabling real-time data processing.

Description

Technical Field

[0001] The present invention belongs to the field of road load identification, and in particular relates to a non-contact road load identification method for multiple regions of interest. Background Art

[0002] Overloaded vehicles pose a serious threat to the road transport industry. They not only increase the probability of traffic accidents but also accelerate the degradation of transportation infrastructure. In extreme cases, they can even cause severe road deformation and sudden bridge collapse. Therefore, identifying overloaded trucks on the road is crucial for ensuring traffic safety and reducing infrastructure maintenance costs. However, static weighing is the most widely used technology for overload enforcement. However, static weighing requires vehicle stops for inspection, often leading to traffic congestion. Furthermore, because static weighing is time-consuming and labor-intensive, overload enforcement currently only covers a limited number of roads, which undoubtedly creates an opportunity for overloaded transport.

[0003] To improve the efficiency of vehicle weight identification, road weigh-in-motion (PWIM) and bridge weigh-in-motion (BWIM) systems have been developed since the 1960s. These systems install sensors on the road surface or bridges to measure changes in physical quantities such as stress and acceleration in the engineering structure caused by passing vehicles, thereby inferring the weight of the moving vehicle. However, installing or replacing sensor networks on roads is extremely expensive and requires temporary closure of the road section. Therefore, the combined installation and maintenance costs of existing PWIM and BWIM systems are very high, making them difficult to promote and popularize.

[0004] Compared to traditional vehicle weight identification technology, emerging non-contact dynamic weighing technology offers advantages such as easy installation and maintenance, low overall cost, and high intelligence. However, existing non-contact dynamic weighing algorithms are often computationally intensive, making it difficult to achieve real-time vehicle weight identification and failing to meet the practical needs of overload enforcement. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the present invention proposes a non-contact road load identification method for multiple regions of interest. The specific technical solution is as follows:

[0006] A non-contact road load identification method for multiple regions of interest, the method comprising the following steps:

[0007] Step 1: Locate the car tire in the camera image and generate multiple regions of interest based on it;

[0008] Step 2: Extract the tire-ground contact line, tire-rim boundary line, and tire model mark on the tire sidewall from multiple regions of interest using computer vision algorithms;

[0009] Step 3: Obtain the inherent parameters of the tire by looking up the table based on the tire model mark identified in step 2; calculate the axle weight of each axle on which the tire is located by combining the inherent parameters of the tire and the tire-ground contact line, and sum all axle weights to obtain the total vehicle weight.

[0010] Furthermore, the step 1 includes the following sub-steps:

[0011] S1.1: Use the YOLO network to locate all car tires in the images captured during the non-contact dynamic weighing process, and record each car tire as an initial bounding box and corresponding class probability;

[0012] S1.2: These initial bounding boxes are cleaned according to their corresponding category probabilities and positions on the image to remove bounding boxes that cannot be used to generate the region of interest;

[0013] S1.3: The cleaned bounding box is converted into the bounding boxes of three regions of interest through different geometric transformation processes: tire-ground contact region of interest, rim region of interest, and tire sidewall text region of interest.

[0014] Furthermore, in step 2, the sub-step of extracting the tire-ground contact line using a computer vision algorithm is as follows:

[0015] (1) using an image resizing operation to proportionally resize the image size of the cropped tire-ground contact interest regions so that all tire-ground contact interest regions have the same image height;

[0016] (2) Increase contrast and suppress noise through pixel-adaptive Wiener filter;

[0017] (3) Using local adaptive thresholding technology to convert the enhanced image into a binary image;

[0018] (4) Morphological operations are used to modify the shape of the region in the binary image to eliminate the interference caused by stains, cracks and concentric stripes on the tire surface;

[0019] (5) The tire-ground contact interval in the morphologically processed binary image is identified by the edge boundary addressing method; the contact edge is obtained by smoothing the tire pattern on the outer edge of the tire in the tire-ground contact interval; and the contact edge is fitted into a straight line segment through a straight line fitting operation, and this line segment is used as the extracted tire-ground contact line.

[0020] Furthermore, the edge boundary addressing method includes:

[0021] (1) Pixel sampling: a binary image of the tire-ground contact area of ​​interest is sampled in a circular manner, and the grayscale value of each sampled column of pixels is taken as a pixel sample;

[0022] (2) Convert the sampled pixel samples into feature vectors;

[0023] (3) Discriminant model classification: Use the discriminant model to classify pixel samples into contact and separation categories based on the feature vector, and update the approximation interval boundary coordinates of each round based on the classification results;

[0024] (4) Repeat steps (1) to (3) until the approximation interval length is 1; the interval [φ2, φ4] at the end of the cycle is the tire-ground contact interval.

[0025] Furthermore, a binary method is used to perform cyclic column sampling in the binary image of the tire-ground contact area of ​​interest, specifically including:

[0026] Assume that the number of sampling rounds is n. When n=1, calculate the column coordinates of the three pixel samples in the first round of sampling in the binary image, and then calculate the two approximate intervals of sampling (φ1, φ2] and (φ3, φ4];

[0027] Among them, the column coordinates of the three pixel samples in the first round of sampling in the binary image are expressed as:

[0028]

[0029]

[0030]

[0031] Where W BW is the width of the binary image, [] represents Gaussian rounding;

[0032] The two approximation intervals (φ1, φ2] and (φ3, φ4] of the first round of sampling are expressed as:

[0033]

[0034]

[0035]

[0036]

[0037] When n≥2, that is, starting from the second round of sampling, two pixel samples are sampled in each round. First, the column coordinates of the selected pixel samples in the binary image of the tire-ground contact area of ​​interest are calculated. Then calculate the coordinates of the approximation interval boundary of the current round;

[0038] Among them, the column coordinates of the nth round of sampling The calculation formula is as follows:

[0039]

[0040]

[0041] Among them, φ m (n-1) represents the mth boundary coordinate of the n-1th round of approximation interval, and 1≤φ1<φ2≤φ3<φ4≤W BW ;

[0042] The pixel samples obtained by the n-th round of sampling are converted into feature vectors, and the pixel samples obtained by the n-th round of sampling are classified according to the feature vectors; the coordinates of the approximate interval boundary of the n-th round are updated according to the classification results. The specific updating method is to use the pixel sample column coordinates of the n-th round Replace the approximation interval boundary coordinates φ(n-1) of the (n-1)th round with the same discrimination type; the approximation interval boundary that is not replaced in the nth round inherits its coordinates in the (n-1)th round.

[0043] Furthermore, in step 2, the sub-step of extracting the tire-rim boundary line using a computer vision algorithm is as follows:

[0044] (1) using an image resizing operation to proportionally resize the image size of the cropped rim regions of interest so that all rim regions of interest have the same image height;

[0045] (2) Increase contrast and suppress noise through pixel-adaptive Wiener filter;

[0046] (3) Using local adaptive thresholding technology to convert the enhanced image into a binary image;

[0047] (4) Morphological operations are used to modify the region shape in the binary image to eliminate the interference caused by stains, cracks and unevenness on the rim surface;

[0048] (5) Extract the edge of the largest area in the binary image of the rim area of ​​interest and fit it with the equation of a circle. The fitted circle is the tire-rim boundary line.

[0049] Furthermore, in step 2, extracting the tire model mark on the tire sidewall using a computer vision algorithm specifically includes:

[0050] (1) The annular tire sidewall text interest region is converted into a rectangular region through polar coordinate transformation, so that the text originally arranged in an arc is arranged in a straight line;

[0051] (2) Sharpen the image and adjust the contrast to make the text more prominent;

[0052] (3) Obtain the tire model mark in the tire sidewall text interest area after processing using the OCR model.

[0053] Furthermore, the step three specifically includes the following sub-steps:

[0054] S3.1: Query the national standard according to the tire model mark and obtain the corresponding wheel rim height G s ;

[0055] S3.2: Calculate the tire-ground contact area, a, according to the following formula:

[0056]

[0057] Where l and d are the length of the tire-ground contact line and the diameter of the tire-rim boundary line, respectively, in pixels; D s and B s are the nominal rim diameter and nominal section width identified from the text and symbols on the tire sidewall, G s The wheel rim height is obtained by looking up the text and symbols on the tire sidewall; D s 、B s and G s Unit unification;

[0058] S3.3: Calculate the axle weight AW of the axle on which the tire is located according to the following formula: k :

[0059]

[0060] Among them, n t is the number of tires on the axle; ξ is the pressure correction factor, which is determined by the tire structure type; g is the local acceleration of gravity; p in Inflate the tire to the pressure identified by the text and symbols on the tire sidewall.

[0061] S3.4: Repeat steps S3.1 to S3.3 to obtain the axle weight of each axle of the vehicle to be tested. Sum all axle weights to obtain the gross vehicle weight (GVW):

[0062]

[0063] Where K is the number of axles of the vehicle to be tested.

[0064] The beneficial effects of the present invention are as follows:

[0065] (1) The present invention uses the YOLO network to convert images collected by non-contact highway dynamic weighing into multiple regions of interest. This technical approach effectively filters out a large number of redundant pixels. Compared with the traditional method of directly extracting information from the entire image, the present invention has significant improvements in computational efficiency and accuracy under complex lighting conditions.

[0066] (2) In terms of the key technical issue of identifying the tire-ground contact zone, the edge boundary addressing method provided by the present invention has higher computational efficiency than the Hough transform technology in traditional image processing technology because it does not require "cumulative voting" in the parameter space for every edge pixel in the image;

[0067] (3) The load identification method provided by the present invention facilitates real-time data processing on edge devices, which can help highway overload law enforcement to achieve rapid response. At the same time, data is processed on the edge device, which reduces the amount of data upload and download and improves data security. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a main flow chart of a non-contact highway dynamic weighing method for multiple interest areas according to an embodiment of the present invention;

[0069] Figure 2 This is the processing result of step 1 of the embodiment of the present invention;

[0070] Figure 3 A schematic diagram of the process of generating the first three rounds of pixel samples using the edge boundary addressing method according to an embodiment of the present invention;

[0071] Figure 4 This is the processing result of step 2-1 of the embodiment of the present invention. DETAILED DESCRIPTION

[0072] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments, and the purpose and effects of the present invention will become more apparent. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0073] like Figure 1 As shown, a non-contact road load identification method for multiple regions of interest according to an embodiment of the present invention specifically includes the following steps:

[0074] Step 1: Locate the car tires in the camera image and generate multiple regions of interest based on them.

[0075] As one implementation method, step 1 includes the following sub-steps:

[0076] Step S1-1: Use the YOLO network to locate all vehicle tires in the image captured during the non-contact dynamic weighing process, and record each vehicle tire as an initial bounding box and corresponding class probability. The YOLO network here is preferably YOLOv5s.

[0077] Step S1-2: These initial bounding boxes are cleaned according to the corresponding category probabilities and positions on the image to remove the bounding boxes that cannot be used to generate the region of interest.

[0078] Preferably, data cleaning can be achieved by filtering bounding boxes that meet any of the following conditions: (i) the category probability is less than 0.80; (ii) the distance to the image boundary is less than 10 pixels.

[0079] All filtered bounding boxes in the t-th image captured by the non-contact dynamic weighing system can be obtained by using the matrix A t To express:

[0080] A t =[A t1 ,A t2 ,A t3 ,…,A tu ](1)

[0081]

[0082] Where u = 1, 2, ..., n, n is the total number of bounding boxes after filtering in the t-th image; class u is the category number of the u-th filtered bounding box, xc u ,yc u ,w u ,h u are the center abscissa, center ordinate, pixel width, and pixel height of the u-th filtered bounding box in the image; W and H are the pixel width and height of the entire image, respectively.

[0083] Steps 1-3: The cleaned bounding box is converted into the bounding boxes of at least three regions of interest through different geometric transformation processes: tire-ground contact region of interest, rim region of interest, and tire sidewall text region of interest.

[0084] The geometric transformation process here can be expressed as:

[0085]

[0086] Among them, the matrix B t Represents the bounding box of the region of interest, S 4×4 is the coefficient matrix of the geometric transformation;

[0087] When generating the tire-ground contact region of interest:

[0088]

[0089] When generating the rim region of interest:

[0090] S 4×4 =diag(1,1,s r ,s r )(5)

[0091] When generating tire sidewall text ROI:

[0092] S 4×4 =diag(1,1,s t ,s t )(6)

[0093] Preferably, the five unknown coefficients of formulas (4) to (6) in this embodiment are: h , s w , s h , s r , s t The value is determined according to the following rules:

[0094]

[0095]

[0096]

[0097] Among them, s r =0.56~0.64,Δ u / h u =0.02.

[0098] In order to make step 1 of the present invention easier to understand, this embodiment selects an image taken by a non-contact dynamic weighing system and takes s r =0.60 and steps 1-1 to 1-3 are performed on it. In this embodiment, the parameter b of the geometric transformation coefficient matrix h , s w , s h , s t are equal to 0.41, 0.80, 0.22, and 0.90 respectively, and the image processing results of steps 1-1 to 1-3 are as follows Figure 2 shown.

[0099] Step 2: extracting the tire-ground contact line in the tire-ground contact region of interest, the tire-rim boundary line in the rim region of interest, and the tire model mark in the tire sidewall text region of interest in parallel using a computer vision algorithm;

[0100] Step 2-1. The computer vision algorithm for extracting the tire-ground contact line in the tire-ground contact region of interest includes five steps: image resizing, image enhancement, local adaptive threshold segmentation, morphological operations, and edge boundary addressing.

[0101] Step 2-1-1, using an image resizing operation to proportionally resize the image size of the cropped tire-ground contact interest regions so that all tire-ground contact interest regions have the same image height;

[0102] Step 2-1-2: Increase contrast and suppress noise through pixel-adaptive Wiener filter;

[0103] Step 2-1-3, convert the enhanced image into a binary image using local adaptive thresholding technology;

[0104] Step 2-1-4: Use morphological operations to modify the shape of the region in the binary image to eliminate the interference caused by stains, cracks and concentric stripes on the tire surface; preferably, the morphological operations include expansion and hole filling.

[0105] Step 2-1-5: Identify the tire-ground contact interval in the morphologically processed binary image using the edge boundary addressing method; obtain the contact edge by smoothing the tire tread pattern on the outer edge of the tire within the contact interval; and fit the contact edge into a straight line segment through a straight line fitting operation, which serves as the extracted tire-ground contact line.

[0106] In most cases, there is no obvious dividing point (or corner point) between the edge where the tire contacts the ground and the edge where the tire separates from the ground. At the same time, the light in the part of the tire close to the ground is usually very uneven, and there are various tire patterns on the outer edge of the tire. Therefore, it is difficult to rely on existing machine vision technology to efficiently and accurately extract the tire-ground contact line from the image, and the extraction of the tire-ground contact line often relies on manual labor. In order to meet the actual needs of overload law enforcement, real-time identification of vehicle weight is achieved. An embodiment of the present invention provides an edge boundary addressing method, which can be used to identify the tire-ground contact interval by relying on only a small number of pixels in the image. Therefore, this method can significantly reduce the recognition delay and the requirements for equipment computing power. Moreover, compared with traditional algorithms such as Hough transform, the method provided by the present invention is not easily affected by tire patterns and uneven light.

[0107] Furthermore, the above-mentioned edge boundary addressing method provided by the present invention can be divided into three steps: pixel sampling, feature vector calculation and discriminant model classification; the above-mentioned pixel sampling is to use a dichotomy method to perform cyclic column sampling in the binary image of the tire-ground contact area of ​​interest, and the grayscale value of each sampled column of pixels is used as a pixel sample; the above-mentioned feature vector calculation is to convert the sampled pixel samples into feature vectors according to a predetermined formula; the above-mentioned discriminant model classification is to use the discriminant model to classify the pixel samples according to the feature vector, and the classification categories here are "contact" and "separation".

[0108] The edge boundary addressing method provided by the present invention may use row sampling in addition to column sampling when performing pixel sampling.

[0109] Specifically, the column coordinates of the three columns of pixels sampled in the first round in the binary image of the tire-ground contact area of ​​interest are They are

[0110]

[0111]

[0112]

[0113] Where W BW is the width of the binary image, and [] represents Gaussian rounding. Through the first round of sampling, we can obtain two approximation intervals (φ1, φ2] and (φ3, φ4].

[0114]

[0115]

[0116]

[0117]

[0118] The column coordinates of the selected column in the binary image of the tire-ground contact area of ​​interest in the nth round of sampling (n≥2) after the second round of sampling for

[0119]

[0120]

[0121] The pixel samples obtained by the n-th round of sampling are converted into feature vectors, and the pixel samples obtained by the n-th round of sampling are classified according to the feature vectors; the coordinates of the approximate interval boundary of the n-th round are updated according to the classification results. The specific updating method is to use the pixel sample column coordinates of the n-th round Replace the approximation interval boundary coordinates φ(n-1) of the (n-1)th round with the same discrimination type; the approximation interval boundary that is not replaced in the nth round inherits its coordinates in the (n-1)th round.

[0122] For example, when The sample type at location φ2(n-1) is the same as the sample type at location φ2(n-1) and When the sample type at position φ is the same as the sample type at position φ3(n-1)

[0123] φ1(n)=φ1(n-1)(19)

[0124]

[0125]

[0126] φ4(n)=φ4(n-1)(22)

[0127] The three steps of pixel sampling, feature vector calculation, and discriminant model classification are continuously repeated until the length of the approximation interval reaches 1. The interval [φ2, φ4] at the end of the cycle is the tire-ground contact interval.

[0128] In order to make the above-mentioned edge boundary addressing method provided by the present invention easier to understand, this embodiment provides a schematic diagram of the process of generating the first three rounds of pixel samples using the edge boundary addressing method ( Figure 3 ).

[0129] Preferably, in this embodiment, the above-mentioned feature vector calculation uses the following predetermined formula to convert pixel samples into three feature vectors F1, F2 and F3:

[0130]

[0131] F2(i)=H BW -min(j), when I(i,j)=0 (24)

[0132]

[0133] Where i = 1, 2, ..., W BW ,j=1,2,…,H BW . H BW is the height of the binary image, W BW is the width. F1(i), F2(i), F3(i) are the i-th elements of the feature vectors F1, F2, and F3, and I(i, j) is the grayscale value of the pixel at the i-th column and j-th row in the binary image. I(i, j) = 0 represents the pixel is black, and I(i, j) = 1 represents the pixel is white.

[0134] Preferably, the above-mentioned discriminant model classification in this embodiment uses an artificial neural network (ANN) model including two hidden layers, each hidden layer is composed of 10 neurons, and the neurons use linear rectification equations as activation functions.

[0135] In order to make step 2-1 of the present invention easier to understand, this embodiment performs steps 2-1-1 to 2-1-5 on a tire-ground contact area of ​​interest generated in step 1 according to the above preferred solution. The final image processing results of each step are shown in FIG. Figure 4 .

[0136] Step 2-2, the computer vision algorithm for extracting the tire-rim boundary line in the rim region of interest also includes five steps: image resizing, image enhancement, local adaptive threshold segmentation, morphological operation and tire-rim boundary line fitting.

[0137] Steps 2-2-1 to 2-2-4 are the same as steps 2-1-1 to 2-1-4;

[0138] Step 2-2-5: Fitting the tire-rim boundary line. Extract the edge of the largest area in the binary image of the rim region of interest and fit it with the equation of a circle. The fitted circle is the tire-rim boundary line.

[0139] Step 2-3: The computer vision algorithm for extracting the sidewall text in the tire sidewall text interest region includes three steps: polar coordinate transformation, image enhancement, and optical character recognition (OCR).

[0140] Step 2-3-1: Convert the annular tire sidewall text interest area into a rectangular area through polar coordinate transformation, so that the text originally arranged in an arc becomes arranged in a straight line;

[0141] Step 2-3-2: Sharpen the image and adjust the contrast to make the text more prominent;

[0142] Step 2-3-3: Use the OCR model to obtain the tire model mark from the tire sidewall text interest area processed by steps 2-3-1 and 2-3-2, and thereby obtain the tire inherent parameters including nominal section width, structure code, rim nominal diameter, and inflation pressure.

[0143] Step 3: Obtain the tire's inherent parameters based on the tire model identification identified in step 2; Calculate the axle weight of each tire's axle by combining the tire's inherent parameters and the tire-ground contact line, and sum all axle weights to obtain the vehicle's gross weight. The specific steps are as follows:

[0144] Step 3-1: According to the national standard "Standards for Truck and Passenger Vehicle Rollers (GB / T 31961-2015)", obtain the corresponding wheel flange height G from the tire model mark. s .

[0145] Step 3-2: The wheel rim height G obtained in step 3-1 s , the length l of the tire-ground contact line extracted in step 2-1, the diameter d of the tire-rim boundary line extracted in step 2-2, and the nominal rim diameter D obtained in step 2-3 s and nominal section width B s , calculate the tire-ground contact area a according to formula (26):

[0146]

[0147] Where, l and d are in pixels; D s , B s and G s Any length unit including inches, millimeters, and centimeters can be used, but the units must be consistent.

[0148] Step 3-3: Based on the tire-ground contact area a obtained in step 3-2 and the tire inflation pressure p obtained in step 2-3 in Calculate the axle weight AW of the axle where the tire is located according to formula (27) k :

[0149]

[0150] Where n t is the number of tires on the axle; ξ is the pressure correction coefficient, which is determined by the tire structure type; g is the local acceleration of gravity. in Inflate the tire to the pressure identified by the text and symbols on the tire sidewall.

[0151] Step 3-4: Repeat steps 1-1 to 3-3 to obtain the axle weight of each axle of the vehicle to be tested, and sum all axle weights to obtain the gross vehicle weight (GVW):

[0152]

[0153] Where K is the number of axles of the vehicle to be tested.

[0154] In particular, all of the aforementioned algorithms, including image acquisition, region of interest generation, tire parameter identification, and axle and vehicle weight estimation, can be deployed and run on a Jetson TX2 edge device. Deploying the technical solution described in this invention on edge devices can better enable real-time highway weigh-in-motion.

[0155] Those skilled in the art will understand that the foregoing descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art will still be able to modify the technical solutions described in the foregoing examples or substitute equivalents for some of the technical features therein. Any modifications, equivalent substitutions, etc. made within the spirit and principles of the invention shall be included within the scope of protection of the invention.

Claims

1. A non-contact road load identification method for multiple regions of interest, characterized in that: The method comprises the following steps: Step 1: Locate the car tire in the camera image and generate multiple regions of interest based on it; Step 2: Extract the tire-ground contact line, tire-rim boundary line, and tire model mark on the tire sidewall from multiple regions of interest using computer vision algorithms; Step 3: Obtaining inherent parameters of the tire based on the tire model identification identified in step 2 by looking up the table; calculating the axle weight of each axle on which the tire is located by combining the inherent parameters of the tire and the tire-ground contact line, and summing all axle weights to obtain the gross vehicle weight; In step 2, the sub-steps of extracting the tire-ground contact line using a computer vision algorithm are as follows: (1) using an image resizing operation to proportionally resize the image size of the cropped tire-ground contact interest regions so that all tire-ground contact interest regions have the same image height; (2) Increase contrast and suppress noise through pixel-adaptive Wiener filter; (3) Using local adaptive thresholding technology to convert the enhanced image into a binary image; (4) Morphological operations are used to modify the shape of the region in the binary image to eliminate the interference caused by stains, cracks and concentric stripes on the tire surface; (5) Identify the tire-ground contact interval in the morphologically processed binary image by using an edge boundary addressing method; obtain the contact edge by smoothing the tire pattern on the outer edge of the tire in the tire-ground contact interval; and fit the contact edge to a straight line segment through a straight line fitting operation, and use this line segment as the extracted tire-ground contact line; The edge boundary addressing method includes: (1) Pixel sampling: a binary image of the tire-ground contact area of ​​interest is sampled in a circular manner, and the grayscale value of each sampled column of pixels is taken as a pixel sample; (2) Convert the sampled pixel samples into feature vectors; (3) Discriminant model classification: Use the discriminant model to classify pixel samples into contact and separation categories based on the feature vector, and update the approximation interval boundary coordinates of each round based on the classification results; (4) Repeat steps (1) to (3) until the approximation interval length is 1; the interval [φ2, φ4] at the end of the cycle is the tire-ground contact interval.

2. The non-contact road load identification method for multiple regions of interest according to claim 1, characterized in that: The step 1 includes the following sub-steps: S1.1: Use the YOLO network to locate all car tires in the images captured during the non-contact dynamic weighing process, and record each car tire as an initial bounding box and corresponding class probability; S1.2: These initial bounding boxes are cleaned according to their corresponding category probabilities and positions on the image to remove bounding boxes that cannot be used to generate the region of interest; S1.3: The cleaned bounding box is converted into the bounding boxes of three regions of interest through different geometric transformation processes: tire-ground contact region of interest, rim region of interest, and tire sidewall text region of interest.

3. The non-contact road load identification method for multiple regions of interest according to claim 1, characterized in that: A binary method is used to perform cyclic column sampling in the binary image of the tire-ground contact area of ​​interest, specifically including: Assume that the number of sampling rounds is n. When n=1, calculate the column coordinates of the three pixel samples in the first round of sampling in the binary image, and then calculate the two approximate intervals of sampling (φ1, φ2] and (φ3, φ4]; Among them, the column coordinates of the three pixel samples in the first round of sampling in the binary image are expressed as: Where W BW is the width of the binary image, [] represents Gaussian rounding; The two approximation intervals (φ1, φ2] and (φ3, φ4] of the first round of sampling are expressed as: When n≥2, that is, starting from the second round of sampling, two pixel samples are sampled in each round. First, the column coordinates of the selected pixel samples in the binary image of the tire-ground contact area of ​​interest are calculated. Then calculate the coordinates of the approximation interval boundary of the current round; Among them, the column coordinates of the nth round of sampling The calculation formula is as follows: Among them, φ m (n-1) represents the mth boundary coordinate of the n-1th round of approximation interval, and 1≤φ1<φ2≤φ3<φ4≤W BW ; The pixel samples obtained by the n-th round of sampling are converted into feature vectors, and the pixel samples obtained by the n-th round of sampling are classified according to the feature vectors; the coordinates of the approximate interval boundary of the n-th round are updated according to the classification results. The specific updating method is to use the pixel sample column coordinates of the n-th round Replace the approximation interval boundary coordinates φ(n-1) of the (n-1)th round with the same discrimination type; the approximation interval boundary that is not replaced in the nth round inherits its coordinates in the (n-1)th round.

4. The non-contact road load identification method for multiple regions of interest according to claim 1, characterized in that: In step 2, the sub-step of extracting the tire-rim boundary line using a computer vision algorithm is as follows: (1) using an image resizing operation to proportionally resize the image size of the cropped rim regions of interest so that all rim regions of interest have the same image height; (2) Increase contrast and suppress noise through pixel-adaptive Wiener filter; (3) Using local adaptive thresholding technology to convert the enhanced image into a binary image; (4) Morphological operations are used to modify the region shape in the binary image to eliminate the interference caused by stains, cracks and unevenness on the rim surface; (5) Extract the edge of the largest area in the binary image of the rim area of ​​interest and fit it with the equation of a circle. The fitted circle is the tire-rim boundary line.

5. The non-contact road load identification method for multiple regions of interest according to claim 1, characterized in that: In the second step, extracting the tire model mark on the tire sidewall using a computer vision algorithm specifically includes: (1) The annular tire sidewall text interest region is converted into a rectangular region through polar coordinate transformation, so that the text originally arranged in an arc is arranged in a straight line; (2) Sharpen the image and adjust the contrast to make the text more prominent; (3) Obtain the tire model mark in the tire sidewall text interest area after processing using the OCR model.

6. The non-contact road load identification method for multiple regions of interest according to claim 1, characterized in that: The step three specifically includes the following sub-steps: S3.1: Query the national standard according to the tire model mark and obtain the corresponding wheel rim height G s ; S3.2: Calculate the tire-ground contact area, a, according to the following formula: Where l and d are the length of the tire-ground contact line and the diameter of the tire-rim boundary line, respectively, in pixels; D s and B s are the nominal rim diameter and nominal section width identified from the text and symbols on the tire sidewall, G s The wheel rim height is obtained by looking up the text and symbols on the tire sidewall; D s 、B s and G s Unit unification; S3.3: Calculate the axle weight AW of the axle on which the tire is located according to the following formula: k : Among them, n t is the number of tires on the axle; ξ is the pressure correction factor, which is determined by the tire structure type; g is the local acceleration of gravity; p in Inflate the tire to the pressure identified by the text and symbols on the tire sidewall; S3.4: Repeat steps S3.1 to S3.3 to obtain the axle weight of each axle of the vehicle to be tested. Sum all axle weights to obtain the gross vehicle weight (GVW): Where K is the number of axles of the vehicle to be tested.

Images