A traffic monitoring image target detection method in adverse weather

Abstract

The application is a traffic monitoring image target detection method under adverse weather. The monitoring image to be recovered is input into a dynamic convolution module for convolution processing. After the dynamic convolution module, the image is input into a dual-tree complex wavelet enhancement module. The information enhanced by the dual-tree complex wavelet transform is input into a residual learning recovery module for residual learning to obtain a clean image. The picture output after the residual learning is sent to a YOLOv3 traffic target detection module for target detection. The application uses the dual-tree complex wavelet transform to combine the frequency characteristics and uses the dynamic convolution layer to combine the structural information to mine the robust features in the rain, snow and fog images. The residual network structure is used to refine the obtained features, and the residual learning operation is used to reconstruct the clean image. The application realizes better adaptability to the complex adverse weather scene and is used for the monitoring image recovery in the rain, snow and fog weather.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, image restoration and target detection, and particularly relates to a traffic monitoring image target detection method under adverse weather. BACKGROUND

[0002] With the vigorous development of artificial intelligence in the monitoring field, improving monitoring accuracy is a common problem faced by the industry. In these applications, clear and high-quality pictures can improve the recognition accuracy and better complete the detection, tracking and recognition of targets in the intelligent monitoring system. Low-quality images hinder the recognition of advanced computer vision tasks, leading to false alarms and unstable risks. In the imaging process, many reasons can cause the degradation of monitoring image quality. The image quality is reduced due to the imperfection of the hardware system, and the degradation is also caused by the influence of the medium in the image shooting process, such as rain, snow, fog and other adverse weather, which can produce atmospheric particles that affect the imaging quality, causing the image to lose important information. Haze, fog and smoke are caused by the absorption and scattering of atmospheric particles. Raindrops are transparent, but they can also produce a veil effect (similar to haze or fog) due to the transmission of light. When the rain is heavy, irregular lines that affect image quality can be produced. Snow is a highly complex atmospheric phenomenon that usually contains snowflakes, snow streaks and veil effects. These problems result in poor performance of existing intelligent traffic monitoring in adverse weather.

[0003] Existing methods are mostly focused on single-task adverse weather target detection, such as Image-Adaptive YOLO (IAYOLO). These algorithms have poor practical use in the monitoring field, with problems such as complicated operation, complex model and poor image restoration effect. Adverse weather is unpredictable, and heavy rain and snow weather is often accompanied by a veil effect. Single-task image restoration cannot meet the efficient needs of the monitoring field. These improved target detection methods are not suitable for the traffic monitoring field, which needs to detect small targets. Traffic monitoring needs clearer and higher quality images for target detection in adverse weather. SUMMARY

[0004] The present application aims to solve the problems of the prior art and provides a traffic monitoring image target detection method under adverse weather. The present application can restore clearer images to be detected with a slight increase in computational load, and can complete the rain, snow and fog removal tasks uniformly without the need for complicated algorithm switching. The detection effect can be better when the image to be detected is input into the YOLOv3 detection network, and the detection performance can be improved with the upgrading and iteration of the YOLO series target detection algorithm.

[0005] To achieve the above-mentioned purpose, the following technical solutions are adopted in the present application:

[0006] The application discloses a traffic monitoring image target detection method in severe weather, and the overall network structure comprises five dynamic convolution modules, two eight-layer dual-tree complex wavelet enhancement modules, ten residual learning recovery modules and a YOLOv3 traffic target detection module.

[0007] The specific steps are as follows:

[0008] S1, input the monitoring image to be recovered into a dynamic convolution module for convolution processing;

[0009] The process is as follows: the dynamic convolution module has five layers, the first layer is a convolution layer, the input monitoring image to be recovered is down-sampled, the convolution kernel size is 5*5, and the output channel is 64; the second to fourth layers are two weight generators and a convolution layer with a convolution kernel size of 5*5 respectively; the fifth layer is a ReLU activation function, features are extracted from the dynamic convolution, and the output channel is 64.

[0010] S2, input the image into a dual-tree complex wavelet enhancement module after the dynamic convolution module;

[0011] The process is as follows: the linear structure information is converted into six frequency features by using a dual-tree complex wavelet transform, the frequency features are enhanced by using a residual learning feature enhancement mechanism to enhance the original image features, so as to suppress the rain and snow stripes, thereby recovering detailed information in the image, and finally, the frequency features are converted into linear structure information by using a dual-tree complex wavelet inverse transform;

[0012] S3, input the information enhanced by the dual-tree complex wavelet transform into a residual learning recovery module for residual learning to obtain a clean image;

[0013] The process is as follows: the information enhanced by the dual-tree complex wavelet transform is subjected to first residual learning by using a combination of four residual dense layers and a ReLU activation function, and the features in image denoising are further refined; the output of the first residual learning is subjected to second residual learning by using a combination of four residual dense layers and a ReLU activation function; the output of the second residual learning is fused with the output of the first residual learning, and the deep information of the image is enhanced; finally, the features are refined by using a convolution layer followed by a ReLU, and the input channel and the output channel thereof are 64 respectively.

[0014] S4, input the output picture of the residual learning into a YOLOv3 traffic target detection module for target detection, and perform target detection on the recovered image;

[0015] The process is as follows: the input picture is processed, features are extracted, the extracted feature map is used for predicting a bounding box, the information of the bounding box is decoded for non-maximum suppression, and finally, the detection result is output, and traffic vehicle information and confidence are given.

[0016] In step S1, the weight generator comprises a pooling layer, a 1x1 convolution, an activation function and a softmax classifier; the specific implementation manner of the dynamic convolution is that four weights are obtained by using the weight generator, the dynamic convolution is obtained by adjusting parameters through the weighted manner acting on the four parallel convolution kernels, and the output of the first convolution layer is convolved by using the dynamic convolution.

[0017] In step S2, the feature enhancement is performed by using four residual dense layers, including three convolution layers plus an activation function and a 1x1 convolution layer, and the convolution kernel size of the three convolution layers is 5x5.

[0018] In step S3, the size of all convolution kernels in the residual learning recovery module is 5x5, the input channel number and the output channel number of the two residual blocks are 64, the input channel number of the last convolution layer is 64, and the mean square error is selected as the loss function to train the network.

[0019] The specific operation of selecting the mean square error as the loss function to train the network is that the mean square error is used in a supervised manner to train the denoising model by using {IC, IL} pairs, wherein IC and IL are defined as a clean image and an image to be recovered respectively, the network is optimized by using an Adam optimizer, and appropriate parameters are obtained.

[0020] In step S4, first, the input image is divided into 13x13 grids by YOLOv3, for each grid, YOLOv3 will predict B bounding boxes, and for each bounding box, YOLOv3 will predict the probability of C classes and the confidence of each bounding box, and if the confidence of a bounding box is lower than a threshold, the bounding box will be filtered out.

[0021] Next, YOLOv3 extracts features of the input image by using a convolutional neural network, YOLOv3 adopts Darknet53 as its backbone network, in Darknet53, first, the input image is subjected to multi-layer convolution and pooling operation, then the residual structure is used to further process the feature map, and finally, high-level feature representation is obtained.

[0022] On the feature map output by the last layer of convolution of Darknet53, YOLOv3 uses a convolution layer to predict the bounding box: YOLOv3 uses a 1x1 convolution layer to reduce the depth of the feature map to Bx(5+C), where B is the number of bounding boxes predicted for each grid, 5 represents the position and confidence information of each bounding box, and C is the number of predicted categories, so for each grid and each bounding box, YOLOv3 will predict a position and a confidence value, and the probability of C categories; YOLOv3 converts the predicted bounding box into a real bounding box, for each grid and each bounding box, YOLOv3 first calculates the center coordinates of the bounding box, and the width and height of the bounding box, where the coordinates are offsets relative to the top left corner of the grid, then YOLOv3 converts the center coordinates and width height into the coordinates of the top left corner and the bottom right corner of the bounding box, and finally obtains a real bounding box;

[0023] The non-maximum suppression algorithm is used to screen the overlapping bounding boxes: for each category, first sort all the bounding boxes with a confidence higher than the threshold in descending order of confidence, then select the bounding box with the highest confidence, and compare it with all other bounding boxes, if the overlap of two bounding boxes is greater than a certain threshold, the bounding box with lower confidence is filtered out, until all the bounding boxes are screened; finally, YOLOv3 outputs the screened bounding boxes and their categories and confidence as detection results.

[0024] YOLOv3 outputs the screened bounding boxes and their categories and confidence as detection results, which can be visualized and displayed on the input image, and the category and confidence of each object are given.

[0025] The beneficial effects of the present application are: the present application utilizes the multi-directional and multi-scale characteristics of dual-tree complex wavelet transform, which can better adapt to complex and severe weather scenes, and improve the effect and accuracy of image restoration; the dynamic convolution layer is combined with structural information to mine robust features in rain, snow and fog images, which improves the processing capability of complex scenes; the residual network structure is used to refine the obtained features, which further improves the quality and accuracy of image restoration; through residual learning operation, a clean image can be reconstructed, and a high-quality input image can be obtained, thereby improving the traffic target detection performance. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 The flowchart of the present application is shown in the figure;

[0027] Figure 2 The flowchart of the dynamic convolution algorithm of the present application is shown in the figure;

[0028] Figure 3A flow chart of a dual-tree complex wavelet texture information enhancement algorithm of the present application is shown in the figure.

[0029] Figure 4 A flow chart of a residual learning image restoration algorithm of the present application is shown in the figure.

[0030] Figure 5 A YOLOv3 network structure diagram of the present application is shown in the figure.

[0031] The present application will be described in detail below with reference to the accompanying drawings. DETAILED DESCRIPTION

[0032] The principles and features of the present application will be described below in conjunction with the accompanying drawings, and the examples are only used to explain the present application and are not intended to limit the scope of the present application. In the following paragraphs, the present application is described in more detail by way of example with reference to the accompanying drawings. The advantages and features of the present application will be more apparent from the following description. It should be noted that the drawings are very simplified and use non-precise proportions, only for the purpose of facilitating and clearly assisting the description of the embodiments of the present application.

[0033] It should be noted that when a component is referred to as being "fixed" to another component, it can be directly on the other component or there can be an intervening component. When a component is referred to as being "connected" to another component, it can be directly connected to the other component or there can be an intervening component. When a component is referred to as being "disposed" on another component, it can be directly disposed on the other component or there can be an intervening component. The terms "vertical", "horizontal", "left", "right", and similar expressions used herein are for illustrative purposes only.

[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terminology used in the description of the present application herein is only for the purpose of describing the specific embodiments and is not intended to limit the present application. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0035] The present application will be further described below in conjunction with the accompanying drawings and examples:

[0036] A traffic monitoring image target detection method under severe weather conditions, the overall network structure includes five layers of dynamic convolution module, two eight-layer dual-tree complex wavelet enhancement module, ten-layer residual learning restoration module and YOLOv3 traffic target detection module.

[0037] The process of the present application is shown in the figure. Figure 1

[0038] The specific steps are: ​

[0039] S1, the input dynamic convolution module of the to be recovered monitoring image is convolved;

[0040] The process is: the dynamic convolution module is five layers, the first layer is a convolution layer, the input to be recovered monitoring image is down-sampled, the convolution kernel size is 5x5, and the output channel is 64; the second to fourth layers are two weight generators and a convolution layer with a convolution kernel size of 5x5 respectively; the fifth layer is a ReLU activation function, the features are extracted from the dynamic convolution, and the output channel is 64, as shown in Figure 2 .

[0041] It should be noted that the weight generator includes a pooling layer, a 1x1 convolution, an activation function, and a softmax classifier; the specific implementation of dynamic convolution is to obtain four weights using the weight generator, adjust the parameters by weighting the four parallel convolution kernels to obtain dynamic convolution, and use dynamic convolution to convolve the output of the first convolution layer.

[0042] Dynamic convolution (Dynamic Convolution) uses multiple parallel dynamic convolutions to adjust the weight parameters of each layer according to different images to improve robustness. Convolutional neural networks train models by sharing the parameters of each convolutional layer. Embedding dynamic convolution in neural networks can improve training efficiency and eliminate redundant features.

[0043] S2, the image is input into the dual number complex wavelet enhancement module after the dynamic convolution module;

[0044] The process is: using dual tree complex wavelet transform to convert linear structure information into six frequency features, using residual learning feature enhancement mechanism to enhance the original image features to suppress rain and snow stripes, so as to recover detailed information in the image, and finally using dual tree complex wavelet inverse transform to convert the frequency features into linear structure information, as shown in Figure 3 .

[0045] It should be noted that the feature enhancement uses four layers of residual dense layers, including three convolution layers plus an activation function and a 1x1 convolution layer, and the convolution kernel size of the three convolution layers is 5x5.

[0046] Dual-Tree Complex Wavelet Transform (DTCWT) is an improved form of Discrete Wavelet Transform (DWT). Discrete Wavelet Transform has been applied in image and signal processing for a long time. Its physical idea is to remove the correlation of signals in different spatial scales, and then use data to represent the original signal, through this operation to obtain the features of interest. After decomposition in this way, the low-frequency part and the high-frequency part can be further decomposed by wavelet transform, thereby providing multi-resolution features for signal processing. However, Discrete Wavelet Transform still has some limitations: shift sensitivity; lack of direction selectivity. Dual-Tree Complex Wavelet Transform overcomes these defects by introducing complex wavelet and binary tree structure. Two-dimensional dual-tree complex wavelet transform can generate six high-pass subbands and one low-pass subband. The high-pass subband contains six directions (i.e. ± 15°, ± 45°, ± 75°) of texture information. This feature is conducive to the extraction of rain and snow information, because the image affected by rain and snow particles usually forms diagonal lines and oblique stripes, and the high-frequency subband in the dual-tree complex wavelet transform can better represent the rain and snow stripes. Therefore, the dual-tree complex wavelet transform is embedded into the network, which can better mine the robust features in the rain and snow image.

[0047] S3, the information enhanced by the dual-tree complex wavelet transform is input into a residual learning recovery module to perform residual learning to obtain a clean image;

[0048] The process is: using the combination of four layers of residual dense layers and ReLU activation function to perform the first residual learning on the information enhanced by the dual-tree complex wavelet transform, further refining the features in image denoising; using the combination of four layers of residual dense layers and ReLU activation function to perform the first residual learning on the output of the first residual learning; fusing the output of the second residual learning and the output of the first residual learning to enhance the deep information of the image; finally using a convolution layer followed by ReLU to refine these features, wherein the input channel and the output channel are 64 respectively, as shown in Figure 4 ;

[0049] It should be noted that the size of all convolution kernels in the residual learning recovery module is 5x5, the input channel number and the output channel number of the two residual blocks are 64, and the input channel number of the last convolution layer is 64. The mean square error (MSE) is selected as the loss function to train the network, and the specific operation is: the mean square error (MSE) is used in a supervised manner to train the denoising model with {IC, IL} pairs, wherein IC and IL are defined as clean images and images to be recovered, and the appropriate parameters are obtained by optimization through the Adam optimizer.

[0050] S4, send the picture output after residual learning into a YOLOv3 traffic target detection module for target detection, and perform target detection on the restored image;

[0051] The process is: processing the input picture, then performing feature extraction, predicting the bounding box of the extracted feature map, decoding the information of the bounding box for non-maximum suppression, and finally outputting the detection result, giving the traffic vehicle information and confidence, as shown in Figure 5

[0052] It should be noted that first, YOLOv3 divides the input image into 13x13 grids, (13 is a hyperparameter, which is set according to the specific feature to be detected). For each grid, YOLOv3 will predict B bounding boxes, and for each bounding box, it will predict the probability of C classes, as well as the confidence of each bounding box. If the confidence of a bounding box is lower than a threshold, it will be filtered out;

[0053] Next, YOLOv3 extracts features from the input image through a convolutional neural network. YOLOv3 uses Darknet53 as its backbone network, which is a high-efficiency CNN structure that can improve detection accuracy and speed. In Darknet53, first, the input image is subjected to multi-layer convolution and pooling operations, and then the residual structure is used to further process the feature map, and finally a high-level feature representation is obtained;

[0054] On the feature map output by the last layer of convolution in Darknet53, YOLOv3 uses a convolutional layer to predict the bounding box: YOLOv3 uses a 1x1 convolutional layer to reduce the depth of the feature map to Bx(5+C), where B is the number of bounding boxes predicted for each grid, 5 represents the position and confidence information of each bounding box, and C is the number of predicted classes. Therefore, for each grid and each bounding box, YOLOv3 will predict a position (four coordinate values) and a confidence value, as well as C class probabilities; YOLOv3 converts the predicted bounding box into a real bounding box. For each grid and each bounding box, YOLOv3 first calculates the center coordinates of the bounding box, as well as the width and height of the bounding box, where the coordinates are offsets relative to the top-left corner of the grid. Then, YOLOv3 converts the center coordinates and width and height into the coordinates of the top-left and bottom-right corners of the bounding box, and finally obtains a real bounding box;

[0055] ​The non-maximum suppression algorithm is used to screen the overlapping bounding boxes: for each class, first, all the bounding boxes with confidence higher than the threshold are sorted in descending order of confidence, then the bounding box with the highest confidence is selected and compared with all other bounding boxes, if the overlap degree (i.e. the ratio of the intersection area to the union area) of the two bounding boxes is greater than a certain threshold (usually 0.5), the bounding box with lower confidence is filtered out, until all the bounding boxes are screened; finally, YOLOv3 outputs the screened bounding boxes, their classes and confidences as detection results, which are visualized and displayed on the input image, and the class and confidence of each object are given. In general, YOLOv3 adopts a series of optimization measures, such as efficient CNN structure, better activation function, data augmentation, etc., so as to improve the detection accuracy and speed.

[0056] YOLO (You Only Look Once) is a classic single-stage object detection algorithm, which does not need to scan the image multiple times or use sliding windows to detect objects, but directly predicts on the image to achieve real-time performance. The core idea is to divide the image into grids and predict bounding boxes and probability scores in each grid, and then classify and locate objects according to the prediction results. The present application uses YOLOv3 with FPN feature pyramid for enhanced feature extraction and YoloHead effective feature layer for prediction, which can further improve the detection performance. Based on the improved version of YOLOv3 algorithm, an image adaptive mechanism is used to automatically adjust the algorithm's hyperparameters according to the quality and weather conditions of the input image, improving the target detection effect in foggy weather.

[0057] The key points of the present application are:

[0058] 1. The dual-tree complex wavelet transform feature enhancement module fully excavates the robust features in the rain and snow images by using the multi-directional selectivity of the dual-tree complex wavelet.

[0059] 2. Dynamic convolution is used to speed up the algorithm, and the parameters of the convolution are dynamically adjusted according to different rain and snow images instead of using the same parameters, which can balance between denoising performance and computational cost.

[0060] 3. An image restoration algorithm for the stripe characteristics of rain and snow images is designed uniformly.

[0061] 4. The severe weather image restoration is applied to traffic monitoring target detection by using the YOLO series algorithm, which improves the usability of the traffic monitoring system in rainy and snowy weather, and can obtain better performance with the upgrade of the YOLO series algorithm.

[0062] The application can better adapt to complex adverse weather scenes and improve the effect and accuracy of image restoration by using the multi-direction and multi-scale characteristics of the dual-tree complex wavelet transform; the processing capability for complex scenes is improved by combining structural information through a dynamic convolution layer to mine robust features in the rain, snow and fog images; the quality and accuracy of image restoration are further improved by using a residual network structure to refine the obtained features; and a clean image is reconstructed through residual learning operation, so that a high-quality input image can be obtained, thereby improving the traffic target detection performance.

[0063] The method of the application is universal and can be applied to monitoring image restoration in various adverse weather scenes.

[0064] The application is described above in conjunction with the drawings, and it is obvious that the specific implementation of the application is not limited by the above manner, and various improvements using the method concept and technical solution of the application or direct application to other occasions without improvement are within the protection scope of the application.

Claims

1. A target detection method for traffic monitoring images under severe weather conditions, characterized in that, The overall network structure includes a five-layer dynamic convolutional module, two eight-layer dual-tree complex wavelet enhancement modules, a ten-layer residual learning recovery module, and a YOLOv3 traffic target detection module; The specific steps are as follows: S1. The monitoring image to be recovered is input into the dynamic convolution module for convolution processing; The process is as follows: The dynamic convolution module has five layers. The first layer is a convolutional layer that downsamples the input monitoring image to be restored. The kernel size is 5×5 and the output channel is 64. The second to fourth layers are two weight generators and a convolutional layer with a kernel size of 5×5, respectively. The fifth layer is a ReLU activation function that extracts features from the dynamic convolution and has 64 output channels. S2. After passing through the dynamic convolution module, the image is input into the dual complex wavelet enhancement module; The process is as follows: the linear structure information is converted into six frequency features using dual-tree complex wavelet transform, and the frequency features are enhanced by residual learning feature enhancement mechanism to suppress rain and snow stripes, thereby recovering detailed information in the image. Finally, the frequency features are converted into linear structure information using dual-tree complex wavelet inverse transform. S3. The information enhanced by dual-tree complex wavelet transform is fed into the residual learning and recovery module for residual learning to obtain a clean image. The process is as follows: The information enhanced by the dual-tree complex wavelet transform is used to perform the first residual learning using a combination of four residual dense layers and the ReLU activation function to further refine the features in image denoising; The output of the first residual learning is then used to perform another residual learning using a combination of four residual dense layers and the ReLU activation function. The output of the second residual learning is fused with the output of the first residual learning to enhance the deep information of the image; finally, a convolutional layer followed by ReLU is used to refine these features, with 64 input channels and 64 output channels respectively. S4. The image output after residual learning is sent to the YOLOv3 traffic target detection module for target detection. Target detection is performed on the restored image. The process is as follows: In step S3, the clean image obtained through residual learning is used as the input image. The input image is processed by YOLOv3, which divides the input image into 13×13 grids. For each grid, YOLOv3 predicts B bounding boxes and predicts the probabilities of C classes for each bounding box. At the same time, it predicts the confidence score of each bounding box. If the confidence score of a bounding box is lower than the threshold, it is filtered out. Then, feature extraction is performed. The extracted feature map is used to predict bounding boxes. The information of the bounding boxes is decoded and non-maximum suppression is applied. Finally, the detection result is output, providing traffic vehicle information and confidence scores.

2. The target detection method for traffic monitoring images under severe weather conditions according to claim 1, characterized in that, In step S1, the weight generator includes a pooling layer, a 1×1 convolution, an activation function, and a softmax classifier. The specific implementation of dynamic convolution is as follows: four weights are obtained using the weight generator, and the parameters are adjusted by applying them to four parallel convolution kernels in a weighted manner to obtain dynamic convolution. The output of the first convolutional layer is then convolved using dynamic convolution.

3. The target detection method for traffic monitoring images under severe weather conditions according to claim 1, characterized in that, In step S2, feature enhancement is performed using four residual dense layers, including three convolutional layers with activation functions and one 1×1 convolutional layer. The kernel size of the three convolutional layers is 5×5.

4. The target detection method for traffic monitoring images under severe weather conditions according to claim 1, characterized in that, In step S3, the size of all convolutional kernels in the residual learning recovery module is 5×5, the number of input channels and output channels of the two residual blocks is 64, the number of input channels of the last convolutional layer is 64, and the mean squared error is selected as the loss function to train the network.

5. A target detection method for traffic monitoring images under severe weather conditions according to claim 4, characterized in that, The specific operation of using mean squared error as the loss function to train the network is as follows: the mean squared error is used in a supervised manner to train the denoising model using the {IC, IL} pair, where IC and IL are the clean image and the image to be restored, respectively. The Adam optimizer is used to optimize the model to obtain appropriate parameters.

6. A target detection method for traffic monitoring images under severe weather conditions according to claim 1, characterized in that, In step S4, YOLOv3 extracts features from the input image through a convolutional neural network. YOLOv3 uses Darknet53 as its backbone network. In Darknet53, the input image is first subjected to multiple convolution and pooling operations, and then the feature map is further processed using residual structures to finally obtain a high-level feature representation. On the feature map output by the last convolutional layer of Darknet53, YOLOv3 uses a convolutional layer to predict bounding boxes: YOLOv3 uses a 1×1 convolutional layer to reduce the depth of the feature map to B×(5+C), where B is the number of bounding boxes predicted for each cell, 5 represents the position and confidence information of each bounding box, and C is the number of predicted classes. Therefore, for each cell and each bounding box, YOLOv3 predicts a position, a confidence value, and the probabilities of C classes. YOLOv3 converts the predicted bounding boxes into ground truth bounding boxes. For each cell and each bounding box, YOLOv3 first calculates the center coordinates, width, and height of the bounding box, where the coordinates are the offsets relative to the top-left corner of the cell. Then, YOLOv3 converts the center coordinates and width and height into the coordinates of the top-left and bottom-right corners of the bounding box, finally obtaining a ground truth bounding box. The non-maximum suppression algorithm is used to filter overlapping bounding boxes: For each category, all bounding boxes with confidence scores higher than a threshold are first sorted from highest to lowest confidence score. Then, the bounding box with the highest confidence score is selected and compared with all other bounding boxes. If the overlap between two bounding boxes is greater than a certain threshold, the bounding box with the lower confidence score is filtered out until all bounding boxes have been filtered. Finally, YOLOv3 outputs the filtered bounding boxes, their categories, and confidence scores as the detection results.

7. A target detection method for traffic monitoring images under severe weather conditions according to claim 6, characterized in that, If the overlap between two bounding boxes is greater than 0.5, the bounding box with the lower confidence level will be filtered out.

8. A target detection method for traffic monitoring images under severe weather conditions according to claim 6, characterized in that, YOLOv3 outputs the filtered bounding boxes along with their categories and confidence scores as detection results. The detection results are then visualized on the input image, and the category and confidence score for each object are provided.

Images