A 4D radar and camera fusion target detection method in low visibility environment

Abstract

A target detection method integrating 4D radar and camera in low-visibility environments is disclosed, relating to the field of autonomous driving perception. Specifically, in the radar branch, 4D radar data is processed through radar feature extraction to obtain radar features, which are then encoded using radar BEV (Browser Energy Activation) to generate radar BEV features. A priori masking module generates candidate region priors based on image BEV features and guides sparse point compensation to compensate for low-density regions of radar BEV, resulting in enhanced radar BEV features. In the image branch, the image is processed through a residual network and a bidirectional feature pyramid network to obtain multi-scale pyramid image features. Simultaneously, 4D radar data is projected onto the image plane to generate a radar sparse depth map. Both are input into a view transformation to obtain image BEV features. Finally, dynamic gating multimodal fusion of the two BEV features yields fused BEV features, and the target detection head outputs the target detection category and a 3D bounding box.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving perception, specifically a target detection method that integrates 4D radar and camera in low visibility environments. Background Technology

[0002] With the rapid development of autonomous driving and intelligent transportation technologies, vehicle environmental perception systems have become a key foundation for achieving safe driving and intelligent decision-making. Currently, autonomous vehicle perception systems typically rely on multiple sensors, such as cameras, LiDAR, and millimeter-wave radar, to acquire information about the surrounding environment and provide data support for target recognition, planning, and control. In urban road scenarios with regular structures and good visibility, relying solely on camera and LiDAR perception devices has achieved relatively ideal application results. However, in low-visibility weather conditions and dynamic and complex traffic environments, existing 3D target detection methods still face significant performance degradation. Specifically, LiDAR is easily affected by adverse weather conditions, resulting in a significant decrease in its effective detection range and measurement accuracy. Furthermore, limited by scanning frequency and beam distribution characteristics, its stable perception capability for distant or complex targets is insufficient. While cameras can provide high-resolution visual information and rich semantic features, their imaging quality is highly dependent on lighting and weather conditions. In low-visibility environments, targets are prone to blurring, occlusion, or even failure, thus affecting the reliability of detection.

[0003] In contrast, 4D radar has stronger environmental adaptability, capable of stably measuring target distance, radial velocity, azimuth, and altitude under weather conditions such as rain and fog, and has high reliability in assessing target motion, making it suitable for low-visibility weather. However, its data is sparse, irregularly distributed, noisy, and lacks sufficient semantic information, making it difficult to accurately characterize target categories and 3D geometric structures independently. Furthermore, radar echo intensity varies between different echoes, and low-confidence echoes may interfere with feature representation, causing false detections or positioning errors. To improve the robustness of 3D target detection in low-visibility weather, existing technologies have proposed various fusion perception schemes. Patent CN110363820B discloses a pre-fusion target detection method based on lidar and image, which performs feature extraction and fusion of point cloud and image information to complete target detection. However, when the quality of images and point clouds fluctuates under low visibility conditions, the correlation sensitivity of cross-modal features increases significantly, leading to a decrease in the stability of fused features. Patent CN117274749B discloses a 3D target detection method based on the fusion of 4D radar and images. It uses 4D radar point cloud to generate a "sparse depth map" to explicitly supervise the monocular depth estimation. However, when the radar noise is stronger, the supervision signal may decrease, which will affect the stability of the fusion and thus affect the final detection accuracy.

[0004] In summary, to address the issues of image feature degradation, sparse radar echoes with enhanced noise echo interference, and fluctuating cross-modal fusion features in low-visibility weather traffic scenarios, it is necessary to propose a 4D radar and camera fusion 3D target detection method that can achieve stable fusion representation in BEV space, so as to improve the reliability and robustness of target detection in low-visibility scenarios. Summary of the Invention

[0005] To address the problems of camera imaging degradation under complex weather conditions, sparse and semantically insufficient point clouds of 4D millimeter-wave radar, and unstable cross-modal alignment leading to a decline in 3D target detection performance in existing technologies, this invention proposes a target detection method that fuses 4D radar and camera in low-visibility environments to improve the reliability and robustness of target detection in low-visibility scenarios.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] This invention is a target detection method that integrates 4D radar and camera in low visibility environments. The method includes a radar branch, an image branch, a priori mask module, dynamic gating multimodal fusion, and a target detection head.

[0008] The method includes the following steps:

[0009] Step 1, Radar Branch:

[0010] Step 1.1, Radar Feature Extraction:

[0011] Radar feature extraction receives 4D radar data, constructs and encodes features, and performs decoupling operations to form decoupled features such as position, velocity, and RCS. Further, based on RCS, a radar confidence gating weight is constructed, and the radar features are weighted using a gated self-attention mechanism to reduce the interference of noise points on the radar features, outputting enhanced radar features. The decoupling operation to form decoupled features such as position, velocity, and RCS is shown in equation (1). The RCS confidence gating weight is constructed, and the radar features are weighted using a gated self-attention mechanism as shown in equation (2), as detailed below:

[0012] (1)

[0013] in Indicates radar characteristics, for 4D radar data at any time This represents decoupling operations. These represent three types of features: position, velocity, and RCS.

[0014] (2)

[0015] in For the credibility-gated weights constructed by RCS, For element-wise multiplication, This represents the self-attention operation. This represents the radar signature after gating and self-attention weighting enhancement. This indicates the radar signature before enhancement.

[0016] Step 1.2, Radar BEV coding:

[0017] Radar BEV encoding receives radar features and generates radar features, which are then mapped onto the bird's-eye view BEV grid space. This mapping forms radar BEV features, and high-level semantic features are extracted from these BEV features. These features are then combined with a multi-scale feature enhancement structure for cross-scale fusion, resulting in stable radar BEV features. .

[0018] Step 1.3, Sparse Point Compensation:

[0019] Radar BEV features generated by sparse point compensated receiving radar BEV encoding Guided by the prior masking module, the compensation region for radar BEV features is determined, and low-density regions are interpolated and densified to output radar BEV enhanced features. .

[0020] Step 2. Image branches:

[0021] Step 2.1, Residual Network:

[0022] The residual network receives images and uses a deep residual network structure to perform layer-by-layer convolution and downsampling operations to extract multi-scale image feature maps containing shallow texture information and deep semantic information.

[0023] Step 2.2, Bidirectional Feature Pyramid Network:

[0024] The bidirectional feature pyramid network receives multi-scale image feature maps generated by the residual network. By introducing a bidirectional cross-scale connection mechanism, it performs bottom-up and top-down feature aggregation on the multi-scale image feature maps, enhancing the semantic interaction between features of different scales and obtaining multi-scale pyramid image features with enhanced expressive power.

[0025] Step 2.3, View Transformation:

[0026] View transformation receives multi-scale pyramid image features generated by a bidirectional feature pyramid network and maps them to the BEV space to form image BEV features. The image BEV features are obtained by projecting the 4D radar onto the image plane to form a sparse depth map and then performing a view transformation.

[0027] Step 2.4, Prior Mask Module:

[0028] The prior mask module receives the image BEV features generated by the view transformation, generates a prior mask, and provides the prior mask to the sparse point compensation to limit the spatial range of the sparse point compensation, thereby reducing invalid compensation in non-target areas and enhancing the density of the target area.

[0029] Step 3: Dynamic Gated Multimodal Fusion:

[0030] BEV features of images generated by dynamic gated multimodal fusion receiving sparse point compensation and view transformation Radar BEV Enhancement Features As input, the dynamic gated multimodal fusion is first based on radar BEV enhancement features. Extracting RCS-related confidence scores and constructing physical prior weights by combining spatial diffusion, further integrating with multi-scale convolutional networks to... and The data-driven offsets learned in the model are used to implement a hybrid gating of physical priors and data-driven information using a two-level nested Sigmoid structure. The resulting gating weights are applied to the... and Adaptive weighted fusion is performed to obtain fused features. Subsequently, a channel attention mechanism with RCS global statistical enhancement is introduced to enhance the fused features across modes. This enables feature-level fusion and information interaction between radar BEV enhanced features and image BEV features in the BEV space, resulting in the output of fused BEV features. Specifically, see equations (3) to (12):

[0031] (3)

[0032] in It is the radar BEV enhancement feature obtained by sparse point compensation of the radar branch; Represents the Sigmoid function; This represents the confidence extraction coefficient, used to extract RCS-related confidence information from radar BEV enhancement features; This represents the original RCS confidence plot.

[0033] (4)

[0034] in Indicates the original RCS confidence level; This represents the spatial diffusion coefficient, which propagates sparse high-confidence echoes to the neighborhood. This represents the confidence level of the RCS after diffusion.

[0035] (5)

[0036] in These are physical prior weights; This represents the adaptive adjustment coefficient; , These represent the weighting coefficients; Indicates average pooling; Represent a constant; Indicates the confidence level of RCS after diffusion;

[0037] (6)

[0038] in Represents the BEV features of an image; , This represents multi-scale convolutional feature extraction, each corresponding to a different receptive field; Represents multi-scale fusion features; This represents the radar BEV enhancement feature obtained by sparse point compensation of the radar branch.

[0039] (7)

[0040] in This represents a gated offset learning network; This represents the hyperbolic tangent function, with the offset restricted to [-1, 1]. Indicates the data-driven gating offset; Multi-scale fusion features.

[0041] , , (8)

[0042] in Indicates radar prior gating; Indicates the radar gating weights; Represents image gating weights; , Indicates hyperparameters; This represents mapping physical priors to the gate space.

[0043] (9)

[0044] in This represents the radar BEV enhancement feature obtained by sparse point compensation of the radar branch. Represents image gating weights; Indicates the radar gating weights; Represents the BEV features of an image; This indicates the splicing characteristics after gating fusion.

[0045] (10)

[0046] in This represents the mean; Indicates standard deviation; Indicates the maximum value; Represents the global statistical vector of RCS; It is the physical prior weight.

[0047] , (11)

[0048] in Indicates global average pooling; This indicates the splicing characteristics after gating fusion; Global descriptive vector of gated fusion features; This represents a multilayer perceptron; This indicates vector concatenation; Represents the Sigmoid function; This represents the channel attention weight.

[0049] (12)

[0050] in Indicates the dimensionality reduction parameters; This indicates the splicing characteristics after gating fusion; Indicates channel attention weights; The final output is the fused BEV feature.

[0051] Step 4, Target Detection Head:

[0052] The target detection head receives fused BEV features generated by dynamically gated multimodal fusion. As input, the target category and 3D bounding box parameters are output, and finally the 3D target detection result is obtained.

[0053] The beneficial effects of this invention are as follows: This invention includes a radar branch, an image branch, and dynamic gating multimodal fusion; wherein the radar branch extracts radar features through radar feature extraction, obtains radar BEV features through radar BEV encoding, and simultaneously, a priori mask module generates a priori mask based on image BEV features, and uses the priori mask as input sparse point compensation to obtain radar BEV enhanced features; the image branch extracts features from the image through a residual network to obtain a multi-scale image feature map, and obtains multi-scale pyramid image features through a bidirectional feature pyramid network; simultaneously, 4D radar is projected onto the image plane to generate a radar sparse depth map, and inputs it together with the multi-scale pyramid image features into view transformation to obtain image BEV features; secondly, dynamic gating multimodal fusion fuses the radar BEV enhanced features and image BEV features to achieve cross-modal information interaction, obtaining fused BEV features; finally, the target detection head outputs the target detection category and 3D bounding box parameters to obtain the final detection result, thereby improving the accuracy and robustness of 3D target detection in low-visibility weather and traffic scenarios. Attached Figure Description

[0054] Figure 1 A target detection method that combines 4D radar and camera in low-visibility environments Detailed Implementation

[0055] The present invention will now be described in detail with reference to the accompanying drawings.

[0056] like Figure 1 As shown,

[0057] Step 1, Radar Branch:

[0058] Step 1.1, 4D Radar:

[0059] This includes time 4D radar point cloud data , It can be represented as a set of points ,in This represents the number of radar point clouds in this frame. Each radar point... Include , where represents the spatial position, radial velocity, and RCS reflection intensity of the point in the vehicle coordinate system, respectively.

[0060] Step 1.2, Radar Feature Extraction:

[0061] Radar feature extraction and reception in 4D radar Then, feature construction and encoding are performed to obtain radar features. The radar point features are decoupled and represented to form three types of features: position, velocity, and RCS. Furthermore, a confidence-gated weight is constructed based on the RCS, and weighted enhancement is achieved by combining a gated self-attention mechanism. The decoupling operation forms decoupled features such as position, velocity, and RCS as shown in equation (1). The RCS is used to construct confidence-gated weights, and the radar features are weighted and enhanced by combining a gated self-attention mechanism as shown in equation (2), as detailed below:

[0062] (1)

[0063] in Indicates radar characteristics, For a moment 4D radar data, This represents decoupling operations. These represent three types of features: position, velocity, and RCS, with decoupling operations... It can include position feature coding, velocity feature coding, and RCS feature coding. Position feature coding is... Mapping to obtain Velocity feature encoding is for radial velocity Mapping to obtain RCS feature encoding pair Mapping to obtain The three encodings consist of multiple layers. This ultimately leads to the formation of point-level radar characteristics. .

[0064] (2)

[0065] in For the credibility-gated weights constructed by RCS, For element-wise multiplication, This represents a self-attention operation. This represents the radar signature after gating and self-attention weighting enhancement. This indicates the enhanced radar signature, which can be represented by the RCS feature. It is obtained by Sigmoid normalization and is used to limit the gating range to the interval [0, 1], thereby realizing the enhancement by adjusting the confidence weight.

[0066] Step 1.3, Radar BEV coding:

[0067] Enhanced radar signature output from step 2 of radar BEV encoding reception And map it onto the bird's-eye view BEV grid space to generate radar BEV features. Specifically, the BEV plane is divided into... Each grid cell is based on the planar position of each radar point. The grid index is determined, and point features within the same grid are aggregated using maximum weighting to obtain grid-level features. Subsequently, the grid-level features are mapped to form a radar BEV feature map through a mapping operation. High-level semantic features are extracted from the BEV features, and cross-scale fusion is performed using a multi-scale feature enhancement structure to finally obtain stable radar BEV features. .

[0068] Step 1.4, Sparse Point Compensation:

[0069] Radar BEV features generated by sparse point compensated receiving radar BEV encoding Guided by the prior masking module, the compensation region for radar BEV features is determined. Low-density regions are interpolated and densified, while invalid features in non-target regions are suppressed, resulting in the output of enhanced radar BEV features. The mask output by the prior mask module is denoted as . ,in High-response regions represent candidate target regions. Sparse point compensation is performed only within the masked guidance region to reduce the spread of non-target regions, erroneous compensation, or false points. Dense compensation is achieved through interpolation denserization: for low-density grids in the BEV feature map, interpolation is performed using the features of their neighboring grids to complete the compensated features. Then, the radar BEV characteristics are enhanced by combining the prior mask output. (See equation (3)).

[0070] (3)

[0071] Step 2. Image branches:

[0072] Step 2.2, Image:

[0073] Get Time Camera image data .

[0074] Step 2.3, Residual Network:

[0075] Image reception via residual network Multi-scale image feature maps are extracted through layer-by-layer convolution and downsampling. These multi-scale image feature maps contain both shallow texture information and deep semantic information, outputting a multi-scale image feature set. As shown in equation (4), it provides input for the subsequent bidirectional feature pyramid network and view transformation.

[0076] (4)

[0077] Step 2.4, Bidirectional Feature Pyramid Network:

[0078] A bidirectional feature pyramid network receives a set of multi-scale image feature maps generated by a residual network. By introducing a bidirectional cross-scale connection mechanism, bottom-up and top-down feature aggregation is performed on the multi-scale image feature maps, enhancing the semantic interaction between features of different scales and obtaining multi-scale pyramid image features with enhanced expressive power. This provides a foundation for subsequent view transformations and candidate region prior generation.

[0079] Step 2.5, View Transformation:

[0080] View transformation receives multi-scale pyramid image features generated by a bidirectional feature pyramid network. Mapping to BEV space to form image BEV features The radar branch projects radar points onto the image plane to construct a sparse depth prior for the radar. It is used to impose constraints on the geometric mapping process of view transformation, reducing BEV mapping deviation caused by image degradation.

[0081] Step 2.6, Prior Mask Module:

[0082] The priori masking module receives the BEV features of the image generated by the view transformation. Generate a priori mask The prior mask is then provided to the sparse point compensation of the radar branch to limit the spatial range of sparse point compensation, thereby reducing invalid compensation in non-target areas and enhancing the density of target areas.

[0083] (5)

[0084] in For the Sigmoid function, For convolution mapping, This refers to the BEV features of the image.

[0085] Step 3: Dynamic Gated Multimodal Fusion:

[0086] BEV features of images generated by dynamic gated multimodal fusion receiving sparse point compensation and view transformation Radar BEV Enhancement Features As input, the dynamic gated multimodal fusion is first based on radar BEV enhancement features. Extracting RCS-related confidence scores and constructing physical prior weights by combining spatial diffusion, further integrating with multi-scale convolutional networks to... and The data-driven offsets learned in the model are used to implement a hybrid gating of physical priors and data-driven information using a two-level nested Sigmoid structure. The resulting gating weights are applied to the... and Adaptive weighted fusion is performed to obtain fused features. Subsequently, a channel attention mechanism with RCS global statistical enhancement is introduced to enhance the fused features across modes. This enables feature-level fusion and information interaction between radar BEV enhanced features and image BEV features in the BEV space, resulting in the output of fused BEV features. Specifically, see equations (6) to (15):

[0087] (6)

[0088] in It is the radar BEV enhancement feature obtained by sparse point compensation of the radar branch; Represents the Sigmoid function; This represents the confidence extraction coefficient, used to extract RCS-related confidence information from radar BEV enhancement features; This represents the original RCS confidence level.

[0089] (7)

[0090] in Indicates the original RCS confidence level; This represents the spatial diffusion coefficient, which propagates sparse high-confidence echoes to the neighborhood. This represents the confidence level of the RCS after diffusion.

[0091] (8)

[0092] in These are physical prior weights; This represents the adaptive adjustment coefficient; , These represent the weighting coefficients; Indicates average pooling; Represent a constant; Indicates the confidence level of RCS after diffusion;

[0093] (9)

[0094] in Represents the BEV features of an image; , This represents multi-scale convolution feature extraction; Represents multi-scale fusion features; This represents the radar BEV enhancement feature obtained by sparse point compensation of the radar branch.

[0095] (10)

[0096] in This represents a gated offset learning network; This represents the hyperbolic tangent function, with the offset restricted to [-1, 1]. Indicates the data-driven gating offset; Multi-scale fusion features.

[0097] , , (11)

[0098] in Indicates radar prior gating; Indicates the radar gating weights; Represents image gating weights; , Indicates hyperparameters; This represents mapping physical priors to the gate space.

[0099] (12)

[0100] in This represents the radar BEV enhancement feature obtained by sparse point compensation of the radar branch. Represents image gating weights; Indicates the radar gating weights; Represents the BEV features of an image; This indicates the splicing characteristics after gating fusion.

[0101] (13)

[0102] in This represents the mean; Indicates standard deviation; Indicates the maximum value; Represents the global statistical vector of RCS; It is the physical prior weight.

[0103] , (14)

[0104] in Indicates global average pooling; This indicates the splicing characteristics after gating fusion; Global descriptive vector of gated fusion features; This represents a multilayer perceptron; This indicates vector concatenation; Represents the Sigmoid function; Indicates channel attention weights;

[0105] (15)

[0106] in Indicates the dimensionality reduction parameters; This indicates the splicing characteristics after gating fusion; Indicates channel attention weights; The final output is the fused BEV feature.

[0107] Step 4, Target Detection Head:

[0108] The target detection head receives the fused BEV features generated by dynamic gated multimodal fusion. As input, the target category and 3D bounding box parameters are output, and finally the 3D target detection result is obtained. The 3D bounding box parameters include information such as the target center position, size and orientation.

[0109] In summary, this invention comprises a radar branch, an image branch, and dynamic gating multimodal fusion. The radar branch extracts radar features through radar feature extraction, obtains radar BEV features through radar BEV encoding, and simultaneously, a priori mask module generates a priori mask based on the image BEV features, using this priori mask as input sparse point compensation to obtain radar BEV enhanced features. The image branch extracts features from the image through a residual network to obtain a multi-scale image feature map, and obtains multi-scale pyramid image features through a bidirectional feature pyramid network. Simultaneously, a 4D radar is projected onto the image plane to generate a radar sparse depth map, which is input into the view transformation along with the multi-scale pyramid image features to obtain image BEV features. Next, dynamic gating multimodal fusion fuses the radar BEV enhanced features and image BEV features to achieve cross-modal information interaction, obtaining fused BEV features. Finally, the target detection head outputs the target detection category and 3D bounding box parameters to obtain the final detection result, thereby improving the accuracy and robustness of 3D target detection in low-visibility weather and traffic scenarios.

Claims

1. A target detection method integrating 4D radar and camera in low-visibility environments, used to improve the target detection accuracy of vehicles in low-visibility weather traffic scenarios, characterized in that... Includes the following steps: S1: Acquire 4D radar data and images of the scene; S2: Radar feature extraction and projection onto the image plane are performed on the 4D radar data to generate a radar sparse depth map, and radar BEV encoding is performed on the radar features to obtain radar BEV features. S3: Extract multi-scale image feature maps from the image using a residual network; input the multi-scale image feature maps into a bidirectional feature pyramid network for feature pyramid fusion to obtain multi-scale pyramid image features; project the multi-scale pyramid image features onto the image plane to generate a radar sparse depth map, and then perform view transformation to obtain image BEV features; S4: Based on the image BEV features, a priori mask is generated through the priori mask module, and under the guidance of the priori mask, sparse point compensation is performed on the radar BEV features to obtain radar BEV enhanced features. S5: Dynamically gated multimodal fusion of the radar BEV enhancement features and the image BEV features is performed to obtain fused BEV features; S6: The fused BEV features are output by the target detection head as the target detection category and 3D bounding box to obtain the target detection result.

2. The target detection method based on 4D radar and camera fusion in low-visibility environments according to claim 1, characterized in that, The radar feature extraction includes feature encoding of 4D radar data to obtain radar features, including position features, velocity features and RCS features.

3. The method according to claim 2, characterized in that, A radar confidence gating weight is constructed based on the RCS features, and the radar features are weighted by a gating self-attention mechanism to reduce the interference of noise points on the radar features.

4. The target detection method based on 4D radar and camera fusion in low-visibility environments according to claim 1, characterized in that, The radar BEV encoding includes: forming a BEV feature map by scattering radar features, extracting high-level semantic features from the BEV feature map, and performing cross-scale fusion by combining a multi-scale feature enhancement structure to obtain stable radar BEV features.

5. The target detection method based on 4D radar and camera fusion in low-visibility environments according to claim 1, characterized in that, The residual network extracts multi-scale image feature maps through layer-by-layer convolution and downsampling. The bidirectional feature pyramid network aggregates the multi-scale image features through bidirectional cross-scale connections to obtain multi-scale pyramid image features. The view transformation includes: introducing the projection onto the image plane to generate a radar sparse depth map as a geometric constraint during the process of mapping multi-scale pyramid image features to BEV space, so as to reduce mapping deviation under image degradation conditions.

6. The target detection method based on 4D radar and camera fusion in low-visibility environments according to claim 1, characterized in that, The prior mask module includes: generating a prior mask for indicating candidate target regions from the image BEV features; The sparse point compensation includes: determining the compensation region under the guidance of the prior mask, and performing interpolation densification on the radar BEV features within the compensation region to obtain radar BEV enhanced features.

7. The target detection method based on 4D radar and camera fusion in low-visibility environments according to claim 1, characterized in that, The dynamic gated multimodal fusion includes: generating physical prior weights through RCS confidence extraction and spatial diffusion; combining data-driven offsets learned by a multi-scale convolutional network; implementing physical-data hybrid gating using a two-level Sigmoid nested structure; adaptively weighting and fusing image BEV features and radar BEV features; and achieving cross-modal feature enhancement through a channel attention mechanism enhanced by RCS global statistics. This results in feature-level fusion of radar BEV enhanced features and image BEV features within the BEV space. The calculation formula is shown below: in It is the radar BEV enhancement feature obtained by the radar branch through the sparse point compensation module; Represents the Sigmoid function; This represents the confidence extraction coefficient, used to extract RCS-related confidence information from radar BEV enhancement features; Indicates the original RCS confidence level; in This represents the spatial diffusion coefficient, which propagates sparse high-confidence echoes to the neighborhood. Indicates the confidence level of RCS after diffusion; in These are physical prior weights; This represents the adaptive adjustment coefficient; , These represent the weighting coefficients; Indicates average pooling; Represent a constant; in Represents the BEV features of an image; , This represents multi-scale convolution feature extraction; Represents multi-scale fusion features; in This represents a gated offset learning network; This represents the hyperbolic tangent function, with the offset restricted to [-1, 1]. Indicates the data-driven gating offset; ， ， in Indicates radar prior gating; Indicates the radar gating weights; Represents image gating weights; , Indicates hyperparameters; This represents mapping physical priors to the gate space; in This indicates the splicing characteristics after gating fusion; in This represents the mean; Indicates standard deviation; Indicates the maximum value; Represents the global statistical vector of RCS; These are physical prior weights; ， in Indicates global average pooling; Global descriptive vector of gated fusion features; This represents a multilayer perceptron; This indicates vector concatenation; Indicates channel attention weights; in Indicates the dimensionality reduction parameters; The final output is the fused BEV feature.

8. The target detection method based on 4D radar and camera fusion in low-visibility environments according to claim 1, characterized in that, The target detection head uses fused BEV features to complete the target detection category and 3D bounding box prediction, and obtains the 3D target detection result.

Images