Intersection poor scene target detection optimization method based on confidence analysis and image quality diagnosis

Abstract

The application relates to an intersection severe scene target detection optimization method based on confidence analysis and image quality diagnosis, relates to the technical field of intelligent traffic, and can realize the following effects: through real-time detection of the confidence of a target detection model, average brightness, contrast, definition and other parameters of a video picture, and through classification of the image into three scenes of normal, night and rainy and foggy days by using a classification algorithm, once the video stream picture reaches the severe scene judgment standard, a specially trained detection model is used to replace the ordinary detection model, the detection precision of the algorithm in the severe scene is improved, and the visual perception algorithm can be stably and high-quality operated.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, specifically to an optimized method for target detection in adverse intersection scenarios based on confidence analysis and image quality diagnosis. Background Technology

[0002] In recent years, deep learning-based computer vision technology has been widely used in the field of intelligent transportation, especially in urban intersection electronic police (ECM) monitoring systems. These technologies automatically detect and track targets such as motor vehicles, non-motor vehicles, and pedestrians by analyzing intersection video streams in real time, providing crucial data support for traffic flow statistics, violation capture, event detection, and intersection status assessment. As a fundamental link in the entire perception chain, the accuracy and stability of target detection directly affect the reliability of all subsequent applications. Currently, under "normal" scenarios with good lighting and clear weather, mainstream target detection models can achieve high accuracy levels, basically meeting the needs of daily traffic management.

[0003] However, urban intersections present complex and ever-changing environments, and traffic enforcement cameras often face various challenging visual conditions. For example, in low-light scenarios such as nighttime, dusk, and dawn, insufficient ambient light leads to blurred target outlines and loss of detail. Rain, fog, and snow further interfere with this process, as raindrops, fog, and snowflakes obstruct the lens and scatter light, resulting in blurred images, reduced contrast, and severe interference with target features. Under these adverse conditions, the performance of conventionally trained target detection models significantly declines, easily misclassifying real targets as background (false positives) or mistaking background noise for targets (false positives). The failure or instability of target detection directly leads to a decrease in the output quality or even failure of the entire intelligent traffic perception system, failing to provide accurate and real-time traffic information for traffic management.

[0004] Currently, the main approach to addressing the challenges of target detection in harsh environments is to improve the robustness of a single model, such as by adding samples of severe weather and low light conditions to the training data, or by designing more complex network structures. However, these methods have significant limitations: First, their generalization ability is limited; a single model struggles to maintain optimal performance in all types (such as darkness, dense fog, heavy rain, heavy snow, etc.) and varying degrees of severity in harsh environments. Attempting to "catch all" scenarios often means that performance will not be optimal under specific harsh conditions, or that unnecessary computational overhead will be introduced in normal scenarios. Second, the model lacks dynamic adaptability; when detection quality deteriorates due to environmental degradation, the system cannot automatically trigger effective optimization or compensation mechanisms based on real-time diagnostic results, and can only passively accept the loss of accuracy, resulting in poor perception capabilities during critical periods of severe weather.

[0005] Therefore, in the actual deployment of intelligent transportation systems, there is an urgent need for a method that can perceive and detect changes in the detection status in real time, intelligently diagnose image quality, and dynamically adjust the detection strategy according to the severity of the scene, so as to ensure that the intersection visual perception algorithm can operate stably and with high quality in various complex environments, and provide a continuous and reliable data foundation for traffic status assessment, signal optimization and control, and emergency response. Summary of the Invention

[0006] To overcome the shortcomings of the above technologies, this invention provides an optimization method for target detection in adverse intersection scenes based on confidence analysis and image quality diagnosis, which improves the detection accuracy of algorithms in adverse scenes and ensures the stable and high-quality operation of visual perception algorithms.

[0007] The technical solution adopted by this invention to overcome its technical problems is: An optimized method for target detection in adverse intersection scenes based on confidence analysis and image quality diagnosis includes: S1. Obtain the historical video stream from the intersection's electronic police camera. Image set consisting of historical intersection road images Using image sets Training yields a target detection model for sunny days. Nighttime target detection model Rain and fog target detection model and scene classification model ; S2. Obtain the real-time video stream from the traffic enforcement camera at the intersection, and get... Image set consisting of intersection road images , , For the first Frame intersection road image, ; S3. In the Frame intersection road image A rectangular region is selected as the region of interest (ROI). S4. The first Frame intersection road image Input into the sunny day target detection model , obtained the Frame intersection road image The Middle List of target detection information for each target , , For the first Frame intersection road image The number of targets will be the first Frame intersection road image Input into scene classification model In the middle, the scene detection results are output. ; S5. Calculate the first... Frame intersection road image Image brightness Image contrast Image clarity ; S6. Based on scene detection results Image brightness Image contrast Image clarity Select a target detection model for nighttime. or rain / fog target detection model This serves as the foundation model for a smart transportation hyper-converged platform. S7. Based on the target detection model at night or rain / fog target detection model Combined with scene detection results Determine whether to switch back to the sunny day target detection model. This serves as the foundation model for a smart transportation hyper-converged platform. S8. Display scene detection results Image brightness Image contrast Image clarity The target detection information list is sent to the intelligent transportation hyper-converged platform.

[0008] Furthermore, step S1 includes the following steps: S1-1. Obtain the historical video stream from the intersection's electronic police camera. A collection of historical intersection road images taken on sunny days. , A collection of nighttime scene images composed of historical intersection road images captured at night. , A collection of historical images of intersections captured in rain and fog, comprising images of rain and fog scenes. Sunny Day Scene Image Set Image collection of nighttime scenes Rain and fog scene image set Composition of image set , ; S1-2. From a collection of images depicting sunny scenes Select Images of historical intersections and nighttime scenes. Select Historical intersection road images, from rain and fog scene image collection Select A YOLOv11s model was trained using road images from the Zhang historical intersection to obtain a sunny day target detection model for detecting motor vehicles, non-motor vehicles, and pedestrians in clear weather scenarios. ; S1-3. From a collection of images depicting sunny scenes Select Images of historical intersections and nighttime scenes. Select Historical intersection road images, from rain and fog scene image collection Select A YOLOv11s model was trained using road images from the Zhangshilu intersection to obtain a nighttime target detection model for detecting motor vehicles, non-motor vehicles, and pedestrians in dark scenes. ; S1-4. From the image set of sunny day scenes Select Images of historical intersections and nighttime scenes. Select Historical intersection road images, from rain and fog scene image collection Select A YOLOv11s model was trained using road images from the Zhangshilu intersection to obtain a rain and fog target detection model for detecting motor vehicles, non-motor vehicles, and pedestrians in rain and fog scenes. ; S1-5. Using Image Sets Train the ResNet18 model to obtain a model for classifying sunny, dark, and rainy / foggy scenes. .

[0009] Preferably, in steps S1-2 The ratio is 7:2:1; in steps S1-3 The ratio is 3:6:1; in steps S1-4 The ratio is 3:1:6.

[0010] Furthermore, in step S2, the real-time video stream from the intersection's electronic police camera is acquired via the ONVIF protocol, and the GPU is used to decode the real-time video stream to obtain... Image of a road at an intersection.

[0011] Furthermore, in step S3, in the first Frame intersection road image A rectangular area is designated, with its lower edge coinciding with the lower stop line of the intersection. The upper edge of this rectangle is defined as a point 40-50 meters above the upper stop line of the intersection. The left edge of this rectangle coincides with the lane line of the leftmost lane at the intersection, and the right edge coincides with the lane line of the rightmost lane. The x-coordinate of the center point of this rectangle is [omitted]. The ordinate of the center point of the rectangular region is The length of the rectangular region is The height of the rectangular region is .

[0012] Furthermore, step S4 includes the following steps: S4-1. The first Frame intersection road image Input into the sunny day target detection model Get the box to select the first Frame intersection road image The bounding box of each target and the first List of target detection information for each rectangle , , , For the first The x-coordinate of the center point of the rectangle For the first The y-coordinate of the center point of the rectangle For the first The width of the rectangle, For the first The height of the rectangle For the first The detection confidence of each target. For the first The categories of targets; S4-2. The first Frame intersection road image Input into scene classification model In the middle, the scene detection results are output. , ,in, A sunny day scene. For a nighttime setting, A scene depicting rain and fog.

[0013] Furthermore, in step S5, the cvtColor function from the OpenCV library is used to... Frame intersection road image Convert to grayscale image Calculate grayscale image The average of the pixel values ​​of all pixels is used as the image brightness. Calculate grayscale image The standard deviation of the pixel values ​​of all pixels is used as the image contrast. Use the Laplacian function from the OpenCV library to calculate grayscale images. The gradient of all pixels, and the variance of all gradients as the image sharpness. .

[0014] Furthermore, step S6 includes the following steps: S6-1. According to the... Frame intersection road image The Middle The x-coordinate of the center point of the rectangle and ordinate Judge the first If the center point of each bounding box is located within the Region of Interest (ROI), then calculate the average detection confidence score of all targets located within the ROI. ; S6-2. If the image set continuous The average detection confidence scores of all frames of intersection road images are less than the threshold. And all scene detection results are If the image brightness is less than 50, then the nighttime target detection model is enabled. As the foundation model of the intelligent transportation hyper-converged platform, step S6-3 is executed. The value is 10. The value is 0.4; S6-3. The first Frame intersection road image Input into the nighttime target detection model Get the box to select the first Frame intersection road image The bounding box of each target and the first List of target detection information for each rectangle , , , For the first The x-coordinate of the center point of the rectangle For the first The y-coordinate of the center point of the rectangle For the first The width of the rectangle, For the first The height of the rectangle For the first The detection confidence of each target. For the first The target categories are determined, and the average detection confidence score of all targets located within the Region of Interest (ROI) is calculated. ,if Greater than or equal to Then the target detection information list As the final target detection result; S6-4. If the image set continuous The average detection confidence scores of all frames of intersection road images are less than the threshold. And all scene detection results are If both image contrast and image sharpness are less than 50, then the rain / fog target detection model is enabled. As the base model of the intelligent transportation hyper-converged platform, step S6-5 is executed; S6-5. The first Frame intersection road image Input into the rain and fog target detection model Get the box to select the first Frame intersection road image The bounding box of each target and the first List of target detection information for each rectangle , , , For the first The x-coordinate of the center point of the rectangle For the first The y-coordinate of the center point of the rectangle For the first The width of the rectangle, For the first The height of the rectangle For the first The detection confidence of each target. For the first The target categories are determined, and the average detection confidence score of all targets located within the Region of Interest (ROI) is calculated. ,if Greater than or equal to Then the target detection information list As the final target detection result.

[0015] Furthermore, step S7 includes the following steps: S7-1. When the intelligent transportation hyper-converged platform uses a nighttime target detection model If continuous Average frame detection confidence All are less than or equal to the average of the detection confidence scores. And all scene detection results are Then switch to the sunny day target detection model. As a foundational model for a smart transportation hyper-converged platform; S7-2. When the intelligent transportation hyperconverged platform uses a rain and fog target detection model If continuous Average frame detection confidence All are less than or equal to the average of the detection confidence scores. And all scene detection results are Then switch to the sunny day target detection model. This serves as a foundational model for a smart transportation hyper-converged platform.

[0016] Furthermore, step S8 includes the following steps: S8-1. Send the scene detection results at the SendTime interval. Image brightness Image contrast Image clarity Send the data to the database in JSON format and save it. The SendTime value is 600 seconds. S8-2. List of target detection information of the base model of the intelligent transportation hyper-converged platform Or target detection information list Or target detection information list Send it to the intelligent transportation hyper-converged platform in JSON format.

[0017] The beneficial effects of this invention are: by real-time detection of the confidence level of the target detection model, parameters such as the average brightness, contrast, and sharpness of the video image, and using a classification algorithm to classify the image into three scenes: normal, nighttime, and rainy / foggy weather. Once the video stream image meets the criteria for a severe scene, a specially trained detection model is used to replace the ordinary detection model, improving the detection accuracy of the algorithm in severe scenes and ensuring the stable and high-quality operation of the visual perception algorithm. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] The following is in conjunction with the appendix Figure 1 The present invention will be further described below.

[0020] An optimized method for target detection in adverse intersection scenes based on confidence analysis and image quality diagnosis includes: S1. Obtain the historical video stream from the intersection's electronic police camera. Image set consisting of historical intersection road images Using image sets Training yields a target detection model for sunny days. Nighttime target detection model Rain and fog target detection model and scene classification model .

[0021] S2. Obtain the real-time video stream from the traffic enforcement camera at the intersection, and get... Image set consisting of intersection road images , , For the first Frame intersection road image, .

[0022] S3. In the Frame intersection road image A rectangular region is selected as the region of interest (ROI).

[0023] S4. The first Frame intersection road image Input into the sunny day target detection model , obtained the Frame intersection road image The Middle List of target detection information for each target , , For the first Frame intersection road image The number of targets will be the first Frame intersection road image Input into scene classification model In the middle, the scene detection results are output. .

[0024] S5. Calculate the first... Frame intersection road image Image brightness Image contrast Image clarity .

[0025] S6. Based on scene detection results Image brightness Image contrast Image clarity Select a target detection model for nighttime. or rain / fog target detection model This serves as the foundation model for a smart transportation hyper-converged platform, enabling the calculation of subsequent traffic indicators.

[0026] S7. Based on the target detection model at night or rain / fog target detection model Combined with scene detection results Determine whether to switch back to the sunny day target detection model. This serves as the foundation model for a smart transportation hyper-converged platform.

[0027] S8. Display scene detection results Image brightness Image contrast Image clarity The target detection information list is sent to the intelligent transportation hyper-converged platform for further calculation of traffic parameters.

[0028] This method enables real-time perception of changes in detection status, intelligent diagnosis of image quality, and dynamic adjustment of detection strategies based on the severity of the scene. It ensures that the intersection visual perception algorithm can operate stably and with high quality in various complex environments, providing a continuous and reliable data foundation for traffic condition assessment, signal optimization and control, and emergency response.

[0029] In one embodiment of the present invention, step S1 includes the following steps: S1-1. Obtain the historical video stream from the intersection's electronic police camera. A collection of historical intersection road images taken on sunny days. , A collection of nighttime scene images composed of historical intersection road images captured at night. , A collection of historical images of intersections captured in rain and fog, comprising images of rain and fog scenes. Sunny Day Scene Image Set Image collection of nighttime scenes Rain and fog scene image set Composition of image set , .

[0030] S1-2. From a collection of images depicting sunny scenes Select Images of historical intersections and nighttime scenes. Select Historical intersection road images, from rain and fog scene image collection Select A YOLOv11s model was trained using road images from the Zhang historical intersection to obtain a sunny day target detection model for detecting motor vehicles, non-motor vehicles, and pedestrians in clear weather scenarios. .

[0031] S1-3. From a collection of images depicting sunny scenes Select Images of historical intersections and nighttime scenes. Select Historical intersection road images, from rain and fog scene image collection Select A YOLOv11s model was trained using road images from the Zhangshilu intersection to obtain a nighttime target detection model for detecting motor vehicles, non-motor vehicles, and pedestrians in dark scenes. .

[0032] S1-4. From the image set of sunny day scenes Select Images of historical intersections and nighttime scenes. Select Historical intersection road images, from rain and fog scene image collection Select A YOLOv11s model was trained using road images from the Zhangshilu intersection to obtain a rain and fog target detection model for detecting motor vehicles, non-motor vehicles, and pedestrians in rain and fog scenes. .

[0033] S1-5. Using Image Sets Train the ResNet18 model to obtain a model for classifying sunny, dark, and rainy / foggy scenes. .

[0034] In one embodiment of the present invention, in steps S1-2 The ratio is 7:2:1; in steps S1-3 The ratio is 3:6:1; in steps S1-4 The ratio is 3:1:6.

[0035] In one embodiment of the present invention, in step S2, the real-time video stream of the intersection electronic police camera is obtained through the ONVIF protocol, and the real-time video stream is decoded using a GPU to obtain... Image of a road at an intersection.

[0036] In one embodiment of the present invention, in step S3, the first Frame intersection road image A rectangular area is designated, with its lower edge coinciding with the lower stop line of the intersection. The upper edge of this rectangle is defined as a point 40-50 meters above the upper stop line of the intersection. The left edge of this rectangle coincides with the lane line of the leftmost lane at the intersection, and the right edge coincides with the lane line of the rightmost lane. The x-coordinate of the center point of this rectangle is [omitted]. The ordinate of the center point of the rectangular region is The length of the rectangular region is The height of the rectangular region is .

[0037] In one embodiment of the present invention, step S4 includes the following steps: S4-1. The first Frame intersection road image Input into the sunny day target detection model Get the box to select the first Frame intersection road image The bounding box of each target and the first List of target detection information for each rectangle , , , For the first The x-coordinate of the center point of the rectangle For the first The y-coordinate of the center point of the rectangle For the first The width of the rectangle, For the first The height of the rectangle For the first The detection confidence of each target. For the first The categories of targets.

[0038] S4-2. The first Frame intersection road image Input into scene classification model In the middle, the scene detection results are output. , ,in, A sunny day scene. For a nighttime setting, A scene depicting rain and fog.

[0039] In one embodiment of the present invention, step S5 uses the cvtColor function from the OpenCV library to... Frame intersection road image Convert to grayscale image Calculate grayscale image The average of the pixel values ​​of all pixels is used as the image brightness. Calculate grayscale image The standard deviation of the pixel values ​​of all pixels is used as the image contrast. Use the Laplacian function from the OpenCV library to calculate grayscale images. The gradient of all pixels, and the variance of all gradients as the image sharpness. .

[0040] In one embodiment of the present invention, step S6 includes the following steps: S6-1. According to the... Frame intersection road image The Middle The x-coordinate of the center point of the rectangle and ordinate Judge the first If the center point of each bounding box is located within the Region of Interest (ROI), then calculate the average detection confidence score of all targets located within the ROI. .

[0041] S6-2. If the image set continuous The average detection confidence scores of all frames of intersection road images are less than the threshold. And all scene detection results are If the image brightness is less than 50, then the nighttime target detection model is enabled. As the foundation model of the intelligent transportation hyper-converged platform, step S6-3 is executed. The value is 10. The value is 0.4.

[0042] S6-3. The first Frame intersection road image Input into the nighttime target detection model Get the box to select the first Frame intersection road image The bounding box of each target and the first List of target detection information for each rectangle , , , For the first The x-coordinate of the center point of the rectangle For the first The y-coordinate of the center point of the rectangle For the first The width of the rectangle, For the first The height of the rectangle For the first The detection confidence of each target. For the first The target categories are determined, and the average detection confidence score of all targets located within the Region of Interest (ROI) is calculated. ,if Greater than or equal to Then the target detection information list As the final target detection result.

[0043] S6-4. If the image set continuous The average detection confidence scores of all frames of intersection road images are less than the threshold. And all scene detection results are If both image contrast and image sharpness are less than 50, then the rain / fog target detection model is enabled. As the base model of the intelligent transportation hyper-converged platform, it performs steps S6-5.

[0044] S6-5. The first Frame intersection road image Input into the rain and fog target detection model Get the box to select the first Frame intersection road image The bounding box of each target and the first List of target detection information for each rectangle , , , For the first The x-coordinate of the center point of the rectangle For the first The y-coordinate of the center point of the rectangle For the first The width of the rectangle, For the first The height of the rectangle For the first The detection confidence of each target. For the first The target categories are determined, and the average detection confidence score of all targets located within the Region of Interest (ROI) is calculated. ,if Greater than or equal to Then the target detection information list As the final target detection result.

[0045] In one embodiment of the present invention, step S7 includes the following steps: S7-1. When the intelligent transportation hyper-converged platform uses a nighttime target detection model If continuous Average frame detection confidence All are less than or equal to the average of the detection confidence scores. And all scene detection results are Then switch to the sunny day target detection model. This serves as a foundational model for a smart transportation hyper-converged platform.

[0046] S7-2. When the intelligent transportation hyperconverged platform uses a rain and fog target detection model If continuous Average frame detection confidence All are less than or equal to the average of the detection confidence scores. And all scene detection results are Then switch to the sunny day target detection model. This serves as a foundational model for a smart transportation hyper-converged platform.

[0047] In one embodiment of the present invention, step S8 includes the following steps: S8-1. Send the scene detection results at the SendTime interval. Image brightness Image contrast Image clarity Send the data to the database in JSON format and save it. The SendTime value is 600 seconds.

[0048] S8-2. List of target detection information of the base model of the intelligent transportation hyper-converged platform Or target detection information list Or target detection information list Send the data in JSON format to the intelligent transportation hyperconverged platform to enable functions such as congestion detection, traffic flow detection, and abnormal traffic event detection.

[0049] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An optimized method for target detection in adverse intersection scenes based on confidence analysis and image quality diagnosis, characterized in that, include: S1. Obtain the historical video stream from the intersection's electronic police camera. Image set consisting of historical intersection road images Using image sets Training yields a target detection model for sunny days. Nighttime target detection model Rain and fog target detection model and scene classification model ; S2. Obtain the real-time video stream from the traffic enforcement camera at the intersection, and get... Image set consisting of intersection road images , , For the first Frame intersection road image, ; S3. In the Frame intersection road image A rectangular region is selected as the region of interest (ROI). S4. The first Frame intersection road image Input into the sunny day target detection model , obtained the Frame intersection road image The Middle List of target detection information for each target , , For the first Frame intersection road image The number of targets will be the first Frame intersection road image Input into scene classification model In the middle, the scene detection results are output. ; S5. Calculate the... Frame intersection road image Image brightness Image contrast Image clarity ; S6. Based on scene detection results Image brightness Image contrast Image clarity Select a target detection model for nighttime. or rain / fog target detection model This serves as the foundation model for a smart transportation hyper-converged platform. S7. Based on the target detection model at night or rain / fog target detection model Combined with scene detection results Determine whether to switch back to the sunny day target detection model. This serves as the foundation model for a smart transportation hyper-converged platform. S8. Display scene detection results Image brightness Image contrast Image clarity The target detection information list is sent to the intelligent transportation hyper-converged platform.

2. The optimized method for target detection in adverse intersection scenes based on confidence analysis and image quality diagnosis according to claim 1, characterized in that, Step S1 includes the following steps: S1-1. Obtain the historical video stream from the intersection's electronic police camera. A collection of historical intersection road images taken on sunny days. , A collection of nighttime scene images composed of historical intersection road images captured at night. , A collection of historical images of intersections captured in rain and fog, comprising images of rain and fog scenes. Sunny Day Scene Image Set Image collection of nighttime scenes Rain and fog scene image set Composition of image set , ; S1-2. From a collection of images depicting sunny scenes Select Images of historical intersections and nighttime scenes. Select Historical intersection road images, from rain and fog scene image collection Select A YOLOv11s model was trained using road images from the Zhang historical intersection to obtain a sunny day target detection model for detecting motor vehicles, non-motor vehicles, and pedestrians in clear weather scenarios. ; S1-3. From a collection of images depicting sunny scenes Select Images of historical intersections and nighttime scenes. Select Historical intersection road images, from rain and fog scene image collection Select A YOLOv11s model was trained using road images from the Zhangshilu intersection to obtain a nighttime target detection model for detecting motor vehicles, non-motor vehicles, and pedestrians in dark scenes. ; S1-4. From the image set of sunny day scenes Select Images of historical intersections and nighttime scenes. Select Historical intersection road images, from rain and fog scene image collection Select A YOLOv11s model was trained using road images from the Zhangshilu intersection to obtain a rain and fog target detection model for detecting motor vehicles, non-motor vehicles, and pedestrians in rain and fog scenes. ; S1-5. Using Image Sets Train the ResNet18 model to obtain a model for classifying sunny, dark, and rainy / foggy scenes. .

3. The optimized method for target detection in adverse intersection scenes based on confidence analysis and image quality diagnosis according to claim 2, characterized in that: In step S1-2 The ratio is 7:2:1; in steps S1-3 The ratio is 3:6:1; in steps S1-4 The ratio is 3:1:

6.

4. The optimized method for target detection in adverse intersection scenes based on confidence analysis and image quality diagnosis according to claim 1, characterized in that: In step S2, the real-time video stream from the intersection's electronic police camera is acquired via the ONVIF protocol, and the GPU is used to decode the real-time video stream to obtain... Image of a road at an intersection.

5. The optimized method for target detection in adverse intersection scenes based on confidence analysis and image quality diagnosis according to claim 1, characterized in that: In step S3, at the Frame intersection road image A rectangular area is designated, with its lower edge coinciding with the lower stop line of the intersection. The upper edge of this rectangle is defined as a point 40-50 meters above the upper stop line of the intersection. The left edge of this rectangle coincides with the lane line of the leftmost lane at the intersection, and the right edge coincides with the lane line of the rightmost lane. The x-coordinate of the center point of this rectangle is [omitted]. The ordinate of the center point of the rectangular region is The length of the rectangular region is The height of the rectangular region is .

6. The optimized method for target detection in adverse intersection scenes based on confidence analysis and image quality diagnosis according to claim 1, characterized in that, Step S4 includes the following steps: S4-1. The first Frame intersection road image Input into the sunny day target detection model Get the box to select the first Frame intersection road image The bounding box of each target and the first List of target detection information for each rectangle , , , For the first The x-coordinate of the center point of the rectangle For the first The y-coordinate of the center point of the rectangle For the first The width of the rectangle, For the first The height of the rectangle For the first The detection confidence of each target. For the first The categories of the targets; S4-2. The first Frame intersection road image Input into scene classification model In the middle, the scene detection results are output. , ,in, A sunny day scene. A nighttime scene. A scene depicting rain and fog.

7. The optimization method for target detection in adverse intersection scenes based on confidence analysis and image quality diagnosis according to claim 1, characterized in that: Step S5 uses the cvtColor function from the OpenCV library to... Frame intersection road image Convert to grayscale image Calculate grayscale image The average of the pixel values ​​of all pixels is used as the image brightness. Calculate grayscale image The standard deviation of the pixel values ​​of all pixels is used as the image contrast. Use the Laplacian function from the OpenCV library to calculate grayscale images. The gradient of all pixels, and the variance of all gradients as the image sharpness. .

8. The optimized method for target detection in adverse intersection scenes based on confidence analysis and image quality diagnosis according to claim 6, characterized in that, Step S6 includes the following steps: S6-1. According to the... Frame intersection road image The Middle The x-coordinate of the center point of each rectangle and ordinate Judge the first If the center point of each bounding box is located within the Region of Interest (ROI), then calculate the average detection confidence score of all targets located within the ROI. ; S6-2. If the image set continuous The average detection confidence scores of all frames of intersection road images are less than the threshold. And all scene detection results are If the image brightness is less than 50, then the nighttime target detection model is enabled. As the foundation model of the intelligent transportation hyper-converged platform, step S6-3 is executed. The value is 10. The value is 0.4; S6-3. The first Frame intersection road image Input into the nighttime target detection model Get the box to select the first Frame intersection road image The bounding box of each target and the first List of target detection information for each rectangle , , , For the first The x-coordinate of the center point of the rectangle For the first The y-coordinate of the center point of the rectangle For the first The width of the rectangle, For the first The height of the rectangle For the first The detection confidence of each target. For the first The target categories are determined, and the average detection confidence score of all targets located within the Region of Interest (ROI) is calculated. ,if Greater than or equal to Then the target detection information list As the final target detection result; S6-4. If the image set continuous The average detection confidence scores of all frames of intersection road images are less than the threshold. And all scene detection results are If both image contrast and image sharpness are less than 50, then the rain / fog target detection model is enabled. As the base model of the intelligent transportation hyper-converged platform, step S6-5 is executed; S6-5. The first Frame intersection road image Input into the rain and fog target detection model Get the box to select the first Frame intersection road image The bounding box of each target and the first List of target detection information for each rectangle , , , For the first The x-coordinate of the center point of the rectangle For the first The y-coordinate of the center point of the rectangle For the first The width of the rectangle, For the first The height of the rectangle For the first The detection confidence of each target. For the first The target categories are determined, and the average detection confidence score of all targets located within the Region of Interest (ROI) is calculated. ,if Greater than or equal to Then the target detection information list As the final target detection result.

9. The optimized method for target detection in adverse intersection scenes based on confidence analysis and image quality diagnosis according to claim 8, characterized in that, Step S7 includes the following steps: S7-1. When the intelligent transportation hyper-converged platform uses a nighttime target detection model If continuous Average frame detection confidence All are less than or equal to the average of the detection confidence scores. And all scene detection results are Then switch to the sunny day target detection model. As a foundational model for a smart transportation hyper-converged platform; S7-2. When the intelligent transportation hyperconverged platform uses a rain and fog target detection model If continuous Average frame detection confidence All are less than or equal to the average of the detection confidence scores. And all scene detection results are Then switch to the sunny day target detection model. This serves as a foundational model for a smart transportation hyper-converged platform.

10. The optimized method for target detection in adverse intersection scenes based on confidence analysis and image quality diagnosis according to claim 8, characterized in that, Step S8 includes the following steps: S8-1. Send the scene detection results at the SendTime interval. Image brightness Image contrast Image clarity Send the data to the database in JSON format and save it. The SendTime value is 600 seconds. S8-2. List of target detection information of the base model of the intelligent transportation hyper-converged platform Or target detection information list Or target detection information list Send it to the intelligent transportation hyper-converged platform in JSON format.

Images