A road target detection method in rain and fog weather environment

Abstract

A road target detection method in rain and fog weather environment comprises: inputting a road scene image into a backbone network of a YOLO26 model for multi-scale feature extraction; wherein a local feature enhancement module is constructed in the backbone network for strengthening local feature extraction and generating local enhanced features; a mixed frequency space enhancement module comprising a high-frequency branch and a low-frequency branch is constructed at the end of the backbone network; the high-frequency branch is used for enhancing the representation ability of detail information, and the low-frequency branch is used for aggregating multi-scale structural information and context information of road targets; the local enhanced features are input into the mixed frequency space enhancement module, the outputs of the high-frequency branch and the low-frequency branch are adaptively fused to generate frequency space enhanced features; an adaptive direction perception feature enhancement module is constructed at the first multi-scale feature fusion stage of the neck network, and direction perception enhanced features are output for subsequent fusion of the neck network, target prediction and generation of road target detection results.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology and is applicable to scenarios such as autonomous driving, intelligent transportation, and road monitoring. In particular, it relates to a method for road target detection in rainy and foggy weather conditions. Background Technology

[0002] In autonomous driving systems, vision-based object detection is a crucial component of environmental perception. It acquires road scene images through onboard cameras, enabling real-time identification and localization of traffic participants such as vehicles and pedestrians, providing critical environmental information for subsequent path planning and vehicle control. Due to its balance between detection accuracy and inference speed, the YOLO series of object detection models has been widely applied in autonomous driving, intelligent transportation, and other scenarios. Among them, YOLO26, as a next-generation object detection architecture, further enhances object detection performance and model deployment capabilities.

[0003] In real-world open road environments, adverse weather conditions such as rain and fog significantly impact visual imaging quality, leading to degradation phenomena such as decreased contrast, blurred target edges, and reduced visibility of distant targets, increasing the risk of missed and false detections. Existing technologies for target detection in rainy and foggy weather mainly fall into two categories. The first category is a two-stage approach based on image preprocessing. This involves first improving the quality of the input image using de-raining, de-fogging, or image enhancement models, and then inputting the processed image into a target detection network for recognition. Since image enhancement and target detection are independent optimization tasks, improving image visual quality does not necessarily lead to improved target detection performance. The second category is a method based on improving the target detection network structure. This involves optimizing the network's feature extraction and feature fusion methods to enhance the detection model's adaptability to rainy and foggy weather scenarios. Compared to the two-stage approach, this type of method enables end-to-end joint optimization of target feature extraction and detection tasks.

[0004] However, existing methods based on improved object detection network structures still have certain shortcomings. On the one hand, rain and fog blur the edges of road targets and weaken local textures, limiting the ability of existing feature extraction structures to collaboratively model local details and large-scale contextual information. On the other hand, during multi-scale feature fusion, features at different levels differ in semantic information and spatial resolution, making it difficult to fully achieve effective complementarity of information at different levels through direct feature fusion. Furthermore, rain and fog degradation further weakens the expressive power of target spatial structure and location information, affecting the multi-scale fusion effect. In addition, some methods improve detection performance by introducing complex feature enhancement modules, but this usually increases the number of model parameters and computational overhead, making it difficult to balance the requirements of accuracy, real-time performance, and lightweight deployment in object detection.

[0005] Therefore, how to improve the ability of detection models to effectively extract and fuse target features in rainy and foggy weather while maintaining a low number of parameters and computational overhead, and to balance the accuracy and real-time performance of detection, remains an urgent problem to be solved. Summary of the Invention

[0006] This invention discloses a road target detection method for rainy and foggy weather conditions. Addressing the problems of blurred road target edges, local texture degradation, and decreased localization accuracy caused by rain and fog, the YOLO26 model is improved. This invention significantly reduces the number of model parameters and computational complexity while effectively improving the accuracy and recall rate of road target detection in rainy and foggy weather, meeting the real-time road target detection requirements of applications such as autonomous driving and intelligent transportation.

[0007] First, a local feature enhancement module is constructed in the backbone network to improve the network's ability to represent local details such as target edges and textures. Second, a hybrid frequency spatial enhancement module is constructed at the end of the backbone network. This module strengthens detailed features through high-frequency branches and aggregates multi-scale structural information and large-scale contextual information using low-frequency branches. It also adaptively fuses high- and low-frequency features, enhancing the network's ability to model complex scene contextual information while improving the ability to express target details. Finally, an adaptive direction-aware feature enhancement module is constructed in the first multi-scale feature fusion stage of the neck network to strengthen the direction awareness and spatial position representation of multi-scale fused features, improving the localization and identification capabilities of road targets with lower parameter quantity and computational overhead.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for road target detection in rainy and foggy weather conditions includes the following steps: The road scene image is input into the backbone network of the YOLO26 model for multi-scale feature extraction; among which, A local feature enhancement module is constructed in the backbone network to enhance local feature extraction and generate local enhanced features. A hybrid frequency space enhancement module including high-frequency and low-frequency branches is constructed at the end of the backbone network. The high-frequency branch is used to enhance the ability to represent detailed information, and the low-frequency branch is used to aggregate multi-scale structural information and contextual information of road targets. The local enhanced features are input into the hybrid frequency space enhancement module, and the outputs of the high-frequency and low-frequency branches are adaptively fused to generate frequency space enhanced features. An adaptive orientation-aware feature enhancement module is constructed in the first multi-scale feature fusion stage of the neck network to output orientation-aware enhanced features for subsequent fusion of the neck network, target prediction, and generation of road target detection results.

[0009] As a preferred option, constructing a local feature enhancement module in the backbone network includes: A multi-branch deep convolutional structure is used to extract local spatial features from the input feature map. The outputs of each branch are summed element by element and then batch normalized to obtain local spatial enhancement features. The local spatial enhancement features are input into a lightweight feedforward network for channel feature encoding to obtain channel encoded features. A learnable feature scaling mechanism is introduced to adaptively adjust the channel encoded features, and a dual residual connection structure is combined to perform feature fusion to obtain local enhancement features.

[0010] Preferably, the lightweight feedforward network consists of two convolutional normalization units, each of which contains... The system consists of convolutional and batch normalization layers; the first convolutional normalization unit is used for channel mapping, and the second convolutional normalization unit is used to restore the feature dimension. The GELU activation function is used between the two layers for non-linear feature transformation.

[0011] As a preferred option, multi-branch depthwise convolutional structures include Depth convolution branches, Depthwise convolution branch and identity mapping branch; The depthwise convolution branch is used to extract spatial structure information of the local neighborhood. The depthwise convolution branch is used to independently map and enhance the features of each channel, while the identity mapping branch is used to preserve the original input features.

[0012] As a preferred embodiment, constructing a hybrid frequency space enhancement module includes: For the input local enhancement features, the high-frequency branch uses two strip depth convolutions with different receptive fields to extract local structural information in the horizontal and vertical directions, respectively; after element-wise addition and batch normalization of the two directional features, the high-frequency detail enhancement features are obtained. The low-frequency branch uses three branches with different expansion rates. The local enhancement features are processed in parallel using deep convolution. After element-wise addition and batch normalization of the features at each scale, low-frequency context enhancement features are obtained.

[0013] Preferably, the output of the adaptive fusion of the high-frequency branch and the low-frequency branch includes: Set learnable parameters and map them to a coefficient mapping function. The interval is used to obtain adaptive fusion weights; based on the adaptive fusion weights, high-frequency detail enhancement features and low-frequency context enhancement features are weighted and fused to obtain fused features; the fused features are then processed... Convolution, batch normalization, and SiLU activation function are used to perform channel mapping and nonlinear transformation to obtain projection features; the projection features are then fused with local enhancement features using residuals to generate frequency space enhancement features.

[0014] As a preferred embodiment, the adaptive orientation-aware feature enhancement module constructed in the first multi-scale feature fusion stage of the neck network includes: performing global average pooling along the width and height dimensions for the features obtained from the first multi-scale feature fusion of the neck network, respectively, to obtain height-direction feature descriptions and width-direction feature descriptions; concatenating the two along the spatial dimension to obtain joint features; inputting the joint features into a shared feature encoding network for unified encoding to obtain shared encoded features; re-dividing the shared encoded features along the spatial dimension according to height and width to obtain height-direction encoded features and width-direction encoded features; setting learnable parameters and converting them into orientation adjustment coefficients through a coefficient mapping function; restoring the height-direction encoded features and width-direction encoded features to the input channel dimension, and adjusting the response intensity using adjustment coefficients to generate orientation response weights in the height and width directions; fusing the two orientation response weights, and adaptively recalibrating the features obtained from the first multi-scale feature fusion of the neck network to obtain orientation-aware enhanced features.

[0015] Compared with the prior art, the beneficial effects of the present invention are reflected in: 1. This invention addresses the problem of easily degrading local details of road targets in rainy and foggy weather by constructing a feature extraction mechanism in the backbone network that integrates local detail enhancement and hybrid frequency information collaborative modeling. Specifically, the local feature enhancement module enhances the extraction capability of road target texture, edge, and other detailed information, while the hybrid frequency spatial enhancement module, through collaborative modeling of high-frequency detailed information and low-frequency multi-scale contextual information, enhances the representation of large-scale contextual information while preserving local details, thereby improving the network's ability to perceive blurred, small-scale, and distant road targets.

[0016] 2. This invention constructs an adaptive direction-aware feature enhancement module in the initial multi-scale feature fusion stage of the neck network. By aggregating directional features along horizontal and vertical directions, it establishes a feature representation with direction-aware capabilities. Furthermore, a lightweight shared feature encoding network is employed to achieve information interaction and joint modeling of features from different directions. The response intensity of directional features is adaptively adjusted using learnable adjustment coefficients, thereby enhancing the spatial structure and positional information representation capabilities of the multi-scale fused features. This module can improve the fusion quality of features at different levels while maintaining low parameter count and computational overhead, thus improving the accuracy of road target detection.

[0017] 3. This invention organically combines a local feature enhancement module, a hybrid frequency spatial enhancement module, and an adaptive direction-aware feature enhancement module, deploying them at different levels of the backbone network and the neck network respectively, constructing a multi-level feature enhancement system of "local enhancement—frequency synergy—direction awareness." Specifically, the local feature enhancement module operates in the middle layer of the backbone network, enhancing the texture, edge, and other detailed information of road targets; the hybrid frequency spatial enhancement module operates in the deep feature extraction stage of the backbone network, collaboratively modeling high-frequency details and low-frequency multi-scale contextual information; and the adaptive direction-aware feature enhancement module operates in the feature fusion stage of the neck network, adaptively enhancing the directional structure information in the multi-scale fused features. Through cross-level information enhancement and feature synergy, these three modules achieve a comprehensive gain effect distinct from a single enhancement module, improving the detection accuracy of road targets in rainy and foggy weather conditions while maintaining low computational overhead. Attached Figure Description

[0018] Figure 1 This is a model structure diagram of Embodiment 1 of the present invention; Figure 2 This is a structural diagram of the local feature enhancement module in Embodiment 1 of the present invention; Figure 3 This is a structural diagram of the hybrid frequency spatial enhancement module according to Embodiment 1 of the present invention; Figure 4 This is a structural diagram of the adaptive direction-aware feature enhancement module of Embodiment 1 of the present invention; Figure 5 This is a comparison chart of the detection results of the method of the present invention and the YOLO26 method in Embodiment 1 of the present invention. Detailed Implementation

[0019] To make the technical means, inventive features, objectives, and effects of the invention readily understandable, the invention is further described below with reference to specific illustrations. However, the invention is not limited to the embodiments described below.

[0020] It should be noted that the structures, proportions, sizes, etc., illustrated in the accompanying drawings of this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0021] This invention's model, based on YOLO26, consists of a backbone network, a neck network, and a detection head. The backbone network extracts multi-scale features step-by-step through convolution (Conv) and the C3k2 feature extraction module. Local Feature Enhancement Blocks (LFE Blocks) are introduced in the middle and later layers to enhance local details such as target edges and textures. A Hybrid-Frequency Spatial Enhancement Module (HFSEM) is added at the end of the backbone to achieve collaborative modeling of high-frequency detail information and low-frequency contextual information. The neck network employs a top-down and bottom-up multi-scale feature fusion structure, embedding an Adaptive Direction-Aware Feature Enhancement Module (ADAFEM) in the initial feature fusion stage to enhance the directional spatial structure and positional information representation of the fused features. Finally, the fused features at the three scales are input into the YOLO26 detection head to achieve the detection of road targets at different scales.

[0022] Example 1: like Figure 1 The method for road target detection in rainy and foggy weather includes the following steps: Step 1: Image Input and Backbone Network Feature Extraction The input image first undergoes multi-layer feature extraction through a backbone network. Specifically, the image sequentially passes through a convolutional module and a C3k2 feature extraction module, forming feature representations at different scales during the progressive downsampling process. As the network depth increases, the model gradually extracts the edges, textures, and high-level semantic features of the road target.

[0023] To address issues such as blurred road target edges, degraded texture details, and insufficient local feature representation under rainy and foggy weather, the local feature extraction structure in the YOLO26 backbone network is optimized, and a local feature enhancement module is constructed. This module improves local feature representation capabilities while maintaining computational efficiency. Its structure is as follows: Figure 2 As shown.

[0024] For the input feature map First, a multi-branch depthwise convolutional structure is used to extract local spatial features, including Depth convolution branches, The depthwise convolution branch and the identity mapping branch. Among them, The depthwise convolution branch is used to extract spatial structure information of the local neighborhood. The depthwise convolution branch is used to independently map and enhance the features of each channel, while the identity mapping branch is used to preserve the original input features. The outputs of the three branches are summed element-wise and then batch normalized to obtain the locally enhanced features. : in, Indicates the kernel size as Channel-wise depthwise convolution, This represents the batch normalization layer. Local feature enhancement modules can be set at different scales in the backbone network.

[0025] Subsequently, the local spatial enhancement features The input is processed by a lightweight feedforward network (FFN) for channel feature encoding. The FFN consists of two convolutional normalization units, each containing... Convolutional and batch normalization layers. The first convolutional normalization unit performs channel mapping, and the second convolutional normalization unit recovers the feature dimension. A GELU activation function is used between them for non-linear feature transformation to obtain the channel-encoded features. : in, and These represent the first convolution normalization unit and the second convolution normalization unit, respectively. This represents the Gaussian error linear unit activation function.

[0026] Finally, a learnable feature scaling mechanism is introduced to... Adaptive adjustment is performed, and feature fusion is achieved by combining a dual residual connection structure. First, a learnable channel scaling factor is used. right Perform channel-wise scaling and compare with input features X Perform the first residual fusion to adaptively adjust The contribution in the residual branch; subsequently, the intermediate features obtained. Again with input features X Residual fusion is performed to enhance local augmentation information while maintaining the stability of the original feature representation. Specifically, this is expressed as: in, This indicates element-wise multiplication.

[0027] To address the degradation of local features of road targets in rainy and foggy weather, a local feature enhancement module integrates multi-branch deep convolution, FFN, learnable channel scaling, and dual residual connections to construct a mutually supportive local feature enhancement mechanism. This mechanism aims to improve the representational ability of degraded features, specifically in the following aspects: First, rain streaks and scattered light cause local texture blurring and loss of edge information, making it difficult for traditional single-branch convolution to simultaneously enhance local details and suppress noise. Therefore, this module is configured with... Depth convolution, The multi-branch parallel structure composed of depthwise convolution and identity mapping enhances local texture and edge features while preserving the original feature information and reducing information loss caused by convolution transformation by fusing complementary information from the three branches.

[0028] Secondly, considering the varying sensitivities of different channels to rain and fog interference, this module introduces learnable channel scaling coefficients after FFN completes channel encoding. Through end-to-end training, it learns the scaling weights of the enhanced features for each channel, thereby differentially adjusting the response intensity of different channels. Based on this, a dual residual connection is used for step-by-step feature fusion: first, the scaled channel-encoded features are fused with the original input using a first residual fusion to form intermediate features; then, the intermediate features are fused with the original input using a second residual fusion to further preserve the original features and stabilize feature propagation.

[0029] After processing by the local feature enhancement module, the backbone network can effectively improve the representation of local texture and edge details of road targets. However, due to the limitation of the local receptive field, the local feature extraction structure is difficult to simultaneously model detailed features and large-scale contextual information. To address this, a hybrid frequency spatial enhancement module is constructed at the end of the backbone network. The high-frequency branch enhances the representation of detailed information, while the low-frequency branch aggregates multi-scale structural and contextual information of the road targets. Through the collaborative modeling and adaptive fusion of high- and low-frequency information, the global contextual feature representation is enhanced while preserving local detailed information.

[0030] The structure of the hybrid frequency space enhancement module is as follows: Figure 3 As shown. For input features The high-frequency branch employs two strip depthwise convolutions with different receptive fields to extract local structural information in the horizontal and vertical directions, respectively. After element-wise addition and batch normalization of the two directional features, the high-frequency detail enhancement features are obtained. : in, and These represent the kernel size as follows: and Depth convolution, This indicates the batch normalization layer.

[0031] The low-frequency branch uses three branches with different expansion rates. Deep convolutions are used to process input features in parallel to construct effective receptive fields at multiple scales. After element-wise summation and batch normalization, low-frequency context-enhanced features are obtained from the features at each scale. : in, Indicates the expansion rate of Depth convolutions with dilation rates of 1, 3, and 5 were used to construct receptive fields of different scales.

[0032] To fully leverage the complementary advantages between high-frequency detail features and low-frequency semantic features, learnable parameters are introduced. And mapped to through the coefficient mapping function The interval is used to obtain adaptive fusion weights. , This can be achieved using the Sigmoid function. Then, respectively using... and The high-frequency and low-frequency features are weighted and fused to obtain the fused features. The fusion features were further processed. Convolution, batch normalization, and SiLU activation function are used to perform channel mapping and nonlinear transformation to obtain projected features. Finally, the projected features are... Input features Perform residual fusion to obtain the module output. .

[0033] Step 2: Multi-scale feature fusion based on orientation-aware enhancement in the neck network After processing by the hybrid frequency spatial enhancement module, the backbone network can obtain deep features that combine local detail information with multi-scale contextual information. However, features at different levels differ in semantic information and spatial resolution, making direct fusion insufficient to fully leverage the complementary advantages of different levels. Therefore, an adaptive direction-aware feature enhancement module is constructed in the first multi-scale feature fusion stage of the neck network. This module aggregates directional features along both the horizontal and vertical directions to construct a feature representation with direction-aware capabilities. It utilizes a lightweight shared feature encoding network to achieve interaction between horizontal and vertical directional information and incorporates learnable adjustment coefficients to adaptively adjust the response intensity of directional features, thereby enhancing the ability to represent the target spatial structure and location information in the fused features. The module structure is as follows: Figure 4 As shown.

[0034] Features obtained from the first multi-scale feature fusion of the neck network Global average pooling is performed along both the width and height dimensions. By averaging along the width dimension, the width dimension is compressed to 1 while retaining the positional information in the height direction, resulting in a height-oriented feature description. By averaging along the height dimension, the height dimension is compressed to 1 while retaining the positional information in the width direction, thus obtaining the feature description in the width direction. : in, C , H and W They represent The corresponding number of channels, height, and width, c , h and w These represent the indices for the corresponding dimensions.

[0035] Subsequently, the width direction features Perform a dimensional transformation to align it with the height-oriented features. It has a unified spatial arrangement and is spliced ​​along the spatial dimension to obtain a joint feature that simultaneously contains information in both the horizontal and vertical directions: Subsequently, the joint features Input shared feature coding network A unified encoding is performed. The shared feature coding network consists of two feature coding layers, each containing, in sequence, a set of features. Convolutional layers, batch normalization layers, and the ReLU activation function reduce channel dimensionality while enabling the interaction of horizontal and vertical information, resulting in shared encoded features. Subsequently, based on the spatial dimensions of the original features, the shared encoded features are... Along spatial dimensions according to height H and width W The spatial arrangement of the width-direction features is redefined and restored to obtain the height-direction encoded features. and width direction encoding features : To enable the module to adaptively adjust the enhancement level of directional features according to the training process, learnable parameters are introduced. And through coefficient mapping function Convert it to direction adjustment coefficient , This can be achieved using the Sigmoid function. Based on this, it can be implemented using two independent... Convolution restores the directional encoded features to the input channel dimension and utilizes adjustment coefficients. The response intensity of the directional features is adjusted to generate directional response weights in the height and width directions. and : in, and Representing the height and width directions respectively Feature mapping convolution, This is the activation function.

[0036] Finally, the response weights from the two directions are fused, and then... Adaptive recalibration is performed to obtain the output of the adaptive orientation-aware feature enhancement module. .in, This indicates element-wise multiplication.

[0037] It should be noted that although both this module and the existing Coordinate Attention (CA) aggregate features along the width and height dimensions, they differ significantly in their directional feature adjustment and bidirectional weight fusion mechanisms. CA directly generates directional attention weights by sharing encoding and branch mapping the pooled features from both directions; this invention, however, introduces a learnable directional adjustment coefficient between the shared encoding and weight generation, adaptively learning the importance of different channels in directional feature representation based on training feedback from the detection task. Furthermore, this invention employs an additive approach to fuse the attention weights in the height and width directions, making the structural information of the two directions complementary and avoiding excessive suppression of the overall features by weak responses in a single direction during multiplicative fusion.

[0038] The above design addresses the non-uniform degradation of road target structural features under rain and fog conditions: directional rain streaks easily interfere with texture and edge information in specific directions, while fog scattering further weakens target contrast and local details. Based on this, the present invention establishes a processing mechanism of "directional feature decomposition—channel response adjustment—bidirectional information complementarity" to improve the structural representation capability of road targets under rain and fog conditions.

[0039] Step 3: Complete road target detection based on multi-scale detection head The multi-scale feature maps, enhanced by fusion through the neck network, are input into three detection branches for target prediction. Different detection branches correspond to feature layers with different downsampling scales, used to detect road targets at different scales. The three detection branches output target prediction results at their respective scales, and the final road target detection result is obtained through end-to-end inference.

[0040] Experimental verification: The experiment used the BDD100K public road scene dataset, from which 5000 images representing typical road scenes such as rainy, foggy, and sunny days were selected. Rainy / foggy scenes accounted for 60%, and sunny scenes accounted for 40%. The dataset was divided into training, validation, and test sets in a 7:1:2 ratio. During model training, data augmentation strategies such as random photometric distortion, random magnification and padding, and random cross-cutting ratio (CCR) cropping were applied to the input images, and the image size was uniformly adjusted to 640×640 pixels. During training, the batch size was set to 16, the SGD optimizer was used for parameter updates, the initial learning rate was set to 0.0001, and the training epochs were set to 300.

[0041] Table 1: Model Performance Comparison

[0042] Table 1 presents the performance comparison results of each model. The method of this invention achieves the best performance in all three metrics: mAP@50, mAP@50-95, and recall. Compared with the baseline model YOLO26-M, the method of this invention reduces the number of parameters and the number of floating-point operations per billion by 19.1% and 11.0%, respectively, while improving mAP@50, mAP@50-95, and recall by 3.1%, 5.3%, and 3.1%, respectively, achieving simultaneous improvement in detection performance and model lightweighting. Compared with YOLOv11-M, which is also a real-time detection architecture, the method of this invention reduces the number of parameters and the number of floating-point operations per billion by 17.5% and 10.8%, respectively, while improving mAP@50, mAP@50-95, and recall by 5.6%, 7.0%, and 6.4%, respectively. Compared to YOLOv11-L, the number of parameters and the number of floating-point operations per billion are reduced by 34.8% and 30.3%, respectively, while mAP@50, mAP@50-95, and recall are improved by 2.7%, 4.0%, and 4.8%, respectively. This indicates that the method of this invention can still achieve better road target detection performance while significantly reducing model complexity. The method of this invention achieves the best performance in recall, indicating that it can more effectively alleviate the problem of missed road targets caused by reduced image contrast and target feature degradation in rainy and foggy weather compared to other comparative methods.

[0043] A comparison of the detection results of the method of this invention and the YOLO26 method is presented. The red boxes in the figure mark targets where YOLO26 missed or false detections, while the method of this invention correctly detected them. The results show that, under rainy and foggy weather conditions, the method of this invention can detect more targets and obtain more accurate classification results. For example, Figure 5 (b1) detected Figure 5 Vehicles, trucks and traffic lights that were missed in (a1). Figure 5 (b2) detected Figure 5 The vehicles in (a2) that were missed in the detection were partially obscured and blurry. Figure 5 (b3) Detected Figure 5 (a3) Large target blurred vehicles and medium target occlusion blurred vehicles that are missed due to feature loss caused by rain degradation effect. Figure 5 (b4) was corrected Figure 5 (a4) contains the classification error of misidentifying a truck as a car and the false detection of a wooden plank as a car. Figure 5 (b5) detected Figure 5 (a5) Large target vehicles with only partial features remaining, as well as small target vehicles, trucks and pedestrians in the distance.

Claims

1. A method for detecting road targets in rainy or foggy weather conditions, characterized in that, Includes the following steps: The road scene image is input into the backbone network of the YOLO26 model for multi-scale feature extraction; among which, A local feature enhancement module is constructed in the backbone network to enhance local feature extraction and generate local enhanced features. A hybrid frequency space enhancement module including high-frequency and low-frequency branches is constructed at the end of the backbone network. The high-frequency branch is used to enhance the ability to represent detailed information, and the low-frequency branch is used to aggregate multi-scale structural information and contextual information of road targets. The local enhanced features are input into the hybrid frequency space enhancement module, and the outputs of the high-frequency and low-frequency branches are adaptively fused to generate frequency space enhanced features. An adaptive orientation-aware feature enhancement module is constructed in the first multi-scale feature fusion stage of the neck network to output orientation-aware enhanced features for subsequent fusion of the neck network, target prediction, and generation of road target detection results.

2. The method for road target detection in rainy and foggy weather conditions according to claim 1, characterized in that, Constructing local feature enhancement modules in the backbone network includes: A multi-branch deep convolutional structure is used to extract local spatial features from the input feature map. The outputs of each branch are summed element by element and then batch normalized to obtain local spatial enhancement features. The local spatial enhancement features are input into a lightweight feedforward network for channel feature encoding to obtain channel encoded features. A learnable feature scaling mechanism is introduced to adaptively adjust the channel encoded features, and a dual residual connection structure is combined to perform feature fusion to obtain local enhancement features.

3. The method for road target detection in rainy and foggy weather conditions according to claim 2, characterized in that, The lightweight feedforward network consists of two convolutional normalization units, each of which contains The system consists of convolutional and batch normalization layers; the first convolutional normalization unit is used for channel mapping, and the second convolutional normalization unit is used to restore the feature dimension. The GELU activation function is used between the two layers for non-linear feature transformation.

4. The method for road target detection in rainy and foggy weather conditions according to claim 2, characterized in that, Multi-branch depthwise convolutional structures include Depth convolution branches, Depthwise convolution branch and identity mapping branch; Depthwise convolution branches are used to extract spatial structure information of local neighborhoods. The depthwise convolution branch is used to independently map and enhance the features of each channel, while the identity mapping branch is used to preserve the original input features.

5. The method for road target detection in rainy and foggy weather conditions according to claim 1, characterized in that, The construction of the hybrid frequency space enhancement module includes: For the input local enhancement features, the high-frequency branch uses two strip depth convolutions with different receptive fields to extract local structural information in the horizontal and vertical directions, respectively; after element-wise addition and batch normalization of the two directional features, the high-frequency detail enhancement features are obtained. The low-frequency branch uses three branches with different expansion rates. The local enhancement features are processed in parallel using deep convolution. After element-wise addition and batch normalization of the features at each scale, the low-frequency context enhancement features are obtained.

6. The method for road target detection in rainy and foggy weather conditions according to claim 1, characterized in that, The output of adaptive fusion of high-frequency and low-frequency branches includes: Set learnable parameters and map them to a coefficient mapping function. The interval is used to obtain adaptive fusion weights; based on the adaptive fusion weights, high-frequency detail enhancement features and low-frequency context enhancement features are weighted and fused to obtain fused features; the fused features are then processed... Convolution, batch normalization, and SiLU activation function are used to perform channel mapping and nonlinear transformation to obtain projection features; the projection features are then fused with local enhancement features using residuals to generate frequency space enhancement features.

7. The method for road target detection in rainy and foggy weather conditions according to claim 1, characterized in that, The adaptive orientation-aware feature enhancement module constructed in the first multi-scale feature fusion stage of the neck network includes: performing global average pooling along the width and height dimensions for the features obtained from the first multi-scale feature fusion of the neck network to obtain height orientation feature descriptions and width orientation feature descriptions, respectively; concatenating the two along the spatial dimension to obtain joint features; inputting the joint features into a shared feature encoding network for unified encoding to obtain shared encoded features; re-dividing the shared encoded features along the spatial dimension according to height and width to obtain height orientation encoded features and width orientation encoded features; setting learnable parameters and converting them into orientation adjustment coefficients through a coefficient mapping function; restoring the height orientation encoded features and width orientation encoded features to the input channel dimension, and adjusting the response intensity using adjustment coefficients to generate orientation response weights in the height and width directions; fusing the two orientation response weights and adaptively recalibrating the features obtained from the first multi-scale feature fusion of the neck network to obtain orientation-aware enhanced features.

Images