Rainy night lane line identification method

Abstract

The invention relates to the technical field of lane line identification, in particular to a rain night lane line identification method, which comprises the following steps of: inputting a rain night lane line picture into a night rain stripe detection module, and processing through a PReNet network progressive recursion mode in the night rain stripe detection module; obtaining rain stripe position information endowed with a higher weight on the rain night lane line picture; inputting the rain stripe position information and the rain night lane line picture into an image restoration network Uform at the same time for restoration, and deleting rain stripes corresponding to the rain stripe position information in the rain night lane line picture to obtain an enhanced image x; and the enhanced image x is input into a lane line detection module, detection is carried out through an improved ERFNet detection network in the lane line detection module, and a lane line detection result is obtained through detection. According to the method, the problems that the lane line image is covered by rain stripes at rainy night and the lane line detection precision is poor are solved, and the lane line detection precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of lane line recognition, and in particular to a method for recognizing lane lines on rainy nights. Background Art

[0002] Lane recognition is a key part of the autonomous driving technology system. The fundamental challenge is to ensure real-time and accurate detection of lanes. The situation is particularly complex when the vehicle is operating in rainy conditions at night. At this time, the sky often appears dim, making the lanes, which are already susceptible to environmental factors, appear even more blurred in the images captured by the camera.

[0003] On rainy nights, the road surface is soaked by rain, and the light from the headlights and street lamps interweaves on the wet road surface, producing a significant light scattering effect. This scattering phenomenon not only disperses the light originally focused on the lane line, but also causes a large number of disordered light spots and reflections in the image, reducing the visual contrast and boundary clarity of the lane line. The resulting complex lighting environment greatly increases the possibility of positioning deviation or missing lane lines in the lane line detection algorithm when processing such images. Summary of the invention

[0004] In view of this, the purpose of the present invention is to propose a lane line recognition method for rainy nights, so as to solve the problem that in the complex environment of rainy nights, due to the obscuration of rain streaks, the possibility of positioning deviation or missing lane lines in the lane line detection algorithm when processing such images is greatly increased.

[0005] Based on the above purpose, the present invention provides a method for lane line recognition in rainy night, comprising the following steps:

[0006] Get lane line image data on rainy nights;

[0007] The rainy night lane line picture is input into the night rain streak detection module, and processed by the progressive recursive method of the PReNet network in the night rain streak detection module to obtain the rain streak position information with a higher weight on the rainy night lane line picture;

[0008] The rain streak position information and the rainy night lane line picture are simultaneously input into the image restoration network Uformer for restoration, and the rain streaks corresponding to the rain streak position information in the rainy night lane line picture are deleted to obtain the enhanced image x;

[0009] The enhanced image x is input into the lane line detection module, and is detected by the improved ERFNet detection network in the lane line detection module, and a lane line detection result is obtained through the detection.

[0010] Optionally, the architecture of the PReNet network consists of an input layer, a residual module, a coordinate attention module, a recurrent layer, and an output layer;

[0011] The input layer integrates a standard convolution with a kernel size of 3×3 and a padding of 1, and cooperates with the ReLU activation function to convert the 6-channel input into a 32-channel feature representation. In the network design, the intermediate image generated in the previous stage is concatenated with the original low-quality input image as the input data for processing in the current stage;

[0012] The residual module consists of two conventional convolutions with ReLU activation function, where the convolution kernel is 3×3 and the padding is 1. The residual module is used to learn the deep features of the image;

[0013] The coordinate attention module is used to give a higher weight to the rain streak position information;

[0014] The output layer is a common convolution with a convolution kernel of 3×3, which is used to reconvert the number of channels of the input feature map to 3.

[0015] Optionally, a mixed dilated convolution is used in the residual module, and the mixed dilated convolution group consists of 4 dilated convolutions with dilated rates of [1, 2, 5]. The mixed dilated convolution is used to expand the receptive field to obtain global information and reduce the checkerboard effect.

[0016] The dilated convolution introduces a parameter: dilation rate. The dilation rate of the standard convolution is set to 1. The dilation rate is equal to the number of zero values ​​inserted between the convolution kernels plus 1. The calculation formula is:

[0017] r = z + 1;

[0018] Where r is the dilation rate; z is the number of zero values ​​inserted between the convolution kernels.

[0019] Optionally, the step of assigning a higher weight to the rain streak position information by the coordinate attention module includes:

[0020] The coordinate attention module includes a coordinate information embedding unit and a coordinate attention generating unit. In the coordinate information embedding unit, by inputting a feature map X1, the feature map X1 represents the feature map obtained after being processed by a convolution module, a long short-term memory module, and a residual module in the PReNet network. The sizes of the two pooling kernels in the PReNet network are (H, 1) and (1, W), respectively. The channel attention is pooled and decomposed into two parallel 1-dimensional vectors in the horizontal and vertical directions through the two pooling kernels. Assume that the feature map x1 of the cth channel c The size of is H×W, and the pooling output is as follows;

[0021]

[0022] in and Represents a 1-dimensional vector containing h and w elements respectively, W is the feature map x c The width of H is the feature map x c The height, x c (h,i) is the value of the pixel at position (h,i) in the input feature map X1 in the cth channel, x c (j,w) is the value of the pixel at position (j,w) in the input feature map X in the cth channel;

[0023] In the coordinate attention generation unit, the spatial information encoding in both horizontal and vertical directions is aggregated into feature maps, and the formula is as follows;

[0024] f=δ(F1([z h ,z w ]))

[0025] Where f is the intermediate feature information encoding spatial information in the horizontal and vertical directions, which is decomposed into f h ,f w ; δ is the activation function, and F1 is the fusion operation in the horizontal and vertical directions;

[0026] Use two 1×1 convolutions to transform f h ,f w Transformed into a new feature vector g with the same channels as the original input feature map X1 h ,g w ;

[0027] g h =σ(F h (f h ))

[0028] g w =σ(F w (f w ))

[0029] Where σ is the Sigmoid function, and the coordinate attention output is as follows;

[0030]

[0031] where y c (i,j) is the output of the coordinate attention module, x c (i, j) is the value at position (i, j) in the input feature map X1;

[0032] By y c (i, j) gives higher weight to the rain streak location information.

[0033] Optionally, the loss function in the image restoration network Uformer uses a combination of MS-SSIM loss and L1 loss, where the calculation formula of MS-SSIM is as follows;

[0034]

[0035]

[0036] Among them, α, β and γ are the importance adjustment parameters of their corresponding sub-functions, μ x and μ y are the means of the enhanced image x and the real image y respectively; σ x and σ y are the variances of the enhanced image x and the real image y respectively; σ xy is the covariance between the enhanced image x and the real image y; l(x,y) is the brightness similarity between the enhanced image x and the real image y; c(x,y) is the contrast similarity between the enhanced image x and the real image y; s(x,y) is the structural similarity between the enhanced image x and the real image y; C1, C2, C3 are constants;

[0037] The formula of MS-SSIM loss function is as follows;

[0038] L MS-SSIM (p) = 1 - [MS-SSIM (p)]

[0039] Where p is the pixel coordinate in the image; the overall loss function of the image restoration network Uformer is as follows;

[0040] L=ε·L MS-SSIM +(1-ε)·L1, where L MS-SSIM is the multi-scale structural similarity loss, L1 is the L1 loss, also known as the mean absolute error (MAE), which is used to measure the absolute difference between the predicted value and the true value, and ε is the weight factor ranging from 0 to 1, which is used to balance L MS-SSIM and the contribution of L1.

[0041] Optionally, the step of inputting the enhanced image x into an improved ERFNet detection network for detection, and obtaining a lane line detection result through detection includes:

[0042] For the improved ERFNet lane detection network, self-attention distillation is introduced, and the attention map of the back layer of the encoding layer is used to distill the attention map of the front layer, so that the attention map of the front layer can learn the contextual information of the attention map of the back layer, thereby improving the network's detection accuracy of lane lines.

[0043] Optionally, the improved ERFNet network expands its encoding layer to 4 layers while adding self-attention distillation, adds an attention map generator after the encoding modules E1, E2, E3, and E4, and extracts the attention maps of the corresponding encoding modules;

[0044] The formula for the attention map generator is as follows:

[0045]

[0046] Where B is the bilinear upsampling operation, Φ is the spatial softmax operation, It is the mapping relationship from the entire image to the area that needs attention, and its formula is as follows:

[0047]

[0048] Among them A i For the set of all elements of the i-th channel, use the later layer attention map such as E2 to distill the feature attention map of E1, so that the attention map of E1 learns the contextual relationship of the attention map of E2, thereby improving the accuracy of lane line detection.

[0049] In this method, a rainy night lane line picture is first obtained, and then the rainy night lane line picture is input into a night rain streak detection module, and processed by a progressive recursive method of a PReNet network in the night rain streak detection module to obtain rain streak position information with a higher weight on the rainy night lane line picture. At this time, the position information of the rain streaks on the rainy night lane line picture can be obtained more accurately, and then the rain streak position information and the rainy night lane line picture are simultaneously input into an image restoration network Uformer for repair, and the rain streaks corresponding to the rain streak position information in the rainy night lane line picture are deleted, thereby obtaining an enhanced image x with the rain streaks removed.

[0050] From the above, it can be seen that this method solves the problem that the lane line image is obscured by rain streaks on rainy nights and the lane line detection accuracy is poor, improves the accuracy of lane line detection, and provides a guarantee for unmanned driving technology on rainy nights. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0052] Figure 1 It is a structural block diagram of the PReNet model in an embodiment of the present invention;

[0053] Figure 2It is a block diagram of the Uformer network structure in an embodiment of the present invention;

[0054] Figure 3 is a block diagram of a network structure for obtaining an enhanced image in an embodiment of the present invention;

[0055] Figure 4 It is a data graph of Night-Rain-Cityscpes in an embodiment of the present invention;

[0056] Figure 5 is a network structure block diagram of a residual module in an embodiment of the present invention;

[0057] Figure 6 is a network structure diagram of a coordinate attention module in an embodiment of the present invention;

[0058] Figure 7 1 is a network block diagram of a lane detection module in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0060] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0061] like Figure 1-Figure 3 As shown, a lane line recognition method for rainy nights includes the following steps:

[0062] Get lane line image data on rainy nights;

[0063] The rainy night lane line picture is input into the night rain streak detection module, and processed by the progressive recursive method of the PReNet network in the night rain streak detection module to obtain the rain streak position information with a higher weight on the rainy night lane line picture;

[0064] The rain streak position information and the rainy night lane line picture are simultaneously input into the image restoration network Uformer for restoration, and the rain streaks corresponding to the rain streak position information in the rainy night lane line picture are deleted to obtain the enhanced image x;

[0065] The enhanced image x is input into the lane line detection module, and is detected by the improved ERFNet detection network in the lane line detection module, and a lane line detection result is obtained through the detection.

[0066] In this method, a rainy night lane line picture is first obtained, and then the rainy night lane line picture is input into a night rain streak detection module, and processed by a progressive recursive method of a PReNet network in the night rain streak detection module to obtain rain streak position information with a higher weight on the rainy night lane line picture. At this time, the position information of the rain streaks on the rainy night lane line picture can be obtained more accurately, and then the rain streak position information and the rainy night lane line picture are simultaneously input into an image restoration network Uformer for repair, and the rain streaks corresponding to the rain streak position information in the rainy night lane line picture are deleted, thereby obtaining an enhanced image x with the rain streaks removed.

[0067] From the above, it can be seen that this method solves the problem that the lane line image is obscured by rain streaks on rainy nights and the lane line detection accuracy is poor, improves the accuracy of lane line detection, and provides a guarantee for unmanned driving technology on rainy nights.

[0068] Among them, the parameters in the night rain streak detection module and the lane line detection module can be obtained through training, and the rainy night lane line picture data set can be established, including: adjusting the contrast of the pictures in the Cityscapes data set to simulate the night rain scene. The original Cityscapes data set contains 9432 training images and 1188 test images. After the above processing, its visual effect is closer to the rainy night scene in the real world. A rainy night lane line picture data set is established based on the pictures in the adjusted Cityscapes data set. The pictures in the rainy night lane line picture data set are input into the night rain streak detection module and the lane line detection module for training through the above method, so as to obtain a detection model that meets the predetermined requirements. Figure 4 As shown, the network model involved in the present invention can also be trained on the Night-Rain-Cityscapes synthetic dataset, and the network performance can be further evaluated using a synthetic test set specially constructed in the dataset.

[0069] In some embodiments, the architecture of the PReNet network consists of an input layer, a residual module, a coordinate attention module, a recurrent layer, and an output layer;

[0070] The input layer integrates a standard convolution with a kernel size of 3×3 and a padding of 1, and cooperates with the ReLU activation function to convert the 6-channel input into a 32-channel feature representation. In the network design, the intermediate image generated in the previous stage is concatenated with the original low-quality input image as the input data for processing in the current stage;

[0071] The residual module consists of two conventional convolutions with ReLU activation function, where the convolution kernel is 3×3 and the padding is 1. The residual module is used to learn the deep features of the image;

[0072] The coordinate attention module is used to give a higher weight to the rain streak position information;

[0073] The output layer is a common convolution with a convolution kernel of 3×3, which is used to convert the number of channels of the input feature map back to 3.

[0074] Optional, such as Figure 5 As shown in FIG. 1 , the residual module uses a mixed dilated convolution. The purpose of introducing the mixed dilated convolution is to expand the receptive field and obtain large-scale image features. The mixed dilated convolution group consists of 4 dilated convolutions with dilation rates of [1, 2, 5].

[0075] Compared with standard convolution, dilated convolution introduces a parameter: dilation rate. The dilation rate of standard convolution is set to 1. The dilation rate is equal to the number of zero values ​​inserted between the convolution kernels plus 1. The calculation formula is:

[0076] r = z + 1;

[0077] Where r is the dilation rate; z is the number of zero values ​​inserted between the convolution kernels.

[0078] Compared with traditional convolution, dilated convolution only introduces one parameter, dilated rate, to expand the receptive field without reducing the resolution of feature maps or increasing the amount of computation. Using dilated convolution alone will produce a checkerboard effect, and a checkerboard-like grid will appear in the image, so mixed dilated convolution is used. Mixed dilated convolution uses a set of dilated convolutions with different dilated rates to form a convolution combination. Its advantage is that it can expand the receptive field to obtain global information, while reducing the checkerboard effect.

[0079] like Figure 6 As shown, in some embodiments, the method of assigning a higher weight to the rain streak position information by the coordinate attention module includes:

[0080] The coordinate attention module includes a coordinate information embedding unit and a coordinate attention generating unit. In the coordinate information embedding unit, a feature map X1 is input, where the feature map X1 represents a feature map obtained after being processed by a convolution module (Conv), a long short-term memory module (LSTM), and a residual module (Res) in a PReNet network, such as Figure 3 As shown in the figure, the two pooling kernels in the PReNet network have sizes of (H, 1) and (1, W) respectively. The two pooling kernels are used to decompose the channel attention into two parallel 1-dimensional vectors in the horizontal and vertical directions. The output of the feature map X1 of the c-th channel at height h and width w is shown below;

[0081]

[0082] Where W is the width of the feature map, H is the height of the feature map, and x c (h,i) is the value of the pixel at position (h,i) in the input feature map X1 in the cth channel, x c (j,w) is the value of the pixel at position (j,w) in the input feature map X in the cth channel;

[0083] In the coordinate attention generation unit, the spatial information encoding in both horizontal and vertical directions is aggregated into feature maps, and the formula is as follows;

[0084] f=δ(F1([z h ,z w ]))

[0085] Where f is the intermediate feature information encoding spatial information in the horizontal and vertical directions, which is decomposed into f h ,f w ; δ is the activation function, and F1 is the fusion operation in the horizontal and vertical directions;

[0086] Use two 1×1 convolutions to transform f h ,f w Transformed into a new feature vector g with the same channels as the original input feature map X1 h ,g w ;

[0087] g h =σ(F h (f h ))

[0088] g w =σ(F w (f w ))

[0089] Where σ is the Sigmoid function, and the coordinate attention output is as follows;

[0090]

[0091] where y c (i,j) is the output of the coordinate attention module, x c (i, j) is the value at position (i, j) in the input feature map X1;

[0092] By y c (i, j) gives higher weight to the rain streak location information.

[0093] As shown in the formula x c (i,j) is the value of the original feature map in channel c, row i, and column j. is the value of the attention module in the i-th row and j-th column, and these values ​​give higher weight to the rain streak position information.

[0094] In order to emphasize the rain streak location information, the coordinate attention module is used after the residual module of the PReNet network to give higher weights to the rain streak location area. The coordinate attention module effectively guides the network to focus on important features by assigning different weights to the extracted features. In the currently widely used attention mechanism, in convolutional neural networks, the output of each convolutional layer is usually multiple feature maps, each feature map corresponds to a channel. For example, the SE module and the ECA module, although emphasizing the information processing at the channel level, ignore the key role of position information. In order to more effectively capture the position information and consider the correlation between channels, the coordinate attention module decomposes the conventional two-dimensional global pooling operation into two independent one-dimensional encoding steps, and integrates the position information into the channel attention calculation. This can accurately locate the position of the feature.

[0095] The Uformer network usually uses the Charbonnier loss function during training. The Charbonnier loss function is a type of L1 loss function, which does not take into account the visual perception effect of humans. To solve this problem, in some embodiments, the loss function in the image restoration network Uformer uses a combination of MS-SSIM loss and L1 loss, where the calculation formula of MS-SSIM is as follows:

[0096]

[0097] Among them, α, β and γ are the importance adjustment parameters of their corresponding sub-functions, μ x and μ y are the means of the enhanced image x and the real image y respectively; σ x and σ y are the variances of the enhanced image x and the real image y respectively; σ xyis the covariance between the enhanced image x and the real image y; l(x,y) is the brightness similarity between the enhanced image x and the real image y; c(x,y) is the contrast similarity between the enhanced image x and the real image y; s(x,y) is the structural similarity between the enhanced image x and the real image y; C1, C2, C3 are constants;

[0098] The formula of MS-SSIM loss function is as follows;

[0099] L MS-SSIM (p) = 1 - [MS-SSIM (p)]

[0100] Where p is the pixel coordinate in the image; the overall loss function of the image restoration network Uformer is as follows;

[0101] L=ε·L MS-SSIM +(1-ε)·L1, where L MS-SSIM is the multi-scale structural similarity loss. MS-SSIM is an indicator used to measure the structural similarity between two images, which can better reflect the perceptual quality of the image. L1 is the L1 loss, also known as the mean absolute error (MAE), which is used to measure the absolute difference between the predicted value and the true value. ε is a weight factor ranging from 0 to 1, which is used to balance L MS-SSIM and the contribution of L1.

[0102] By combining MS-SSIM loss and L1 loss, human visual perception can be taken into account, thereby improving the image quality of the enhanced image x.

[0103] In some embodiments, inputting the enhanced image x into the improved ERFNet detection network for detection, and obtaining the lane line detection result through detection includes:

[0104] For the improved ERFNet lane detection network, self-attention distillation is introduced, and the attention map of the back layer of the encoding layer is used to distill the attention map of the front layer, so that the attention map of the front layer can learn the contextual information of the attention map of the back layer, thereby improving the network's detection accuracy of lane lines.

[0105] The enhanced image is input into the improved lane detection module network ERFNet. The improved ERFNet network is as follows Figure 7 shown.

[0106] The improved ERFNet network adds self-attention distillation and expands its encoding layer to 4 layers. An attention map generator is added after the encoding modules E1, E2, E3, and E4 to extract the attention map of the corresponding encoding module. Self-attention distillation is a method that combines the self-attention mechanism with knowledge distillation. Its purpose is to guide the model to focus on more important feature areas through the attention mechanism during training, thereby improving the performance of the model. Figure 7 The input image passes through encoding modules E1, E2, E3 and E4 in sequence. After each encoding module, a corresponding attention map is generated and weighted fused with the feature map. The formula of the attention map generator is as follows:

[0107]

[0108] Where B is the bilinear upsampling operation and Φ is the spatial softmax operation. It is the mapping relationship from the entire image to the area that needs attention, and its formula is as follows:

[0109]

[0110] Among them A i is the set of all elements of the i-th channel. Use the later layer attention map such as E2 to distill the feature attention map of E1. Make the attention map of E1 learn the contextual relationship of the attention map of E2, thereby improving the accuracy of lane line detection.

[0111] It should be understood by those skilled in the art that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present invention (including the claims) is limited to these examples. Within the scope of the present invention, the technical features in the above embodiments or in different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the above aspects of the present invention, which are not provided in detail for the sake of simplicity.

[0112] The present invention is intended to cover all such substitutions, modifications and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for lane line recognition on rainy nights, characterized in that: The following steps are involved: Get lane line image data on rainy nights; The rainy night lane line picture is input into the night rain streak detection module, and processed by the progressive recursive method of the PReNet network in the night rain streak detection module to obtain the rain streak position information with a higher weight on the rainy night lane line picture; The rain streak position information and the rainy night lane line picture are simultaneously input into the image restoration network Uformer for restoration, and the rain streaks corresponding to the rain streak position information in the rainy night lane line picture are deleted to obtain the enhanced image x; The enhanced image x is input into the lane line detection module, and is detected by the improved ERFNet detection network in the lane line detection module, and a lane line detection result is obtained through the detection.

2. The method for identifying lane lines on rainy nights according to claim 1, characterized in that: The architecture of the PReNet network consists of an input layer, a residual module, a coordinate attention module, a recurrent layer, and an output layer; The input layer integrates a standard convolution with a kernel size of 3×3 and a padding of 1, and cooperates with the ReLU activation function to convert the 6-channel input into a 32-channel feature representation. In the network design, the intermediate image generated in the previous stage is concatenated with the original low-quality input image as the input data for processing in the current stage; The residual module consists of two conventional convolutions with ReLU activation function, where the convolution kernel is 3×3 and the padding is 1. The residual module is used to learn the deep features of the image; The coordinate attention module is used to give a higher weight to the rain streak position information; The output layer is a common convolution with a convolution kernel of 3×3, which is used to convert the number of channels of the input feature map back to 3.

3. The method for identifying lane lines on rainy nights according to claim 2, characterized in that: A mixed dilated convolution is used in the residual module. The mixed dilated convolution group consists of 4 dilated convolutions with dilated rates of [1, 2, 5]. The mixed dilated convolution is used to expand the receptive field to obtain global information and reduce the checkerboard effect. The dilated convolution introduces a parameter: dilation rate. The dilation rate of the standard convolution is set to 1. The dilation rate is equal to the number of zero values ​​inserted between the convolution kernels plus 1. The calculation formula is: r＝z+1； Where r is the dilation rate; z is the number of zero values ​​inserted between the convolution kernels.

4. The method for identifying lane lines on rainy nights according to claim 2, characterized in that: The method of using the coordinate attention module to assign a higher weight to the rain streak position information includes: The coordinate attention module includes a coordinate information embedding unit and a coordinate attention generating unit. In the coordinate information embedding unit, by inputting a feature map X1, the feature map X1 represents the feature map obtained after being processed by a convolution module, a long short-term memory module, and a residual module in the PReNet network. The sizes of the two pooling kernels in the PReNet network are (H, 1) and (1, W), respectively. The channel attention is pooled and decomposed into two parallel 1-dimensional vectors in the horizontal and vertical directions through the two pooling kernels. Assume that the feature map x1 of the cth channel c The size of is H×W, and the pooling output is as follows; in and Represents a 1-dimensional vector containing h and w elements respectively, W is the feature map x c The width of H is the feature map x c The height, x c (h,i) is the value of the pixel at position (h,i) in the input feature map X1 in the cth channel, x c (j,w) is the value of the pixel at position (j,w) in the input feature map X in the cth channel; In the coordinate attention generation unit, the spatial information encoding in both horizontal and vertical directions is aggregated into feature maps, and the formula is as follows; f=δ(F1([z h ,z w ])) Where f is the intermediate feature information encoding spatial information in the horizontal and vertical directions, which is decomposed into f h ,f w ; δ is the activation function, and F1 is the fusion operation in the horizontal and vertical directions; Use two 1×1 convolutions to transform f h ,f w Transformed into a new feature vector g with the same channels as the original input feature map X1 h ,g w ; g h =σ(F h (f h )) g w =σ(F w (f w )) Where σ is the Sigmoid function, and the coordinate attention output is as follows; where y c (i,j) is the output of the coordinate attention module, x c (i, j) is the value at position (i, j) in the input feature map X1; By y c (i, j) gives higher weight to the rain streak location information.

5. The method for identifying lane lines on rainy nights according to claim 1, characterized in that: The loss function in the image restoration network Uformer uses a combination of MS-SSIM loss and L1 loss, where the calculation formula of MS-SSIM is as follows: Among them, α, β and γ are the importance adjustment parameters of their corresponding sub-functions, μ x and μ y are the means of the enhanced image x and the real image y respectively; σ x and σ y are the variances of the enhanced image x and the real image y respectively; σ xy is the covariance between the enhanced image x and the real image y; l(x,y) is the brightness similarity between the enhanced image x and the real image y; c(x,y) is the contrast similarity between the enhanced image x and the real image y; s(x,y) is the structural similarity between the enhanced image x and the real image y; C1, C2, C3 are constants; The formula of MS-SSIM loss function is as follows; L MS-SSIM (p)=1-[MS-SSIM(p)] Where p is the pixel coordinate in the image; the overall loss function of the image restoration network Uformer is as follows; L=ε·L MS-SSIM +(1-ε)·L1, where L MS-SSIM is the multi-scale structural similarity loss, L1 is the L1 loss, which is used to measure the absolute difference between the predicted value and the true value, and ε is the weight factor, ranging from 0 to 1, which is used to balance L MS-SSIM and the contribution of L1.

6. The method for identifying lane lines on rainy nights according to claim 1, characterized in that: The enhanced image x is input into the improved ERFNet detection network for detection, and the lane line detection result obtained by detection includes: For the improved ERFNet lane detection network, self-attention distillation is introduced, and the attention map of the back layer of the encoding layer is used to distill the attention map of the front layer, so that the attention map of the front layer can learn the contextual information of the attention map of the back layer, thereby improving the network's detection accuracy of lane lines.

7. The method for identifying lane lines on rainy nights according to claim 1, characterized in that: The improved ERFNet network expands its encoding layer to 4 layers while adding self-attention distillation, adds an attention map generator after the encoding modules E1, E2, E3, and E4, and extracts the attention maps of the corresponding encoding modules; The formula for the attention map generator is as follows: Where B is the bilinear upsampling operation, Φ is the spatial softmax operation, It is the mapping relationship from the entire image to the area that needs attention, and its formula is as follows: Among them A i For the set of all elements of the i-th channel, use the later layer attention map such as E2 to distill the feature attention map of E1, so that the attention map of E1 learns the contextual relationship of the attention map of E2, thereby improving the accuracy of lane line detection.

Images