Road surface illumination detection system and method based on machine learning and visual identification

Abstract

The invention provides a road surface illumination detection system and method based on machine learning and visual identification. The system comprises: an image acquisition module obtains an image to be analyzed by using a machine learning method; the semantic segmentation module performs semantic segmentation on the to-be-analyzed image to obtain a road region image; and the illumination prediction module performs multi-dimensional feature extraction by using a feature extraction model, and inputs the extracted image features into a self-defined multi-layer neural network regression model for training and prediction to obtain the average illumination and illumination uniformity of the road surface. The method comprises the following steps: S1, acquiring a road image in real time; s2, performing semantic segmentation on the to-be-analyzed image to obtain a road region image; and S3, inputting the extracted image features into a self-defined multi-layer neural network regression model for training and prediction to obtain the average illumination and the illumination uniformity of the road surface. According to the method, the existing road monitoring system or image acquisition equipment is utilized, and an advanced machine learning algorithm is combined, so that efficient and accurate road surface illumination detection is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of road surface illumination testing, and in particular to a road surface illumination detection system and method based on machine learning and visual recognition. Background Art

[0002] At present, road illumination detection mainly adopts traditional methods, which have many problems.

[0003] In terms of procedures, traditional road illumination detection methods often require the installation of special equipment such as illuminance meters or reflectivity boards. The installation process requires professional operation, and sometimes the road has to be closed. When using reflectivity boards, the position and angle of the stickers must be precisely adjusted. Some methods also rely on vehicles or drones to collect data, which makes the operation complicated.

[0004] In terms of real-time performance, many traditional methods cannot detect in real time and often need to be measured in a specific time period or environment, such as when there are no cars or during nighttime aerial photography. Data collection and processing take a long time and cannot reflect the road's immediate illumination conditions in a timely manner, and cannot meet the needs of dynamic lighting adjustment application scenarios such as smart city lighting.

[0005] In terms of cost, traditional methods require the purchase of expensive measuring equipment, such as high-precision illuminance meters, drones, or special reflectivity panels, which are costly and require regular maintenance and calibration. The installation and maintenance of these devices require professionals, which has high labor costs, and equipment wear and tear and updates also require additional investment. Some methods also require the installation of physical markers or sensors on the road surface, which have high construction and maintenance costs.

[0006] Patent document CN202311351123.5 discloses a road lighting illuminance test method based on a standard reflectivity plate. The patent measures the illuminance of the standard reflectivity plate under different light outputs of dimming lamps and obtains the illuminance of the tested road section by fitting. However, the patent has the disadvantages that the standard reflectivity sticker may cause glare, needs to be replaced and maintained regularly, and has high cost for large-scale laying.

[0007] Patent document CN202310974104.1 discloses a method for detecting illumination of aerial road lighting, which is achieved by obtaining aerial data of road lighting, including road lighting images and positioning information, dividing the road lighting images into grids, and constructing a neural network model to evaluate illumination. However, this patent requires blocking roads, unstable night flights, and safety risks, and the equipment is expensive and the data processing is complex.

[0008] In summary, the existing road illumination detection method has complex procedures, poor real-time performance and high cost, and a new road illumination detection method and system is urgently needed. Summary of the invention

[0009] In view of the defects in the prior art, the purpose of the present invention is to provide a road illumination detection method and system based on machine learning and visual recognition.

[0010] A road illumination detection system based on machine learning and visual recognition provided by the present invention includes:

[0011] An image acquisition module is used to acquire road images in real time, use machine learning methods to detect interference, and obtain images to be analyzed;

[0012] Semantic segmentation module, which performs semantic segmentation on the image to be analyzed to obtain a road area image;

[0013] The illumination prediction module preprocesses the road area image, then uses the feature extraction model to extract multi-dimensional features, and inputs the extracted image features into a custom multi-layer neural network regression model for training and prediction to obtain the average illumination and illumination uniformity of the road surface.

[0014] Preferably, in the image acquisition module, the road image is acquired in real time through an existing road monitoring system or image acquisition equipment.

[0015] Preferably, in the image acquisition module, a road image is selected every second, and interference items are detected on the road image using a machine learning method. If a vehicle interference item is detected, the road image is discarded; if no vehicle interference item is detected, the road image is retained, and the time point and detection result are marked as the image to be analyzed.

[0016] Preferably, in the semantic segmentation module, a machine learning model is used or a region of interest is manually set according to the image acquisition position to perform semantic segmentation on the image to be analyzed, identify road areas and non-road areas, and obtain a road area image by processing the non-road area.

[0017] The present invention also provides a road illumination detection method based on machine learning and visual recognition, comprising the following steps:

[0018] Step S1, acquiring a road image in real time, using a machine learning method to perform interference detection, and obtaining an image to be analyzed;

[0019] Step S2, performing semantic segmentation on the image to be analyzed to obtain a road area image;

[0020] Step S3, preprocess the road area image, then use the feature extraction model to perform multi-dimensional feature extraction, input the extracted image features into a custom multi-layer neural network regression model for training and prediction, and obtain the average illumination and illumination uniformity of the road surface.

[0021] Preferably, step S1 includes the following sub-steps:

[0022] Step M1: Acquire road images in real time through an existing road monitoring system or image acquisition equipment, by selecting one road image per second, and each road image includes an acquisition time point;

[0023] Step M2: According to the requirements of the YOLOv8 model for the input image, the collected road image is resized and the resolution is set, followed by color space conversion, normalization, and standardization to obtain a preprocessed road image;

[0024] Step M3: Use the pre-trained YOLOv8 target detection model to detect interference items on the preprocessed road image. If a vehicle interference item is detected in the road image, the current road image is annotated and removed, and the annotated label is: time point + vehicle passing; if no vehicle interference item is detected, the current road image is annotated and output as the image to be analyzed, and the annotated label is: time point + no vehicle passing.

[0025] Preferably, step S2 adopts the following steps:

[0026] Step U1: construct a segmentation model to distinguish the road area and non-road area in the road image;

[0027] Step U2: The image to be analyzed is provided as input to the segmentation model. The segmentation model performs semantic segmentation on the analysis image to identify the road area and non-road area. The identified non-road area is covered with a white mask to obtain a road area image containing only the road area.

[0028] Preferably, step S2 adopts the following steps: if the position and shooting angle of the camera are fixed, a manual road area segmentation method is adopted to manually delineate the road area as the area of ​​interest, and the non-road area is covered with a white mask to obtain a road area image containing only the road area.

[0029] Preferably, constructing the segmentation model includes the following sub-steps:

[0030] Step P1: Use a designated camera on a designated urban road to collect road images during the time when street lights are on, and then use the labelme tool to perform pixel-level annotation on the road area and non-road area in the collected road images to generate a corresponding annotation map;

[0031] Step P2: The size of the annotation image is uniformly adjusted to the size of the model input image, and then data enhancement is used to increase the diversity of the image to obtain a preprocessed image set. The data enhancement method includes rotating the image at multiple angles, adjusting the brightness and color of the image, and cropping the image.

[0032] Step P3: Select the DeepLabV3+ model, load the pre-trained MobileNetV2 as the backbone network, and optimize the model using your own annotated dataset;

[0033] Step P4: Divide the preprocessed image set into a ratio of 8:2, with 80% of the data used as the training set and 20% of the data used as the validation set. After training the model on the training set, the performance of the trained model is evaluated on the validation set. The accuracy of the model is measured by calculating the average intersection-over-union ratio indicator. At the same time, road images that were not involved in the training are used to further verify the effect of the model.

[0034] Preferably, step S3 includes the following sub-steps:

[0035] Step R1: Use the midpoint method to measure the illuminance value of each point on the specified road section, and calculate the average illuminance and illuminance uniformity of the road section. Collect and shoot road images during the time when street lights are on, and make a one-to-one correspondence between road images and road lighting data to create a road lighting dataset. At the same time, use data enhancement technology to expand the road lighting dataset.

[0036] Step R2: remove the white area in the road area image, then convert it into the HSV color space, and then resize and standardize the converted image to obtain a preprocessed road area image;

[0037] Step R3: Input the preprocessed road area image into the ResNet50 model, extract the high-level features of the image through multi-layer convolution and pooling operations, and perform adaptive average pooling on the extracted feature map to obtain the extracted image features;

[0038] Step R4: Input the extracted image features into the regression model of the custom multi-layer neural network. The regression model is constructed by a custom multi-layer perceptron (MLP). The multi-layer perceptron includes multiple linear regression layers, batch normalization layers, ReLU activation function layers, and Dropout layers. The regression model trains and predicts the extracted image features to obtain the average illumination and illumination uniformity of the road.

[0039] Step R5: Train the regression model. First, create a dataset and a data loader, and divide the dataset into a training set and a test set. Then initialize the regression model, loss function, and optimizer. Optimize the parameters of the regression model through the training data. Use the mean square error (MSE) as the loss function. Perform back propagation and weight update through the Adam optimizer. The training process includes setting the learning rate, batch size, and training rounds to obtain the trained regression model.

[0040] Step R6: Divide the road lighting dataset into a training set and a validation set with a ratio of 8:2. Evaluate the performance of the trained regression model on the validation set by calculating the mean square error (MSE) between the predicted value and the true value to ensure that the model performs well on the validation set and verify the performance of the regression model.

[0041] Step R7: Input the road image to be detected into the trained regression model, predict the average illumination and illumination uniformity, output the prediction results and record them, and at the same time, calculate the error between the predicted value and the actual value for subsequent analysis and practical application.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. In terms of process, the present invention makes full use of existing road monitoring systems or other image acquisition equipment, without the need to install additional special equipment, reducing the workload of equipment installation and debugging. A machine learning model is used for automated processing to automatically detect vehicle passage conditions and select appropriate image frames for analysis, reducing reliance on manual operations and simplifying the operating process. Illumination detection is achieved through software processing without the need for physical changes to the road or installation of additional equipment, reducing operational complexity and thereby improving detection efficiency. The present invention uses semantic segmentation technology to accurately distinguish between road areas and other areas, thereby improving the accuracy and effectiveness of data processing. The present invention uses preprocessing and standardization processing to reduce the impact of environmental changes, improve the diversity, consistency and accuracy of feature extraction, and use a custom multi-layer neural network model for high-precision illumination prediction, which can output key indicators of road illumination in real time and provide reliable illumination data support.

[0044] 2. In terms of real-time performance, the present invention acquires road images in real time and uses machine learning models for real-time analysis to achieve real-time road illumination detection. The continuously updated image data reflects the real-time road illumination at any time. It automatically detects vehicle passing conditions and selects images without vehicle passing for processing to ensure data accuracy and real-time performance. It can detect anytime and anywhere without waiting for a specific time period or blocking roads. It helps to dynamically adjust the intensity of street lamp lighting, optimize energy use, and improve road safety and driving comfort.

[0045] 3. In terms of cost, the present invention makes full use of existing monitoring systems or image acquisition equipment, without the need to purchase additional expensive equipment, thus reducing hardware costs. Illuminance detection is achieved through software algorithms, reducing dependence on physical equipment. Performance is improved through software upgrades and model updates, reducing long-term use costs. There is no need to make physical changes or install markings on the road surface, thus avoiding construction and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0047] Figure 1 is a schematic diagram of a real-time road illumination detection system based on machine learning and visual recognition in an embodiment of the present invention;

[0048] Figure 2 A schematic diagram of interference item detection and frame selection steps in an embodiment of the present invention;

[0049] Figure 3 Schematic diagram of the MobileNetV2-DeeplabV3+ model in an embodiment of the present invention;

[0050] Figure 4 Schematic diagram of a prediction network model in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0052] Terminology explanation:

[0053] 1. Semantic segmentation: A task in the field of computer vision, which aims to assign each pixel in an image to a specific category, thereby achieving accurate recognition and segmentation of different objects or regions in the image and marking the precise boundaries of the objects.

[0054] 2. YOLO model: YOLO (You Only Look Once) is a real-time target detection system.

[0055] 3. MobileNetV2-DeeplabV3+ model: The MobileNetV2-DeeplabV3+ model is a combination of MobileNetV2 as the backbone network and DeeplabV3+ as the semantic segmentation architecture.

[0056] 4. MLP: MLP (Multilayer Perceptron) is a feedforward neural network. MLP realizes the mapping from input to output through a multi-layer neuron structure and is widely used in classification, regression and other tasks.

[0057] The present invention discloses a real-time road surface illumination detection system and method based on machine learning and visual recognition, aiming to utilize existing road monitoring systems or image acquisition equipment in combination with advanced machine learning algorithms to achieve efficient and accurate road surface illumination detection.

[0058] Embodiment 1:

[0059] Figure 1 Schematic diagram of a real-time road illumination detection system based on machine learning and visual recognition in an embodiment of the present invention.

[0060] like Figure 1 As shown, an embodiment of the present invention provides a road illumination detection system based on machine learning and visual recognition, including: an image acquisition module, a semantic segmentation module and an illumination prediction module.

[0061] The image acquisition module is used to acquire road images in real time, use machine learning methods to detect interference, and obtain images to be analyzed.

[0062] The road images can be acquired in real time through the existing road monitoring system or image acquisition equipment, directly based on the already installed shooting equipment.

[0063] In this embodiment, road images are acquired by an image acquisition device, which is a camera. The best effect is achieved by the camera looking down at the road surface at an angle of 30 to 45 degrees. In order to ensure the stability and consistency of image quality, the automatic enhancement function of the camera is turned off, and the video stream picture is set to be output one picture per second or one picture per period of time.

[0064] Furthermore, by selecting a road image every second, a machine learning method is used to detect interference items in the road image. If a vehicle interference item is detected, the road image is discarded. If no vehicle interference item is detected, the road image is retained, and the standard time point and detection result are used as the image to be analyzed.

[0065] The semantic segmentation module performs semantic segmentation on the image to be analyzed to obtain a road area image.

[0066] Specifically, a machine learning model is used or a region of interest is manually set according to the image acquisition location to perform semantic segmentation on the image to be analyzed, identify road areas and non-road areas, and obtain a road area image by processing the non-road area.

[0067] In this embodiment, MobileNetV2-DeeplabV3+ is used to implement image segmentation and identify road areas and non-road areas.

[0068] The illumination prediction module preprocesses the road area image, then uses the feature extraction model to extract multi-dimensional features, and inputs the extracted image features into a custom multi-layer neural network regression model for training and prediction to obtain the average illumination and illumination uniformity of the road surface.

[0069] In this embodiment, ResNet50 is used to extract features from the preprocessed road area image.

[0070] Embodiment 2:

[0071] The embodiment of the present invention provides a road illumination detection method based on machine learning and visual recognition, comprising the following steps:

[0072] Step S1, acquiring a road image in real time, using a machine learning method to perform interference detection, and obtaining an image to be analyzed.

[0073] Figure 2 Schematic diagram of interference item detection and frame selection steps in an embodiment of the present invention.

[0074] like Figure 2 As shown, step S1 includes the following sub-steps:

[0075] Step M1: Image acquisition: Specifically, the road image is acquired in real time through the existing road monitoring system or image acquisition equipment, and one road image is selected every second, and each road image includes the acquisition time point, so as to establish a visualization platform later.

[0076] Step M2: Image preprocessing. Specifically, according to the requirements of the YOLOv8 model for the input image, the collected road image is resized and the resolution is set, followed by color space conversion, normalization, and standardization to obtain a preprocessed road image.

[0077] Step M3: Interference item detection and labeling. Specifically, the pre-trained YOLOv8 target detection model is used to detect interference items on the pre-processed road image. If a vehicle interference item is detected in the road image, the current road image is labeled and removed, and the labeled label is: time point + vehicle passing; if no vehicle interference item is detected, the current road image is labeled and output as the image to be analyzed, and the labeled label is: time point + no vehicle passing.

[0078] Specifically, the YOLOv8 model has efficient detection capabilities and can quickly identify distractors such as vehicles, pedestrians, and animals in images. The model outputs detection results, including the bounding box coordinates and category labels of each distractor. By identifying and labeling distractors in the image, it provides necessary information for subsequent processing.

[0079] Step S2, performing semantic segmentation on the image to be analyzed to obtain a road area image.

[0080] In this embodiment, MobileNetV2-DeeplabV3+ is used to achieve image segmentation. First, the deep network is deployed, features are extracted through MobileNetV2, multi-scale information is captured through the hole convolution module, and then the fusion feature map is upsampled to restore the resolution to generate the final semantic analysis result. This embodiment is based on the retraining of the pre-trained model to ensure its good generalization ability.

[0081] Figure 3 Schematic diagram of the MobileNetV2-DeeplabV3+ model in an embodiment of the present invention.

[0082] like Figure 3 As shown, this embodiment adopts the MobileNetV2-DeeplabV3+ model, and the MobileNetV2-DeeplabV3+ network adopted by this model is a lightweight semantic segmentation model. MobileNetV2 Atrous Conv uses the method of deep separable convolution to decompose the image into smaller feature blocks, which is mainly used to extract the underlying features of the image; then the atrous convolution method is used to capture contextual information of different scales through multi-scale feature extraction, that is, to simultaneously fuse the image after feature extraction of multiple convolution layers and image pooling layers with different expansion rates; then through the splicing operation, that is, Concat fuses the deep separable convolution image and the atrous convolution image to fuse the multi-scale features; finally, upsampling is used, that is, Upsample restores the image to the original image resolution to complete the segmentation task.

[0083] Specifically, step S2 adopts the following steps:

[0084] Step U1: Construct a segmentation model to distinguish road areas and non-road areas in road images.

[0085] Specifically, building a segmentation model includes the following sub-steps:

[0086] Step P1: Data collection and labeling. Specifically, a designated camera is used on a designated urban road to collect road images during the time when street lights are on. Then, the labelme tool is used to label the road areas and non-road areas in the collected road images at the pixel level to generate the corresponding labeling map.

[0087] Step P2: Data preprocessing. Specifically, the size of the annotation image is uniformly adjusted to the size of the model input image, and then data enhancement is used to increase the diversity of the image to obtain the preprocessed image set.

[0088] Data augmentation methods include rotating images at multiple angles, adjusting image brightness and color, and cropping images to increase the robustness of the model.

[0089] Step P3: Model training. Specifically, select the DeepLabV3+ model, load the pre-trained MobileNetV2 as the backbone network, and use your own annotated dataset to optimize the model.

[0090] Specifically, the model convergence speed is accelerated by loading the pre-trained model, the initial performance of the model is improved, and then the model is optimized and adjusted on its own labeled data set. Specifically, the performance of the model is improved by configuring the training parameters (learning rate, batch size, training rounds, etc.).

[0091] Step P4: Model verification. Specifically, the preprocessed image set is divided into 8:2 ratios, with 80% of the data used as the training set and 20% of the data used as the validation set. After the model is trained on the training set, the performance of the trained model is evaluated on the validation set. The accuracy of the model is measured by calculating the mean intersection-over-union (MIOU) indicator, and the effect of the model is further verified using road images that were not involved in the training.

[0092] Preferably, the average intersection-over-union ratio index is greater than 0.95.

[0093] Step U2: semantic segmentation and result output. Specifically, the image to be analyzed is provided as input to the segmentation model, and the segmentation model performs semantic segmentation on the analysis image to identify the road area and non-road area. The identified non-road area is covered with a white mask to obtain a road area image containing only the road area.

[0094] Step S2 can also adopt the following steps: if the position and shooting angle of the camera are fixed, the road area is manually segmented, the road area is manually demarcated as the area of ​​interest, and the non-road area is covered with a white mask to obtain a road area image containing only the road area. This method can improve the speed of semantic segmentation of the road area and serve as an alternative method for MobileNetV2-DeeplabV3+ to achieve image segmentation.

[0095] Step S3, preprocess the road area image, then use the feature extraction model to perform multi-dimensional feature extraction, input the extracted image features into a custom multi-layer neural network regression model for training and prediction, and obtain the average illumination and illumination uniformity of the road surface.

[0096] In this embodiment, ResNet50 and a custom multi-layer perceptron regression model are used to achieve illumination prediction. First, image preprocessing is performed, and then image features are extracted through the ResNet50 model to capture important information in the image. Finally, these features are input into a custom multi-layer fully connected neural network to generate prediction results, including average illumination and illumination uniformity as indicators of road lighting conditions.

[0097] Figure 4 Schematic diagram of a prediction network model in an embodiment of the present invention.

[0098] like Figure 4 As shown, the prediction network model used in this embodiment first receives the input RGB image and converts it into an HSV image and preprocesses the image, and then uses the convolution layer (CONV), pooling layer (MAXPOOL) and residual block (BTNK) of ResNet50 to extract the underlying features and high-level features of the image. The image after feature extraction is adjusted to a fixed size through an adaptive average pooling layer. Then, these features are input into a multi-layer fully connected neural network, each layer of which contains linear transformation, batch normalization (BN), ReLU activation (RL) and Dropout, and finally outputs the predicted average illumination and illumination uniformity.

[0099] Specifically, step S3 includes the following sub-steps:

[0100] Step R1: Use the midpoint method to measure the illuminance value of each point on the specified road section, and calculate the average illuminance and illuminance uniformity of the road section. Collect and shoot road images during the time period when street lights are on, and make a road lighting dataset by matching the road images and road lighting data one by one. At the same time, use data enhancement technology to expand the road lighting dataset to make the dataset more diverse and representative, thereby improving the robustness and generalization ability of the model.

[0101] Step R2: Remove the white area in the road area image, then convert it into the HSV color space, and then resize and standardize the converted image to obtain a preprocessed road area image.

[0102] Step R3: Input the preprocessed road area image into the ResNet50 model, extract the high-level features of the image through multi-layer convolution and pooling operations, perform adaptive average pooling on the extracted feature map, and then scale the feature map to a fixed size to obtain the extracted image features.

[0103] Step R4: Input the extracted image features into the regression model of the custom multi-layer neural network. The regression model is constructed by a custom multi-layer perceptron (MLP). The multi-layer perceptron includes multiple linear regression layers, batch normalization layers, ReLU activation function layers, and Dropout layers. The regression model trains and predicts the extracted image features to obtain the average illumination and illumination uniformity of the road.

[0104] Step R5: Train the regression model. First, create a dataset and a data loader, and divide the dataset into a training set and a test set. Then initialize the regression model, loss function, and optimizer. Optimize the parameters of the regression model through the training data. Use the mean square error (MSE) as the loss function. Perform back propagation and weight update through the Adam optimizer. The training process includes setting the learning rate, batch size, and training rounds to obtain the trained regression model.

[0105] Step R6: Divide the road lighting dataset into a training set and a validation set with a ratio of 8:2. Evaluate the performance of the trained regression model on the validation set by calculating the mean square error (MSE) between the predicted value and the true value to ensure that the model performs well on the validation set and verify the performance of the regression model.

[0106] Step R7: Input the road image to be detected into the trained regression model, predict the average illumination and illumination uniformity, output the prediction results and record them, and at the same time, calculate the error between the predicted value and the actual value for subsequent analysis and practical application.

[0107] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0108] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A road illumination detection system based on machine learning and visual recognition, characterized in that: include: An image acquisition module is used to acquire road images in real time, use machine learning methods to detect interference, and obtain images to be analyzed; A semantic segmentation module performs semantic segmentation on the image to be analyzed to obtain a road area image; The illumination prediction module preprocesses the road area image, then uses a feature extraction model to extract multi-dimensional features, inputs the extracted image features into a custom multi-layer neural network regression model for training and prediction, and obtains the average illumination and illumination uniformity of the road surface.

2. A road illumination detection system based on machine learning and visual recognition according to claim 1, characterized in that: In the image acquisition module, the road image is acquired in real time through an existing road monitoring system or image acquisition equipment.

3. A road illumination detection system based on machine learning and visual recognition according to claim 2, characterized in that: In the image acquisition module, a road image is selected every second, and interference items are detected on the road image using a machine learning method. If a vehicle interference item is detected, the road image is discarded; if no vehicle interference item is detected, the road image is retained, and the standard time point and detection result are used as the image to be analyzed.

4. The road illumination detection system based on machine learning and visual recognition according to claim 1, characterized in that: In the semantic segmentation module, a machine learning model is used or a region of interest is manually set according to the image acquisition position to perform semantic segmentation on the image to be analyzed, identify road areas and non-road areas, and obtain a road area image by processing the non-road area.

5. A road illumination detection method based on machine learning and visual recognition, characterized in that: The steps include: Step S1, acquiring a road image in real time, using a machine learning method to perform interference detection, and obtaining an image to be analyzed; Step S2, performing semantic segmentation on the image to be analyzed to obtain a road area image; Step S3, preprocessing the road area image, then using a feature extraction model to perform multi-dimensional feature extraction, inputting the extracted image features into a custom multi-layer neural network regression model for training and prediction, and obtaining the average illumination and illumination uniformity of the road surface.

6. The method for detecting road illumination based on machine learning and visual recognition according to claim 5, characterized in that: The step S1 includes the following sub-steps: Step M1: Acquire road images in real time through an existing road monitoring system or image acquisition equipment, by selecting one road image per second, and each road image includes an acquisition time point; Step M2: According to the requirements of the YOLOv8 model for the input image, the collected road image is resized and the resolution is set, followed by color space conversion, normalization, and standardization to obtain a preprocessed road image; Step M3: Use the pre-trained YOLOv8 target detection model to detect interference items on the preprocessed road image. If a vehicle interference item is detected in the road image, the current road image is marked and removed, and the marked label is: time point + vehicle passing; if no vehicle interference item is detected, the current road image is marked and output as the image to be analyzed, and the marked label is: time point + no vehicle passing.

7. The method for detecting road illumination based on machine learning and visual recognition according to claim 5, characterized in that: The step S2 adopts the following steps: Step U1: construct a segmentation model to distinguish the road area and non-road area in the road image; Step U2: The image to be analyzed is provided as input to the segmentation model, and the segmentation model performs semantic segmentation on the analysis image to identify road areas and non-road areas. The identified non-road areas are covered with a white mask to obtain a road area image containing only the road area.

8. The method for detecting road illumination based on machine learning and visual recognition according to claim 5, characterized in that: The step S2 adopts the following steps: if the position and shooting angle of the camera are fixed, a method of manually segmenting the road area is adopted, the road area is manually demarcated as the area of ​​interest, and the non-road area is covered with a white mask to obtain a road area image containing only the road area.

9. The method for detecting road illumination based on machine learning and visual recognition according to claim 7, characterized in that: The construction of the segmentation model includes the following sub-steps: Step P1: Use a designated camera on a designated urban road to collect road images during the time when street lights are on, and then use the labelme tool to perform pixel-level annotation on the road area and non-road area in the collected road images to generate a corresponding annotation map; Step P2: uniformly adjusting the size of the labeled image to the size of the model input image, and then using data enhancement to increase the diversity of the image to obtain a preprocessed image set. The data enhancement method includes rotating the image at multiple angles, adjusting the brightness and color of the image, and cropping the image. Step P3: Select the DeepLabV3+ model, load the pre-trained MobileNetV2 as the backbone network, and optimize the model using your own annotated dataset; Step P4: The preprocessed image set is divided into a ratio of 8:2, wherein 80% of the data is used as a training set and 20% of the data is used as a validation set. After the model is trained on the training set, the performance of the trained model is evaluated on the validation set. The accuracy of the model is measured by calculating the average intersection-over-union ratio indicator. At the same time, road images that were not involved in the training are used to further verify the effect of the model.

10. The method for detecting road illumination based on machine learning and visual recognition according to claim 5, characterized in that: The step S3 includes the following sub-steps: Step R1: Use the midpoint method to measure the illuminance value of each point on the specified road section, calculate the average illuminance and illuminance uniformity of the road section, collect and shoot road images during the time period when street lights are on, match the road images with road lighting data one by one, create a road lighting dataset, and use data enhancement technology to expand the road lighting dataset; Step R2: removing the white area in the road area image, then converting it into the HSV color space, and then resizing and standardizing the converted image to obtain a preprocessed road area image; Step R3: Input the preprocessed road area image into the ResNet50 model, extract high-level features of the image through multi-layer convolution and pooling operations, and perform adaptive average pooling on the extracted feature map to obtain extracted image features; Step R4: inputting the extracted image features into a regression model of a custom multi-layer neural network, wherein the regression model is constructed by a custom multi-layer perceptron, wherein the multi-layer perceptron comprises a plurality of linear regression layers, a batch normalization layer, a ReLU activation function layer, and a Dropout layer, wherein the regression model trains and predicts the extracted image features to obtain the average illumination and illumination uniformity of the road; Step R5: training the regression model, first creating a data set and a data loader, and dividing the data set into a training set and a test set, then initializing the regression model, the loss function and the optimizer, optimizing the parameters of the regression model through the training data, using the mean square error as the loss function, and performing back propagation and weight update through the Adam optimizer. The training process includes setting the learning rate, batch size and training rounds to obtain the trained regression model; Step R6: Divide the road lighting dataset into a training set and a validation set in a ratio of 8:2, evaluate the performance of the trained regression model on the validation set, calculate the mean square error between the predicted value and the true value, ensure that the model performs well on the validation set, and perform performance verification on the regression model; Step R7: Input the road image to be detected into the trained regression model, predict the average illumination and illumination uniformity, output and record the prediction results, and at the same time, calculate the error between the predicted value and the actual value for subsequent analysis and practical application.

Images