Traffic identification model construction method and model suitable for low-illumination snow scene

Abstract

The application discloses a traffic recognition model construction method and model suitable for a low-illumination snow scene. The technical problem of a traffic recognition model suitable for a low-illumination snow scene and used for a polarized image is solved in the field of traffic recognition. A YOLOv11 model is improved to obtain a traffic recognition model suitable for a low-illumination snow scene. By introducing a high-efficiency feature selection module, high-frequency snow noise can be effectively suppressed in a low-illumination snow scene, while the structural edge information of a target and the detection ability of a small target are maintained. A C2PSA_EDFFN module can effectively suppress non-structured snow noise and enhance high-frequency details with structural significance. A displacement-based interactive channel mixing module explicitly enhances the interaction ability between channels at a very low calculation cost, can more effectively retain fine-grained structural information, improve the feature fusion quality of small targets and low-contrast targets, and meanwhile maintains the lightweight characteristics of the overall calculation overhead.

Description

Technical Field

[0001] This invention belongs to the field of detection model construction technology, specifically relating to a traffic recognition model construction method and model suitable for low-light snowy weather scenarios. Background Technology

[0002] Object detection is a crucial research task in computer vision, and road environment perception is one of the fundamental key technologies in autonomous driving systems. Most existing object detection methods are based on visible light RGB images for modeling and recognition, relying primarily on apparent information such as light intensity, color, and texture. However, in low-light snowy weather or harsh environmental conditions, RGB images are prone to reduced contrast, increased noise, and loss of detail due to insufficient light, snow particle scattering, and high-reflection from the ground. This leads to a decline in the ability to represent target features, significantly reducing detection accuracy and stability. Polarization imaging, on the other hand, can characterize the polarization state of light waves while acquiring light intensity information. Its polarization characteristics have a strong ability to distinguish between scattered light and high-reflection light, maintaining a relatively stable structural representation even in low-light conditions. This information is more helpful for better target detection in harsh conditions. Currently, in the field of traffic recognition, there is no traffic recognition model applicable to low-light snowy scenes using polarization images. Summary of the Invention

[0003] To address the technical problem that there is no traffic recognition model applicable to low-light snowy scenes using polarized images in the field of traffic recognition, this invention provides a method and model for constructing a traffic recognition model suitable for low-light snowy scenes.

[0004] The first aspect of this invention provides a method for constructing a traffic recognition model suitable for low-light snowy weather scenarios, specifically: S1. Select the YOLOv11 model as the base model and build the model based on the YOLOv11 model; S2. Improve the feature extraction part of the YOLOv11 model by replacing the C3K2 module. Specifically, replace the C3K2 module with an efficient feature selection module. S3. Improve the feature extraction part of the YOLOv11 model and improve the C2PSA module. Specifically, improve the C2PSA module by using an effective frequency domain discrimination-based feedforward network EDFFN. S4. Improve the feature fusion part of the YOLOv11 model by adding a displacement-based interactive channel mixing module before the two tensor stitching modules in the upsampling stage; and add an efficient feature selection module after the last tensor stitching module in the downsampling stage; thus obtaining a detection model suitable for low-light snow scenes. S5. Collect traffic scene images in snowy weather, construct a polarized image dataset, divide the dataset into training set, validation set and test set, use the polarized image dataset to train and test the detection model suitable for low-light snowy weather scenes, and obtain a target recognition model suitable for low-light snowy weather scenes that can be applied.

[0005] Furthermore, when replacing the C3K2 modules, the specific replacement targets are the third and fourth C3K2 modules in the feature extraction part from the input to the output direction.

[0006] Furthermore, the efficient feature selection module includes: an input convolution module, a channel splitting module, a dual-branch parallel structure, a tensor concatenation module, and an output convolution module connected in sequence; the dual-branch parallel structure includes a direct branch and a main branch, the main branch containing N cascaded C3K_ESFSConv modules, each C3K_ESFSConv module consisting of n cascaded Bottleneck_ESFSConv modules, and each Bottleneck_ESFSConv module consisting of two cascaded ESFSConv modules; the ESFSConv module is built on the existing spatial-frequency selected convolution SFSConv, and performs spatial-frequency selected convolution SFSConv operations on some channel features.

[0007] Furthermore, in the C3K_ESFSConv module, the received feature map is divided into two branches. One branch passes through n concatenated Bottleneck_ESFSConv modules in sequence, and the other branch is a direct connection branch. The outputs of the two branches are concatenated by the tensor concatenation module and then input into the convolution module. In the convolution module, channel fusion and channel compression are performed before output.

[0008] Furthermore, the efficient feature selection module includes: an input convolution module, a channel splitting module, a dual-branch parallel structure, a tensor concatenation module, and an output convolution module connected in sequence; the dual-branch parallel structure includes a direct branch and a main branch, the main branch consisting of N cascaded Bottleneck_ESFSConv modules, each Bottleneck_ESFSConv module consisting of two cascaded ESFSConv modules; the ESFSConv module is built on the existing spatial-frequency selected convolution SFSConv, and performs spatial-frequency selected convolution SFSConv operations on some channel features.

[0009] Furthermore, in the Bottleneck_ESFSConv module, the received feature map is divided into two branches. One branch passes through two cascaded ESFSConv modules in sequence, and the other branch is a direct connection branch. The outputs of the two branches are added element by element and then output.

[0010] Furthermore, the ESFSConv module is built on the existing spatial-frequency selected convolution SFSConv. The specific structure is as follows: the input feature map is split in the channel dimension, a part of the channel features are processed by spatial-frequency selected convolution SFSConv, and the other part of the channel features are directly passed while maintaining the identity mapping. Finally, the two parts of features are concatenated by the tensor concatenation module and output.

[0011] Furthermore, the C2PSA module is improved by using an effective frequency-domain discrimination-based feedforward network EDFFN. Specifically, the FFN unit of the PSABlock module in the C2PSA module is replaced with the frequency-domain discrimination-based feedforward network EDFFN.

[0012] Furthermore, the displacement-based interactive channel mixing module sequentially includes an upsampling module, a depthwise convolution module, a batch normalization layer, a ReLU activation module, a channel shuffling and channel displacement mixing module, and a 1*1 convolution module.

[0013] In another aspect, the present invention provides a traffic recognition model suitable for low-light snowy weather scenarios, wherein the model is constructed using the traffic recognition model construction method for low-light snowy weather scenarios described above.

[0014] The beneficial effects of the method and the constructed model described in this invention are as follows: The YOLOv11 model was improved to obtain a traffic recognition model suitable for low-light snowy scenes. By introducing the Efficient Feature Selection Block (EFS Block), high-frequency snow noise can be effectively suppressed in low-light snowy scenes while preserving the structural edge information of targets and the ability to detect small targets. The C2PSA_EDFFN module effectively suppresses unstructured snow noise while enhancing structurally significant high-frequency details. The Shift-based Interaction and Channel Mixing Block (SICM Block) explicitly enhances the interaction capability between channels with extremely low computational cost, more effectively preserving fine-grained structural information and improving the feature fusion quality for small and low-contrast targets, while maintaining the overall lightweight computational overhead. The improved network model was trained, and road target detection in polarized images was performed using the trained network model, resulting in higher detection performance and suitability for scene analysis of autonomous driving in low-light snowy conditions. Attached Figure Description

[0015] Figure 1 This is a structural diagram of a detection model applicable to low-light snowy weather scenarios in an embodiment of the present invention; Figure 2 This is a structural diagram of the efficient feature selection module in an embodiment of the present invention; Figure 3 This is a structural diagram of the C3K_ESFSConv module in an embodiment of the present invention; Figure 4 This is a structural diagram of the Bottleneck_ESFSConv module in an embodiment of the present invention; Figure 5 This is a structural diagram of the ESFSConv module in an embodiment of the present invention; Figure 6 This is a structural diagram of the PSABlock_EDFFN module in an embodiment of the present invention; Figure 7 This is a structural diagram of the C2PSA_EDFFN module in an embodiment of the present invention; Figure 8 This is a structural diagram of the displacement-based interactive channel hybrid module in an embodiment of the present invention; Figure 9 This is a schematic diagram of the highest brightness within the detectable brightness range during model testing as described in this embodiment of the invention; Figure 10 This is a schematic diagram of the lowest brightness within the range detectable during model testing in this embodiment of the invention. Detailed Implementation

[0016] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0017] Example 1 This implementation provides a method for constructing a traffic recognition model suitable for low-light snowy weather scenarios, the method including: S1. Select the YOLOv11 model as the base model and build the model based on the YOLOv11 model; S2. Improve the feature extraction part of the YOLOv11 model by replacing the C3K2 module. Specifically, replace the C3K2 module with the Efficient Feature Selection Block (EFS Block). The replacement targets are the 3rd and 4th C3K2 modules in the feature extraction part from input to output. S3. Improve the feature extraction part of the YOLOv11 model and improve the C2PSA module. Specifically, improve the C2PSA module by using an effective frequency domain discrimination-based feedforward network EDFFN; the improved C2PSA module is called the C2PSA_EDFFN module. S4. Improve the feature fusion part of the YOLOv11 model by adding a shift-based interaction and channel mixing block (SICM Block) before the two tensor concatenation (Concat) modules in the upsampling stage; and add an efficient feature selection module after the last tensor concatenation module in the downsampling stage; thus obtaining a detection model suitable for low-light snow scenes. S5. Collect traffic scene images in snowy weather, construct a polarized image dataset, divide the dataset into training set, validation set and test set, use the polarized image dataset to train and test the detection model suitable for low-light snowy weather scenes, and obtain an applicable target detection model suitable for low-light snowy weather scenes.

[0018] In step S1, a polarization image dataset is acquired and divided into a training set, a validation set, and a test set according to a ratio. The training set is then preprocessed. The polarization image dataset consists of polarization images of road targets under low-light snowy weather conditions.

[0019] Optionally, offline enhancement of the training set images can be performed using CLAHE (contrast-limited adaptive histogram equalization) to improve the model's learning of features under low light conditions, and Mosaic data augmentation strategy can be used during training to improve the model's generalization ability.

[0020] In one implementation, natural light is unpolarized, but becomes partially polarized after reflection from the surface of an artificial target. The polarization characteristics of this light can be described using the Stokes vector. The four parameters of the Stokes vector are I, Q, U, and V (also expressed as S0, S1, S2, and S3). I / S0 represents the total light intensity, Q / S1 represents the intensity difference between the horizontal and vertical linearly polarized components, U / S2 represents the intensity difference between the 45° and 135° linearly polarized components, and V / S3 represents the circularly polarized component. In natural environments, the circularly polarized component is negligible within a certain measurement range; therefore, the Stokes formula simplifies to: ; in, These represent the linearly polarized light intensities measured at polarization directions of 0°, 45°, 90°, and 135°, respectively.

[0021] The formula for degree of polarization is: ; The formula for the polarization angle is: ; The encoding format of the polarization image dataset used in this embodiment is determined by... , , The dataset consists of three channels: R, G, and B, with each channel corresponding to a polarization direction image. The polarization image dataset is then divided into training, validation, and test sets in a 7:2:1 ratio.

[0022] In one implementation, the YOLOv11 model's target feature extraction network extracts features from the input image, and the quality of feature extraction directly impacts the accuracy of subsequent network detection. An efficient feature selection module, EFS Block, is introduced to effectively suppress high-frequency snow noise in low-light snowy scenes while preserving the structural edge information of the target and the ability to detect small targets. The C2PSA_EDFFN module effectively suppresses unstructured snow noise while enhancing structurally significant high-frequency details. The YOLOv11 model's target feature fusion network fuses feature maps from different receptive fields to obtain the final feature map, which is then passed to the detection layer. A shift-based interaction and channel mixing module, SICM Block, is introduced to explicitly enhance the interaction capabilities between channels with extremely low computational cost. This more effectively preserves fine-grained structural information, improves the feature fusion quality for small and low-contrast targets, and maintains the lightweight nature of the overall computational overhead.

[0023] The structure of the improved network model is as follows Figure 1 As shown, the input image is passed through a convolutional (Conv) module and then connected to two sets of convolutional modules + C3K2 module. Next, it is connected to two sets of convolutional modules + EFS module and finally connected to SPPF module + C2PSA_EDFFN module to form the target feature extraction network. The convolutional module structure is Conv2d + BN + SiLU (two-dimensional convolutional layer + batch normalization layer + SiLU activation function).

[0024] After the C2PSA_EDFFN module, the features are concatenated with the same feature map size output from the C3K2 and EFS modules in the target feature extraction network via the SICM and convolution modules, and then output as a feature map via the EFS module, forming the target feature fusion network. The detection network receives multi-scale feature maps from the feature fusion network. For each layer of feature maps, the network uses a structurally isomorphic but parameter-independent decoupled detection head for processing. Each detection head contains two parallel branches consisting of multiple layers of 3*3 convolutions and 1*1 convolutions: one for classification, predicting class probabilities at the grid point level; the other for regression, predicting the discrete distance distribution from the target boundary to the grid points in the four directions (up, down, left, and right) within an anchor-free framework. This distribution is decoded into continuous offsets by DFL, and combined with IoU quality constraints, the accurate target bounding box is finally recovered.

[0025] The EFS module is as follows Figure 2 As shown, it includes: an input convolution module, a channel splitting (Split) module, a dual-branch parallel structure (direct connection branch + main branch), a feature fusion module, and an output convolution module connected in sequence. The main branch contains N cascaded C3K_ESFSConv modules or Bottleneck_ESFSConv modules (controlled by C3K=True / False).

[0026] The input feature map is passed through a 1x1 convolutional module with a stride of 1 and no padding. The output is then split into two feature maps with the same number of channels. One feature map passes through the main branch and is then concatenated with the other feature map (which passes through a direct connection branch) using a tensor concatenation operation. Finally, the output passes through a convolutional module. The specific structure of the convolutional module is Conv2d+BN+SiLU.

[0027] like Figure 3 and 4 The figures shown are the structure diagrams of the C3K_ESFSConv module and the Bottleneck_ESFSConv module, respectively.

[0028] The C3K_ESFSConv module is internally composed of n cascaded Bottleneck_ESFSConv modules, and each Bottleneck_ESFSConv module is internally composed of two cascaded ESFSConv modules. In the C3K_ESFSConv module, after the feature map is input, it splits into two branches. One branch consists of a convolutional module and n cascaded Bottleneck_ESFSConv modules, while the other branch is a direct connection branch containing a convolutional module. The outputs of the two branches are concatenated using tensors, and finally passed through a convolutional module. Channel fusion and compression are performed in the Conv convolutional module before output. The specific structure of the convolutional module is Conv2d+BN+SiLU.

[0029] The Bottleneck_ESFSConv module consists of two cascaded ESFSConv modules. Within the Bottleneck_ESFSConv module, the received feature map is divided into two branches. One branch passes sequentially through the two cascaded Bottleneck_ESFSConv modules, while the other branch is a direct connection. The outputs of the two branches are summed element-wise before outputting the result.

[0030] The Bottleneck_ESFSConv and C3K_ESFSConv modules together constitute the core feature transformation unit of EFS. The Bottleneck_ESFSConv module is a lightweight residual feature extraction unit. Its structure typically consists of 1×1 convolutions for channel compression and expansion, ESFSConv for spatial-frequency feature modeling, and residual connections for feature reuse. Its main function is to achieve joint modeling of local spatial structure information and spectral texture information while maintaining low computational complexity, thereby improving the diversity of feature representation and gradient propagation efficiency. The C3K_ESFSConv module is a feature extraction unit that adopts a cross-stage partially connected structure. It organizes multiple Bottleneck_ESFSConv units through channel splitting, multi-branch convolutional paths, and feature concatenation and fusion mechanisms. This allows some features to be directly transmitted across stages while others enter the depth transformation path, thus achieving multi-path feature propagation and cross-channel information interaction. Its main function is to enhance the semantic expressive power of deep features, improve feature reuse efficiency, and reduce redundant gradient calculations. From the perspective of structural differences, Bottleneck_ESFSConv is a single-path residual structure that focuses on local feature extraction and parameter efficiency optimization, while C3K_ESFSConv is a multi-branch cross-stage fusion structure that focuses on high-level semantic modeling and feature fusion capability improvement. The C3K_ESFSConv module has a larger computational scale and can significantly improve the richness and stability of feature representation. In the shallow stages of the network (typically corresponding to higher spatial resolution and lower channel count, such as 256 or below), the model emphasizes computational efficiency and local texture feature extraction. Therefore, the Bottleneck_ESFSConv module is preferred as the basic feature transformation unit to achieve basic spatial-frequency feature modeling with lower computational overhead. In the mid-to-late stages of the network (such as when the number of channels increases to 512 or 1024 and the spatial resolution is significantly reduced), the feature maps have stronger semantic information and higher channel expressive power. At this time, the network focuses more on high-level semantic modeling and cross-channel information interaction. Therefore, by setting c3k to True to enable the C3K_ESFSConv module, multiple Bottleneck_ESFSConv modules are organized in a multi-branch cross-stage feature fusion manner to enhance the deep feature expressive power and improve information propagation efficiency.

[0031] like Figure 5The diagram shows the structure of the ESFSConv module. The ESFSConv (Efficient Space-Frequency Selective Convolution) module is built on the basis of the existing Space-Frequency Selective Convolution (SFSConv), and performs Space-Frequency Selective Convolution (SFSConv) operations on some channel features. Figure 5 In the diagram, B, C, H, and W represent the batch, channel, height, and width of the feature map, respectively; n_div represents the channel partitioning ratio parameter.

[0032] In the ESFSConv module, the input features are first divided along the channel dimension. Convolution operations are performed on only a portion of the channels, while the remaining channels are directly passed with an identity mapping, thereby reducing redundant computation and preserving the original information.

[0033] Within the sub-channels involved in convolution, a spatial-frequency selected convolution (SFSConv) operation is performed, which includes a feature splitting structure and a channel selection structure. The feature splitting structure further divides the input features into spatial feature branches and frequency feature branches, used to capture the spatial structure information and texture spectrum information of the target, respectively. In the spatial branch, the spatial layout and contextual relationships of the target are modeled through convolution operations, while in the frequency branch, frequency-aware operations enhance the response to high-frequency texture details, thereby improving the model's ability to distinguish targets under complex backgrounds and noise interference. Subsequently, the channel selection structure adaptively filters and fuses spatial and frequency features based on the response intensity of different channel features to suppress redundant information and enhance the expression of key features. Finally, the features processed by spatial-frequency selected convolution and the unconvolutioned through features are concatenated and reconstructed along the channel dimension, thereby achieving a more complete and efficient feature expression while significantly reducing computational complexity.

[0034] In polarization images, the detected target exhibits high contrast with the background, meaning the target's outline is relatively clear. However, real-world road environments are complex, especially under adverse weather conditions (low-light snowy weather), which introduce significant redundant information and noise. Furthermore, the low contrast of low-light images undoubtedly affects the accuracy of target detection. Utilizing EDFFN's local frequency sensing and selective frequency enhancement can effectively suppress unstructured snow noise while enhancing structurally significant high-frequency details.

[0035] For the C2PSA module in the feature extraction part of the YOLOv11 model, which includes the PSABlock module, the PSABlock module is improved by replacing the FFN module in the PSABlock module with the EDFFN module, resulting in the improved PSABlock module, called the PSABlock_EDFFN module. Figure 6 As shown, the working principle of PSABlock_EDFFN is as follows: First, it expands the channels using 1*1 pointwise convolution (PConv), then captures local spatial features using depthwise convolution (DConv), and achieves channel-level nonlinear selective enhancement through GELU gating. Next, it reconstructs the channels using 1*1 pointwise convolution (PConv), and then uses patch unfolding to divide and rearrange the two-dimensional feature map into local feature vector representations (patchtokens), thereby constructing a more flexible feature representation while preserving local structural information. Then, it performs a Fast Fourier Transform (FFT) on the rearranged features, mapping them from the spatial domain to the frequency domain, allowing image features to be represented in the form of different frequency components. In the frequency domain, a learnable quantization matrix W is introduced to perform frequency-selective modulation of the spectral features. By weighting and filtering different frequency components, it enhances the more critical high-frequency details in the image while suppressing redundant or noisy frequency components, achieving efficient frequency domain feature selection. Subsequently, the enhanced frequency domain features are transformed back to the spatial domain representation using Inverse Fast Fourier Transform (IFFT), allowing the details and structural information after spectral modulation to be remapped to their corresponding spatial locations. Finally, the processed patch tokens are restored to the original two-dimensional feature map structure through patch folding, thereby achieving local frequency awareness, structural edge enhancement, and non-structural noise suppression, ultimately outputting enhanced high-frequency structured features. Its structure is as follows: Figure 6 As shown.

[0036] The improved C2PSA module is called the C2PSA_EDFFN module, and its structure diagram is as follows: Figure 7 As shown.

[0037] like Figure 8The diagram shows the SICM module structure. SICM first expands the spatial scale of the feature map through upsampling, and then performs channel-wise local feature extraction using depthwise convolution (DConv), thereby improving resolution while maintaining low computational complexity. Subsequently, batch normalization (BN) and rectified linear units (ReLU) are applied sequentially to stabilize the feature distribution, accelerate network convergence, and introduce non-linear mapping, further enhancing the network's feature representation capabilities. Next, channel shuffling is performed on the obtained features to break the independence between channels and promote information interaction between different channel groups. To further enhance the spatial modeling capability of the features, a shift channel mixing mechanism is used, dividing the feature map into four sub-features by channel and cyclically shifting them in the horizontal and vertical directions, thus achieving cross-spatial information propagation without additional convolution parameters. Finally, 1×1 point convolutions are used to linearly combine and fuse the features from each channel to obtain the final output features.

[0038] The formula for SICM is:

[0039] This indicates an upsampling operation (magnification factor of 2). Represents depthwise convolution; Indicates batch normalization; Indicates the activation function; This represents a 1x1 convolution.

[0040] Example 2 This embodiment further defines Embodiment 1. In one implementation, the above improved method can be applied to any model size of YOLOv11. In this embodiment, the YOLOv11m model size is used as the training example. When an improved network model is obtained based on the YOLOv11m model, the SGD optimizer can be used with training parameters of lr0 (initial learning rate) = 0.01, lrf (final learning rate) = 0.01, Batchsize = 32, mosaic (Mosaic data augmentation probability / intensity) = 0.3, close_mosaic (number of iterations to close mosaic augmentation) = 100, and a cosine annealing learning rate reduction method is used. The number of iterations is 300, and the IoU threshold is 0.5. The optimized weight parameter model is obtained through multiple rounds of iterative training as the target network model.

[0041] The test set is input into the target network model for testing, and the road target detection results are obtained through the tested target network model.

[0042] Optionally, after inputting the test set into the trained optimal target network model for testing and obtaining the test results, the performance of the tested target network model is evaluated using preset performance evaluation metrics, including IoU (Intersection over Union), precision P, recall R, mAP@50, mAP@50:95, and the number of parameters.

[0043] In one implementation, the Intersection over Union (IoU) refers to the ratio of the intersection to the union of the predicted bounding box A and the ground truth bounding box B generated by the model. The IoU value ranges from 0 to 1, where 1 represents complete overlap between the predicted and ground truth bounding boxes, and 0 represents complete separation. IoU reflects the accuracy of localization; a higher ratio indicates that the predicted result is closer to the true value. .

[0044] Precision P refers to the probability that the predicted class is the correct one, that is, the proportion of correctly predicted road targets out of the total number of targets predicted by the detection network as road targets. .

[0045] Recall R refers to the probability that the correct class is predicted in the positive sample, that is, the proportion of correctly predicted road targets out of all true road targets: .

[0046] Wherein: TP is a true positive, meaning a positive sample is correctly identified as a positive sample; FN is a false negative, meaning a positive sample is incorrectly identified as a negative sample; TN is a true negative, meaning a negative sample is correctly identified as a negative sample; FP is a false positive, meaning a negative sample is incorrectly identified as a positive sample.

[0047] mAP@50 measures whether an object detection model can effectively detect and correctly classify objects by calculating the average accuracy of each category at an IoU=0.5 threshold.

[0048] mAP@50:95 comprehensively measures the overall performance of the target detection model in terms of classification accuracy and boundary localization precision by averaging the AP at multiple IoU thresholds.

[0049] The number of parameters represents the total number of learnable weights and biases in the model, and is a key indicator for measuring model size, storage overhead, and lightweightness.

[0050] In one application scenario, to test the improvement in accuracy of the method of this application for road target detection in adverse weather conditions, relevant indicators were calculated for both the existing YOLOv11m and the improved target network model of this application. The comparison results of relevant indicators between the prior art and the method of this application are shown in Table 1. The detectable brightness range and average brightness of the method of this application for low-light images are shown in Table 2. Figure 9 and Figure 10 These correspond to the brightest and darkest traffic scenarios in Table 2, respectively.

[0051] Image brightness is measured by calculating the average intensity value of grayscale pixels, which ranges from 0 to 255.

[0052] Table 1 Comparison of relevant indicators between prior art and the methods of the embodiments of this application

[0053] Table 2. Detectable brightness range and average brightness of low-light images

[0054] As can be seen, the target network model in this embodiment improves on every metric and slightly reduces the number of parameters. Thanks to the introduced lightweight module, the method in this embodiment is more feasible and superior to the prior art. This embodiment can detect low-light images with brightness ranging from 17.1 to 55.1 with good detection accuracy.

[0055] As can be seen from the above, the traffic recognition model construction method for low-light snowy scenes provided by this application improves YOLOv11. The improved network model introduces an efficient feature selection module (EFS Block), which effectively suppresses high-frequency snow noise in low-light snowy scenes while maintaining the structural edge information of targets and the ability to detect small targets. The C2PSA_EDFFN module effectively suppresses unstructured snow noise while enhancing structurally significant high-frequency details. The shift-based interaction and channel mixing module (SICMBlock) explicitly enhances the interaction capability between channels with extremely low computational cost, more effectively preserving fine-grained structural information and improving the feature fusion quality for small and low-contrast targets, while maintaining the lightweight characteristics of overall computational overhead. Training the improved network model and performing road target detection on polarized images using the trained network model results in higher detection performance, making it suitable for scene analysis of traffic target recognition in low-light snowy scenes.

Claims

1. A method for constructing a traffic recognition model suitable for low-light snowy weather scenarios, characterized in that, The method is specifically as follows: S1. Select the YOLOv11 model as the base model and build the model based on the YOLOv11 model; S2. Improve the feature extraction part of the YOLOv11 model by replacing the C3K2 module. Specifically, replace the C3K2 module with an efficient feature selection module. S3. Improve the feature extraction part of the YOLOv11 model and improve the C2PSA module. Specifically, improve the C2PSA module by using an effective frequency domain discrimination-based feedforward network EDFFN. S4. Improve the feature fusion part of the YOLOv11 model by adding a displacement-based interactive channel mixing module before the two tensor splicing modules in the upsampling stage; and add an efficient feature selection module after the last tensor splicing module in the downsampling stage. A detection model suitable for low-light snowy weather scenarios was obtained; S5. Collect traffic scene images in snowy weather, construct a polarized image dataset, divide the dataset into training set, validation set and test set, use the polarized image dataset to train and test the detection model suitable for low-light snowy weather scenes, and obtain a target recognition model suitable for low-light snowy weather scenes that can be applied.

2. The traffic recognition model construction method for low-light snowy weather scenarios according to claim 1, characterized in that, When replacing the C3K2 modules, the specific replacement targets are the third and fourth C3K2 modules in the feature extraction part from the input to the output direction.

3. The traffic recognition model construction method for low-light snowy weather scenarios according to claim 2, characterized in that, The efficient feature selection module includes: an input convolution module, a channel splitting module, a dual-branch parallel structure, a tensor concatenation module, and an output convolution module connected in sequence; the dual-branch parallel structure includes a direct branch and a main branch, the main branch containing N cascaded C3K_ESFSConv modules, each C3K_ESFSConv module consisting of n cascaded Bottleneck_ESFSConv modules, and each Bottleneck_ESFSConv module consisting of two cascaded ESFSConv modules; the ESFSConv module is built on the existing spatial-frequency selected convolution SFSConv, and performs spatial-frequency selected convolution SFSConv operations on some channel features.

4. The traffic recognition model construction method for low-light snowy weather scenarios according to claim 3, characterized in that, In the C3K_ESFSConv module, the received feature map is divided into two branches. One branch passes through n cascaded Bottleneck_ESFSConv modules in sequence, and the other branch is a direct connection branch. The outputs of the two branches are concatenated by the tensor concatenation module and then input into the convolution module. In the convolution module, channel fusion and channel compression are performed before output.

5. The method for constructing a traffic recognition model suitable for low-light snowy weather scenarios according to claim 2, characterized in that, The efficient feature selection module includes: an input convolution module, a channel splitting module, a dual-branch parallel structure, a tensor concatenation module, and an output convolution module connected in sequence; the dual-branch parallel structure includes a direct branch and a main branch, the main branch consisting of N cascaded Bottleneck_ESFSConv modules, and each Bottleneck_ESFSConv module consisting of two cascaded ESFSConv modules; the ESFSConv module is built on the existing spatial-frequency selected convolution SFSConv, and performs spatial-frequency selected convolution SFSConv operations on some channel features.

6. The traffic recognition model construction method for low-light snowy weather scenarios according to claim 5, characterized in that, In the Bottleneck_ESFSConv module, the received feature map is divided into two branches. One branch passes through two cascaded ESFSConv modules in sequence, and the other branch is a direct connection branch. The outputs of the two branches are output after element-wise addition.

7. The method for constructing a traffic recognition model suitable for low-light snowy weather scenarios according to any one of claims 4 or 6, characterized in that, The ESFSConv module is built on the existing Spatial-Frequency Selective Convolution (SFSConv). The specific structure is as follows: the input feature map is split in the channel dimension, a part of the channel features are processed by Spatial-Frequency Selective Convolution (SFSConv), and the other part of the channel features are directly passed while maintaining the identity mapping. Finally, the two parts of features are concatenated by the tensor concatenation module and output.

8. The method for constructing a traffic recognition model suitable for low-light snowy weather scenarios according to claim 7, characterized in that, The C2PSA module is improved by using an effective frequency-domain discriminant-based feedforward network EDFFN. Specifically, the FFN unit of the PSABlock module in the C2PSA module is replaced with the frequency-domain discriminant-based feedforward network EDFFN.

9. The method for constructing a traffic recognition model suitable for low-light snowy weather scenarios according to claim 8, characterized in that, The displacement-based interactive channel mixing module sequentially includes an upsampling module, a depthwise convolution module, a batch normalization layer, a ReLU activation module, a channel shuffling and channel displacement mixing module, and a 1*1 convolution module.

10. A traffic recognition model suitable for low-light snowy weather scenarios, characterized in that, The model is constructed using the method described in claim 9.

Images