Vehicle three-dimensional over-limit intelligent detection method and system

Abstract

The invention belongs to the technical field of traffic monitoring, and discloses a vehicle three-dimensional overrun intelligent detection method and system, and the method comprises the steps: obtaining the image information of a road region where a target vehicle is located, and recognizing a static road reference object with a standard physical size in the image information; constructing a scale field model of the road area based on the pixel size of the static road reference object in the image information and the corresponding standard physical size; extracting a feature point coordinate of the target vehicle in the image information, and determining a direction angle of the target vehicle; calculating the three-dimensional physical size of the target vehicle according to the scale field model, the feature point coordinates and the direction angle; acquiring a statistical size parameter corresponding to the vehicle type of the target vehicle, and comparing the three-dimensional physical size with the statistical size parameter to determine an over-limit state of the target vehicle; therefore, the detection precision and reliability can be improved, the environmental adaptability is enhanced, the misjudgment rate is reduced, and a calibration object does not need to be preset.

Description

Technical Field

[0001] This application relates to the field of traffic monitoring technology, and more specifically, to a three-dimensional intelligent detection method and system for vehicles exceeding limits. Background Technology

[0002] Oversized and overweight transport is a serious problem that endangers road safety and damages infrastructure. Three-dimensional oversized transport refers to exceeding limits in length, width, and height. Current three-dimensional oversized transport detection technologies mainly face the following technical bottlenecks: While fixed-calibration video measurement technology can achieve a certain degree of dimensional measurement, this method has significant limitations. It relies on calibration patterns specifically set on the road surface within the monitored area for camera calibration, and these pre-set calibration objects are easily affected by dirt, obstruction, and environmental changes. More seriously, any displacement or change in angle of the camera necessitates recalibration, resulting in high system maintenance costs.

[0003] While deep learning regression-based video measurement techniques avoid the need for pre-defined calibrations, their technical limitations are equally prominent. This method heavily relies on the quality and quantity of training data, exhibiting poor generalization ability in practical applications. Due to the "black box" nature of deep learning models, their decision-making process lacks interpretability, making it difficult for measurement results to provide convincing evidence in law enforcement.

[0004] While methods based on a single natural reference attempt to utilize existing road elements such as lane width as a reference, they are prone to failure when the reference is missing, blurred, or occluded. More importantly, these methods fail to adequately account for the spatial variations in perspective distortion, resulting in measurement accuracy that is insufficient for practical needs.

[0005] Current technologies generally suffer from the following core defects: incomplete detection dimensions, particularly poor accuracy and reliability in detecting extra-long and extra-wide vehicles; weak environmental adaptability, easily affected by changes in lighting, inclement weather, and obstructions; high system fragility, over-reliance on preset, fixed, and single measurement benchmarks, which are prone to failure in real-world road environments; persistently high false positive rates, lacking intelligent recognition of the correlation between vehicle type and size, leading to compliant large vehicles frequently being falsely reported as oversized; limited functionality, with most systems only providing simple "yes / no" judgments and unable to output detailed reports with quantifiable evidence; and poor deployment flexibility, with stringent requirements for camera installation parameters and difficulty in adapting to dome cameras with variable viewing angles.

[0006] Of particular concern is the general lack of effective mechanisms for maintaining measurement stability under complex interference conditions, such as dynamic occlusion, drastic changes in lighting, and severe weather. Furthermore, these technologies fail to fully utilize the diverse existing road marking resources for intelligent complementary measurements, resulting in significant deficiencies in the system's reliability and adaptability in practical applications.

[0007] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0008] The purpose of this application is to provide a three-dimensional intelligent detection method and system for vehicles that can improve detection accuracy and reliability, enhance environmental adaptability, reduce the false judgment rate, and eliminate the need for pre-set calibration objects.

[0009] Firstly, this application provides a three-dimensional intelligent detection method for vehicles exceeding weight limits, comprising the following steps: A1. Acquire image information of the road area where the target vehicle is located, and identify static road reference objects with standard physical dimensions in the image information; A2. Based on the pixel size of the static road reference in the image information and its corresponding standard physical size, a scale field model of the road area is constructed; the scale field model reflects the mapping relationship between the pixel size and physical size of each pixel on the road surface in the road area. A3. Extract the coordinates of feature points of the target vehicle in the image information, and determine the direction angle of the target vehicle; A4. Calculate the three-dimensional physical dimensions of the target vehicle based on the scale field model, the coordinates of the feature points, and the direction angle; the three-dimensional physical dimensions include height, length, and width; A5. Obtain the statistical dimension parameters corresponding to the vehicle model of the target vehicle, and compare the three-dimensional physical dimensions with the statistical dimension parameters to determine the over-limit status of the target vehicle.

[0010] Preferably, after step A5, the following step is also included: A6. Perform graded alarm operations based on the over-limit status of the target vehicle.

[0011] Secondly, this application provides a three-dimensional intelligent detection system for overloaded vehicles, including a camera and a control system; The camera is used to collect image information of the road area where the target vehicle is located; The control system is used to identify static road references with standard physical dimensions in the image information; based on the pixel dimensions of the static road references in the image information and their corresponding standard physical dimensions, a scale field model of the road area is constructed; the scale field model reflects the mapping relationship between the pixel dimensions and physical dimensions of each pixel on the road surface in the road area; the feature point coordinates of the target vehicle in the image information are extracted, and the orientation angle of the target vehicle is determined; according to the scale field model, the feature point coordinates, and the orientation angle, the three-dimensional physical dimensions of the target vehicle are calculated; the three-dimensional physical dimensions include height, length, and width; statistical dimension parameters corresponding to the vehicle type of the target vehicle are obtained, and the three-dimensional physical dimensions are compared with the statistical dimension parameters to determine the over-limit status of the target vehicle.

[0012] Beneficial Effects: This application provides a three-dimensional intelligent detection method and system for vehicles exceeding limits. It acquires image information of the road area where the target vehicle is located and identifies static road reference objects with standard physical dimensions in the image information. Based on the pixel dimensions of the static road reference objects in the image information and their corresponding standard physical dimensions, a scale field model of the road area is constructed. Feature point coordinates of the target vehicle in the image information are extracted, and the orientation angle of the target vehicle is determined. The three-dimensional physical dimensions of the target vehicle are calculated according to the scale field model, the feature point coordinates, and the orientation angle. The three-dimensional physical dimensions include height, length, and width. Statistical dimension parameters corresponding to the vehicle model are obtained, and the three-dimensional physical dimensions are compared with the statistical dimension parameters to determine the over-limit status of the target vehicle. This improves detection accuracy and reliability, enhances environmental adaptability, reduces the false judgment rate, and eliminates the need for pre-set calibration objects. Attached Figure Description

[0013] Figure 1 A flowchart of a three-dimensional intelligent detection method for vehicles exceeding limits provided in this application.

[0014] Figure 2 This is a schematic diagram of a vehicle three-dimensional over-limit intelligent detection system provided in this application.

[0015] Figure 3 This is a schematic diagram of the bounding box of the target vehicle.

[0016] Labeling explanation: 1. Camera; 2. Control system. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0018] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0019] Please refer to Figure 1 A vehicle three-dimensional over-limit intelligent detection method according to some embodiments of this application includes the following steps: A1. Acquire image information of the road area where the target vehicle is located, and identify static road reference objects with standard physical dimensions in the image information; A2. Based on the pixel size of the static road reference in the image information and its corresponding standard physical size, a scale field model of the road area is constructed; the scale field model reflects the mapping relationship between the pixel size and physical size of each pixel on the road surface in the road area. A3. Extract the coordinates of feature points of the target vehicle in the image information, and determine the direction angle of the target vehicle; A4. Calculate the three-dimensional physical dimensions of the target vehicle based on the scale field model, the coordinates of the feature points, and the direction angle; the three-dimensional physical dimensions include height, length, and width; A5. Obtain the statistical dimension parameters corresponding to the vehicle model of the target vehicle, and compare the three-dimensional physical dimensions with the statistical dimension parameters to determine the over-limit status of the target vehicle.

[0020] This application proposes a three-dimensional intelligent detection method for vehicles exceeding limits. This method aims to solve the problems of existing three-dimensional over-limit detection technologies, such as reliance on preset calibration objects, poor environmental robustness, low detection accuracy, high false judgment rate, limited functionality, and insufficient deployment flexibility.

[0021] Specifically, in step A1, image information can be acquired through various methods such as vehicle-mounted cameras and roadside surveillance cameras. Identifying static road reference objects can employ image processing and pattern recognition technologies. For example, algorithms such as edge detection, shape matching, and color segmentation can be used to detect traffic markings, traffic signs, or fixed road facilities in the image. For instance, a database containing features of various static road reference objects can be pre-set, and static road reference objects can be identified by comparing the visual features in the image with the features in the database.

[0022] Further, in step A2, the scale field model reflects the mapping relationship between the pixel size and physical size of each pixel on the road surface in the road area. Various methods can be used to construct the scale field model. For example, by observing multiple static road references, obtaining their pixel size and corresponding physical size in the image, and then using mathematical methods such as interpolation and regression to establish the mapping relationship between pixel size and physical size. For instance, multiple traffic marking segments at different depths and locations in the image can be selected, their pixel width or length in the image can be measured, and combined with their national standard physical width or length, the scale factor at different locations can be calculated, thereby fitting a scale field model covering the entire road area (mainly the road surface). This scale field model can dynamically correct the influence of perspective distortion and road geometric changes on size measurement, providing spatial scale information for subsequent accurate calculation of the three-dimensional physical dimensions of vehicles.

[0023] Subsequently, in step A3, the extraction of feature point coordinates can utilize deep learning models for vehicle detection and key point recognition. For example, a convolutional neural network can be trained to automatically identify the pixel coordinates of key parts of the vehicle, such as the front, rear, roof, and wheels. The orientation angle of the target vehicle can be estimated by analyzing the vehicle's contour, the contact points between the wheels and the road surface, or the vehicle's trajectory in consecutive frames. For example, the direction of the vehicle's bottom edge line can be detected, and the vehicle's movement direction on the ground (road surface) after scale field correction obtained from short-time multi-frame trajectory tracking can be fused to jointly estimate the vehicle's driving direction in the current local coordinate system of the road surface.

[0024] Next, in step A4, specifically, the pixel coordinates of feature points can be mapped to the physical world coordinate system using a scale field model. Then, based on the mapped 3D coordinates and the vehicle's orientation angle, the actual height, length, and width of the vehicle are obtained through geometric calculations. For example, for height measurement, the pixel coordinates of the highest point of the roof and the bottom ground contact point can be used, combined with the camera model and ground equations, to calculate the absolute height through geometric relationships. For length and width measurement, the local scale factors of the corresponding positions of the front, rear, left, and right points in the scale field can be queried, and the pixel coordinates of the key points can be transformed to the ground physical coordinate system. Then, the projected distance between the front and rear points along the vehicle's longitudinal axis is calculated as the length, and the projected distance between the left and right points in the direction perpendicular to the longitudinal axis is calculated as the width. This process effectively corrects the effects of perspective distortion and road geometry by mapping the image coordinates to physical space and projecting them along the correct local direction, thus obtaining the orthographic projection size of the vehicle in the real world.

[0025] Finally, in step A5, obtaining the statistical dimensional parameters can be achieved by querying a pre-established vehicle size statistics database, which stores the standard size ranges or statistical distributions of various vehicle types. The comparison process compares the measured three-dimensional physical dimensions with the mean and standard deviation of the corresponding vehicle in the database to determine whether the measured dimensions exceed a preset confidence interval, thereby determining whether the vehicle exceeds the limits. For example, a threshold can be set; when the measured dimensions exceed this threshold, it is determined to be out of limit.

[0026] The aforementioned technical solution utilizes existing static road references to construct a dynamic scale model (i.e., a scale field model), combines vehicle feature points and orientation information to accurately calculate three-dimensional dimensions, and compares these with vehicle model statistics. This achieves highly robust over-limit detection that requires no pre-calibration and adapts to environmental changes. Compared to traditional video measurement techniques that rely on pre-calibrated patterns, this application eliminates the need for pre-calibrated objects on the road surface (referring to calibration objects specifically used for scale detection), avoiding the problems of calibration objects being susceptible to dirt, occlusion, and environmental changes. Furthermore, it eliminates the need for recalibration after camera displacement, significantly reducing maintenance costs. Compared to video measurement techniques based on deep learning regression, this application provides quantifiable physical size evidence through the construction of an interpretable scale field model and geometric solutions, enhancing the credibility of the detection results rather than operating as a "black box." In addition, this application utilizes multi-source static road references to construct a dense scale field, effectively solving the problem of failure when a single natural reference is missing, blurred, or occluded, and fully considers the spatial variations of perspective distortion, significantly improving measurement accuracy. By introducing statistical dimensional parameters of vehicle models for comparison, this application can distinguish between compliant large vehicles and oversized vehicles, effectively reducing the false positive rate and solving the problem of traditional methods lacking understanding of the correlation between vehicle type and size. Overall, the method of this application demonstrates significant advantages in terms of detection dimensions, environmental robustness, system vulnerability, false positive rate, functional diversity, and deployment flexibility, providing a more advanced and reliable solution for intelligent 3D oversized vehicle detection.

[0027] In some preferred embodiments, after step A5, the following step is further included: A6. Perform graded alarm operations based on the over-limit status of the target vehicle.

[0028] Among them, graded alarm operation refers to the system taking different levels and content of response measures based on the severity and potential risk of the detected over-limit situation of the target vehicle. This operation aims to avoid a one-size-fits-all alarm approach and instead provide more accurate and effective feedback based on the actual situation. Specifically, graded alarm operation can be implemented in several ways. For example, it can be based on preset thresholds, that is, comparing the over-limit value of any one or more of the target vehicle's height, length, and width with preset thresholds of different levels (such as statistical thresholds, legal limits, safety thresholds, etc.) to trigger different levels of alarms. Another implementation method is to use a risk assessment model. This model not only considers the absolute value of the over-limit, but also combines other contextual information, such as road type, weather conditions, vehicle load distribution, etc., to conduct a comprehensive risk assessment of the over-limit event and dynamically adjust the alarm level and response strategy based on the assessment results. In addition, machine learning methods can be used to train models to identify the correlation between different over-limit patterns and corresponding risk levels, thereby achieving intelligent graded alarms.

[0029] This application's solution introduces a tiered alarm mechanism, making the response mechanism to vehicle over-limit situations more refined and intelligent. After calculating the target vehicle's three-dimensional physical dimensions (including height, length, and width) and comparing them with corresponding statistical dimensional parameters to determine its over-limit status, the system no longer simply provides a "yes / no" judgment. Instead, it executes differentiated alarm actions based on the degree and nature of the over-limit situation. For example, for minor over-limit situations (such as exceeding the statistical range but not reaching the legal limit), the system can simply record the information as a basis for subsequent analysis or evidence collection, avoiding unnecessary interference. For severe over-limit situations that reach or exceed the legal limit, a higher-level alarm will be triggered, such as audible and visual alarms and information push notifications, along with a detailed evidence package to ensure timely and effective intervention. Furthermore, for over-limit situations that may lead to serious safety hazards (such as combined over-limit situations involving excessive length, height, and width), the system can further assess the vehicle's anti-rollover stability, thereby issuing a high-risk warning before the potential risk occurs. This tiered processing mechanism enables the entire detection system to provide appropriate responses based on the actual risk level, significantly improving the efficiency and accuracy of road traffic safety management.

[0030] The following is a concrete example to illustrate this. After calculating the three-dimensional physical dimensions of the target vehicle (including height, length, and width) and comparing them with the statistical dimensional parameters corresponding to the vehicle model, the system determines the vehicle's over-limit status based on the comparison results and executes tiered alarm operations accordingly. For example, if at least one of the target vehicle's height, length, and width exceeds the limit, but none of the over-limit values ​​exceed the legal limits, the system will trigger a Level 1 warning, recording only image information as evidence for later review or as evidence of minor violations. If at least one of the target vehicle's height, length, and width exceeds the limit, and at least one of the over-limit values ​​exceeds the legal limits, the system will trigger a Level 2 alarm. In this case, an audible and visual alarm will be immediately triggered to alert on-site personnel, and the alarm information will be pushed to the relevant management platform, attaching an evidence package containing the location and precise dimensions of the over-limit violation, providing direct and quantifiable evidence for law enforcement. Based on this Level 2 alarm, if the system further determines that the over-limit situation may lead to serious safety risks (e.g., simultaneous occurrence of exceeding height and width limits), it will estimate the vehicle's center of gravity position and, in conjunction with road curvature information, calculate the vehicle's rollover stability coefficient. When this rollover stability coefficient is lower than a preset safety threshold, the system will trigger a high-risk warning to indicate the potential rollover risk, thereby enabling more urgent intervention measures to be taken.

[0031] Through the above technical solution, this application can provide refined, tiered alarm operations based on the over-limit status of the target vehicle, effectively solving the problems of imprecise alarm response mechanisms and the inability to provide differentiated processing based on the degree of over-limit and risk level in traditional detection methods. This tiered alarm mechanism not only avoids overreacting to minor over-limit situations and improves system operating efficiency, but also provides timely and accurate early warnings and detailed evidence for severe over-limit and high-risk scenarios, significantly improving the effectiveness of road traffic safety management and the credibility of law enforcement, especially in preventing high-risk accidents such as vehicle rollovers.

[0032] Specifically, the static road reference includes at least one of traffic markings, traffic signs, and fixed road facilities.

[0033] Traffic markings may include lane edge lines, lane dividers, guide lines, deceleration markings, and distance confirmation lines; traffic signs may include circular prohibitory signs, circular instruction signs, triangular warning signs, square directional signs, rectangular directional signs, kilometer markers, and 100-meter markers; fixed road facilities may include corrugated beam guardrails, gantries, bridge expansion joints, and median strip anti-glare panels. These static road reference structures are all set up according to relevant national standards and must be installed strictly according to their corresponding standard dimensions.

[0034] Specifically, static road references refer to objects or markers with fixed, known physical dimensions in the road environment. They play a crucial role in intelligent 3D vehicle over-limit detection methods, acting as a bridge between the physical world and the pixel world of images, providing necessary standard size information for constructing scale field models. By identifying the pixel dimensions of these references in the image and combining them with their known standard physical dimensions, the system can calculate the scale factor of different regions in the image, thereby achieving a precise mapping from pixel size to physical size.

[0035] Traffic markings, as static road references, are characterized by their wide distribution, diverse types, and clearly defined national standard dimensions. For example, the standard widths of lane edge lines and lane dividers, the specific geometric dimensions of guide lines and deceleration markings, and the fixed intervals of vehicle distance confirmation lines all provide reliable physical dimensional benchmarks. These markings are typically drawn directly on the road surface and reflect close-range scale information about the road surface.

[0036] Traffic signs are another important static road reference point, characterized by their location above or to the side of the road, high visibility, and standardized geometry and dimensions. Examples include circular prohibitory signs, circular instruction signs, triangular warning signs, square directional signs, and rectangular directional signs, all with strictly defined diameters or side lengths. Milestone markers and 100-meter markers also have standard sizes and font specifications. These signs provide dimensional information at different heights and distances.

[0037] Fixed road infrastructure serves as a supplementary static road reference, characterized by structural stability, fixed dimensions, and typically large physical size. For example, the standard corrugated beam length and post spacing of corrugated beam guardrails, the post diameter, beam height, and span of gantry cranes, the regular arrangement and spacing of bridge expansion joints, and the standard dimensions and spacing of median strip anti-glare panels can all serve as reliable dimensional references. These facilities can provide dimensional information over longer distances or for specific areas.

[0038] This application's solution significantly enhances the robustness and accuracy of the vehicle 3D over-limit intelligent detection method by introducing diverse static road references, emphasizing that they are all set according to relevant national standards and strictly adhere to corresponding standard dimensions. When constructing a scale field model of the road area, the system no longer relies on a single type of reference. When a certain type of reference (e.g., traffic markings) is difficult to accurately identify or measure due to occlusion, wear, or changes in lighting, the system can flexibly switch to or integrate other types of references (e.g., traffic signs or fixed road facilities) to obtain the required standard dimension information. This parallel detection and intelligent optimization fusion strategy using multi-source references ensures that a reliable dimension benchmark can always be found under complex and changing environmental conditions, thereby continuously and stably constructing a high-precision scale field model. Furthermore, the strict standardization requirements for reference object dimensions guarantee the accuracy of the measurement benchmark from the source, avoiding systematic errors introduced by reference object size deviations, further improving the overall detection accuracy and reliability.

[0039] In some implementations, step A2 includes: A201. Obtain an image sequence composed of multiple consecutive frames of image information; A202. For each observation point on the static road reference, obtain the standard physical size corresponding to the observation point and the pixel size of the static road reference corresponding to the observation point in the image sequence, in order to calculate the scale observation value sequence corresponding to the observation point in the image sequence; the scale observation value represents the pixel size corresponding to a unit physical length at the observation point; A203. The scale observation sequence is corrected, and valid scale observations are obtained from the corrected scale observation sequence; the correction process includes time-domain filtering, outlier handling, and missing value imputation. A204. Based on the effective scale observations of each observation point, a scale field model of the road area is constructed using the Gaussian process regression method.

[0040] In step A201, the image sequence refers to a collection of multiple image frames captured consecutively over a period of time. This can be achieved by continuously capturing images at a certain frame rate using a vehicle-mounted camera, roadside surveillance camera, or other visual sensors. This continuity provides information in the temporal dimension, which is crucial for temporal analysis and stability improvement in subsequent processing. For example, a camera can continuously capture video streams at a fixed frame rate (e.g., 30 frames / second or 60 frames / second), decode the video stream into a series of independent image frames, and then sort these images according to their timestamps to form an image sequence.

[0041] Step A202 aims to provide foundational data for constructing the scale field model. Observation points are specific locations on static road references with clearly defined physical dimensions. Standard physical dimensions are the known, fixed dimensions of these references in the real world, and the pixel dimensions of the static road references are their corresponding pixel lengths or widths in the image. Scale observations are the observed ratios of physical dimensions to pixel dimensions (i.e., scale values), reflecting the number of pixels occupied per unit physical length at a specific location in the image, and are a key factor in the conversion from pixel space to physical space. For example, for traffic markings, their edges can be identified, their pixel width in the image measured, and their corresponding physical width obtained from national standards; or, for circular traffic signs, their pixel diameter in the image detected, their physical diameter obtained from national standards, and then the scale observations calculated.

[0042] In step A203, the correction process aims to eliminate noise, errors, or incomplete data introduced during image acquisition and feature extraction, ensuring high accuracy and reliability of the scale observations used for model construction. Temporal filtering smooths short-term fluctuations in the time series, outlier handling identifies and removes data points that significantly deviate from the normal range, and missing value imputation fills in data gaps caused by occlusion or other reasons. For example, temporal filtering can use a moving average filter to smooth the scale observation sequence, or a Kalman filter can be used to dynamically estimate and correct observations through prediction and update mechanisms; outlier handling can employ statistical methods, such as the 3σ principle, marking and removing observations exceeding three standard deviations of the mean, or methods based on historical trends; missing value imputation can be achieved through temporal extrapolation, using the scale observation sequence of the observation point in previous frames for prediction, or through spatial interpolation, using the scale observations of neighboring observation points for interpolation estimation.

[0043] In step A204, Gaussian Process Regression (GPR) is a nonparametric regression method that models functions and provides estimates of prediction uncertainty. In this step, GPR is used to map discrete, calibrated effective scale observations to the entire road area, thereby constructing a continuous scale field model. This model can predict the scale value at any pixel location on the road and provide the confidence level of that prediction. For example, the pixel coordinates of each observation point can be used as input, and the corresponding effective scale observation as output. Training can be performed using a standard Gaussian Process Regression library (the specific training process is existing technology and will not be detailed here). A suitable kernel function (such as the Radial Basis Function (RBF) kernel or the Matérn kernel) can be selected to capture the spatial correlation of the scale field. Alternatively, the quality parameter of the observation points can be introduced into the GPR model as a weight; for example, the quality parameter can be inversely mapped to the observation noise variance, making high-quality observations have a greater impact on the model, thereby improving the model's accuracy and robustness.

[0044] Through the synergistic effect of the above steps, this method overcomes the limitations of single-frame images or single reference objects being susceptible to interference in complex environments. The introduction of consecutive multi-frame images enables the system to utilize information in the temporal dimension for data correction and prediction, significantly enhancing its robustness against interference such as instantaneous occlusion and illumination changes. Temporal filtering, outlier handling, and missing value imputation work together to ensure that the scale observations input to the Gaussian process regression model are stable, accurate, and complete. The Gaussian process regression method further utilizes these high-quality observations to construct a scale field model that accurately reflects the mapping relationship between pixel size and physical size in the road area, and provides uncertainty in prediction, thus providing a more reliable and accurate foundation for subsequent calculation of the target vehicle's 3D physical dimensions. This optimized scale field model construction method significantly improves the overall accuracy and environmental adaptability of the vehicle 3D over-limit intelligent detection method, solving the problem of insufficient model accuracy and robustness caused by inaccurate scale observations in complex environments in traditional methods.

[0045] Preferably, step A204 may include: Obtain the pixel coordinates and quality parameters of each observation point; the quality parameters include the confidence level, completeness, and type of the static road reference corresponding to the observation point; Calculate the fusion weight of each observation point based on the quality parameters; Based on the effective scale observations, pixel coordinates, and fusion weights of each observation point, a scale field model of the road area is constructed using the Gaussian process regression method. The Gaussian process regression method uses the Matérn 3 / 2 kernel function as the covariance function and inversely maps the fusion weights to the observation noise variance in the Gaussian process regression.

[0046] Specifically, when constructing a scale-field model of a road area, the first step is to obtain the pixel coordinates and quality parameters of each observation point. Pixel coordinates are used to accurately identify the location of the observation point in the image, forming the basis for spatial modeling. Quality parameters are used to evaluate the reliability of the scale observations provided by each observation point. Confidence reflects the clarity and reliability of the identification of the static road reference corresponding to the observation point in the image. For example, it can be determined by evaluating the edge sharpness and contrast of the reference using image processing algorithms, or by determining the classification confidence output by a deep learning model. Integrity reflects whether the static road reference corresponding to the observation point is occluded or damaged. For example, it can be evaluated by calculating the proportion of occluded pixels using image segmentation algorithms, or by detecting the degree of breakage of the reference using morphological operations. Type refers to the specific category of the static road reference to which the observation point belongs, such as traffic markings, traffic signs, or fixed road facilities. This helps in subsequent differentiated processing based on the characteristics of different types of references.

[0047] After obtaining the quality parameters, the fusion weights for each observation point need to be calculated based on these parameters. The fusion weight is a quantitative indicator representing the degree to which each observation point should contribute to the construction of the scale-field model. This weight can comprehensively consider various factors such as confidence, completeness, and type, and can be calculated, for example, through weighted summation. The fusion weight aims to enhance the influence of high-quality observation points on the model while reducing the interference from low-quality observation points (such as occluded, blurred, or unrecognized references).

[0048] Subsequently, a scale field model of the road area is constructed using Gaussian process regression, combining the effective scale observations, pixel coordinates, and calculated fusion weights of each observation point. Gaussian process regression is a non-parametric probabilistic model capable of modeling functions and providing estimates of prediction uncertainty, making it well-suited for constructing continuous scale fields. In this method, the Matérn 3 / 2 kernel function is used as the covariance function, which effectively captures spatial correlations and assumes moderate smoothness of the scale field, consistent with the scale variation characteristics of actual road areas. Furthermore, the fusion weights are inversely mapped to the observation noise variance in Gaussian process regression. This means that observation points with lower fusion weights have larger corresponding observation noise variances, thus automatically weakening their impact on the model during regression and reducing their contribution to the final scale field model, effectively suppressing noise and outliers.

[0049] This application's solution, by introducing quality parameters and a fusion weight mechanism, combined with a Gaussian process regression method with a specific configuration, enables the construction of a scale-field model for road areas to fully consider the reliability differences of different observation points. First, the data quality of each observation point is meticulously evaluated by acquiring quality parameters such as confidence, completeness, and type. Second, fusion weights are calculated based on these quality parameters, ensuring that high-quality observation points dominate the model construction, while low-quality observation points are assigned lower weights. Finally, when constructing the scale-field model using the Gaussian process regression method, the fusion weights are inversely mapped to the observation noise variance, and the Matérn 3 / 2 kernel function is used to capture spatial correlation. This approach allows the model to adaptively adjust the level of trust in different observation data, effectively suppressing noise interference caused by occlusion, blurring, or recognition uncertainty, thereby constructing a more accurate and robust scale-field model. This scale-field model provides a more accurate pixel-to-physical-size mapping, laying a solid foundation for the subsequent accurate calculation of the target vehicle's 3D physical dimensions, thus improving the accuracy and reliability of the entire intelligent 3D over-limit vehicle detection method.

[0050] The following is a concrete example to illustrate this. After obtaining the pixel coordinates of each observation point, image analysis can be performed on the static road reference object corresponding to each observation point to obtain quality parameters. For example, for confidence, image processing techniques can be used to calculate the gradient strength and sharpness of the reference object's edges, or a pre-trained deep learning model can be used to classify the detected reference objects and output their confidence scores; a higher score indicates higher confidence. For completeness, image segmentation algorithms can be used to identify the complete region of the reference object and calculate its overlap with the expected complete shape or the proportion of occlusion. For example, if the reference object is occluded by more than 50%, the completeness score is low. For type, image recognition algorithms can be used to accurately classify the reference object into "lane edge line," "circular prohibition sign," or "wave beam guardrail," etc. The following formula can be used when calculating the fusion weights: ,in, Let i be the fusion weight for the i-th observation point. Let be the confidence level of the static road reference corresponding to the i-th observation point. The integrity of the static road reference point corresponding to the i-th observation point. The type of static road reference corresponding to the i-th observation point. for The stability of historical scale observations of static road reference objects (for example, it can be calculated based on the reciprocal of the variance of historical scale observations of this type of reference object; the smaller the variance, the higher the stability). for Priority coefficients for static road reference objects (can be preset as needed). For the i-th observation point, the temporal stability factor reflects the stability of the scale observations of that observation point in the image sequence (for example, it can be calculated based on the reciprocal of the standard deviation of the scale observations of that observation point in multiple consecutive frames of images; the smaller the standard deviation, the higher the temporal stability). , , , , This is the scaling factor (which can be preset as needed). When constructing the scale field model, Gaussian process regression can be implemented using machine learning libraries such as GPy or scikit-learn in Python. The corrected effective scale observations are used as the output of the Gaussian process regression, and the pixel coordinates of the observation points are used as the input. In the model configuration, the covariance function is set to the Matérn 3 / 2 kernel function. Simultaneously, the calculated fusion weights are... The reciprocal of the value (or a preset multiple thereof) is used as the observation noise variance for each observation point, and this variance is passed to the Gaussian process regression model for training. In this way, when fitting the scale field, the model automatically reduces the influence of low-weight observation points, thus obtaining a more realistic and robust scale field model.

[0051] In some implementations, step A3 includes: A301. Extract the pixel coordinates of the feature points of the target vehicle in the current frame image information, and use them as the feature point coordinates; the feature points include the foremost point of the front of the vehicle, the last point of the rear of the vehicle, the outermost point of the left side, the outermost point of the right side, and the highest point of the roof; A302. Detect the contact points between the front and rear wheels of the target vehicle and the road surface in the current frame image information, and map the contact points to the road surface physical coordinate system based on the scale field model, and calculate the direction angle of the target vehicle based on the mapped contact points to obtain the first direction angle estimate; A303. Track the contact points between the front or rear wheels of the target vehicle and the road surface in multiple consecutive frames of image information to obtain the trajectory point sequence in the image coordinate system. Based on the scale field model, map the trajectory point sequence to the road surface physical coordinate system, and calculate the direction angle of the target vehicle based on the mapped trajectory point sequence to obtain the second direction angle estimate. A304. Combine the first azimuth angle estimate and the second azimuth angle estimate to obtain the final azimuth angle.

[0052] Step A301 aims to accurately identify and locate key geometric feature points of the target vehicle from the current frame image information. These feature points are the basis for subsequent calculations of the vehicle's three-dimensional physical dimensions and orientation angles. Extracting the pixel coordinates of these feature points can employ various image processing and computer vision techniques. For example, pre-trained deep learning models, such as object detection and keypoint detection models based on convolutional neural networks (CNNs), can be used. This model can first identify the bounding boxes of the target vehicle (e.g., the bounding box of the entire vehicle, the bounding box of the front or rear side of the vehicle, and the bounding box of the longitudinal side of the vehicle body, where the longitudinal side refers to the vertical side located between the front and rear sides of the vehicle). Figure 3 In the diagram, the largest red rectangle represents the bounding box of the entire vehicle, the smaller red rectangle on the left represents the bounding box of the front side of the vehicle, and the blue rectangle on the right represents the bounding box of the longitudinal side of the vehicle. Then, specific key points inside the vehicle are predicted, such as the foremost point of the front (e.g., the lower corner of the bounding box of the longitudinal side of the vehicle near the front), the rearmost point (e.g., the lower corner of the bounding box of the longitudinal side of the vehicle near the rear), the outermost point on the left (e.g., the lower left corner of the bounding box of the front or rear side of the vehicle), the outermost point on the right (e.g., the lower right corner of the bounding box of the front or rear side of the vehicle), and the highest point of the roof (e.g., a point where the top of the bounding box of the entire vehicle connects to the vehicle). Alternatively, these key points can be determined by combining the vehicle's geometric model and edge detection algorithms, analyzing the salient features of the vehicle's outline. For example, identifying the vehicle's extreme points or inflection points can determine the highest points of the front, rear, left and right sides, and roof.

[0053] Step A302 aims to use the instantaneous information of the current frame to preliminarily estimate the vehicle's heading angle by identifying the contact points between the vehicle and the road surface. Detecting the contact points between the front and rear wheels of the target vehicle and the road surface can be achieved using image segmentation techniques or specialized contour detection algorithms to identify the junction between the bottom edge of the wheel and the road surface. Once the pixel coordinates of these contact points are detected, they can be mapped from the image coordinate system to the road surface physical coordinate system using the previously constructed scale field model. The scale field model provides a mapping relationship between pixel size and physical size, thereby enabling the conversion of two-dimensional pixels into three-dimensional physical coordinates (e.g., for each contact point (u,v), where u and v are pixel coordinates, the local scale factor k of that point is obtained by querying the scale field model; combined with the camera's intrinsic and extrinsic parameters, and this local scale factor k, the pixel coordinates of the contact points between the front and rear wheels and the road surface are converted into three-dimensional coordinates in the road surface physical coordinate system through inverse perspective mapping (IPM) or direct geometric triangulation). In the physical coordinate system of the road surface, by connecting the contact points between the front and rear wheels and the road surface, a line segment representing the longitudinal direction of the vehicle bottom (which can be called the bottom edge line) can be formed, and then the estimated value of the vehicle's first direction angle can be calculated.

[0054] Step A303 aims to enhance the robustness of the orientation angle estimation using time-series information. The contact points between the front or rear wheels of the target vehicle and the road surface are tracked using optical flow, Kalman filtering, or deep learning tracking algorithms. The pixel coordinates of these contact points are continuously identified and recorded across multiple frames of image information, forming a trajectory point sequence. Similar to single-frame processing, each point in this trajectory point sequence is mapped to the road surface physical coordinate system based on a scale field model (mapping method described above). By analyzing the trajectory point sequence mapped to the road surface physical coordinate system, the vehicle's motion trend and direction can be more stably reflected. For example, the trajectory point sequence can be fitted to obtain the vehicle's average motion direction over a period of time, thereby calculating the estimated second orientation angle. This multi-frame tracking method effectively smooths out any instantaneous noise or detection errors that may exist in single-frame images.

[0055] Step A304 aims to combine single-frame instantaneous information and multi-frame temporal information to obtain a more accurate and robust vehicle heading angle. Various data fusion techniques can be employed to fuse the first and second heading angle estimates. For example, a weighted average method can be used, assigning different weights based on the confidence or stability of the two estimates. Through fusion, the real-time nature of single-frame information and the stability of multi-frame information can be fully utilized, effectively overcoming the limitations of a single estimation method and improving the accuracy of heading angle estimation and its adaptability to complex environments.

[0056] This application further proposes a step for determining the orientation angle of a target vehicle. This step overcomes the limitations of traditional methods, which rely heavily on environmental interference, dynamic occlusion, or changes in illumination, by combining instantaneous information from a single frame with temporal information from multiple consecutive frames. First, the system accurately extracts the pixel coordinates of key feature points of the target vehicle from the current frame image information. These feature points include the foremost point of the vehicle's front, the rearmost point of its rear, the outermost point of its left side, the outermost point of its right side, and the highest point of its roof, providing a basic geometric reference for subsequent size calculations and orientation angle estimation. Based on this, to obtain the instantaneous orientation angle of the vehicle, the system detects the contact points between the front and rear wheels of the target vehicle and the road surface in the current frame image. These contact points directly reflect the interaction between the vehicle and the road surface, and their positional information is crucial for determining the vehicle's attitude on the road. Using a previously constructed scale field model of the road area, the pixel coordinates of these contact points are accurately mapped to the road surface's physical coordinate system. The scale field model provides a dynamic mapping relationship between pixels and physical dimensions, enabling two-dimensional image points to be converted into three-dimensional coordinates with actual physical meaning. Based on these mapped physical coordinates, a first heading angle estimate of the vehicle can be calculated, reflecting the vehicle's instantaneous driving direction at the current moment. However, the instantaneous estimate from a single frame image may be affected by noise or local occlusion. To improve the stability and robustness of the heading angle estimate, this application further introduces multi-frame temporal analysis. The system tracks the contact points between the front or rear wheels of the target vehicle and the road surface in consecutive multi-frame image information, thereby obtaining a series of trajectory point sequences in the image coordinate system. Similar to single-frame processing, these trajectory point sequences are also mapped to the road surface physical coordinate system through a scale field model. By analyzing these mapped trajectory point sequences, a second heading angle estimate of the vehicle can be calculated. This second heading angle estimate, through accumulation and smoothing in the time dimension, effectively suppresses the instantaneous errors that may exist in a single frame image, and more stably reflects the vehicle's motion trend. Finally, to obtain an optimal heading angle that combines real-time performance and stability, the system fuses the first and second heading angle estimates. This fusion mechanism fully utilizes the instantaneous response capability of the single-frame estimate and the anti-interference capability of the multi-frame estimate, generating a more accurate and reliable final heading angle through comprehensive consideration. This final orientation angle not only accurately reflects the vehicle's direction of travel (yaw angle) in the current local coordinate system of the road surface, but also demonstrates greater robustness compared to methods relying solely on single-frame image contours in complex scenarios such as curves. This approach is closely integrated with fundamental intelligent 3D vehicle over-limit detection methods, particularly utilizing a constructed scale field model of the road region. The scale field model provides precise depth information and local geometric correction capabilities for the physical coordinate mapping of the vehicle-road contact point, enabling accurate conversion of contact points from pixel space to physical space, whether in a single frame or multiple frames.This precise physical coordinate transformation is fundamental to calculating vehicle heading angles, especially under non-planar road conditions such as curves or slopes. The scale field model can effectively correct the impact of perspective distortion and road surface geometric deformation on heading angle estimation. In this way, this scheme not only provides accurate heading angles but also provides high-precision input for subsequent calculations of the target vehicle's three-dimensional physical dimensions (height, length, and width), thereby significantly improving the accuracy and environmental adaptability of the entire three-dimensional over-limit detection method.

[0057] In some implementations, step A4 includes: A401. Construct the road surface equations based on the aforementioned scale field model; A402. Using the pixel coordinates of the highest point of the vehicle roof and the contact points between the front and rear wheels of the target vehicle and the road surface, combined with camera parameters and the road surface equation, the height of the target vehicle is calculated through geometric relationships; A403. Based on the pixel coordinates of the foremost point of the front of the vehicle, the last point of the rear of the vehicle, the outermost point of the left side, and the outermost point of the right side, and the scale field model, determine the scale values ​​of the foremost point of the front of the vehicle, the last point of the rear of the vehicle, the outermost point of the left side, and the outermost point of the right side. A404. Based on the scale values ​​of the foremost point of the vehicle's front, the last point of the vehicle's rear, the outermost point of the left side, and the outermost point of the right side, the corresponding feature points are mapped to the road surface physical coordinate system to obtain the three-dimensional coordinates of the foremost point of the vehicle's front, the last point of the vehicle's rear, the outermost point of the left side, and the outermost point of the right side. A405. Based on the three-dimensional coordinates of the foremost point of the vehicle's front end and the rearmost point of the vehicle's rear end, and in conjunction with the final direction angle, calculate the projected distance between the foremost point of the vehicle's front end and the rearmost point of the vehicle's rear end along the longitudinal axis of the target vehicle to obtain the length of the target vehicle. A406. Based on the three-dimensional coordinates of the outermost points on the left and right sides, and in conjunction with the final direction angle, calculate the projected distance of the outermost points on the left and right sides in the transverse direction of the target vehicle to obtain the width of the target vehicle.

[0058] In step A401, the road surface equation can be constructed in the following ways: for example, each pixel on the road surface can be mapped to the road surface physical coordinate system (the mapping method is described above), and then the three-dimensional coordinates of the mapped points can be used to fit the road surface equation; or, the road surface equation can be constructed or updated in real time by combining the road elevation data or lidar point cloud data collected in advance with vehicle positioning information.

[0059] Secondly, step A402 aims to accurately calculate the vertical height of the vehicle. By combining the image coordinates of key points of the vehicle, the intrinsic and extrinsic parameters of the camera (camera parameters), and the constructed road surface equations, the two-dimensional information in the image can be converted into three-dimensional physical height using the principles of geometric projection and back projection. The height calculation can be specifically as follows: For example, based on the pixel coordinates of the contact points between the front and rear wheels of the target vehicle and the road surface, a bottom edge line passing through these two contact points is determined, and the projection point of the highest point of the vehicle roof onto this bottom edge line along the longitudinal direction of the image is determined. Mapping this projection point to the road surface physical coordinate system yields its three-dimensional coordinates (the mapping method is described above). Based on the pixel coordinates of the highest point of the vehicle roof and the camera parameters, the ray equation of the ray passing through the highest point of the vehicle roof from the camera's optical center can be determined. Based on the three-dimensional coordinates of the projection point and the road surface equation, the normal equation of the road surface normal line passing through the projection point can be determined. With the highest point of the vehicle roof simultaneously located at both the ray and the road surface normal as a constraint, the three-dimensional coordinates of the highest point of the vehicle roof are calculated by simultaneously solving the ray equation and the normal equation (i.e., calculating the intersection of the ray and the normal). Finally, based on the three-dimensional coordinates of the highest point of the vehicle roof and the road surface equation, the distance from the highest point of the vehicle roof to the road surface is calculated, thus obtaining the height of the target vehicle.

[0060] Next, step A403 involves obtaining the local scale information of the vehicle's key feature points in the image. The scale value represents the number of pixels per unit physical length at a specific pixel location. Due to perspective effects, the scale values ​​differ at different locations in the image; therefore, it is necessary to query the precise local scale for each key point based on the scale field model. The scale value of the feature points can be determined as follows: for example, the scale value of the road surface point that is closest to the foremost point of the vehicle's front, rearmost point, outermost point on the left, and outermost point on the right in the vertical direction of the image can be approximated as the scale value of the corresponding feature point.

[0061] Subsequently, in step A404, the two-dimensional image coordinates (i.e., pixel coordinates) of the vehicle's key feature points are converted into three-dimensional physical coordinates using their corresponding local scale values. This is a crucial transformation from image space to real-world space, providing accurate three-dimensional positional information for subsequent length and width calculations. The method for mapping feature points to the road surface physical coordinate system can be found in the preceding text.

[0062] Then, step A405 aims to accurately calculate the vehicle's length while considering its actual direction of travel. By projecting the three-dimensional coordinates of the front and rear points onto the vehicle's longitudinal axis, the effect of perspective distortion caused by vehicle yaw (direction angle) on length measurement can be eliminated, obtaining the vehicle's true physical length. The length calculation can be specifically as follows: for example, calculating the projected distance between the front and rear points on the vehicle's longitudinal axis (a unit vector defined by the yaw angle); or, alternatively, first calculating the Euclidean distance between the foremost point of the front and the rearmost point of the rear, and then decomposing this distance onto the vehicle's longitudinal axis based on the final direction angle to obtain the length.

[0063] Finally, step A406 aims to accurately calculate the vehicle's width, also considering the vehicle's actual driving direction. By projecting the three-dimensional coordinates of the outermost points on the left and right sides onto the vehicle's transverse axis (perpendicular to the longitudinal axis and parallel to the road surface), the effects of vehicle yaw and perspective distortion on width measurement can be eliminated, obtaining the vehicle's true physical width. The width calculation can be specifically as follows: for example, calculate the projected distance between the left and right points in a direction perpendicular to the longitudinal axis and parallel to the road surface (unit vector); or, the Euclidean distance between the outermost points on the left and right sides can be calculated first, and then decomposed onto the vehicle's transverse axis based on the final direction angle to obtain the width.

[0064] This application's solution first constructs the road surface equation using a scale field model, laying a precise geometric foundation for all subsequent 3D dimension calculations. Based on this, the vehicle height is calculated through geometric relationships, effectively avoiding the impact of perspective distortion on vertical dimension measurements. For length and width, the solution further utilizes the scale field model to provide accurate local scale values ​​for key vehicle feature points, accurately mapping these points from 2D image space to the 3D physical coordinate system. More importantly, when calculating length and width, the solution incorporates the target vehicle's final orientation angle, projecting the distance between feature points onto the vehicle's longitudinal and lateral axes. This projection mechanism effectively corrects measurement errors caused by vehicle yaw attitude and perspective effects, ensuring that the obtained length and width are the orthographic projection dimensions of the vehicle in the real world.

[0065] In some implementations, step A5 includes: A501. Identify the model of the target vehicle; A502. Based on the vehicle model, retrieve the corresponding statistical dimension parameters from the vehicle model size statistics database; the statistical dimension parameters include the mean and standard deviation of height, length, and width; A503. Based on the mean and standard deviation of the height, length, and width obtained from the query, determine the valid confidence intervals for the height, length, and width; A504. Compare the height, length, and width of the target vehicle with the corresponding valid confidence intervals to determine the over-limit status of the target vehicle.

[0066] First, step A501 identifies the vehicle model of the target vehicle. The purpose of this step is to determine the specific model or category of the vehicle for targeted size comparison later. Specifically, a lightweight convolutional neural network (CNN) can be used to classify the vehicle image, thereby outputting a predefined vehicle model. Alternatively, vehicle model identification can be achieved by analyzing vehicle exterior features, such as body outline, headlight shape, or logo, combined with image processing and pattern recognition techniques.

[0067] Secondly, step A502 serves to provide statistical basis for subsequent over-limit judgments. This vehicle size statistical database can dynamically maintain a historical measurement sample set of the length, width, and height for each vehicle model, and calculate the corresponding mean and standard deviation in real time. Alternatively, these statistical dimensional parameters can be pre-obtained from official data provided by vehicle manufacturers, industry standards, or large-scale vehicle census data, and stored in a structured database, such as a relational or non-relational database, where each vehicle model record includes the mean and standard deviation of its height, length, and width.

[0068] Furthermore, step A503 defines the normal fluctuation range of vehicle dimensions, thereby effectively distinguishing between normal variation and exceeding limits. For example, the effective confidence interval can be set as [μ−2σ,μ+2σ], where μ is the mean and σ is the standard deviation. Additionally, depending on actual needs and statistical principles, different confidence levels can be selected to determine the confidence interval; for example, [μ−1.96σ,μ+1.96σ] corresponds to a 95% confidence level, or [μ−3σ,μ+3σ] corresponds to a 99.7% confidence level, to accommodate different testing stringency requirements.

[0069] Finally, step A504 determines whether the actual dimensions of the target vehicle exceed limits based on statistical principles. Specifically, if all three dimensions are within their corresponding valid confidence intervals, the vehicle is considered within limits; if a dimension exceeds the upper limit of its corresponding valid confidence interval, it is considered to exceed limits. In addition to determining whether the upper limit is exceeded, a lower limit can also be set. For example, if the dimension is below the lower limit of the valid confidence interval, it may indicate a measurement error or a vehicle type identification error, thus triggering a further review mechanism.

[0070] This application's solution employs statistical distribution modeling and a confidence interval mechanism to refine the assessment of vehicle over-limit status. After obtaining the target vehicle's three-dimensional physical dimensions, the vehicle model is first identified. This ensures the specificity of the size comparison and avoids generalization errors caused by ignoring model differences. Subsequently, based on the identified model, the corresponding statistical size parameters, including the mean and standard deviation of height, length, and width, are retrieved from a vehicle size statistical database. These statistical parameters provide the data foundation for subsequent confidence interval calculations. The mean represents the typical size benchmark for the vehicle model, while the standard deviation quantifies the natural range of size variation, thus capturing the distribution characteristics of actual dimensions. Based on this, effective confidence intervals for height, length, and width are determined according to the mean and standard deviation. This step defines the normal fluctuation boundary of dimensions based on statistical principles, ensuring that only dimensions significantly exceeding this boundary are considered abnormal. Finally, the measured height, length, and width of the target vehicle are compared with the corresponding effective confidence intervals to determine the target vehicle's over-limit status. When the size exceeds the upper limit of the confidence interval, an over-limit judgment is made. This mechanism combines vehicle model characteristics and statistical distribution to ensure the accuracy and reliability of the judgment, while adapting to the size variation patterns of different vehicle models. Based on the above scheme and the fundamental calculation of vehicle three-dimensional physical dimensions, this application further provides an intelligent judgment mechanism based on vehicle model statistical characteristics. This enables the system to distinguish between normal fluctuations in vehicle size and actual over-limit behavior, significantly reducing the risk of false alarms for compliant large vehicles, thereby improving the accuracy and robustness of the entire intelligent vehicle three-dimensional over-limit detection method.

[0071] Preferably, after step A504, the following step may be included: A505. If the target vehicle does not exceed the limit, then update the statistical dimension parameters of the corresponding vehicle model in the vehicle model size statistical database according to the three-dimensional physical dimensions of the target vehicle.

[0072] This step aims to dynamically adjust and optimize the vehicle size statistics stored in the vehicle size statistics database, ensuring it reflects the true distribution and changing trends of vehicle sizes on the road. The update formula can be in the form of an exponentially weighted moving average (EWMA) to calculate the new mean and standard deviation online. The specific formula is as follows: , , This is the mean before the update. This is the updated mean. The size of the target vehicle. This is a preset forgetting factor, for example, 0.95. The standard deviation before the update. This is the updated standard deviation.

[0073] This conditional update mechanism ensures that only compliant and non-abnormal vehicle data is included in the statistical model's learning process, effectively preventing contamination of the statistical database by overloaded vehicles or measurement errors, and maintaining data purity and reliability. The update process uses the Exponentially Weighted Moving Average (EWMA) formula to calculate new means and standard deviations for the height, length, and width dimensions respectively. The forgetting factor is also included. The introduction of this technology allows the database to smoothly adapt to slow changes in vehicle dimensions, such as the launch of new models or minor adjustments to existing models, without drastic changes caused by single measurement fluctuations. This continuous, adaptive updating ensures that the vehicle size statistical database remains up-to-date and highly accurate, providing a solid foundation for subsequent over-limit judgments. This solution is closely integrated with the aforementioned intelligent vehicle 3D over-limit detection method, forming a closed-loop feedback system. Precise 3D vehicle dimension measurements provide high-quality input for database updates; vehicle model recognition and over-limit judgment provide strict filtering conditions for data updates. In this way, the accuracy of the database is continuously improved, which in turn improves the accuracy of over-limit judgments, enabling the entire detection system to maintain high precision and robustness when facing constantly changing vehicle types and sizes.

[0074] refer to Figure 2 This application provides a three-dimensional intelligent detection system for overloaded vehicles, including a camera 1 and a control system 2; The camera 1 is used to collect image information of the road area where the target vehicle is located; The control system 2 is used to identify static road reference objects with standard physical dimensions in the image information; based on the pixel dimensions of the static road reference objects in the image information and their corresponding standard physical dimensions, a scale field model of the road area is constructed; the scale field model reflects the mapping relationship between the pixel dimensions and physical dimensions of each pixel on the road surface in the road area; the feature point coordinates of the target vehicle in the image information are extracted, and the orientation angle of the target vehicle is determined; according to the scale field model, the feature point coordinates, and the orientation angle, the three-dimensional physical dimensions of the target vehicle are calculated; the three-dimensional physical dimensions include height, length, and width; statistical dimension parameters corresponding to the vehicle type of the target vehicle are obtained, and the three-dimensional physical dimensions are compared with the statistical dimension parameters to determine the over-limit state of the target vehicle (the specific process can be referred to steps A1-A5 above).

[0075] In some implementations, the control system 2 is also used to perform: graded alarm operations based on the over-limit status of the target vehicle (the specific process can be referred to step A6 above).

[0076] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A three-dimensional intelligent detection method for vehicles exceeding limits, characterized in that, Including the following steps: A1. Acquire image information of the road area where the target vehicle is located, and identify static road reference objects with standard physical dimensions in the image information; A2. Based on the pixel size of the static road reference in the image information and its corresponding standard physical size, a scale field model of the road area is constructed; the scale field model reflects the mapping relationship between the pixel size and physical size of each pixel on the road surface in the road area. A3. Extract the coordinates of feature points of the target vehicle in the image information, and determine the direction angle of the target vehicle; A4. Calculate the three-dimensional physical dimensions of the target vehicle based on the scale field model, the coordinates of the feature points, and the direction angle; the three-dimensional physical dimensions include height, length, and width; A5. Obtain the statistical dimension parameters corresponding to the vehicle model of the target vehicle, and compare the three-dimensional physical dimensions with the statistical dimension parameters to determine the over-limit status of the target vehicle.

2. The intelligent three-dimensional over-limit vehicle detection method according to claim 1, characterized in that, Following step A5, the following steps are also included: A6. Perform graded alarm operations based on the over-limit status of the target vehicle.

3. The intelligent three-dimensional over-limit vehicle detection method according to claim 1, characterized in that, The static road reference includes at least one of traffic markings, traffic signs, and fixed road facilities.

4. The intelligent three-dimensional over-limit vehicle detection method according to claim 1, characterized in that, Step A2 includes: A201. Obtain an image sequence composed of multiple consecutive frames of image information; A202. For each observation point on the static road reference, obtain the standard physical size corresponding to the observation point and the pixel size of the static road reference corresponding to the observation point in the image sequence, in order to calculate the scale observation value sequence corresponding to the observation point in the image sequence; the scale observation value represents the pixel size corresponding to a unit physical length at the observation point; A203. The scale observation sequence is corrected, and valid scale observations are obtained from the corrected scale observation sequence; the correction process includes time-domain filtering, outlier handling, and missing value imputation. A204. Based on the effective scale observations of each observation point, a scale field model of the road area is constructed using the Gaussian process regression method.

5. The intelligent three-dimensional over-limit vehicle detection method according to claim 4, characterized in that, Step A204 includes: Obtain the pixel coordinates and quality parameters of each observation point; the quality parameters include the confidence level, completeness, and type of the static road reference corresponding to the observation point; Calculate the fusion weight of each observation point based on the quality parameters; Based on the effective scale observations, pixel coordinates, and fusion weights of each observation point, a scale field model of the road area is constructed using the Gaussian process regression method. The Gaussian process regression method uses the Matérn 3 / 2 kernel function as the covariance function and inversely maps the fusion weights to the observation noise variance in the Gaussian process regression.

6. The intelligent three-dimensional over-limit vehicle detection method according to claim 1, characterized in that, Step A3 includes: A301. Extract the pixel coordinates of the feature points of the target vehicle in the current frame image information, and use them as the feature point coordinates; the feature points include the foremost point of the front of the vehicle, the last point of the rear of the vehicle, the outermost point of the left side, the outermost point of the right side, and the highest point of the roof; A302. Detect the contact points between the front and rear wheels of the target vehicle and the road surface in the current frame image information, and map the contact points to the road surface physical coordinate system based on the scale field model, and calculate the direction angle of the target vehicle based on the mapped contact points to obtain the first direction angle estimate; A303. Track the contact points between the front or rear wheels of the target vehicle and the road surface in multiple consecutive frames of image information to obtain the trajectory point sequence in the image coordinate system. Based on the scale field model, map the trajectory point sequence to the road surface physical coordinate system, and calculate the direction angle of the target vehicle based on the mapped trajectory point sequence to obtain the second direction angle estimate. A304. Combine the first azimuth angle estimate and the second azimuth angle estimate to obtain the final azimuth angle.

7. The intelligent three-dimensional over-limit vehicle detection method according to claim 6, characterized in that, Step A4 includes: A401. Construct the road surface equations based on the aforementioned scale field model; A402. Using the pixel coordinates of the highest point of the vehicle roof and the contact points between the front and rear wheels of the target vehicle and the road surface, combined with camera parameters and the road surface equation, the height of the target vehicle is calculated through geometric relationships; A403. Based on the pixel coordinates of the foremost point of the front of the vehicle, the last point of the rear of the vehicle, the outermost point of the left side, and the outermost point of the right side, and the scale field model, determine the scale values ​​of the foremost point of the front of the vehicle, the last point of the rear of the vehicle, the outermost point of the left side, and the outermost point of the right side. A404. Based on the scale values ​​of the foremost point of the vehicle's front, the last point of the vehicle's rear, the outermost point of the left side, and the outermost point of the right side, the corresponding feature points are mapped to the road surface physical coordinate system to obtain the three-dimensional coordinates of the foremost point of the vehicle's front, the last point of the vehicle's rear, the outermost point of the left side, and the outermost point of the right side. A405. Based on the three-dimensional coordinates of the foremost point of the vehicle's front end and the rearmost point of the vehicle's rear end, and in conjunction with the final direction angle, calculate the projected distance between the foremost point of the vehicle's front end and the rearmost point of the vehicle's rear end along the longitudinal axis of the target vehicle to obtain the length of the target vehicle. A406. Based on the three-dimensional coordinates of the outermost points on the left and right sides, and in conjunction with the final direction angle, calculate the projected distance of the outermost points on the left and right sides in the transverse direction of the target vehicle to obtain the width of the target vehicle.

8. The intelligent three-dimensional over-limit vehicle detection method according to claim 1, characterized in that, Step A5 includes: A501. Identify the model of the target vehicle; A502. Based on the vehicle model, retrieve the corresponding statistical dimension parameters from the vehicle model size statistics database; the statistical dimension parameters include the mean and standard deviation of height, length, and width; A503. Based on the mean and standard deviation of the height, length, and width obtained from the query, determine the valid confidence intervals for the height, length, and width; A504. Compare the height, length, and width of the target vehicle with the corresponding valid confidence intervals to determine the over-limit status of the target vehicle.

9. The intelligent three-dimensional over-limit vehicle detection method according to claim 8, characterized in that, Following step A504, the following steps are also included: A505. If the target vehicle does not exceed the limit, then update the statistical dimension parameters of the corresponding vehicle model in the vehicle model size statistical database according to the three-dimensional physical dimensions of the target vehicle.

10. A three-dimensional intelligent detection system for overloaded vehicles, characterized in that, Including cameras and control systems; The camera is used to collect image information of the road area where the target vehicle is located; The control system is used to identify static road reference objects with standard physical dimensions in the image information; based on the pixel dimensions of the static road reference objects in the image information and their corresponding standard physical dimensions, a scale field model of the road area is constructed; the scale field model reflects the mapping relationship between the pixel dimensions and physical dimensions at each pixel of the road surface in the road area. The coordinates of feature points of the target vehicle in the image information are extracted, and the orientation angle of the target vehicle is determined. Based on the scale field model, the coordinates of the feature points, and the orientation angle, the three-dimensional physical dimensions of the target vehicle are calculated. The three-dimensional physical dimensions include height, length, and width. Statistical dimension parameters corresponding to the vehicle model of the target vehicle are obtained, and the three-dimensional physical dimensions are compared with the statistical dimension parameters to determine the over-limit state of the target vehicle.

Images