Methods and systems for evaluating the night visibility of traffic safety facilities

Abstract

This application discloses a method and system for evaluating the night visibility of traffic safety facilities, relating to the field of image processing technology. The method includes: acquiring an image of a traffic safety facility to be evaluated; if the image is a traffic sign image, determining the retroreflection coefficient of each color region of the traffic sign; evaluating the night visibility of the traffic sign based on the retroreflection coefficient and the uniformity of gray values ​​in the specific distribution area of ​​the traffic sign; if the image is a traffic marking image, determining the retroreflection brightness coefficient of the traffic marking and the uniformity of gray values ​​in the specific image region where the traffic marking is located; and evaluating the night visibility of the traffic marking based on the retroreflection brightness coefficient and the uniformity of gray values ​​in the specific distribution area of ​​the traffic marking. The night visibility detection method for traffic safety facilities in this application has high detection accuracy and strong reliability.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method and system for evaluating the night visibility of traffic safety facilities. Background Technology

[0002] Traffic markings and signs are crucial components of road traffic safety facilities. Their visibility directly affects drivers' ability to recognize road conditions and traffic information, especially at night. Public data shows that the incidence of traffic accidents is significantly higher at night than during the day. Drivers' visual abilities decline in low ambient light conditions, making traffic signs and markings primarily reliant on the retroreflective properties of vehicle headlights for effective recognition. Therefore, the nighttime visibility performance of traffic safety facilities, particularly the level of retroreflective brightness and the uniformity of reflection, has become an important indicator for evaluating their safety performance.

[0003] Current technologies primarily rely on manual visual inspection, manual testing equipment, and comparison with standard samples for evaluation. These methods generally suffer from high subjectivity, low quantification, low efficiency, and poor consistency, making it difficult to accurately reflect the visibility of facilities under real-world nighttime driving conditions. Furthermore, traditional inspections typically require manual operation on the road, resulting in high safety risks, high costs, limited sampling, and insufficient coverage, hindering large-scale, rapid, and continuous inspection of road markings and signs. Therefore, there is an urgent need for an automated, objective, and quantitative technology for nighttime visibility performance detection and evaluation. Summary of the Invention

[0004] The main objective of this application is to provide a method and system for evaluating the night visibility of traffic safety facilities, aiming to solve the technical problem of low reliability in the evaluation of night visibility of traffic safety facilities in related technologies.

[0005] To achieve the above objectives, this application provides a method for evaluating the night visibility of traffic safety facilities, the method comprising:

[0006] Obtain images of the traffic safety facilities to be evaluated;

[0007] Determine the type of the traffic safety facility image to be evaluated. If the traffic safety facility image to be evaluated is a traffic sign image to be evaluated, then determine the retroreflection coefficient of each color region of the traffic sign in the traffic sign image to be evaluated.

[0008] Based on the retroreflection coefficient and the uniformity of gray values ​​in the specific distribution area of ​​traffic signs, the night visibility of traffic signs is evaluated, and the night visibility evaluation results of traffic signs are obtained.

[0009] If the image of the traffic safety facility to be evaluated is the image of the traffic marking to be evaluated, then determine the retroreflection brightness coefficient of the traffic marking in the image of the traffic marking to be evaluated and the uniformity of gray value of the specific image area where the traffic marking is located.

[0010] Based on the retroreflective luminance coefficient and the uniformity of gray values ​​in the specific distribution area of ​​traffic markings, the night visibility of traffic markings is evaluated, and the night visibility evaluation results of traffic markings are obtained.

[0011] In one embodiment, the steps for evaluating the night visibility of traffic signs based on the retroreflection coefficient and the uniformity of gray values ​​in the specific distribution area of ​​the traffic signs, and obtaining the night visibility evaluation results of the traffic signs, include:

[0012] For each color region in the specific distribution area of ​​traffic signs, the color region is divided into several small regions of equal area, and the standard deviation of the gray value of each small region is calculated to obtain a uniform dataset.

[0013] Calculate the mean and standard deviation of the homogeneous dataset, and determine the coefficient of variation based on the mean and standard deviation;

[0014] The uniformity of grayscale values ​​in each color region of a traffic sign is determined based on the coefficient of variation.

[0015] Based on the range of grayscale uniformity and the range of retroreflection coefficient, the night visibility of traffic signs is evaluated, and the night visibility evaluation results of traffic signs are obtained.

[0016] In one embodiment, the step of determining the retroreflection coefficient of each color region of a traffic sign in an image to be evaluated includes:

[0017] Identify the location of traffic signs in the image of traffic signs to be evaluated, mark them with a sign box, and determine whether the traffic signs to be evaluated are within the evaluation range.

[0018] If the traffic sign to be evaluated is within the evaluation area, then the image area within the sign frame is segmented by color to determine the various color areas of the traffic sign in the image of the traffic sign to be evaluated.

[0019] Based on the grayscale values ​​of each color region, the retroreflection coefficient of the traffic sign is calculated using the corresponding retroreflection coefficient calculation model.

[0020] In one embodiment, the steps of acquiring the traffic sign image to be evaluated through a traffic sign detection camera installed on a detection vehicle, identifying the position of the traffic sign in the image, marking it with an identification box, and determining whether the traffic sign to be evaluated is within the evaluation range include:

[0021] Traffic sign dot cloud data is obtained by detecting radar installed on the vehicle;

[0022] Identify traffic signs in the traffic sign image to be evaluated, mark them with bounding boxes, and determine the pixel coordinates of the traffic signs within the bounding boxes;

[0023] Based on the pixel coordinates and point cloud data of traffic signs, radar-visual fusion ranging is performed to determine whether the distance between the traffic sign in the image to be evaluated and the detection vehicle is within the evaluation range.

[0024] In one embodiment, the step of identifying traffic signs in a traffic sign image to be evaluated and marking them with bounding boxes includes:

[0025] The traffic sign image to be evaluated is input into the traffic sign recognition model for recognition, the position of the traffic sign in the image is determined, and it is marked with an identification box;

[0026] The traffic sign recognition model consists of a feature extraction unit, a feature fusion unit, and a detection output unit. The feature extraction unit is a backbone network with CSPDarknet-53 as the main backbone and RetinexNet as a lightweight branch. The backbone network also embeds the CoAtNet attention enhancement mechanism. The feature fusion unit uses a BiFPN bidirectional feature pyramid network. The detection output unit includes a classification head and a regression head, and uses the CIoU loss function. The classification head consists of three 1×1 convolutional layers and a Softmax function, and the regression head consists of three 1×1 convolutional layers and a Sigmoid function.

[0027] In one embodiment, the step of performing color segmentation processing on the image region within the sign frame to determine each color region of the traffic sign in the traffic sign image to be evaluated includes:

[0028] The image region within the bounding box is input into the traffic sign color segmentation model to determine the color regions of the traffic sign in the image to be evaluated. The traffic sign color segmentation model includes an input layer, a hidden layer, and an output layer.

[0029] The input layer includes a color space conversion unit and a color feature extraction unit. The color space conversion unit is used to convert the image region from RGB to LAB space. The color feature extraction unit includes a 1×1 convolutional layer and a 2×2 average pooling layer. The 1×1 convolutional layer is used to extract local color association features in LAB space, and the 2×2 average pooling layer is used to reduce the resolution of the image region.

[0030] The hidden layer includes a first fully connected layer, a second fully connected layer, and a third fully connected layer. The SENet channel attention mechanism is embedded between the first and second fully connected layers, and a residual branch is introduced between the second and third fully connected layers.

[0031] The output layer consists of six output neurons, each activated by a Softmax activation function to output the color category probability of a traffic sign.

[0032] In one embodiment, the retroreflection coefficient calculation model is determined by fitting the retroreflection coefficients of multiple calibration plates and the corresponding average gray level of the calibration plates.

[0033] In one embodiment, the road traffic marking image to be evaluated is acquired by a marking detection camera at a fixed angle on the detection vehicle. The fixed angle of the marking detection camera is determined based on the brightness distribution of the high beams of the detection vehicle.

[0034] To achieve the above objectives, this application further provides a traffic safety facility night vision evaluation system, the system comprising:

[0035] The road marking detection camera is mounted on the detection vehicle via a first stabilizer and is used to acquire images of the road markings to be evaluated.

[0036] A traffic sign detection camera, mounted on the detection vehicle via a second stabilizer, is used to capture images of traffic signs to be evaluated.

[0037] Radar, installed on the inspection vehicle, is used to collect point cloud data of traffic signs;

[0038] The controller is connected to the road marking detection camera, traffic sign detection camera and radar respectively, and is used to send the road marking image to be evaluated, the traffic sign image to be evaluated and the traffic sign point cloud data to the host computer.

[0039] The host computer, connected to the controller, includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the above-described method for evaluating the night visibility of traffic safety facilities.

[0040] One or more technical solutions proposed in this application have at least the following technical effects:

[0041] This application achieves automated, objective, and quantitative detection and evaluation of the night vision performance of traffic signs and markings through vehicle-mounted multi-sensor collaboration and intelligent algorithms. Compared with traditional methods that rely on manual visual inspection and static sampling, this application can continuously, rapidly, and extensively detect traffic safety facilities under real nighttime driving conditions, effectively avoiding the problems of strong subjectivity, high safety risks, and insufficient coverage of manual inspection. At the same time, by introducing a joint evaluation mechanism of grayscale uniformity analysis and retroreflection performance, it not only improves the refinement and consistency of night vision performance evaluation results, but also more realistically reflects the actual level of facility reflectivity uniformity and recognizability, thereby significantly improving the reliability, accuracy, and engineering application value of night vision evaluation of traffic safety facilities. Attached Figure Description

[0042] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating the night visibility evaluation method for traffic safety facilities in this application.

[0045] Figure 2 This is a schematic diagram of the night visibility evaluation system for traffic safety facilities in this application.

[0046] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0047] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0048] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0049] The main solution of this application embodiment is: to acquire nighttime images of traffic signs and traffic markings by a detection vehicle; for traffic signs, to calculate the retroreflection coefficient based on a calibration model and to evaluate night vision performance by combining grayscale uniformity; for traffic markings, to calculate their retroreflection brightness coefficient and to conduct a comprehensive evaluation by combining grayscale uniformity.

[0050] Specifically, this application provides a method for evaluating the night visibility of traffic safety facilities, referring to... Figure 1 , Figure 1 This is a schematic diagram of the overall process of an embodiment of the night visibility evaluation method for traffic safety facilities in this application.

[0051] In this embodiment, the method for evaluating the night visibility of traffic safety facilities includes steps S10 to S50:

[0052] Step S10: Obtain images of the traffic safety facilities to be evaluated.

[0053] Step S20: Determine the type of the traffic safety facility image to be evaluated. If the traffic safety facility image to be evaluated is a traffic sign image to be evaluated, then determine the retroreflection coefficient of each color region of the traffic sign in the traffic sign image to be evaluated.

[0054] Step S30: Based on the retroreflection coefficient and the uniformity of gray values ​​in the specific distribution area of ​​the traffic sign, the night visibility of the traffic sign is evaluated to obtain the night visibility evaluation result of the traffic sign.

[0055] Step S40: If the image of the traffic safety facility to be evaluated is the image of the traffic marking to be evaluated, then determine the retroreflection brightness coefficient of the traffic marking in the image of the traffic marking to be evaluated and the uniformity of gray value of the specific image area where the traffic marking is located.

[0056] Step S50: Based on the retroreflective brightness coefficient and the uniformity of gray values ​​in the specific distribution area of ​​the traffic markings, the night visibility of the traffic markings is evaluated to obtain the night visibility evaluation results of the traffic markings.

[0057] Specifically, the images of traffic safety facilities to be evaluated are either traffic sign images or traffic marking images. In this embodiment, different evaluation methods are used for the traffic sign images and the traffic marking images to be evaluated:

[0058] For traffic sign images to be evaluated, considering that traffic signs are usually of various colors and that the reflectivity of different color regions varies greatly, in order to determine the specific location of the traffic sign in the image, it is necessary to divide the traffic sign at that location by color, segment it into different color regions, and perform retroreflection coefficient detection for each color region.

[0059] For a traffic marking image to be evaluated, a traffic marking is usually one color. Therefore, after determining the specific location of the traffic marking in the image, it is necessary to identify the color of the traffic marking and determine the detection method of the retroreflective brightness coefficient for that color.

[0060] Understandably, traffic signs are typically multi-colored reflective components mounted on facades, and their nighttime visibility primarily depends on the reflectivity characteristics of different colored materials under vehicle headlight illumination. The retroreflection coefficient characterizes the reflectivity of sign materials to incident light under unit incident illuminance conditions and is closely related to color, material type, and microstructure characteristics. It effectively eliminates the influence of external factors such as shooting distance and light intensity, making it suitable for comparing and evaluating the reflectivity of different color areas of traffic signs. On the other hand, traffic markings are linear or planar structures attached to the road surface, and their nighttime visibility is directly reflected in the brightness within the driver's field of vision, influenced by a combination of factors including headlight illumination angle, road surface roughness, and wear condition. The retroreflection luminance coefficient directly reflects the luminance level presented to the driver by the markings under actual lighting conditions, more closely aligning with human subjective perception and engineering testing standards, making it suitable for describing the nighttime recognizability of traffic markings. Therefore, this implementation method takes into account the differences in geometric shape, reflective mechanism and visibility evaluation focus of traffic signs and traffic markings, and selects retroreflection coefficient and retroreflection luminance coefficient as core evaluation parameters respectively, so that the night vision performance test results not only conform to the actual use scenario, but also take into account the scientific nature, accuracy and engineering applicability of the evaluation.

[0061] Furthermore, when evaluating traffic safety facilities, this embodiment also considers the uniformity of grayscale values ​​in the area where the traffic safety facility is located in the image of the facility to be evaluated. Its core function is to supplement the evaluation of the nighttime visibility performance of traffic safety facilities, avoiding evaluation bias caused by relying solely on retroreflection parameters. While the retroreflection coefficient or retroreflection brightness coefficient mainly reflects the overall reflective intensity level of traffic signs or markings, it is difficult to characterize the consistency of reflective space distribution. In actual engineering, even if the retroreflection index meets the standard, uneven reflective distribution may still result in localized overbrightness or underbrightness, making it difficult for drivers to recognize the facilities. By introducing grayscale uniformity, the spatial consistency of reflective performance in different areas can be effectively reflected, compensating for the shortcomings of a single intensity index.

[0062] In one feasible implementation, step S20 includes steps A10 to A30:

[0063] Step A10: Identify the position of the traffic sign in the image of the traffic sign to be evaluated, and determine whether the traffic sign to be evaluated is within the evaluation range.

[0064] Step A20: If the traffic sign to be evaluated is within the evaluation range, then perform color segmentation processing on the image area within the sign frame to determine each color area of ​​the traffic sign in the image of the traffic sign to be evaluated.

[0065] Specifically, when acquiring images of traffic signs to be detected, the detection vehicle is required to maintain a certain distance from the sign to ensure that the traffic sign is within the evaluable range. In one feasible implementation, the image of the traffic sign to be evaluated is acquired by a traffic sign detection camera mounted on the detection vehicle. Step A10 includes steps A11 to A13:

[0066] Step A11: Obtain traffic sign point cloud data by detecting the radar installed on the vehicle.

[0067] Step A12: Identify the traffic signs in the traffic sign image to be evaluated, mark them with a bounding box, and determine the pixel coordinates of the traffic signs within the bounding box.

[0068] Step A13: Perform radar-visual fusion ranging based on the pixel coordinates of the traffic sign and the traffic sign point cloud data to determine whether the distance between the traffic sign in the image to be evaluated and the detection vehicle is within the evaluation range.

[0069] For example, the traffic sign detection camera and LiDAR are first jointly calibrated to obtain a transformation matrix. Then, using the transformation matrix, the actual distance information of the traffic sign relative to the detection vehicle is calculated based on the pixel coordinates of the detected traffic sign in the image. LiDAR-visual fusion is a relatively mature technology; only a brief overview of the execution steps is provided here:

[0070] The relative positions of the fixed radar and traffic sign detection camera are determined. A calibration board is prepared, and 2D images and 3D point cloud data of the calibration board at different positions and tilt angles are captured. The corner coordinates (x, y) of the calibration board in the 2D image are detected, and the corresponding point cloud coordinates (X, Y, Z) in the 3D point cloud data are also detected. Multiple sets of corner coordinates and point cloud coordinates corresponding to the corner points of the calibration board are obtained. Using a fusion algorithm, the transformation matrix that maps the 2D pixel coordinates (u, v) to the world coordinates (X, Y, Z) is calculated, which includes the selection matrix R and the translation matrix T (R: 3×3 rotation matrix, T: 3×1 translation vector).

[0071] Using a transformation matrix, whenever a sign is detected in an image, the pixel coordinates of the sign are calculated and transformed into world coordinates. Point cloud clusters located 55-65 meters away from the detection vehicle and 4.5-7 meters high are retained to verify whether they fall within the range of traffic signs. This allows for the selection of the relative distance, relative height, and relative lateral distance between the sign to be detected and the detection vehicle, thus determining the range to be evaluated.

[0072] Furthermore, step A12 includes step A121:

[0073] Step A121: Input the traffic sign image to be evaluated into the traffic sign recognition model for recognition, determine the position of the traffic sign in the image, and mark it with a bounding box.

[0074] The traffic sign recognition model consists of a feature extraction unit, a feature fusion unit, and a detection output unit. The feature extraction unit is a backbone network with CSPDarknet-53 as the main backbone and RetinexNet as a lightweight branch. The backbone network also embeds the CoAtNet attention enhancement mechanism. The feature fusion unit uses a BiFPN bidirectional feature pyramid network. The detection output unit includes a classification head and a regression head, and uses the CIoU loss function. The classification head consists of three 1×1 convolutional layers and a Softmax function, and the regression head consists of three 1×1 convolutional layers and a Sigmoid function.

[0075] Specifically, to address the issues of low contrast, noise interference, and blurred details in nighttime traffic signs, this implementation method achieves high-precision detection through structural innovation and training strategy optimization, solving the pain points of traditional models such as low detection accuracy, inaccurate positioning, high false detection rate, high false negative rate, and color classification being affected by lighting.

[0076] This implementation uses an improved Yolov8 network for feature enhancement and localization optimization in low-light scenes to construct a traffic sign recognition model:

[0077] Based on YOLOv8 (Backbone-Neck-Head three-stage architecture), the following three key improvements are made to address issues such as weak nighttime features and difficulty in locating small targets:

[0078] Backbone: Based on CSPDarknet-53, a lightweight RetinexNet branch (3×3 convolution to extract L channels + 2 residual blocks) is added to enhance dark area features, and CoAtNet attention is embedded to enhance the signature features and suppress noise;

[0079] Neck: Replace PANet with BiFPN, set up a bidirectional fusion path of "top-down + bottom-up" (8×8 scale, initial weight 0.7) to enhance small target features, and add a bilinear interpolation feature alignment module to prevent edge misalignment;

[0080] Head: The "classification + regression shared head" is split into independent heads (each with 3 layers of 1×1 convolution, using Softmax for classification and Sigmoid for regression); CIoU loss (with 3 constraints such as center point distance) is used instead of IoU, which significantly reduces the localization error.

[0081] By employing FP16 quantization and TensorRT optimization, the model size is further compressed and deployment efficiency is improved. TensorRT's FP16 hybrid quantization preserves FP16 accuracy in the convolutional and attention layers of the Backbone (avoiding feature loss), while INT8 quantization is used for the BiFPN fusion layers, compressing the model size to nearly half of its original size. TensorRT is used for layer fusion and dynamic shape optimization (adapting to different input resolutions), resulting in inference speed improvements of more than 2x.

[0082] Sign detection and recognition model dataset creation and training strategy: Special optimization for nighttime scenes, collecting 30,000 images of various signs in low-light scenes (dim outdoor areas without streetlights), and using a combination algorithm of "RetinexNet denoising + adaptive histogram equalization" to improve the brightness of dark areas while preserving sign details;

[0083] LabelImg is used to annotate the bounding boxes and categories to generate an annotation file; during the training phase, random augmentation (probability 0.5) is performed in real time, including geometric augmentation: random flipping (horizontal / vertical), rotation (-15°~15°), scaling (0.8~1.2 times) and illumination augmentation: random brightness jitter (-0.3~0.3), contrast adjustment (0.7~1.3), and Gaussian noise addition (variance 0.01~0.03) to simulate different nighttime illumination fluctuations; the ratio of training set, validation set, and test set is 8:1:1.

[0084] A phased convergence optimization strategy is employed during training. Pre-training initialization: Backbone pre-trains the CSPDarknet-53 weights using COCO, and initializes the weights using He normal distributions for improved modules such as CoAtNet and BiFPN to prevent gradient vanishing in the early stages of training.

[0085] Phased Training: Phase 1 (Freeze Training, 1-20 rounds): Freeze the first 3 layers of the backbone, training only the attention module, BiFPN, and detector head. The optimizer uses AdamW (weight decay of 0.001) to quickly converge basic localization capabilities. Phase 2 (Unfreeze Training, 21-80 rounds): Unfreeze all network layers, using cosine annealing for learning rate scheduling, focusing on optimizing feature matching in low-light scenes. Phase 3 (Fine-tuning Training, 81-100 rounds): Introduce hard sample mining, increasing the training weights for hard samples with IoU < 0.3 (such as blurred or occluded signs) to reduce the false negative rate. Combined Loss Function: To address "class imbalance" (e.g., more prohibition signs than instruction signs), CIoU Loss is used for regression loss, and Focal Loss (α = 0.25, γ = 2) is used for classification loss. The total loss is the weighted sum of the two (weight ratio 1:1.5).

[0086] Furthermore, step A20 includes step A21:

[0087] Step A21: Input the image region within the bounding box into the traffic sign color segmentation model to determine the color regions of the traffic sign in the image to be evaluated; the traffic sign color segmentation model includes an input layer, a hidden layer, and an output layer; wherein,

[0088] The input layer includes a color space conversion unit and a color feature extraction unit. The color space conversion unit is used to convert the image region from RGB to LAB space. The color feature extraction unit includes a 1×1 convolutional layer and a 2×2 average pooling layer. The 1×1 convolutional layer is used to extract local color association features in LAB space, and the 2×2 average pooling layer is used to reduce the resolution of the image region.

[0089] The hidden layer includes a first fully connected layer, a second fully connected layer, and a third fully connected layer. The SENet channel attention mechanism is embedded between the first and second fully connected layers, and a residual branch is introduced between the second and third fully connected layers.

[0090] The output layer consists of six output neurons, each activated by a Softmax activation function to output the color category probability of a traffic sign.

[0091] Specifically, after the traffic sign image to be evaluated is acquired, the traffic sign recognition model identifies the position of the traffic sign in the image and marks its location using a bounding box. Then, it is determined whether the relative distance between the traffic sign and the detection vehicle meets the detection distance requirements. If it does, the color segmentation model determines each color region of the traffic sign as the region to be evaluated.

[0092] In this embodiment, the strategy for creating and training the traffic sign color segmentation model dataset is as follows:

[0093] The improved multilayer perceptron MLP-Softmax multiclass network was used to segment the color regions of the sign.

[0094] The collected nighttime sign images were cropped into different color regions according to color, and then unified into an image dataset with consistent width and height. After preprocessing, this dataset was used to train the sign color classification model.

[0095] Creating a color classification model dataset

[0096] 30,000 color regions were manually cut out from 30,000 low-light images of the sign detection and recognition model (each region is a color block of an independent sign, such as the red region of a red circular prohibition sign and the blue region of a blue rectangular instruction sign).

[0097] Unified preprocessing: All color regions are resized to 224×224 (to ensure consistent input size), converted to LAB space, and then normalized (mapped to the [-1,1] interval) to reduce the impact of brightness fluctuations on color.

[0098] Labeling method: Label according to "color category + confidence level". Each color area is labeled as one of the 6 categories (such as "red"), and the confidence level of human judgment is labeled (0.8-1.0, used for weight allocation during training).

[0099] Data augmentation: Specific enhancements are performed on color features by adding a ±5 random offset to the A / B channels (simulating color shift under nighttime lighting), a 3×3 Gaussian filter (variance 0.02) to suppress color noise, and random cropping of color areas by 90%~100% to enhance the robustness of local colors.

[0100] Data partitioning: The ratio of training set, validation set, and test set is 7:2:1.

[0101] Training strategy (optimizing the stability of color classification)

[0102] Initialization and optimizer: The network weights are initialized using the Xavier normal distribution, and the optimizer uses SGD (momentum 0.9, weight decay 0.0005) to avoid overfitting AdamW on small sample color data;

[0103] Learning rate scheduling: A "stepped learning rate" is adopted, with an initial learning rate of 1e-3, which is reduced to 1e-4 after 50 training rounds, and then to 1e-5 after 80 training rounds, for a total of 100 training rounds, to ensure stable convergence in the later stages;

[0104] The loss function uses weighted cross-entropy loss, with weights set according to the number of samples in each category (e.g., if there are few yellow samples, the weight is set to 1.5; if there are many red samples, the weight is set to 0.8). At the same time, the label confidence is introduced as a loss weight (samples with high confidence have a weight of 1.0, and samples with low confidence have a weight of 0.6) to improve classification accuracy.

[0105] Regularization measures: Dropout (probability 0.3) is added to the first and second hidden layers to suppress overfitting; an "early stop strategy" is adopted during training (training is stopped if the accuracy on the validation set does not improve for 5 consecutive rounds) to save the optimal model.

[0106] Post-segmentation processing (improving engineering application results) Region connectivity filtering: For the color probability map output by the model, 8-neighborhood connectivity analysis is used to remove small noise points with an area of ​​<10 pixels (such as pseudo-color regions caused by noise) and retain continuous marker color regions.

[0107] Color consistency verification: Combine the detection box of the sign recognition model to perform consistency verification on the segmented color areas (e.g., "the proportion of red area in the detection box of the prohibition sign should be ≥80%)", further reducing the missegmentation rate.

[0108] Improved color segmentation model of MLP-Softmax:

[0109] To address the issues of traditional MLP (fully connected network) networks, such as "loss of image spatial information and susceptibility to noise interference in color classification," an improved MLP structure combining "local feature extraction + attention enhancement" is designed to achieve accurate segmentation of marker colors.

[0110] Input layer: Color space conversion + preprocessing

[0111] Color space optimization: RGB to LAB space (L is the luminance channel, A / B are the color channels), separating luminance and color information to avoid the influence of L channel fluctuations on A / B channel color judgment at night;

[0112] Local Feature Module: A new "1×1 convolution (32 kernels, stride 1) + 2×2 average pooling" module has been added: 1×1 convolution extracts local color-related features in the LAB space (such as the high-value area of ​​the A channel marked in red), avoiding the loss of spatial information caused by image flattening; the pooling layer reduces the resolution from 224×224 to 112×112, reducing the number of parameters by 75% and improving training efficiency;

[0113] Input vector: The feature map after pooling is flattened into a 401408-dimensional vector, which is then input into the MLP network.

[0114] Hidden layer: The basic structure of channel attention + residual connection is 3 fully connected layers (number of neurons 2048→1024→512), and the activation function is LeakyReLU (negative slope 0.01), which solves the problem of "dead neurons" in traditional ReLU;

[0115] Embedded SENet channel attention: Between the first and second hidden layers, the 2048-dimensional vector is compressed into 1 dimension through global average pooling, and channel weights are generated through 2 fully connected layers (128→2048) + Sigmoid to enhance color-related channels (such as the red A channel).

[0116] Add a residual branch: Use a 1×1 convolution between the 2nd and 3rd hidden layers to adjust the dimension, alleviate gradient vanishing in deep MLPs, and improve the efficiency of color feature transfer.

[0117] Output layer: multi-class classification + confidence filtering.

[0118] Output dimension: Based on traffic sign color categories (red, yellow, blue, green, white, brown, 6 categories), there are 6 output neurons, and the output probability distribution is activated by Softmax.

[0119] Confidence constraint: A new filtering module has been added. When the maximum probability is less than 0.7, it is judged as a "fuzzy region" (such as nighttime glare) to avoid misclassification and improve the reliability of segmentation.

[0120] After determining the various color regions of the traffic sign, step A30 involves calculating the retroreflection coefficient of the traffic sign based on the grayscale values ​​of each color region using the corresponding retroreflection coefficient calculation model.

[0121] For example, the reflective properties of different colored marking films vary significantly, requiring calibration according to each color. The marking film is then pasted onto a smooth, flat substrate to create a marking calibration board. The selection of the substrate considers factors such as weight, surface flatness, durability, and resistance to deformation; suitable materials include wood, acrylic, and honeycomb aluminum. Before pasting, the substrate should be cleaned with alcohol. During pasting, at least two people should work together, using a scraper to slowly push the marking film firmly against the substrate, preventing air bubbles. If air bubbles do form, they can be popped with a pin and smoothed out with a scraper.

[0122] The calibration plate is assigned a value using a calibrated handheld retroreflection coefficient testing device.

[0123] Road signs commonly come in six colors: red, white, yellow, green, blue, and brown. At least five signs of each color with different retroreflectance coefficient levels were selected as calibration plates. Taking white as an example, multiple calibration plates with different retroreflectance values ​​were selected, and the size of the calibration plates was kept as close as possible to the size of the sign to be tested. The retroreflectance coefficient data is shown in Table 1.

[0124] Table 1 Retroreflection Coefficient Data Table

[0125]

[0126] This example uses a 1m x 1m reflective film as the calibration plate. At night, under ideal detection distance (60 meters from the sign detection camera), the sign is installed at a fixed height to simulate the installation height of signs on actual roads. The installation height of the lower edge of signs on ordinary roads is generally greater than 4.5 meters. Images of the calibration plate are captured using the sign camera, and the retroreflection coefficient corresponding to the calibration plate is matched one-to-one with the calculated grayscale values ​​of the calibration plate in the images.

[0127] The relationship between the retroreflection coefficient and grayscale value of different colored calibration plates is calculated to obtain the retroreflection coefficient calculation model. The coefficients of the retroreflection coefficient calculation model differ depending on the height and angle of the traffic sign detection camera. Mathematical models are generally represented by functional relationships, which can be linear, power, exponential, logarithmic, quadratic, or cubic functions, etc. Multiple calibration plates are selected...

[0128] Taking a white calibration plate as an example, the calculation model for the retroreflection coefficient is as follows:

[0129] Y=0.0449X 2 -10.026X+702.1, R 2 = 0.9995

[0130] Y represents the retroreflection coefficient of the traffic sign, and X represents the grayscale value of the calibration plate.

[0131] In actual testing, the average gray value of the white area of ​​the traffic sign in the detected image to be evaluated is substituted into the retroreflection coefficient calculation model to obtain the retroreflection coefficient corresponding to the white area.

[0132] Subsequently, the traffic sign image to be evaluated is evaluated according to step S30. In this embodiment, step S30 includes steps B10 to B30:

[0133] Step B10: For each color region in the specific distribution area of ​​the traffic sign, divide the color region into several small regions of equal area, calculate the standard deviation of the gray value of each small region, and obtain a uniformity dataset.

[0134] Step B20: Calculate the mean and standard deviation of the homogeneous dataset, and determine the coefficient of variation based on the mean and standard deviation.

[0135] Step B30: Determine the uniformity of grayscale values ​​in each color region of the traffic sign based on the coefficient of variation.

[0136] Step B40: Based on the range of grayscale uniformity and the range of retroreflection coefficient, the night visibility of the traffic sign is evaluated to obtain the night visibility evaluation result of the traffic sign.

[0137] Specifically, the calculation of grayscale uniformity is based on the coefficient of variation. The core idea is to use the standard deviation divided by the mean, and then convert it in reverse to a uniformity score in the 0-1 interval, thus eliminating the influence of numerical magnitude.

[0138] The detection area of ​​the traffic sign or marking to be tested is divided into multiple small areas of equal area. The standard deviation of the gray value of each area is calculated to obtain the dataset for uniformity calculation.

[0139] The specific calculation steps are as follows:

[0140] The first step is to calculate the mean of the dataset:

[0141] μ=(x1+x2+...+x n ) / n

[0142] Where μ is the mean of the dataset, x n This represents a sample in the dataset, where n is the number of data points.

[0143] The second step is to calculate the standard deviation:

[0144]

[0145] Where σ is the standard deviation of the dataset, μ is the mean of the dataset, and n is the number of data points.

[0146] The third step is to calculate the coefficient of variation:

[0147] C V =σ / μ

[0148] Step 4: Calculate the uniformity score:

[0149] U = 1 - C V

[0150] Where U is the uniformity score.

[0151] In one example, a uniformity score ≥ 0.85 indicates relatively concentrated values ​​and good uniformity; a uniformity score ≥ 0.5 and < 0.85 indicates low numerical dispersion and moderate uniformity; and a uniformity score < 0.5 indicates high numerical dispersion and poor uniformity. Based on the range of grayscale value uniformity and the range of retroreflection coefficient of the traffic sign, the night visibility evaluation result of the corresponding traffic safety sign can be determined.

[0152] It should also be noted that when conducting traffic marking detection, it is also necessary to select an appropriate evaluation range to ensure the quality of the obtained traffic marking images. In this embodiment, the road traffic marking images to be evaluated are acquired by a marking detection camera at a fixed angle on the detection vehicle. The fixed angle of the marking detection camera is determined based on the brightness distribution of the high beams of the detection vehicle.

[0153] Understandably, once the setting angle of the road marking detection camera is determined, its shooting range can be determined. Since it is shooting the road markings at an angle, the ideal detection range can be obtained simply by adjusting the camera's setting angle. For example, the road marking detection range is the area 13 to 15 meters in front of the detection vehicle. This is an area where the brightness distribution of the detection vehicle's high beams is relatively uniform, and it is suitable for road markings of various colors and sizes.

[0154] After obtaining the traffic marking image to be evaluated, image enhancement processing is performed on the traffic marking image to be evaluated. Then, the YOLOv8 algorithm is used to identify the traffic marking image to be evaluated, and the image to be evaluated is marked using bounding boxes.

[0155] Then, DeepLabv3 was used to determine the color distribution area of ​​the road markings in the road marking image to be evaluated.

[0156] Traffic markings are thin and elongated, with uneven shapes, and uneven lighting in nighttime images makes them difficult to capture, hindering effective detection using common methods. In this embodiment, the traffic marking images undergo preprocessing algorithms such as enhancement, noise reduction, and filtering to create a training set. The YOLOv8 algorithm is then used for target detection, and DeepLabv3 is used to extract the marking regions. Traffic marking image enhancement commonly employs methods such as the Deep Generative Network EnhanceGAN or autoencoder-based techniques.

[0157] After determining the color area of ​​the traffic markings, the retroreflective brightness coefficient and grayscale uniformity of the area are calculated. The retroreflective brightness coefficient is also calibrated.

[0158] For example, a calibration plate with a relatively dispersed retroreflective luminance coefficient is selected as the calibration plate, with a value range of 22-500 millicandelas per square meter per lux (mcd·m). -2 ·lx -1 At least five calibration plates should be selected to obtain better calibration results. The specific distribution is shown in Table 2, the retroreflection brightness coefficient data table.

[0159] Table 2 Retroreflection Luminance Coefficient Data Table

[0160]

[0161] The data shown in Table 1 is for reference only and should not be used as a limitation in actual use.

[0162] The selection of calibration plates is not limited to five; this is merely an example. The more uniform the distribution of retroreflectance luminance coefficient values ​​on the calibration plates, and the more values ​​present, the better the model fit. This is the optimal choice balancing operational complexity and performance. The values ​​on the calibration plates are obtained using a calibrated handheld retroreflectance gauge.

[0163] This example also proposes a method for obtaining the retroreflective brightness coefficient of a calibration plate, and the operation steps are as follows.

[0164] Taking a calibration plate with a length of 1 meter and a width of 15 cm as an example, the specific selection method is explained as follows:

[0165] In a dark room, the effective detection area of ​​the light emitted by a calibrated handheld retroreflective marker was determined. For example, this invention selects the calibrated Easylux Mini handheld retroreflective luminance coefficient detector from Brazil. When it is working, the area on the flat marker board illuminated by the light emitted by it is approximately 7 cm wide horizontally and 30 cm long vertically.

[0166] Multiple measurements are taken and the average value is calculated. The range of light emitted by the detector should cover the entire calibration plate area. At least 6 measurements are required to cover the entire calibration plate area. The average value of the 6 retroreflection brightness coefficients is then taken as the retroreflection brightness coefficient of the calibration plate.

[0167] The detection area of ​​the handheld road marking retroreflective detector starts 30 centimeters in front of it. A road marking plate of the same type and thickness as the road marking to be measured needs to be placed under the handheld device as a base plate in order to stably measure area 1 and area 2.

[0168] When using handheld devices for measurement, clearly define the measurement direction of the markings; do not mix directions.

[0169] In-service road markings have different color characteristics (yellow, white), width characteristics (10 cm, 15 cm, and 20 cm), and length characteristics (2 m, 3 m, etc.). These different characteristics all affect the nighttime visibility performance of the road markings. The selected road marking calibration plate should be suitable for most usage scenarios. The selected calibration plate is 1 meter long and 15 cm wide.

[0170] During calibration, park the test vehicle on a straight road with a lane width of 3.5 meters or 3.75 meters, and place the calibration plate 14 meters to the right front of the test vehicle. The horizontal position of the calibration plate should be consistent with the lane lines on the road surface. It can cover the original road markings or be placed in the intervals between the original road markings.

[0171] Turn on the car headlights, open the detection software, and obtain the average grayscale value of each calibration plate from the image. Fit the retroreflection brightness coefficient of the calibration plate to the obtained average grayscale value to establish a mathematical relationship model. The mathematical relationship model can be a linear function model, a quadratic function model, a cubic function model, a logarithmic function model, or a power function model, etc.

[0172] The method of obtaining the average gray value of the calibration board is not limited to calculating the average gray value of any channel or combination of channels in the RGB color space or HSV color space.

[0173] Mathematical models were derived for the white and yellow road markings, and R² was used as the basis for selecting the optimal model.

[0174] The calculation model for the retroreflection brightness coefficient obtained by fitting the white marker line is shown below:

[0175] Y = 0.0003X2 +0.1915X+26.472, R²=0.9954

[0176] The calculation model for the retroreflection brightness coefficient obtained by fitting the yellow marker line is shown below:

[0177] Y=0.0014X 2 +0.2476X+100.6, R²=0.9971

[0178] In the above, Y represents the retroreflection brightness coefficient of the calibrator, and X represents the gray value of the calibrator plate in the image.

[0179] In actual detection, the area of ​​the marking to be detected in the marking image is extracted and its average gray value is calculated. The retroreflection luminance coefficient of the marking segment is calculated by the fitted retroreflection luminance coefficient calculation model.

[0180] In practical applications, different combinations of marking features require different retroreflection brightness coefficient calculation models.

[0181] The evaluation method for traffic markings can be exemplified by, for example, using white markings, where the minimum retroreflective luminance coefficient required to ensure road safety is 80 (mcd·m²). -2 ·lx -1 Substituting this into the mathematical model of white road markings, the average grayscale value corresponding to the minimum retroreflective brightness coefficient required to ensure road safety is 43.712. If the average grayscale of the marking area in the image to be inspected is lower than 43.712, it is judged as unqualified, the location information is recorded, and a repair is reported; if the average grayscale of the marking area in the image to be inspected is not lower than 43.712, it is judged as qualified, and normal maintenance and periodic inspection are carried out.

[0182] The retroreflectance value is 80 (mcd·m -2 ·lx -1 A uniformly reflective white marking board can be used as a white standard calibration board, with a retroreflection value of 50 (mcd·m). -2 ·lx -1 A uniformly reflective yellow marking board can be used as a yellow standard calibration board.

[0183] The uniformity of retroreflective brightness of the markings is included in the evaluation index.

[0184] Subsequently, data measurements at multiple test points can reflect the uniformity of the retroreflective brightness coefficient of the road markings to a certain extent, ensuring that the reflective performance of the road markings is relatively consistent at different locations, thereby guaranteeing the safety and visibility of nighttime driving.

[0185] Understandably, even on the same road, different road conditions and drivers' habits can lead to varying degrees of wear and tear on road markings. Common problems include: partial damage to the markings, markings appearing intact but with most of the glass beads detached, and markings covered by foreign objects. All of these can cause the marking segment to lose its reflectivity or exhibit uneven reflectivity. This characteristic manifests in the image as the uniformity of grayscale values ​​in the area being inspected. Only a small portion of the area reflects light normally, which, especially at higher speeds, fails to provide sufficient safety.

[0186] Based on the grayscale distribution of the area to be tested, the uniformity of grayscale values ​​of traffic markings can also be determined using the method described above for calculating the retroreflection coefficient of traffic signs. This will not be elaborated upon further here. The traffic markings are evaluated by combining the retroreflection luminance coefficient and the uniformity of grayscale values.

[0187] Furthermore, this embodiment also provides a traffic safety facility night vision evaluation system, the system comprising:

[0188] The road marking detection camera is mounted on the detection vehicle via a first stabilizer and is used to acquire images of the road markings to be evaluated.

[0189] A traffic sign detection camera, mounted on the detection vehicle via a second stabilizer, is used to capture images of traffic signs to be evaluated.

[0190] Radar, installed on the inspection vehicle, is used to collect point cloud data of traffic signs;

[0191] The controller is connected to the road marking detection camera, traffic sign detection camera and radar respectively, and is used to send the road marking image to be evaluated, the traffic sign image to be evaluated and the traffic sign point cloud data to the host computer.

[0192] The host computer, connected to the controller, includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the above-described method for evaluating the night visibility of traffic safety facilities.

[0193] Specifically, refer to Figure 2The system consists of two parts: data acquisition and data processing. Data acquisition comprises a host computer, road marking detection cameras, sign detection cameras, a controller, a stabilizer, a distance sensor, and a LiDAR. Road marking detection cameras acquire road marking images and send the data to the host computer via Ethernet. Sign detection cameras acquire sign images and send them to the host computer via Ethernet. The distance sensor sends data to the controller via I / O ports. The LiDAR acquires point cloud data of signs above and to the right of the road and sends the data to the host computer via Ethernet. The stabilizer stabilizes the cameras to prevent image shake during vehicle movement. The controller receives information from the distance sensor and sends the distance information to the host computer via Ethernet. Data transmission is not limited to Ethernet; USB or wireless data transmission can also be used. The host computer receives and categorizes the acquired data for later use by the data processing system. The data processing unit is deployed on the host computer to execute the aforementioned night vision evaluation method for traffic safety facilities.

[0194] It should be noted that all the examples above are only for understanding this application and do not constitute a limitation on the night visibility evaluation method of traffic safety facilities in this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0195] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0196] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0197] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the night visibility evaluation method for traffic safety facilities in the above embodiments.

[0198] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0199] The aforementioned computer-readable storage medium may be included in the night vision evaluation system for traffic safety facilities; or it may exist independently and not be installed in the night vision evaluation system for traffic safety facilities.

[0200] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0201] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0202] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0203] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described method for evaluating the night visibility of traffic safety facilities. This solves the technical problem of low reliability in night visibility evaluation of traffic safety facilities in related technologies. Compared with related technologies, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the night visibility evaluation method for traffic safety facilities provided in the above embodiments, and will not be repeated here.

[0204] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the night visibility evaluation method for traffic safety facilities as described above.

[0205] The computer program product provided in this application can solve the technical problem of low reliability in the night vision evaluation of traffic safety facilities in related technologies. Compared with related technologies, the beneficial effects of the computer program product provided in this application are the same as those of the night vision evaluation method for traffic safety facilities provided in the above embodiments, and will not be repeated here.

[0206] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for evaluating the night visibility of traffic safety facilities, characterized in that, The method includes: Obtain images of the traffic safety facilities to be evaluated; Determine the type of the traffic safety facility image to be evaluated. If the traffic safety facility image to be evaluated is a traffic sign image to be evaluated, then determine the retroreflection coefficient of each color region of the traffic sign in the traffic sign image to be evaluated. Based on the retroreflection coefficient and the uniformity of gray values ​​in the specific distribution area of ​​the traffic sign, the night visibility of the traffic sign is evaluated to obtain the night visibility evaluation result of the traffic sign. If the image of the traffic safety facility to be evaluated is an image of traffic markings to be evaluated, then determine the retroreflection brightness coefficient of the traffic markings in the image of the traffic markings to be evaluated and the uniformity of gray values ​​of the specific image area where the traffic markings are located. Based on the retroreflective brightness coefficient and the uniformity of gray values ​​in the specific distribution area of ​​the traffic markings, the night visibility of the traffic markings is evaluated to obtain the night visibility evaluation results of the traffic markings. The specific distribution area of ​​the traffic signs includes the various color areas of the traffic signs within the specific distribution area; The step of determining the retroreflection coefficient of each color region of the traffic sign in the image to be evaluated includes: The location of the traffic sign in the image of the traffic sign to be evaluated is identified, and a bounding box is used to mark it. It is then determined whether the traffic sign to be evaluated is within the evaluation range; the evaluation range includes the distance between the traffic sign in the image of the traffic sign to be evaluated and the detection vehicle. If the traffic sign to be evaluated is within the evaluation range, then the image area within the sign frame is subjected to color segmentation processing to determine each color area of ​​the traffic sign in the image of the traffic sign to be evaluated; Based on the grayscale values ​​of each color region, the retroreflection coefficient of the traffic sign is calculated using the corresponding retroreflection coefficient calculation model. The step of performing color segmentation processing on the image region within the identification frame to determine each color region of the traffic sign in the traffic sign image to be evaluated includes: The image region within the defined bounding box is input into the traffic sign color segmentation model to determine the color regions of the traffic sign in the image to be evaluated. The traffic sign color segmentation model includes an input layer, a hidden layer, and an output layer. The input layer includes a color space conversion unit and a color feature extraction unit. The color space conversion unit is used to convert the image region from RGB to LAB space. The color feature extraction unit includes a 1×1 convolutional layer and a 2×2 average pooling layer. The 1×1 convolutional layer is used to extract local color association features in LAB space, and the 2×2 average pooling layer is used to reduce the resolution of the image region. The hidden layer includes a first fully connected layer, a second fully connected layer, and a third fully connected layer. The SENet channel attention mechanism is embedded between the first fully connected layer and the second fully connected layer, and a residual branch is introduced between the second fully connected layer and the third fully connected layer. The output layer includes six output neurons, each of which is activated by a Softmax function to output the color category probability of a traffic sign.

2. The method for evaluating the night visibility of traffic safety facilities as described in claim 1, characterized in that, The step of evaluating the night visibility of the traffic sign based on the retroreflection coefficient and the uniformity of gray values ​​in the specific distribution area of ​​the traffic sign, and obtaining the night visibility evaluation result of the traffic sign, includes: For each color region in the specific distribution area of ​​the traffic sign, the color region is divided into several small regions of equal area, and the standard deviation of the gray value of each small region is calculated to obtain a uniform dataset. Calculate the mean and standard deviation of the homogeneous dataset, and determine the coefficient of variation based on the mean and standard deviation; The uniformity of grayscale values ​​in each color region of the traffic sign is determined based on the coefficient of variation. Based on the range of grayscale uniformity and the range of retroreflection coefficient, the night visibility of the traffic sign is evaluated to obtain the night visibility evaluation result of the traffic sign. The steps of calculating the mean and standard deviation of the uniformity dataset, determining the coefficient of variation based on the mean and standard deviation, and determining the uniformity of gray values ​​of each color region of the traffic sign based on the coefficient of variation include: Calculate the mean of the homogeneous dataset: µ=(x1+x2+…+x n ) / n Where μ is the mean of the uniform dataset, x n This represents a sample in a uniform dataset, where n is the number of data points. Calculate the standard deviation: Where σ is the standard deviation of the uniform dataset, μ is the mean of the uniform dataset, and n is the number of data points; Calculate the coefficient of variation: C V =s / m Calculate the uniformity score to obtain the uniformity of the grayscale values: U = 1 - C V Where U is the uniformity score.

3. The method for evaluating the night visibility of traffic safety facilities as described in claim 1, characterized in that, The step of identifying the position of the traffic sign in the image of the traffic sign to be evaluated, marking it with an identification box, and determining whether the traffic sign to be evaluated is within the evaluation range by the identification box includes: Traffic sign dot cloud data is obtained by detecting radar installed on the vehicle; Identify traffic signs in the traffic sign image to be evaluated, mark them with an identification box, and determine the pixel coordinates of the traffic signs within the identification box; Based on the pixel coordinates of the traffic sign and the point cloud data of the traffic sign, a radar-visual fusion ranging is performed to determine whether the distance between the traffic sign in the image to be evaluated and the detection vehicle is within the evaluation range.

4. The method for evaluating the night visibility of traffic safety facilities as described in claim 3, characterized in that, The step of identifying traffic signs in the traffic sign image to be evaluated and marking them with bounding boxes includes: The traffic sign image to be evaluated is input into the traffic sign recognition model for recognition, the position of the traffic sign in the traffic sign image to be evaluated is determined, and it is marked with an identification box; The traffic sign recognition model comprises a feature extraction unit, a feature fusion unit, and a detection output unit. The feature extraction unit uses a backbone network with CSPDarknet-53 as the main backbone and RetinexNet as a lightweight branch, and the backbone network also embeds a CoAtNet attention enhancement mechanism. The feature fusion unit employs a BiFPN bidirectional feature pyramid network. The detection output unit includes a classification head and a regression head, and uses the CIoU loss function. The classification head consists of three 1×1 convolutional layers and a Softmax function, while the regression head consists of three 1×1 convolutional layers and a Sigmoid function.

5. The method for evaluating the night visibility of traffic safety facilities as described in claim 1, characterized in that, The retroreflection coefficient calculation model is determined by fitting the retroreflection coefficients of multiple calibration plates and the corresponding average gray level of the calibration plates.

6. The method for evaluating the night visibility of traffic safety facilities as described in claim 1, characterized in that, The traffic marking image to be evaluated is acquired by a marking detection camera at a fixed angle on the detection vehicle. The fixed angle of the marking detection camera is determined based on the brightness distribution of the high beams of the detection vehicle.

7. A night vision evaluation system for traffic safety facilities, characterized in that, The system includes: The road marking detection camera is mounted on the detection vehicle via a first stabilizer and is used to acquire images of the road markings to be evaluated. A traffic sign detection camera, mounted on the detection vehicle via a second stabilizer, is used to capture images of traffic signs to be evaluated. Radar, installed on the inspection vehicle, is used to collect point cloud data of traffic signs; The controller is connected to the road marking detection camera, the traffic sign detection camera and the radar respectively, and is used to send the road marking image to be evaluated, the traffic sign image to be evaluated and the traffic sign point cloud data to the host computer. A host computer, connected to the controller, includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the night visibility evaluation method for traffic safety facilities as described in any one of claims 1 to 6.

Images