A ground marking detection method, device and medium

Abstract

The application provides a ground mark detection method and device and a medium, and relates to the technical field of computer vision. The method comprises the following steps: collecting an image in front of a vehicle to obtain a preprocessed image; inputting the preprocessed image into a feature extraction network to output different scale feature maps; presetting a 2D grid in a BEV space, determining the corresponding coordinates of each grid point in the image, and generating a relationship lookup table; determining the corresponding sub-features of each grid point in the different scale feature maps; determining the weight of the different scale feature maps corresponding to each grid point based on the distance between the first position of each grid point and the vehicle; obtaining the weight features of each grid point based on the corresponding sub-features of each grid point in the different scale feature maps and the weight of the different scale feature maps corresponding to each grid point, filling the weight features into the grid, and generating a BEV feature map; and inputting the BEV feature map into a detection module to obtain an identification result. The present scheme improves the detection accuracy and inference speed.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method, device and medium for detecting ground markings. Background Technology

[0002] In the fields of autonomous driving and high-precision map construction, accurately perceiving and understanding the vehicle's surrounding environment, especially ground markings such as lane lines, stop lines, and arrows, is crucial. Existing solutions generally extract this information from visual data and convert it into a bird's-eye view (BEV) space.

[0003] Patent CN120472048A provides a semi-automatic annotation method for ground elements based on 3D point clouds. This scheme first constructs a dense 3D point cloud of the target scene using multi-frame time-series data, then generates a bird's-eye view and elevation map from the point cloud, displaying them to the user for manual annotation. Finally, the system combines the manual annotation results with elevation information to calculate the final 3D annotation result for the ground elements. However, this scheme heavily relies on offline preprocessing and extensive manual intervention. It requires the pre-construction of a dense 3D point cloud, a process that involves processing large amounts of time-series data, resulting in significant computational overhead and failing to meet the requirements of real-time perception. Furthermore, its annotation process is centered on human-computer interaction and is not a fully automated solution, limiting the method to offline data production and ground truth annotation, making it unsuitable for vehicle-mounted systems requiring real-time perception.

[0004] Patent CN120279527A stitches together multiple top-view images using inverse perspective transformation to create a bird's-eye view, then extracts semantic elements through semantic segmentation and converts them into a point cloud. Subsequently, it integrates IMU and wheel speed sensor data to construct and optimize a semantic point cloud map. Finally, through post-processing steps such as clustering, boundary generation, and curve fitting, it outputs a vector map containing lane lines and parking space information. This technical approach suffers from significant computational complexity and hardware cost issues. This solution deeply integrates sensor data such as IMU and wheel speed sensors to achieve odometer calculation and loop closure detection, which not only increases the system's hardware cost and calibration complexity, but its overall process also includes multiple serial modules, such as image stitching, segmentation, point cloud construction, clustering, and optimization. The computational chain is long and the latency is high, making it difficult to meet the real-time requirements of vehicle environmental perception. Summary of the Invention

[0005] This invention aims to at least solve the aforementioned technical problems existing in the prior art. To this end, a first aspect of this invention proposes a ground marking detection method, the method comprising:

[0006] Acquire an RGB image of the front of the vehicle and preprocess the RGB image to obtain a preprocessed image;

[0007] The preprocessed image is input into a feature extraction network, which outputs feature maps at different scales; the feature extraction network includes a backbone neural network and an asymmetric spatial attention module.

[0008] A 2D mesh in the preset BEV space is used. Based on the camera's intrinsic and extrinsic parameter matrices, and using a projection geometry algorithm, the corresponding coordinates of each mesh point in the 2D mesh in the RGB image are determined, and a relation lookup table is generated. Each mesh point in the 2D mesh corresponds to the ground position in the real world.

[0009] Based on the relationship lookup table, determine the corresponding sub-feature of each grid point in the 2D grid in the feature maps at different scales;

[0010] The weights of the feature maps of different scales corresponding to each grid point are determined based on the distance between the first position of each grid point and the vehicle; the first position corresponds to the real-world ground position of each grid point.

[0011] Based on the corresponding sub-features of each grid point in the feature maps at different scales and the weights of the feature maps at different scales corresponding to each grid point, the weight features of each grid point are obtained, and the weight features are filled into the 2D grid to generate a BEV feature map.

[0012] The BEV feature map is input into the OBB detection module to obtain the recognition result.

[0013] Optionally, the asymmetric spatial attention module includes:

[0014] The multi-scale output map of the backbone neural network is input into the asymmetric spatial attention module. First, feature extraction and channel compression are performed on the multi-scale output map through 3×3 convolution and 1×1 convolution. Then, the obtained feature map is divided into an upper feature map and a lower feature map along the horizontal direction.

[0015] The upper feature map is convolved with a 1×1 convolution to generate an upper attention map;

[0016] The lower feature map is enhanced by 3×3 convolution and 1×1 convolution to generate the lower attention map;

[0017] The upper attention map and the lower attention map are stitched together to form an asymmetric spatial attention mask;

[0018] The asymmetric spatial attention mask is multiplied by the multi-scale output map to obtain the feature maps at different scales.

[0019] Optionally, the expression for the weights of the feature maps at different scales corresponding to each grid point is as follows:

[0020]

[0021]

[0022] Where D is the distance between the first position of the grid point and the vehicle, and the first position corresponds to the real-world ground position of each grid point. Let be the probability density of the grid point at a distance D from the main vehicle on the i-th feature map, where i represents the i-th feature map corresponding to the grid point. Let the initial preference distance center be the i-th feature map. Let be the initial preference distance width for the i-th feature map. The weights of the i-th feature map corresponding to each grid point. Let be the global bias of the i-th feature map. This represents the global bias of the j-th feature map. Let represent the probability density of a grid point at a distance D from the main vehicle on the j-th feature map, and k represent the number of feature maps of k scales corresponding to each grid point.

[0023] Optionally, inputting the BEV feature map into the OBB detection module to obtain the recognition result includes:

[0024] The BEV feature map is input into the OBB detection module to obtain the vertex coordinates and confidence level of at least one rotating box for each ground marker;

[0025] The target rotation box for each ground marker is determined based on the vertex coordinates and confidence level of at least one rotation box for each ground marker.

[0026] Optionally, determining the target rotation box for each ground marker based on the vertex coordinates and confidence level of at least one rotation box for each ground marker includes:

[0027] Based on the vertex coordinates of at least one rotated frame of each ground marker, obtain at least one rotated frame position for each ground marker;

[0028] The nonmaximum suppression method is used, and the position of at least one rotation box for each ground marker is filtered based on the confidence of the rotation box to obtain the target rotation box for each ground marker.

[0029] Optionally, the backbone neural network is ResNet.

[0030] Optionally, the preprocessing of the RGB image to obtain a preprocessed image includes:

[0031] The RGB image is subjected to size normalization, channel normalization, brightness and contrast enhancement, cropping, and rotation to obtain the preprocessed image.

[0032] A third aspect of the present invention provides an electronic device comprising a processor and a memory, wherein the memory stores at least one instruction or at least one program, the at least one instruction or at least one program being loaded and executed by the processor to implement the ground marking detection method as proposed in the first aspect.

[0033] A fourth aspect of the present invention provides a computer-readable storage medium storing at least one instruction or at least one program, wherein the at least one instruction or at least one program is loaded and executed by a processor to implement the ground marking detection method as proposed in the first aspect.

[0034] The beneficial effects of the ground marking detection method, device, and medium are as follows: In the feature extraction stage, this application introduces an asymmetric spatial attention mechanism, enabling feature maps of different scales to adaptively focus on regions in the image that contribute more to ground marking detection. This achieves a high degree of synergy between feature extraction and BEV task requirements, generating superior feature images at different scales. In the OBB detection module, this application proposes a distance-adaptive multi-scale feature fusion strategy in BEV space. This strategy dynamically calculates the weights of feature maps at each scale using the distance between the BEV grid and the vehicle, overcoming the shortcomings of traditional methods in scale processing and significantly improving the model's detection performance across the entire distance range, especially its ability to detect small targets at a distance. This application can determine the relation lookup table offline and perform feature sampling and network forward propagation online, avoiding the huge computational overhead of complex models, significantly improving inference speed, and meeting the real-time requirements of autonomous driving. Based on the asymmetric spatial attention mechanism and multi-scale feature fusion strategy introduced in this application, the geometric and semantic information of the scene can be more fully utilized when generating BEV features and performing target detection. The asymmetric spatial attention mechanism ensures the quality of multi-scale features of the image, while the distance-adaptive feature weighting fusion ensures that more effective BEV features are input to the detector. The combination of the two improves the detection accuracy, especially in complex scenes and the detection of small targets at long distances, where the detection accuracy is significantly improved. Attached Figure Description

[0035] Figure 1 A flowchart of a ground marking detection method provided in an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of the asymmetric spatial attention module in an embodiment of the present invention;

[0037] Figure 3 This is a diagram showing the correspondence between the 2D grid and the ground position in an embodiment of the present invention. Detailed Implementation

[0038] 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.

[0039] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more. Furthermore, the use of "based on" or "according to" implies openness and inclusiveness, because processes, steps, calculations, or other actions "based on" or "according to" one or more of the stated conditions or values ​​may in practice be based on additional conditions or beyond the stated values.

[0040] This invention provides a method for detecting ground markings, such as... Figure 1 As shown, the method may include the following steps:

[0041] Step 101: Acquire an RGB image of the front of the vehicle and preprocess the RGB image to obtain a preprocessed image.

[0042] In one possible implementation, preprocessing the RGB image to obtain a preprocessed image includes:

[0043] The RGB image is subjected to size normalization, channel normalization, brightness and contrast enhancement, cropping, and rotation to obtain the preprocessed image.

[0044] Specifically, RGB images are acquired using an onboard front-view camera. Size normalization, channel normalization, and brightness and contrast enhancement are applied to improve the model's robustness to different lighting and weather conditions and adapt it to the input requirements of subsequent networks. The preprocessing stage also supports image enhancement strategies, such as random cropping, rotation, and other processing, to further improve the model's generalization ability.

[0045] Step 102: Input the preprocessed image into the feature extraction network and output feature maps at different scales; the feature extraction network includes a backbone neural network and an asymmetric spatial attention module.

[0046] In one possible implementation, the backbone neural network is ResNet.

[0047] In one possible implementation, the asymmetric spatial attention module includes:

[0048] The multi-scale output map of the backbone neural network is input into the asymmetric spatial attention module. First, feature extraction and channel compression are performed on the multi-scale output map through 3×3 convolution and 1×1 convolution. Then, the obtained feature map is divided into an upper feature map and a lower feature map along the horizontal direction.

[0049] The upper feature map is convolved with a 1×1 convolution to generate an upper attention map;

[0050] The lower feature map is enhanced by 3×3 convolution and 1×1 convolution to generate the lower attention map;

[0051] The upper attention map and the lower attention map are stitched together to form an asymmetric spatial attention mask;

[0052] The asymmetric spatial attention mask is multiplied by the multi-scale output map to obtain the feature maps at different scales.

[0053] One or more images can be input into the network. Specifically, one or more preprocessed images are first input into the backbone neural network. Since the backbone neural network includes downsampling at different scales, multi-scale output images are obtained. For example, downsampling can be performed on the backbone neural network at 8x, 16x, and 32x, and the width and height of the obtained feature maps are respectively one-eighth, one-sixteenth, and one-thirty-second of the original feature map. The backbone neural network selected in this embodiment of the invention has high performance. This network structure has a deep receptive field design, which can effectively capture information from low-level edges and textures to high-level semantic features.

[0054] To further improve the quality of subsequent BEV spatial features, this embodiment of the invention introduces an asymmetric spatial attention module. This module generates a spatial weight mask for each multi-scale output image through a lightweight attention sub-network. This mask guides the network to focus more on the lower half of the image while appropriately suppressing areas such as the sky and distant objects that contribute less to ground sign detection, thereby optimizing feature extraction and laying the foundation for generating high-quality BEV features. The lower half of the image generally corresponds to the nearby ground and key road signs.

[0055] The structure of the asymmetric spatial attention module is as follows: Figure 2As shown, after inputting the multi-scale output map of the backbone neural network into the asymmetric spatial attention module, feature extraction and channel compression are performed on the C×H×W feature map through 3×3 convolution and 1×1 convolution, while maintaining the spatial resolution. Specifically, after the 3×3 convolutional layer, a batch normalization layer and a ReLU layer are added; after the 1×1 convolutional layer, a ReLU layer is added. After the above processing, the feature map is segmented along the horizontal direction to obtain an upper feature map and a lower feature map, which are then differentiated. The upper feature map is the feature map of the upper half of the image, and the lower feature map is the feature map of the lower half of the image. The upper feature map is processed by 1×1 convolution and a sigmoid activation function to generate attention weights, resulting in the upper attention map, which corresponds to the sky and other secondary regions. The lower feature map is enhanced by a more complex 3×3 convolution, batch normalization, ReLU activation function, and 1×1 convolution and a sigmoid activation function to generate the lower attention map, which focuses on key information such as the ground and roads. Then, the upper attention map and the lower attention map are spliced ​​together to form a 1×H×W asymmetric spatial attention mask. This mask is then multiplied element-wise with the multi-scale output map obtained through the backbone neural network, i.e., the original feature map, to obtain feature maps of different scales, thereby achieving selective enhancement and suppression of features in different regions.

[0056] In this embodiment of the invention, the feature extraction network ultimately outputs multiple feature maps at different scales. These feature maps are optimized feature maps; large-scale feature maps have smaller receptive fields and focus on capturing detailed information, while small-scale feature maps have larger receptive fields and focus on extracting high-level semantic information, thus providing more expressive feature representations for subsequent viewpoint switching. This module can also select other backbone networks according to actual needs, achieving a balance between accuracy and efficiency.

[0057] Step 103: Preset a 2D mesh in BEV space, and based on the camera's intrinsic and extrinsic parameter matrices, use a projection geometry algorithm to determine the corresponding coordinates of each mesh point in the 2D mesh in the RGB image, and generate a relation lookup table; wherein, each mesh point in the 2D mesh corresponds to the ground position in the real world.

[0058] This section describes the FastRay module in an embodiment of the present invention, such as... Figure 3 As shown, a 2D grid G_bev is preset in the BEV space. Each grid point (X, Y, 0) in this grid corresponds to a ground location in the real world. Here, it is assumed that the ground is an XY plane, and Z=0. The range and resolution of this grid can be set according to actual perception requirements. For example, this grid can cover an area of ​​40 meters × 40 meters in front of the vehicle, with a grid resolution of 0.1 meters / pixel.

[0059] For each target coordinate (X, Y, 0) in the grid, using the camera's intrinsic parameter matrix K and extrinsic parameter matrix... The coordinates (u, v) of the corresponding pixel in the image are calculated using the principles of projection geometry, as shown in the following formula:

[0060]

[0061] The above process can be viewed as a ray emitted from the optical center of the camera, passing through the pixel (u, v), and intersecting the ground plane Z=0 at the point (X, Y).

[0062] The FastRay module pre-calculates a lookup table that maps the coordinates of each grid point in the 2D mesh to the coordinates of sampling points in the image, based on the aforementioned projection relationships. Since the camera's intrinsic and extrinsic parameters remain fixed after installation, this lookup table can be calculated and stored offline. During the online inference phase, only efficient lookup operations are required, avoiding repetitive matrix calculations and significantly improving operational efficiency.

[0063] Step 104: Based on the relationship lookup table, determine the corresponding sub-feature of each grid point in the 2D grid in the feature maps at different scales.

[0064] Based on the lookup table generated by the FastRay module, a multi-scale feature adaptive fusion mechanism is designed. First, the corresponding sub-features of each grid point in the feature maps at different scales need to be determined. Specifically, features at each scale are sampled using differentiable bilinear interpolation.

[0065] Step 105: Determine the weights of the feature maps of different scales corresponding to each grid point based on the distance between the first position of each grid point and the vehicle; the first position is the real-world ground position corresponding to each grid point.

[0066] In one possible implementation, the weights of the feature maps at different scales corresponding to each grid point are expressed as follows:

[0067]

[0068]

[0069] Where D is the distance between the first position of the grid point and the vehicle, and the first position corresponds to the real-world ground position of each grid point. Let be the probability density of the grid point at a distance D from the main vehicle on the i-th feature map, where i represents the i-th feature map corresponding to the grid point. Let the initial preference distance center be the i-th feature map. Let be the initial preference distance width for the i-th feature map. The weights of the i-th feature map corresponding to each grid point. Let be the global bias of the i-th feature map. This represents the global bias of the j-th feature map. Let represent the probability density of a grid point at a distance D from the main vehicle on the j-th feature map, and k represent the number of feature maps of k scales corresponding to each grid point.

[0070] For feature maps at different scales, the weights of each scale feature map are dynamically calculated based on the distance between the corresponding grid points of the 2D mesh and the main vehicle, achieving distance-adaptive feature weighting fusion. Here, grid points can also be referred to as cells. Specifically, the module uses a differentiable weighted network and a soft allocation strategy based on parameterized Gaussian kernels to prioritize high-resolution detail features in near-distance regions and low-resolution semantic features in far-distance regions. Finally, features at corresponding locations are sampled from the feature maps at each scale using differentiable bilinear interpolation, weighted and aggregated, and then filled into the corresponding cells of the BEV mesh, ultimately forming a feature map in the BEV space. This feature map retains the advantages of multi-scale visual features and enhances the continuity of feature representation across distance scenarios, thus providing a more robust and structured feature representation for subsequent object detection tasks.

[0071] The core of the parameterized Gaussian kernel soft assignment strategy is to define a "distance preference" for the feature map at each scale and use a smooth Gaussian kernel to describe the strength of this preference as a function of distance. The entire weight generation process is differentiable, facilitating end-to-end learning. Specifically, initial preference distance center, preference distance width, and global bias are set for the optimized feature maps at different scales. Based on the distance D between the corresponding cell in the 2D grid and the main vehicle, the Gaussian response of each scale feature is first calculated, which is the formula above. Secondly, the fusion weights are generated by combining the global bias and normalizing with Softmax. .

[0072] Step 106: Based on the corresponding sub-features of each grid point in the feature maps at different scales and the weights of the feature maps at different scales corresponding to each grid point, obtain the weight features of each grid point, and fill the weight features into the 2D grid to generate a BEV feature map.

[0073] After obtaining the fusion weights, the features at each scale obtained by differentiable bilinear interpolation sampling are aggregated according to the weights and filled into the grid G_bev to form a BEV feature map that combines detail-semantic balance and spatial continuity.

[0074] Step 107: Input the BEV feature map into the OBB detection module to obtain the recognition result.

[0075] In one possible implementation, inputting the BEV feature map into the OBB detection module to obtain the recognition result includes:

[0076] The BEV feature map is input into the OBB detection module to obtain the vertex coordinates and confidence level of at least one rotating box for each ground marker;

[0077] The target rotation box for each ground marker is determined based on the vertex coordinates and confidence level of at least one rotation box for each ground marker.

[0078] In one possible implementation, determining the target rotation box for each ground marker based on the vertex coordinates and confidence level of at least one rotation box for each ground marker includes:

[0079] Based on the vertex coordinates of at least one rotated frame of each ground marker, obtain at least one rotated frame position for each ground marker;

[0080] The nonmaximum suppression method is used, and the position of at least one rotation box for each ground marker is filtered based on the confidence of the rotation box to obtain the target rotation box for each ground marker.

[0081] Specifically, the obtained BEV feature map is input into the Rotated Object Detector (OBB). This detector further extracts high-level semantic features through a multi-layer convolutional neural network and predicts the coordinates of the four vertices, class probability, and confidence score of each rotated bounding box on the feature map. The detection head employs a rotated bounding box parameterization method, which can more accurately describe the actual contours of ground markings. Finally, the system outputs the rotated bounding box parameters and their class labels for all ground markings detected in the BEV space, completing the end-to-end inference process from image to structured BEV output.

[0082] Table 1 below shows the average precision and small target recall of models using different structures. As shown in Table 1, on the Apolloscape lane_segmentation dataset, the embodiments of the present invention significantly improve the detection accuracy compared to the baseline version, especially in complex scenes and for distant targets, where the improvement in detection accuracy is more significant, including: overall accuracy improvement and enhanced small target detection capability. Specifically, the average precision (mAP) of the embodiments of the present invention is improved by 5%~8%; for distant targets more than 50 meters away, the detection recall is improved by more than 15%. The model structure used in the baseline version is ResNet50+FastRay+OBB.

[0083] Table 1

[0084] Model Structure Average accuracy Small target recall rate Baseline (ResNet50 + FastRay + OBB) 36.8% 22.9% Baseline + Asymmetric Spatial Attention Module 42.1% 29.6% Baseline + Multi-scale Feature Adaptive Fusion Mechanism 43.7% 35.1% Baseline + Asymmetric Spatial Attention Module + Multi-Scale Feature Adaptive Fusion Mechanism 45.3% 38.5%

[0085] In the feature extraction stage, this invention introduces an asymmetric spatial attention mechanism, enabling feature maps of different scales to adaptively focus on regions in the image that contribute more to ground marker detection. This achieves a high degree of synergy between feature extraction and BEV task requirements, generating superior feature images at different scales. In the OBB detection module, this invention proposes a distance-adaptive multi-scale feature fusion strategy in BEV space. This strategy dynamically calculates the weights of feature maps at each scale using the distance between the BEV grid and the vehicle, overcoming the shortcomings of traditional methods in scale processing and significantly improving the model's detection performance across the entire distance range, especially its ability to detect small targets at a distance. This invention can also be determined offline. The relation lookup table enables online feature sampling and network forward propagation, avoiding the huge computational overhead of complex models and significantly improving inference speed to meet the real-time requirements of autonomous driving. Based on the asymmetric spatial attention mechanism and multi-scale feature fusion strategy introduced in this embodiment, the geometric and semantic information of the scene can be more fully utilized when generating BEV features and performing target detection. The asymmetric spatial attention mechanism ensures the quality of multi-scale features of the image, while the distance-adaptive feature weighting fusion ensures that more effective BEV features are input to the detector. The combination of the two improves the detection accuracy, especially in complex scenes and the detection of small targets at long distances, where the detection accuracy is significantly improved.

[0086] This invention allows for end-to-end training, from feature extraction and mapping to the detection module. The generation of BEV features is specifically optimized for subsequent detection tasks, achieving a high degree of consistency between feature representation and task objectives. This invention has low hardware computing power requirements, eliminating the need for additional equipment such as multiple sensors or high-precision IMUs. Reliable BEV perception can be achieved solely through monocular visual input, reducing system costs and facilitating large-scale applications.

[0087] In another embodiment of the present invention, an electronic device is also provided, the electronic device including a processor and a memory, the memory storing at least one instruction or at least one program, the at least one instruction or at least one program being loaded and executed by the processor to implement the ground marking detection method proposed in the embodiment of the present invention.

[0088] In another embodiment of the present invention, a computer-readable storage medium is also provided, wherein at least one instruction or at least one program is stored in the storage medium, and the at least one instruction or at least one program is loaded and executed by a processor to implement the ground marking detection method proposed in the embodiment of the present invention.

[0089] The foregoing primarily describes the solutions provided by the embodiments of the present invention from the perspective of the device. It is understood that, in order to achieve the above functions, the device includes corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the algorithmic steps of the various examples described in the embodiments disclosed herein, the present invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0090] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions within the technical scope disclosed in the present invention should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting ground markings, characterized in that, include: Acquire an RGB image of the front of the vehicle and preprocess the RGB image to obtain a preprocessed image; The preprocessed image is input into a feature extraction network, which outputs feature maps at different scales. The feature extraction network includes a backbone neural network and an asymmetric spatial attention module; The asymmetric spatial attention module includes: The multi-scale output map of the backbone neural network is input into the asymmetric spatial attention module. First, feature extraction and channel compression are performed on the multi-scale output map through 3×3 convolution and 1×1 convolution. Then, the obtained feature map is divided into an upper feature map and a lower feature map along the horizontal direction. The upper feature map is convolved with a 1×1 convolution to generate an upper attention map; The lower feature map is enhanced by 3×3 convolution and 1×1 convolution to generate the lower attention map; The upper attention map and the lower attention map are stitched together to form an asymmetric spatial attention mask; The asymmetric spatial attention mask is multiplied by the multi-scale output map to obtain the feature maps at different scales; A 2D mesh in the preset BEV space is used. Based on the camera's intrinsic and extrinsic parameter matrices, and using a projection geometry algorithm, the corresponding coordinates of each mesh point in the 2D mesh in the RGB image are determined, and a relation lookup table is generated. Each mesh point in the 2D mesh corresponds to the ground position in the real world. Based on the relationship lookup table, determine the corresponding sub-feature of each grid point in the 2D grid in the feature maps at different scales; The weights of the feature maps of different scales corresponding to each grid point are determined based on the distance between the first position of each grid point and the vehicle; the first position corresponds to the real-world ground position of each grid point. Based on the corresponding sub-features of each grid point in the feature maps at different scales and the weights of the feature maps at different scales corresponding to each grid point, the weight features of each grid point are obtained, and the weight features are filled into the 2D grid to generate a BEV feature map. The BEV feature map is input into the OBB detection module to obtain the recognition result.

2. The ground marking detection method according to claim 1, characterized in that, The expression for the weights of the feature maps at different scales corresponding to each grid point is as follows: Where D is the distance between the first position of the grid point and the vehicle, and the first position corresponds to the real-world ground position of each grid point. Let be the probability density of a grid point at a distance D from the main vehicle on the i-th feature map, where i represents the i-th feature map corresponding to the grid point. Let the initial preference distance center be the i-th feature map. Let be the initial preference distance width for the i-th feature map. The weights of the i-th feature map corresponding to each grid point. Let be the global bias of the i-th feature map. This represents the global bias of the j-th feature map. Let represent the probability density of a grid point at a distance D from the main vehicle on the j-th feature map, and k represent the number of feature maps of k scales corresponding to each grid point.

3. The ground marking detection method according to claim 1, characterized in that, The step of inputting the BEV feature map into the OBB detection module to obtain the recognition result includes: The BEV feature map is input into the OBB detection module to obtain the vertex coordinates and confidence level of at least one rotating box for each ground marker; The target rotation box for each ground marker is determined based on the vertex coordinates and confidence level of at least one rotation box for each ground marker.

4. The ground marking detection method according to claim 3, characterized in that, Determining the target rotation box for each ground marker based on the vertex coordinates and confidence level of at least one rotation box for each ground marker includes: Based on the vertex coordinates of at least one rotated frame of each ground marker, obtain at least one rotated frame position for each ground marker; The nonmaximum suppression method is used, and the position of at least one rotation box for each ground marker is filtered based on the confidence of the rotation box to obtain the target rotation box for each ground marker.

5. The ground marking detection method according to claim 1, characterized in that, The backbone neural network is ResNet.

6. The ground marking detection method according to claim 1, characterized in that, The step of preprocessing the RGB image to obtain a preprocessed image includes: The RGB image is subjected to size normalization, channel normalization, brightness and contrast enhancement, cropping, and rotation to obtain the preprocessed image.

7. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing at least one instruction or at least one program, the at least one instruction or at least one program being loaded and executed by the processor to implement the ground marking detection method as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction or at least one program, which is loaded and executed by a processor to implement the ground marking detection method as described in any one of claims 1-6.

Images