A night road traffic key information real-time edge detection system and method

Abstract

The application discloses a kind of night road traffic key information real-time edge detection system and method, the method first improves night image brightness problem through night image enhancement module, so that target feature information is significantly enhanced, then build lightweight target detection network suitable for night detection, the best weight data is obtained by training target detection network to the image data of pre-processing, finally the best weight data is loaded into target detection network, and the input image data to be detected is detected, and the real-time edge detection of road traffic key information under the condition of night low brightness is realized by fusing image enhancement and improving YOLO11.The method model structure of the application is simple, and has the ability of real-time detection and has good detection stability, detection precision and speed can satisfy the requirement of target detection task under low brightness night scene, has very practical engineering application value in road traffic safety.

Description

Technical Field

[0001] This invention belongs to the field of intelligent traffic management and control technology, and specifically relates to a real-time edge detection system and method for key information on nighttime road traffic. Background Technology

[0002] With the rapid development of intelligent transportation and autonomous driving technologies, the improved real-time information perception capabilities at the edge can effectively ensure road safety. The incidence of traffic accidents at night is significantly higher than during the day, and traditional detection solutions relying on cloud computing suffer from high data transmission latency and bandwidth limitations, making it difficult to meet the real-time requirements of millisecond-level response. Against this backdrop, the collaborative innovation of lightweight models and edge computing has become a key technological breakthrough. By deploying computing resources to roadside terminal devices and combining them with deep learning methods, low-power, high-efficiency localized inference capabilities can be achieved, providing instant decision support for complex traffic scenarios.

[0003] However, the current technology system faces two major contradictions: First, the conflict between the computational complexity of low-brightness enhancement algorithms and the limited resources of edge devices: Existing enhancement methods mostly rely on training with a large amount of labeled data. Although they can improve image quality, their models are large in scale and have poor hardware adaptability, making them difficult to deploy on terminals with limited computing power. At the same time, the scarcity of labeled data leads to insufficient generalization ability of the algorithms, making it difficult to cope with dynamic interference such as sudden changes in lighting on roads at night and glare from vehicle lights. Second, the dilemma of balancing detection accuracy and real-time performance: Target features degrade severely in low-light environments, and traditional detection models are prone to false negatives. While introducing complex network structures can improve accuracy, it exacerbates the computational load and cannot meet the real-time processing needs of edge devices.

[0004] Therefore, providing a real-time edge detection system and method for key nighttime road traffic information is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time edge detection system and method for key nighttime road traffic information, in order to solve the problems of difficulty in extracting nighttime road traffic feature information and insufficient detailed information.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A real-time edge detection system for key nighttime road traffic information includes: The acquisition module is used to acquire image data in nighttime scenes and preprocess the acquired image data; The input module is used to input the preprocessed image data into the pre-trained nighttime multi-object detection model; The detection module is used for target detection and localization.

[0007] Furthermore, the nighttime multi-target detection model uses the YOLO11 network as the base model, including: an image enhancement network, a backbone network, and a head network. The image enhancement network includes ordinary convolutional layers and a symmetric skip connection structure; the backbone network includes convolutional modules and feature enhancement modules; and the head network includes a bidirectional weighted feature fusion network structure.

[0008] A real-time edge detection method for key nighttime road traffic information, based on the real-time edge detection system for key nighttime road traffic information described above, includes the following steps: 1) Acquire images of low-light scenes at night and preprocess the acquired image data; 2) Input the preprocessed image data into the pre-trained nighttime multi-object detection model; 3) Based on the feature map calculated by the nighttime multi-target detection model, the detection box localization is optimized and calibrated by combining the SIOU loss function with the output layer, and the output includes the bounding box coordinates, category label and confidence score.

[0009] Furthermore, in step 2), the image enhancement network is a low-brightness image enhancement algorithm based on a no-reference curve, and the specific steps include: Multiple sets of curve parameter maps are generated using a depth curve estimation network. The brightness enhancement curve is calculated from the curve parameter maps. The low-brightness image is enhanced through multiple iterations of the brightness enhancement curve until the final low-brightness enhanced image is obtained.

[0010] Furthermore, the input to the depth curve estimation network is a low-brightness image, and the output is a curve parameter map of the brightness enhancement curve corresponding to the input image. The depth curve estimation network is a lightweight fully convolutional neural network, which includes ordinary convolutional layers and symmetric skip connection structures. Each ordinary convolutional layer consists of 32 convolutional kernels with a kernel size of 3×3 and a stride of 1.

[0011] Furthermore, the brightness enhancement curve is defined as: In the formula, For pixel coordinates, For a parameter mapping map with the same size as the input image, Let be the number of iterations, when The brightness enhancement curve can be converted into the following formula: In the formula, for Brightness enhancement curve at time, Low-brightness image; are trainable parameters and It is used to adjust the curvature of the brightness enhancement curve.

[0012] Furthermore, the convolutional module of the backbone network includes four consecutive convolutional layers for initial feature extraction; the feature enhancement module C3k2-DIMB expands the receptive field through dynamic hybrid convolution and enhances multi-scale feature extraction capabilities by combining residual structures.

[0013] Furthermore, the feature enhancement module C3k2-DIMB includes a multi-kernel dynamic depthwise convolution module, a dual-group dynamic convolution hybrid module, and a separable convolution-gated residual module. The feature enhancement steps are as follows: In the multi-core dynamic depthwise convolution module, the following is adopted Standard depth convolution, and Three kernel structures of strip depthwise convolution are used to collect features in parallel, and adaptive global pooling is employed. Convolution generates a three-channel weight matrix to dynamically adjust each convolution kernel, thereby capturing multi-scale and diverse feature information; Subsequently, a dual-group dynamic convolution hybrid module is constructed using the multi-core dynamic depthwise convolution module as a component. Fine-grained feature enhancement is achieved through channel-dimensional splitting and dual-branch dynamic convolution: first, the feature map is divided into two groups according to channels, and each group is input into an independently configured dynamic convolution module. The dual outputs are then concatenated by channels. Convolutional fusion enables cross-branch feature interaction; Finally, training stability is improved by using a separable convolutional-gated residual module: the outputs of the mixer and the gated linear unit are respectively adjusted by a learnable layer scaling factor and combined with random path drop-off, so that the network can adaptively balance the information flow of the original features and the enhanced features.

[0014] Furthermore, the specific calculation steps for feature enhancement are as follows: Let the input features be The output features are , Among them, the multi-core dynamic depthwise convolution module:

[0015] In the formula: W dkw For the dynamic weights of the three-channel convolution kernel, For adaptive average pooling operation, For activation function, For the first Depthwise separable convolution operations, among which The value in this invention is set to 11. The final output is F, which is the Hadamard product. conv Multi-branch convolutional fusion features are used to construct a dual-group dynamic convolutional hybrid structure by treating the above modules as components:

[0016] In the formula: The features are input to the dual-group dynamic convolutional hybrid structure. The input features are grouped for the slicing operation. For the feature grouping processing results, Further multi-kernel dynamic depthwise convolution module is used to obtain enhanced features. , For channel-dimensional splicing operations, the final result will be... Channel-dimensional concatenation is performed to obtain fused features, which are then input into the separable convolutional-gated residual module.

[0017] In the formula: Features are input to the separable convolution-gated residual module. For activation function, Scaling the parameter vector for trainable layers. , For random depth discarding operation, For batch normalization operation;

[0018] final, These correspond to the composite functions of hybrid feature enhancement and MLP transform, respectively.

[0019] Furthermore, the bidirectional weighted feature fusion network structure combines top-down and bottom-up bidirectional paths through a bidirectional cross-scale connection mechanism to achieve dynamic interactive fusion of multi-level feature maps. It also introduces a learnable weight allocation module to assign normalized weight parameters to the input feature maps of different scales and dynamically adjust the importance of features through a fast normalization weight method. By enhancing the representation ability of multi-scale targets through cross-scale feature fusion, it supports deep interaction of feature maps of different resolutions, ultimately forming an efficient and robust network structure.

[0020] Furthermore, the specific steps of the fast normalized weighting method are as follows: BiFPN introduces learnable weight factors into the input features.

[0021]

[0022] In the formula, These are input features. It is an intermediate feature. It is the output feature. It is an adaptive weighting factor and a stability parameter. ; Then, the fusion weight formula is calculated using the weight normalization formula:

[0023] In the above formula, It is an adaptive weighting factor. For learnable parameters, This is a stability parameter.

[0024] Furthermore, the SIOU loss function is used to calculate the detection box. and real frame The traditional loss term for mismatches takes into account the direction factor, and an angle loss is added. The angle loss will... Leading to The X-axis or Y-axis direction is used to move closer to the bounding box, and then the movement continues along that direction to reduce the time complexity of the bounding box distance loss. Thus, SIoU Loss includes angular loss. Distance loss Shape loss and Among them, angle loss Defined as:

[0025] if Then calculate In the formula, c is the distance between the center points of the ground truth bounding box and the predicted bounding box. h The height difference between the center points of the ground truth bounding box and the predicted bounding box. These are the x and y coordinates of the center points of the ground truth bounding box and the detection bounding box, respectively. Distance loss Defined as:

[0026] In the formula, As a weighting factor, it controls the decay rate of the exponential function, dynamically adjusts the sensitivity of the loss to the error, and depends on the angle difference; It is the relative error, which measures the relative deviation between the center points of the predicted bounding box and the ground truth bounding box; c w ,c h The width and height of the minimum bounding rectangle between the ground truth bounding box and the predicted bounding box; loss Defined as:

[0027] In the formula, ω w ,ω h These represent the width error and the height error, w,h,w gt ,h gt These are the width and height of the predicted bounding box and the ground truth bounding box, respectively. This is the shape attention parameter, whose value is calculated by a genetic algorithm; Defined as:

[0028] In the formula, These are the detection bounding box and the ground truth bounding box, respectively. The formula for the total loss function is as follows:

[0029] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a real-time edge detection system and method for key road traffic information at night. By integrating dynamic illumination restoration technology and lightweight target detection algorithm, dual collaborative optimization is achieved at the edge: the illumination enhancement module improves the discernibility of dark area details through an adaptive adjustment mechanism, while the detection module reduces the computational load through lightweight design, effectively solving the problem of feature degradation and real-time performance in low-light environments. Attached Figure Description

[0030] Figure 1 The flowchart shows the real-time edge detection method for key nighttime road traffic information provided by the present invention.

[0031] Figure 2 A diagram illustrating the specific calculation steps of the real-time edge detection method for key nighttime road traffic information provided by this invention. Figure 3 A flowchart of the image enhancement network provided by the present invention; Figure 4 A structural diagram of the feature enhancement module C3k2-DIMB provided by this invention; Figure 5 A comparison diagram of the bidirectional weighted feature fusion network structure and the PAFPN structure provided by this invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Depend on Figure 1-5 As shown, this embodiment of the invention discloses a real-time edge detection system for key nighttime road traffic information, comprising: The acquisition module is used to acquire image data in nighttime scenes and preprocess the acquired image data; The input module is used to input preprocessed image data into a pre-trained nighttime multi-object detection model. This model uses the YOLO11 network as its base model, with the core objective of improving the visualization quality of low-light nighttime images to provide clearer input features for subsequent object detection. It includes an image enhancement network, a backbone network, and a head network. The image enhancement network comprises ordinary convolutional layers and a symmetric skip connection structure. The backbone network includes convolutional modules and a feature enhancement module. The backbone network is responsible for extracting multi-scale, highly semantic features from the enhanced image, including the feature enhancement module C3k2-DIMB, the spatial pyramid pooling module SPPF, and the attention module C2PSA. The head network includes a bidirectional weighted feature fusion network structure for further processing and fusing features, fusing and enhancing feature information at different scales to support the final detection results. The detection module is used for accurate detection and positioning of targets.

[0034] This invention employs a bidirectional feature pyramid structure, which can fully extract and fuse multi-scale feature information and enhance the information interaction between feature maps. By applying SIOU as the bounding box regression loss function of the target detection network, it accelerates model convergence and improves the accuracy of prediction boxes, thereby achieving efficient and accurate detection of key information on nighttime road traffic.

[0035] This invention also discloses a real-time edge detection method for key nighttime road traffic information, based on the real-time edge detection system for key nighttime road traffic information described above, and includes the following steps: 1) Acquire images of low-light scenes at night and preprocess the acquired image data; 2) Input the preprocessed image data into the pre-trained nighttime multi-object detection model; In this step, after the input image enters the image enhancement network, a depth curve estimation network first generates multiple sets of curve parameter maps. From the curve parameter maps, the brightness enhancement curve can be calculated. The low-brightness image is enhanced multiple times through the LE curve until the final low-brightness enhanced image is obtained. Then, the enhanced image is input into the backbone network. It first passes through a series of convolutional modules and feature enhancement modules to extract features at different scales. Then, it combines Spatial Pyramid Pooling (SPPF) to enhance the feature expression ability of key regions through dynamic feature adjustment and multi-scale receptive field expansion. Finally, a C2PSA attention mechanism is introduced to obtain the final feature map of the backbone network. Then, the output feature map of the backbone network is input into the head network and a bidirectional feature fusion network is used to fuse high-dimensional features and low-dimensional features to retain feature information at different scales to the maximum extent. 3) Based on the feature map calculated by the nighttime multi-target detection model, the detection box localization is optimized and calibrated by combining the SIOU loss function with the output layer, and the output includes the bounding box coordinates, category label and confidence score.

[0036] Specifically, the image enhancement network is a low-brightness image enhancement algorithm based on a no-reference curve, and the specific steps include: Multiple sets of curve parameter maps are generated using a depth curve estimation network. The brightness enhancement curve is calculated from the curve parameter maps. The low-brightness image is enhanced through multiple iterations of the brightness enhancement curve until the final low-brightness enhanced image is obtained.

[0037] Specifically, the input to the depth curve estimation network is a low-brightness image, and the output is a curve parameter map of the brightness enhancement curve corresponding to the input image. The depth curve estimation network is a lightweight fully convolutional neural network, which includes ordinary convolutional layers and symmetric skip connection structures. Each ordinary convolutional layer consists of 32 convolutional kernels with a kernel size of 3×3 and a stride of 1.

[0038] Specifically, the brightness enhancement curve is defined as:

[0039] In the formula, For the first The brightness enhancement curve of the next iteration. For pixel coordinates, For a parameter mapping map with the same size as the input image, Let be the number of iterations, when The brightness enhancement curve can be converted into the following formula:

[0040] In the formula, for Brightness enhancement curve at time, Low-brightness image; are trainable parameters and It is used to adjust the curvature of the brightness enhancement curve.

[0041] Specifically, the convolutional module of the backbone network includes four consecutive convolutional layers for initial feature extraction; the feature enhancement module C3k2-DIMB expands the receptive field through dynamic hybrid convolution and enhances multi-scale feature extraction capabilities by combining residual structures.

[0042] Specifically, the feature enhancement module C3k2-DIMB includes a multi-kernel dynamic depthwise convolution module, a dual-group dynamic convolution hybrid module, and a separable convolution-gated residual module. The feature enhancement steps are as follows: In the multi-core dynamic depthwise convolution module, the following is adopted Standard depth convolution, and Three kernel structures of strip depthwise convolution are used to collect features in parallel, and adaptive global pooling is employed. Convolution generates a three-channel weight matrix to dynamically adjust each convolution kernel, thereby effectively capturing multi-scale and diverse feature information; Subsequently, a dual-group dynamic convolution hybrid module is constructed using the multi-core dynamic depthwise convolution module as a component. Fine-grained feature enhancement is achieved through channel-dimensional splitting and dual-branch dynamic convolution: first, the feature map is divided into two groups according to channels, and each group is input into an independently configured dynamic convolution module. The dual outputs are then concatenated by channels. Convolutional fusion enables cross-branch feature interaction; this process uses a parameterized grouping strategy to reduce computational cost while preserving the complementarity of multi-scale information. Finally, training stability is improved by using a separable convolutional-gated residual module: the outputs of the mixer and the gated linear unit are respectively adjusted by a learnable layer scaling factor and combined with random path drop-off, so that the network can adaptively balance the information flow of the original features and the enhanced features.

[0043] Specifically, the calculation steps for feature enhancement are as follows: Let the input features be The output features are , Among them, the multi-core dynamic depthwise convolution module:

[0044] In the formula: W dkw For the dynamic weights of the three-channel convolution kernel, For adaptive average pooling operation, For activation function, For the first Depthwise separable convolution operations, among which The value in this invention is set to 11. The final output is F, which is the Hadamard product. conv Multi-branch convolutional fusion features are used to construct a dual-group dynamic convolutional hybrid structure by treating the above modules as components:

[0045] In the formula: The features are input to the dual-group dynamic convolutional hybrid structure. The input features are grouped for the slicing operation. For the feature grouping processing results, Further multi-kernel dynamic depthwise convolution module is used to obtain enhanced features. , For channel-dimensional splicing operations, the final result will be... Channel-dimensional concatenation is performed to obtain fused features, which are then input into the separable convolutional-gated residual module.

[0046] In the formula: Features are input to the separable convolution-gated residual module. For activation function, Scaling the parameter vector for trainable layers. , For random depth discarding operation, For batch normalization operation;

[0047] final, These correspond to the composite functions of hybrid feature enhancement and MLP transform, respectively.

[0048] Specifically, the bidirectional weighted feature fusion network structure combines top-down and bottom-up bidirectional paths through a bidirectional cross-scale connection mechanism to achieve dynamic interactive fusion of multi-level feature maps. It also introduces a learnable weight allocation module to assign normalized weight parameters to input feature maps of different scales and dynamically adjust the importance of features through a fast normalized weight method. By enhancing the representation ability of multi-scale targets through cross-scale feature fusion, it supports deep interaction of feature maps of different resolutions, ultimately forming an efficient and robust network structure.

[0049] Specifically, the steps of the fast normalized weighting method are as follows: BiFPN introduces learnable weight factors into the input features.

[0050]

[0051] In the formula, These are input features. It is an intermediate feature. It is the output feature. It is an adaptive weighting factor and a stability parameter. ; Then, the fusion weight formula is calculated using the weight normalization formula:

[0052] In the above formula, w i It is an adaptive weighting factor, λ i ω is a learnable parameter, and ò is a stability parameter.

[0053] Specifically, the SIOU loss function is used to calculate the detection box. and real frame The traditional loss term for mismatches takes into account the direction factor, and an angle loss is added. The angle loss will... Leading to The X-axis or Y-axis direction is used to move closer to the bounding box, and then the movement continues along that direction to reduce the time complexity of the bounding box distance loss. Thus, SIoU Loss includes angular loss. Distance loss Shape loss and Among them, angle loss Defined as:

[0054] if Then calculate In the formula, σ is the distance between the center points of the ground truth bounding box and the predicted bounding box, and c h The height difference between the center points of the ground truth bounding box and the predicted bounding box. These are the x and y coordinates of the center points of the ground truth bounding box and the detection bounding box, respectively. Distance loss Defined as:

[0055] In the formula, As a weighting factor, it controls the decay rate of the exponential function, dynamically adjusting the sensitivity of the loss to the error, and depends on the angle difference; ρ t It is the relative error, which measures the relative deviation between the center points of the predicted bounding box and the ground truth bounding box; c w ,c h The width and height of the minimum bounding rectangle between the ground truth bounding box and the predicted bounding box; loss Defined as:

[0056] In the formula, ω w ,ω h These represent the width error and the height error, w,h,w gt ,h gt These are the width and height of the predicted bounding box and the ground truth bounding box, respectively. This is the shape attention parameter, whose value is calculated by a genetic algorithm; Defined as:

[0057] In the formula, These are the detection bounding box and the ground truth bounding box, respectively. The formula for the total loss function is as follows:

[0058] This invention discloses a method that uses the YOLO11 network as the base network and combines it with specific optimization modules to effectively solve the problem of difficult detection in nighttime scenes due to low brightness and limited target features. The model performs well in extracting features in low-light environments and has excellent edge real-time performance, thus enhancing the ability to identify key road traffic information under low-light conditions at night.

[0059] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for real-time edge detection of key road traffic information at night, based on a real-time edge detection system for key road traffic information at night, characterized in that, A real-time edge detection system for key nighttime road traffic information includes: The acquisition module is used to acquire image data in nighttime scenes and preprocess the acquired image data; The input module is used to input the preprocessed image data into the pre-trained nighttime multi-object detection model; The detection module is used for target detection and localization. The nighttime multi-target detection model uses the YOLO11 network as its base model, including: an image enhancement network, a backbone network, and a head network. The image enhancement network includes ordinary convolutional layers and a symmetrical skip connection structure; the backbone network includes convolutional modules and feature enhancement modules; and the head network includes a bidirectional weighted feature fusion network structure. Includes the following steps: 1) Acquire images of low-light scenes at night and preprocess the acquired image data; 2) Input the preprocessed image data into the pre-trained nighttime multi-object detection model; 3) Based on the feature map calculated by the nighttime multi-target detection model, the detection box localization is optimized and calibrated by combining the SIOU loss function with the output layer, and the output includes the detection results containing the bounding box coordinates, category labels and confidence scores; The backbone network's convolutional module includes four consecutive convolutional layers for initial feature extraction; the feature enhancement module C3k2-DIMB expands the receptive field through dynamic hybrid convolutions and enhances multi-scale feature extraction capabilities by combining residual structures. The feature enhancement module C3k2-DIMB includes a multi-kernel dynamic depthwise convolution module, a dual-group dynamic convolution hybrid module, and a separable convolution-gated residual module. The feature enhancement steps are as follows: In the multi-kernel dynamic depth convolution module, three kernel structures, namely 3×3 standard depth convolution, 1×11 and 11×1 strip depth convolution, are used to collect features in parallel. Adaptive global pooling and 1×1 convolution generate a three-channel weight matrix to dynamically adjust each convolution kernel, thereby capturing multi-scale and diverse feature information. Subsequently, a dual-group dynamic convolution hybrid module is constructed using the multi-core dynamic depthwise convolution module as a component. Fine-grained feature enhancement is achieved through channel-dimensional splitting and dual-branch dynamic convolution: first, the feature map is divided into two groups according to channels, and each group is input into an independently configured dynamic convolution module. The dual outputs are then concatenated by channels. Convolutional fusion enables cross-branch feature interaction; Finally, training stability is improved by using a separable convolutional-gated residual module: the outputs of the mixer and the gated linear unit are respectively adjusted by a learnable layer scaling factor and combined with random path drop-off, so that the network can adaptively balance the information flow of the original features and the enhanced features.

2. The method for real-time edge detection of key nighttime road traffic information according to claim 1, characterized in that, In step 2), the image enhancement network is a low-brightness image enhancement algorithm based on a no-reference curve, and the specific steps include: Multiple sets of curve parameter maps are generated using a depth curve estimation network. The brightness enhancement curve is calculated from the curve parameter maps. The low-brightness image is enhanced through multiple iterations of the brightness enhancement curve until the final low-brightness enhanced image is obtained.

3. The method for real-time edge detection of key nighttime road traffic information according to claim 2, characterized in that, The input to the depth curve estimation network is a low-brightness image, and the output is a curve parameter map of the brightness enhancement curve corresponding to the input image. The depth curve estimation network is a lightweight fully convolutional neural network, which includes ordinary convolutional layers and symmetric skip connection structures. Each ordinary convolutional layer consists of 32 convolutional kernels with a kernel size of 3×3 and a stride of 1.

4. A real-time edge detection method for key nighttime road traffic information according to claim 2 or 3, characterized in that, The brightness enhancement curve is defined as: In the formula, For pixel coordinates, For a parameter mapping map with the same size as the input image, Let be the number of iterations, when The brightness enhancement curve can be converted into the following formula: In the formula, for Brightness enhancement curve at time, Low-brightness image; are trainable parameters and It is used to adjust the curvature of the brightness enhancement curve.

5. The method for real-time edge detection of key nighttime road traffic information according to claim 1, characterized in that, The specific calculation steps for feature enhancement are as follows: Let the input features be The output features are , Among them, the multi-core dynamic depthwise convolution module: In the formula: W dkw For the dynamic weights of the three-channel convolution kernel, For adaptive average pooling operation, For activation function, For the first Depthwise separable convolution operations, among which The value is set to 11, ⊙ represents the Hadamard product, and the final output is F. conv Multi-branch convolutional fusion features are used to construct a dual-group dynamic convolutional hybrid structure by treating the above modules as components: In the formula: F in The features are input to the dual-group dynamic convolutional hybrid structure. The input features are grouped for the slicing operation. For the feature grouping processing results, Further multi-kernel dynamic depthwise convolution module is used to obtain enhanced features. Concat is a concatenation operation based on channel dimensions, and finally... Channel-dimensional concatenation is performed to obtain fused features; furthermore, the specific steps of the separable convolution-gated residual module, which runs parallel to the dual-group dynamic convolutional hybrid structure, are as follows: In the formula: Features are input to the separable convolution-gated residual module. For activation function, Scaling the parameter vector for trainable layers. For gated linear units, For random depth discarding operation, For batch normalization operation; Ultimately, T1 and T2 correspond to the composite functions of hybrid feature enhancement and MLP transform, respectively.

6. The method for real-time edge detection of key nighttime road traffic information according to claim 1, characterized in that, The bidirectional weighted feature fusion network structure combines top-down and bottom-up bidirectional paths through a bidirectional cross-scale connection mechanism to achieve dynamic interactive fusion of multi-level feature maps; and introduces a learnable weight allocation module to assign normalized weight parameters to the input feature maps of different scales, and dynamically adjust the importance of features through a fast normalized weight method. By fusing cross-scale features, the representation capability of multi-scale targets is enhanced, supporting deep interaction of feature maps at different resolutions, and ultimately forming an efficient and robust network structure. The specific steps of the fast normalized weighting method are as follows: BiFPN introduces learnable weight factors into the input features. In the formula, It is the first The input features of the feature layer, It is the first Intermediate features of the layer feature layer It is the first The output features of the feature layer, It is an adaptive weighting factor and a stability parameter. ; Then, the fusion weight formula is calculated using the weight normalization formula: In the above formula, It is an adaptive weighting factor. For learnable parameters, This is a stability parameter.

Images