A method, system, electronic device and vehicle for intelligent driving static perception

Abstract

The present application relates to the technical field of automatic driving environment perception, and particularly relates to a static perception method and system for intelligent driving, an electronic device and a vehicle, comprising constructing a rasterized space to generate an aerial view grid and initializing an aerial view query; extracting road scene elements based on a semantic segmentation network to generate a scene complexity heat map; obtaining a resolution scaling factor by splicing features of the heat map elements and the initial query, and obtaining a dynamic resolution query set by interpolation processing; training the query set to generate a feature map, assigning different resolution heterogeneous features after grouping, applying channel attention weighting and splitting to obtain multi-scale features; combining the query and the multi-scale features to generate a sampling deviation, extracting and fusing features to obtain an aligned aerial view feature, and outputting a static perception result after analysis. The present application balances the perception accuracy and the consumption of on-board hard resources, avoids frequent adjustment leading to detail loss, solves the problem of missing feature information, and provides support for the landing of intelligent driving safety.

Description

Technical Field

[0001] This invention relates to the field of environmental perception technology for autonomous driving, and in particular to a static perception method, system, electronic device, and vehicle for intelligent driving. Background Technology

[0002] In the field of autonomous driving environmental perception, the construction of bird's-eye view perception models is the core foundation for understanding complex road structures and achieving high-precision positioning and planning. The construction of existing bird's-eye view perception models mainly relies on how to construct bird's-eye view query grids and fuse road element features to generate accurate bird's-eye views. However, in complex scenarios, the fixed-resolution bird's-eye view query grids commonly used in existing technologies are difficult to maintain both detailed representation and computational efficiency when dividing complex road scenarios (such as multi-way intersections and construction enclosure areas). In simple road scenarios (such as straight lanes and lane lines without obstacles), excessive division can lead to computational redundancy. Moreover, the single-scale feature interaction mechanism causes problems such as loss of zebra crossing details and interruption of guide lines, ultimately resulting in insufficient accuracy of environmental perception data.

[0003] Therefore, it is necessary to propose a static perception method, system, electronic device, and vehicle for intelligent driving to solve the above problems. Summary of the Invention

[0004] The purpose of this invention is to provide a static perception method, system, electronic device and vehicle for intelligent driving, to solve the problems that fixed resolution bird's-eye view query grid is difficult to adapt to the dynamic changes of road topology, and simple scenarios will lead to computational redundancy due to query density. Furthermore, the single-scale feature interaction mechanism sever the macro and micro connections of road elements, resulting in the loss of zebra crossing details and the interruption of guide lines.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A static perception method for intelligent driving, the method comprising: Construct a rasterized space to generate a bird's-eye view grid, initialize a bird's-eye view query, and ensure that the bird's-eye view query corresponds one-to-one with the grid cells within the bird's-eye view grid; Based on a pre-trained semantic segmentation network, road scene elements are extracted to obtain a scene complexity heatmap, wherein the elements of the scene complexity heatmap represent the semantic density corresponding to the grid in the bird's-eye view. By using the splicing features of elements in the scene complexity heatmap and the initial bird's-eye view query, a resolution scaling factor is obtained. The bird's-eye view grid is then interpolated using the resolution scaling factor to obtain a query set with dynamic resolution. The query set is trained to obtain a feature map. The feature map is grouped and heterogeneous features of different scales are assigned to obtain a spliced ​​feature group. Channel attention is applied to the spliced ​​feature group to obtain weights. The spliced ​​feature group is then weighted to obtain multiple weighted feature groups of different sizes. The weighted feature groups are then split to obtain multi-scale features. Based on the query set and multi-scale features, a multi-scale sampling deviation is generated, and features are extracted from the multi-scale features through the multi-scale sampling deviation to obtain aligned bird's-eye view features. Static perception results are obtained by extracting and analyzing the features of the bird's-eye view.

[0006] Preferably, the method of extracting road scene elements based on a pre-trained semantic segmentation network to obtain a scene complexity heatmap includes the following steps: Lane line density map and obstacle distribution entropy map are extracted using a semantic segmentation network; The lane line density map and obstacle distribution entropy map are linearly weighted to generate a scene complexity heatmap, which is then output using the semantic segmentation network.

[0007] Preferably, the elements of the scene complexity heatmap represent the semantic density at the corresponding bird's-eye view grid, and the semantic density includes a linear combination of the lane line density map and the obstacle distribution entropy map at the corresponding bird's-eye view grid.

[0008] Preferably, the feature maps are obtained by training the query set, and the feature maps are grouped and assigned heterogeneous features of different scales to obtain concatenated feature groups, including the following steps: The input feature map is divided into multiple groups according to the channel dimension to obtain multiple groups of feature maps; Different scales of heterogeneous features are assigned to the multiple sets of feature maps to obtain spliced ​​feature groups.

[0009] Preferably, the channel dimension includes feature information of each spatial location within the bird's-eye view grid.

[0010] Preferably, heterogeneous features of different scales are assigned to the multiple sets of feature maps, including high resolution, medium resolution and low resolution.

[0011] This invention provides a static perception system for intelligent driving, the system comprising: The building module is used to construct the rasterized space of the bird's-eye view and initialize the bird's-eye view query that corresponds one-to-one with the raster. The analysis module is used to extract static elements of the road scene; The generation module is used to generate a scene complexity heatmap based on the static elements of the road scene in the analysis module. The adjustment module is used to stitch the initial bird's-eye view query with the scene complexity heatmap to obtain a resolution scaling factor. The bird's-eye view grid is then processed using the resolution scaling factor to obtain a query combination with dynamic resolution. The fusion module is used to process feature maps, generate multi-scale heterogeneous features and weight them, and then split the weighted feature groups to obtain multi-scale features. The alignment and fusion module is used to generate multi-scale sampling deviations, extract features, and fuse them to obtain aligned bird's-eye view features; The parsing module is used to analyze the bird's-eye view features from the alignment and fusion module, extract static element information, and output the final static perception result.

[0012] Preferably, the system further includes: The data acquisition module is used to receive raw data from multiple sensor sources in intelligent driving vehicles; The preprocessing module is used to perform format unification, noise filtering, and spatial calibration on the raw data received by the acquisition module.

[0013] The present invention provides an electronic device, comprising: a processor, and a memory communicatively connected to the processor, the memory storing computer-executable instructions, the processor executing the computer-executable instructions stored in the memory to implement the steps of any of the above-described intelligent driving static perception methods.

[0014] The present invention provides a vehicle comprising any of the above-described static perception systems for intelligent driving.

[0015] The technical effects and advantages of the present invention in the above technical solution are as follows: This invention generates a dynamic resolution query set through a scene complexity heatmap. It can allocate high-resolution grids to semantically dense scenes such as multi-way intersections and construction zones to accurately capture lane line intersections and obstacle distribution details, while reducing grid resolution for simple scenes such as straight roads to reduce unnecessary computation. This balances perception accuracy with the consumption of onboard hardware resources and avoids aging caused by frequent unit adjustments. By grouping and allocating heterogeneous features of high, medium and low resolution, combined with channel attention to enhance key features, it completely solves the problem of information loss in traditional single-scale features, improves feature quality, and provides core perception assurance for the safe implementation of intelligent driving. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a static perception method for intelligent driving according to the present invention.

[0017] Figure 2 This is a flowchart of a static perception system for intelligent driving according to the present invention. Detailed Implementation

[0018] 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.

[0019] Example 1: The present invention provides as follows Figure 1 The method shown includes the following steps: Static perception for intelligent driving Step 101: Construct a rasterized space to generate a bird's-eye view grid, initialize the bird's-eye view query, and ensure that the bird's-eye view query corresponds one-to-one with the grid cells within the bird's-eye view grid.

[0020] As a spatial benchmark construction step in the static perception method for intelligent driving, a two-dimensional bird's-eye view coordinate system is first established with the rear axle center of the autonomous vehicle as the coordinate origin, with the x-axis (left and right direction of the vehicle) as the horizontal axis and the y-axis (direction of vehicle travel) as the vertical axis. The physical space range is determined in combination with the perception requirements. Then, the space is evenly divided into bird's-eye view grids with height H × width W according to the preset resolution (such as 0.5 meters / grid, matching the lane line positioning accuracy requirements), forming a precise mapping between the grid and the physical area.

[0021] As a further preferred option, an initial bird's-eye view query set with dimensions H×W×C (where H is the height, W is the width, and C is the channel dimension) is initialized. Each query is both fixedly bound to the physical coordinates of the corresponding grid to anchor its spatial location, and the learnable feature parameters are initialized with a random normal distribution to ensure that each grid corresponds one-to-one with a unique bird's-eye view query, providing a unified spatial coordinate system to support subsequent scene analysis, dynamic resolution adjustment, and multimodal feature aggregation.

[0022] Step 102: Extract road scene elements based on the pre-trained semantic segmentation network to obtain a scene complexity heatmap, where the elements of the scene complexity heatmap represent the semantic density at the corresponding grid in the bird's-eye view.

[0023] The process of extracting road scene elements based on a pre-trained semantic segmentation network to obtain a scene complexity heatmap includes the following steps: Lane line density map and obstacle distribution entropy map are extracted using a semantic segmentation network; The scene complexity heatmap is generated by linearly weighting the lane line density map and the obstacle distribution entropy map, and then output using a semantic segmentation network.

[0024] The elements of the scene complexity heatmap represent the semantic density at the corresponding bird's-eye view grid. The semantic density is a linear combination of the lane line density map and the obstacle distribution entropy map at the corresponding bird's-eye view grid.

[0025] Based on the bird's-eye view mesh with the established grid-physical space mapping constructed in step 101, the multimodal sensor data of the intelligent driving vehicle (such as road images collected by surround-view cameras and point cloud data generated by LiDAR) is adapted to the established bird's-eye view coordinate system and then input into a pre-trained semantic segmentation network (such as the Mask2Former network trained on nuScenes, KITTI public datasets, and real-world road datasets). This network accurately extracts the core static elements in the road scene, including the distribution and density of lane lines, the continuity and material boundaries of road edges, and the spatial distribution of static obstacles. Subsequently, based on these elements, the raster semantic density of each bird's-eye view mesh is calculated. For example, the lane line density is obtained by statistically analyzing the percentage of lane line pixels within a single raster. The obstacle distribution entropy map is obtained by calculating the dispersion of obstacle distribution in the grid and surrounding areas using the Shannon entropy formula. These metrics are then fused together to generate a scene complexity heatmap that corresponds one-to-one with the grid elements in the bird's-eye view. Each element in the heatmap... The numerical value represents the semantic density at the corresponding grid cell. The semantic density includes a linear combination of features such as lane curvature variance and obstacle distribution entropy. The formula for calculating the scene complexity heatmap is: , in, , .

[0026] It is important to understand that the denser the semantic information, the higher the value. For example, the semantic density value of intersection grids is significantly higher than that of open straight road grids because lane lines intersect and signs are concentrated.

[0027] Step 103: Using the splicing features of elements in the scene complexity heatmap and the initial bird's-eye view query, obtain the resolution scaling factor. Use the resolution scaling factor to perform interpolation processing on the bird's-eye view grid to obtain a query set with dynamic resolution.

[0028] Based on the initial bird's-eye view query generated in step 101 and the scene complexity heatmap output in step 102, for each grid cell in the bird's-eye view grid, the corresponding scene complexity heatmap element is compared with the initial bird's-eye view query for that grid cell. The features are dimensionally concatenated, specifically by incorporating semantic density information into the C-dimensional features of the initial bird's-eye view query, forming concatenated features that ensure that the concatenated features of each grid simultaneously contain spatial query information and scene semantic density information.

[0029] The stitching features of all rasters are input into a pre-built gating network, which calculates a unique resolution scaling factor for each raster. , Among them, the higher the semantic density of the raster, The closer to 1.0, the higher the required resolution; the lower the semantic density of the raster, the better. The closer the resolution is to 0.2, the lower the required resolution. Then, using the initial raster resolution set in step 101 (e.g., 0.5 m / raster) as a baseline, a nonlinear interpolation algorithm is used to perform interpolation transformation on the bird's-eye view grid. , ,in, , This is the base grid step size (default 0.5), and (x, y) are the original query coordinates. () represents the query coordinates under adaptive resolution; for larger rasters, the physical size of the raster is reduced after interpolation, for example... When =1.0, maintain 0.5m. When the value is 0.5, it shrinks to 0.25m, thus improving the perception of detail. For smaller grids, the physical size of the grid increases after interpolation, for example... When the value is 0.2, it is expanded to 1.0m to reduce computational redundancy. At the same time, the bird's-eye view query parameters corresponding to each grid are updated synchronously. Finally, a query set with dynamic resolution adapted to the scene complexity is obtained, which provides a spatial query foundation that balances accuracy and efficiency for subsequent multi-scale feature fusion.

[0030] It's important to understand that the gated network is lightweight, with a specific structure of a 2-layer fully connected network. The hidden layer dimension is set to 64, the activation function is ReLU, and the output layer uses the Sigmoid function combined with a linear transformation to constrain the output value to the range [0.2, 1.0].

[0031] As a further preferred approach, obtaining a query set with dynamic resolution will result in denser features in high semantic density areas (such as intersections) and relatively sparser features in low semantic density areas (such as straight roads) in the bird's-eye view feature map. This can easily lead to redundancy, with overlapping features in high-density areas and scattered features in low-density areas. Feature clustering using the K-means algorithm is employed to aggregate features in the bird's-eye view feature map into feature clusters based on semantic attributes (such as lane lines, curbs, and traffic signs) or spatial correlation (such as continuous feature points in the same lane). This makes the expression of similar static features more concentrated. Top-K attention further calculates the association weight between each bird's-eye view query and the center of the feature cluster, selecting the Top-K features that contribute the most to static perception and reducing invalid feature input in subsequent grouping steps.

[0032] Based on improvement The algorithm performs semantically constrained clustering on multi-view image features and optimizes the objective function, which is: , Among them, S ( () represents the probability output of the pre-trained semantic segmentation network. To balance geometric and semantic losses.

[0033] For each bird's-eye view query, calculate its relationship with the cluster center { The association weights are used to filter the first M=3 clusters, and then the process is performed within each cluster. Feature selection: , in, To query the similarity between a bird's-eye view and the k-th feature in the m-th cluster, the computational complexity is from... Down to .

[0034] Step 104: Obtain feature maps by training the query set, group the feature maps and assign heterogeneous features of different scales to obtain spliced ​​feature groups, apply channel attention to the spliced ​​feature groups to obtain weights, and weight the spliced ​​feature groups by weights to obtain multiple weighted feature groups of different sizes. After splitting the weighted feature groups, obtain multi-scale features.

[0035] Based on the query set with dynamic resolution output in step 103, this dynamic resolution query set is fused and trained with the multimodal sensor features of the intelligent driving vehicle. Through a spatial cross-attention mechanism, each dynamic query accurately aggregates the sensor features within the corresponding bird's-eye view grid area, thereby generating a bird's-eye view feature map that matches the dynamic resolution grid. This feature map inherits the scene adaptability of step 103 and contains rich static environmental semantic information. Next, this bird's-eye view feature map is divided into multiple groups according to the channel dimension, and heterogeneous features of different scales are assigned according to the perception function positioning of different groups. Channel attention is applied to each group of spliced ​​feature groups, and the importance information of features within the group is extracted through global average pooling. After compression and activation by a fully connected layer, channel attention weights are generated, and these weights are used to weight the spliced ​​feature groups to strengthen key static features and suppress redundant information, resulting in multiple weighted feature groups of different sizes. Finally, each group of weighted feature groups is split according to the original channel grouping logic to obtain multi-scale features of micro-details, meso-level completion, and macro-topology.

[0036] Feature maps are obtained by training the query set. The feature maps are then grouped and assigned heterogeneous features of different scales to obtain concatenated feature groups, including the following steps: The input feature map is divided into multiple groups according to the channel dimension to obtain multiple groups of feature maps. The channel dimension includes the feature information of each spatial location within the bird's-eye view grid.

[0037] The input feature map is the bird's-eye view feature map generated by training the query set with dynamic resolution output in step 103. Its dimension is the number of rows of dynamic resolution raster × the number of columns of dynamic resolution raster × the number of channels C. The channel dimension encodes the feature information of each raster's spatial location in the bird's-eye view grid (such as lane line edge texture, curb material properties, intersection guide strip topology contour, etc.). During the classification, the channels are first classified according to the feature semantic type carried by the channel dimension. For example, channels that encode pixel-level details within the raster (such as zebra crossing texture, lane line solid and dashed edges) are classified into one category, and channels that encode the relationship structure between rasteres (such as curb continuity, lane connection relationship) are classified into another category. Channels encoding global topological features (such as intersection shape and multi-lane distribution) are grouped into one category. Then, the total number of channels C is evenly divided into N groups (e.g., 4 groups, with C / 4 channels in each group) according to the classification results, ensuring that each group of channels focuses on one type of feature information. At the same time, combined with the dynamic resolution attribute, channel features corresponding to high semantic density rasters are preferentially assigned to the group carrying fine-grained information, while channel features corresponding to low semantic density rasters are assigned to the group carrying macroscopic topological information. Finally, N groups of feature maps with clear feature emphasis and even number of channels are obtained, and the spatial dimension of each group of feature maps is completely matched with the dynamic resolution rasters of the input bird's-eye view feature map.

[0038] Heterogeneous features of different scales are assigned to multiple sets of feature maps to obtain spliced ​​feature groups, with scales including high resolution, medium resolution and low resolution.

[0039] High-resolution heterogeneous features are downsampled at 1 / 8 and use a 3×3 standard convolutional kernel (for extracting fine-grained static features, such as lane line sub-pixel offset and traffic sign edge details). Medium-resolution heterogeneous features are downsampled at 1 / 16 and use a dilated convolutional kernel with a dilation rate of 3 (for completing discontinuous static features, such as broken curbs and obscured zebra crossings). Low-resolution heterogeneous features are downsampled at 1 / 32 and use a 7×7 large convolutional kernel (for parsing macroscopic static topology, such as the shape of intersection guide strips and multi-lane merging relationships). During allocation, precise matching is performed based on the feature emphasis of multiple feature maps. For example, high-resolution heterogeneous features are allocated to feature groups that carry pixel-level details within the grid, medium-resolution heterogeneous features are allocated to feature groups that carry inter-grid relational structural features, and low-resolution heterogeneous features are allocated to feature groups that carry global topological features. After allocation, each feature map is fused with its matching heterogeneous features through channel-dimensional concatenation to obtain a concatenated feature group.

[0040] Step 105: Based on the query set and multi-scale features, generate multi-scale sampling deviation, and extract features from the multi-scale features through multi-scale sampling deviation to obtain aligned bird's-eye view features.

[0041] Establish a spatial mapping relationship between bird's-eye view query and multi-scale features. Specifically, take the raster physical coordinates corresponding to the dynamic resolution bird's-eye view query as the benchmark, and calculate the projection position of the coordinates in the high, medium and low resolution feature maps respectively. Since the resolution of feature maps at different scales is different, the projection position needs to be converted by the scale scaling factor. For example, the scaling factor of high resolution feature map is 1 / 8, medium resolution is 1 / 16 and low resolution is 1 / 32.

[0042] For each bird's-eye view query, the offset of the projection position on the feature map at each scale is calculated based on the scene complexity of its corresponding raster. The sampling deviation of high semantic density raster corresponding to high resolution features is smaller, while the sampling deviation of low semantic density raster corresponding to low resolution features can be appropriately increased, forming a sampling deviation set that corresponds one-to-one with multi-scale features. Through a deformable attention mechanism, features are extracted at the projection offset position of the feature map at the corresponding scale based on the sampling deviation at each scale. The features extracted at each scale are spatially aligned and fused. Using the physical coordinates of the raster in the bird's-eye view query as a unified anchor point, the spatial misalignment of features at different scales is eliminated through feature weighted fusion. Finally, aligned bird's-eye view features that accurately match the dynamic resolution bird's-eye view query and have strong spatial consistency are obtained.

[0043] Step 106: Extract and analyze the features of the bird's-eye view to obtain the static perception results.

[0044] Example 2: The present invention provides as follows Figure 2 The illustrated static perception system for intelligent driving includes: The building module is used to construct the rasterized space of the bird's-eye view and initialize the bird's-eye view query that corresponds one-to-one with the raster.

[0045] The analysis module is used to extract static elements of the road scene.

[0046] The generation module is used to generate a scene complexity heatmap based on the static elements of the road scene in the analysis module.

[0047] The adjustment module is used to stitch the initial bird's-eye view query with the scene complexity heatmap to obtain the resolution scaling factor, and then process the bird's-eye view grid with the resolution scaling factor to obtain a query set with dynamic resolution.

[0048] The fusion module is used to process feature maps, generate multi-scale heterogeneous features and weight them, and then split the weighted feature groups to obtain multi-scale features.

[0049] The alignment and fusion module is used to generate multi-scale sampling deviations, extract features, and fuse them to obtain aligned bird's-eye view features.

[0050] The parsing module is used to analyze the bird's-eye view features from the alignment and fusion module, extract static element information, and output the final static perception result.

[0051] The adjustment module calculates the resolution scaling factor by fusing the semantic density of the scene complexity heatmap with the initial bird's-eye view query. For complex scenes with dense semantics, such as multiple intersections and construction barriers, the grid scaling factor is made close to 1.0 to ensure that detailed information such as lane line intersections and static obstacle distribution is not lost. For simple scenes such as straight lanes and obstacle-free areas, the grid scaling factor is reduced to about 0.2, which significantly reduces the computational redundancy caused by invalid grid division and avoids the waste of onboard hardware resources.

[0052] Specifically, the fusion module groups the bird's-eye view feature maps generated by the dynamic resolution query set according to the channel dimension, and selectively allocates heterogeneous features of high resolution (1 / 8 downsampling rate, 3×3 convolution kernel, focusing on fine-grained details such as lane line sub-pixel offset), medium resolution (1 / 16 downsampling rate, dilated convolution, to complete discontinuous features such as broken road edges), and low resolution (1 / 32 downsampling rate, 7×7 convolution kernel, to analyze macroscopic information such as intersection guide strip topology). This effectively avoids the problems of zebra crossing detail loss and guide line interruption caused by traditional single-scale features, and achieves full-dimensional feature coverage of micro-details, meso-level completion, and macro-level topology.

[0053] The system also includes an acquisition module and a preprocessing module. The acquisition module is used to receive raw data from multiple sources of sensors in the intelligent driving vehicle, and the preprocessing module is used to perform format unification, noise filtering and spatial calibration on the raw data received by the acquisition module.

[0054] The collected data includes self-built datasets, simulation data, and publicly available data. The self-built datasets include multi-source heterogeneous data from over 500,000 vehicles, accumulating over 5,000 kilometers of real-world road data from cities in the Yangtze River Delta region. This data covers extreme conditions such as rain, fog, backlight, and nighttime, and high-precision annotation (annotation error <5cm) is achieved using semi-automatic annotation tools (such as CVAT). The simulation data includes 2,000 virtual scenes generated by the CARLA engine, simulating real-world challenging conditions such as glare from icy and snowy roads, sudden changes in light and dark at tunnel entrances, and worn road markings. Domain randomization (lighting, weather, and viewing angle) is used to improve the model's generalization ability. The publicly available data includes 1,000 scene data segments covering complex urban roads and multi-lane highways, covering six lighting conditions including sunny, rainy, and nighttime. Static elements such as lane lines, curbs, and guide strips are highlighted. In addition, 500 intersection, roundabout, and construction zone scenes were collected, and high-precision LiDAR point clouds (accuracy ±2cm) were used to generate ground truth values ​​of lane line topology from a bird's-eye view.

[0055] Data processing requires data labeling and standardization. For lane lines / curves, geometric attributes (curvature, slope, heading angle) and topological connectivity are defined according to the OpenDrive standard, with lateral labeling error ≤5cm. For drivable areas, centimeter-level grid maps are generated based on LiDAR point clouds and RTK trajectories to distinguish standard lanes, emergency lanes, shoulders, and other areas.

[0056] Furthermore, the consistency of annotations is verified by randomly selecting 10% of the annotation data and cross-validating it by three independent annotators. The Kappa coefficient must be ≥0.95; otherwise, re-annotation is initiated.

[0057] When assessing coverage under extreme conditions, ensure that the proportion of data from rain, fog, backlight, and nighttime in the test set is ≥30%, and the proportion of long-tail scenarios (construction areas, irregular intersections) is ≥15%.

[0058] During annotation, a combination of semi-automatic annotation and ground truth fusion is selected. Semi-automatic annotation uses pre-trained models (such as LaneNet and PointPillars) to generate initial annotations, which are then manually verified (using the annotation platform CVAT) to correct topological errors. Ground truth fusion uses LiDAR point cloud projection and multi-frame RTK trajectory optimization to solve the scale ambiguity problem of monocular visual annotation and ensure the consistency of annotations under the bird's-eye view.

[0059] Preferably, the data also needs to be divided into a training set (70%), a validation set (15%), and a test set (15%). The training set includes publicly available data and self-collected data, covering 80% of common scenarios. The validation set includes self-collected data, focusing on complex structures and common extreme conditions. The test set includes self-collected data and simulation data, specifically for evaluating long-tail scenarios and extreme conditions.

[0060] The data is preprocessed, including image distortion correction, spatiotemporal synchronization, and bird's-eye view ground truth generation. Image distortion correction is based on camera intrinsic parameter calibration results and uses bilinear interpolation to ensure geometric alignment of multi-view images. Spatiotemporal synchronization is achieved by aligning multiple camera frames (deviation <1ms) through hardware trigger signals and interpolating and fusing GNSS / IMU data to generate 6DoF pose for each frame. Bird's-eye view ground truth generation is achieved by projecting LiDAR point clouds and manually annotated vector maps onto a bird's-eye view grid (resolution 0.1m) to generate dense semantic raster and instance mask.

[0061] Example 3: Based on Example 1, this example provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program runs, it implements the specific steps 101-106 in Example 1.

[0062] The electronic device provided in this embodiment has a hardware architecture with a processor as the computing core and a memory as the data and program storage carrier. When the computer program pre-stored in the memory runs on the processor, it can realize the static perception method for intelligent driving of the present invention. At the same time, the memory can synchronously store intermediate data and final results in the static perception process. The processor can transmit the static perception results to the vehicle's decision module in real time according to the requirements of the intelligent driving system. If remote monitoring or collaborative work is required, data synchronization with a remote terminal can also be achieved through the vehicle communication interface. It supports remote viewing of static perception data and intermediate processing, and assists multiple people in collaboratively debugging static perception algorithm parameters. This ensures that the electronic device can not only efficiently execute the entire static perception process in intelligent driving scenarios, but also improve the practicality and intelligence level of static perception through data interconnection.

[0063] Example 4: Based on Example 2, the present invention provides a vehicle that includes the intelligent driving static perception system of Example 2.

[0064] The vehicle provided in Example 4 integrates the intelligent driving static perception system of Example 2 into its hardware architecture. It establishes a data transmission channel with the system's acquisition module through its onboard multi-source sensors (such as surround-view cameras, LiDAR, and millimeter-wave radar), transmitting raw road environment data to the acquisition module in real time. The vehicle's onboard computing unit (such as an onboard chip) serves as the system's hardware carrier, supporting the operation of the construction module, analysis module, generation module, adjustment module, fusion module, alignment and fusion module, and parsing module. It transmits the static perception results to the vehicle's decision-making and planning module in real time, providing precise static environment awareness support for lane keeping, path selection, and other driving control functions. Simultaneously, the vehicle can synchronize data between the static perception system and a remote terminal via the onboard communication module, facilitating remote monitoring of the static perception status and further enhancing the safety and reliability of intelligent driving.

[0065] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A static perception method for intelligent driving, characterized in that, The method includes: Construct a rasterized space to generate a bird's-eye view grid, initialize a bird's-eye view query, and ensure that the bird's-eye view query corresponds one-to-one with the grid cells within the bird's-eye view grid; Based on a pre-trained semantic segmentation network, road scene elements are extracted to obtain a scene complexity heatmap, wherein the elements of the scene complexity heatmap represent the semantic density corresponding to the grid in the bird's-eye view. By using the splicing features of elements in the scene complexity heatmap and the initial bird's-eye view query, a resolution scaling factor is obtained. The bird's-eye view grid is then interpolated using the resolution scaling factor to obtain a query set with dynamic resolution. The query set is trained to obtain a feature map. The feature map is grouped and heterogeneous features of different scales are assigned to obtain a spliced ​​feature group. Channel attention is applied to the spliced ​​feature group to obtain weights. The spliced ​​feature group is then weighted to obtain multiple weighted feature groups of different sizes. The weighted feature groups are then split to obtain multi-scale features. Based on the query set and multi-scale features, a multi-scale sampling deviation is generated, and features are extracted from the multi-scale features through the multi-scale sampling deviation to obtain aligned bird's-eye view features. Static perception results are obtained by extracting and analyzing the features of the bird's-eye view.

2. The static perception method for intelligent driving according to claim 1, characterized in that, The process of extracting road scene elements based on a pre-trained semantic segmentation network to obtain a scene complexity heatmap includes the following steps: Lane line density map and obstacle distribution entropy map are extracted using a semantic segmentation network; The lane line density map and obstacle distribution entropy map are linearly weighted to generate a scene complexity heatmap, which is then output using the semantic segmentation network.

3. The static perception method for intelligent driving according to claim 1, characterized in that, The elements of the scene complexity heatmap represent the semantic density at the corresponding bird's-eye view grid. The semantic density includes a linear combination of the lane line density map and the obstacle distribution entropy map at the corresponding bird's-eye view grid.

4. The static perception method for intelligent driving according to claim 1, characterized in that, The feature maps are obtained by training the query set. The feature maps are then grouped and assigned heterogeneous features of different scales to obtain concatenated feature groups, including the following steps: The input feature map is divided into multiple groups according to the channel dimension to obtain multiple groups of feature maps; Different scales of heterogeneous features are assigned to the multiple sets of feature maps to obtain spliced ​​feature groups.

5. A static perception method for intelligent driving according to claim 4, characterized in that: The channel dimension includes feature information of each spatial location within the bird's-eye view grid.

6. A static perception method for intelligent driving according to claim 4, characterized in that: The multiple sets of feature maps are assigned heterogeneous features of different scales, including high resolution, medium resolution and low resolution.

7. A static perception system for intelligent driving, characterized in that, The system includes: The building module is used to construct the rasterized space of the bird's-eye view and initialize the bird's-eye view query that corresponds one-to-one with the raster. The analysis module is used to extract static elements of the road scene; The generation module is used to generate a scene complexity heatmap based on the static elements of the road scene in the analysis module. The adjustment module is used to stitch the initial bird's-eye view query with the scene complexity heatmap to obtain a resolution scaling factor. The bird's-eye view grid is then processed using the resolution scaling factor to obtain a query combination with dynamic resolution. The fusion module is used to process feature maps, generate multi-scale heterogeneous features and weight them, and then split the weighted feature groups to obtain multi-scale features. The alignment and fusion module is used to generate multi-scale sampling deviations, extract features, and fuse them to obtain aligned bird's-eye view features; The parsing module is used to analyze the bird's-eye view features from the alignment and fusion module, extract static element information, and output the final static perception result.

8. A static perception system for intelligent driving according to claim 7, characterized in that, The system also includes: The data acquisition module is used to receive raw data from multiple sensor sources in intelligent driving vehicles; The preprocessing module is used to perform format unification, noise filtering, and spatial calibration on the raw data received by the acquisition module.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor, the memory storing computer-executable instructions, the processor executing the computer-executable instructions stored in the memory to implement the method as described in any one of claims 1-6.

10. A vehicle, characterized in that, The vehicle includes a static perception system for intelligent driving as described in any one of claims 7-8.

Images