Methods and apparatus for extracting map elements

Abstract

This application provides a method and apparatus for extracting map elements. The method includes: acquiring a point cloud top-down view of a target area; performing feature extraction on the point cloud top-down view based on a preset multi-task network model to obtain an image feature map, and performing semantic segmentation processing on the image feature map to obtain a semantic segmentation result; performing vector feature extraction processing on the image feature map in the row and column directions respectively based on the preset multi-task network model to obtain the row vector point sequence of row vector targets and the column vector point sequence of column vector targets; and fusing the row vector point sequence of the row vector targets and the column vector point sequence of the column vector targets according to the semantic segmentation result to obtain the vector point sequence of map elements in the target area. This method aims to improve the efficiency, accuracy, and completeness of map element extraction.

Description

Technical Field

[0001] This application relates to the field of map technology, and in particular to a method and apparatus for extracting map elements. Background Technology

[0002] Lane lines, pedestrian crossings, road arrows, and other map elements are key components of a map. Accurately and completely extracting map elements within a target area can provide basic data support for vehicle positioning, route planning, driving decisions, and map updates, which is of great significance for improving the reliability and accuracy of autonomous driving systems and navigation services.

[0003] In related technologies, map feature extraction methods typically rely on manual annotation or subsequent fitting processing based on semantic segmentation results. Manual annotation requires manually drawing and editing map features point by point, resulting in low efficiency, high cost, and difficulty in ensuring consistency. Semantic segmentation-based extraction methods usually involve first classifying the image at the pixel level to obtain segmentation results, and then fitting these results through post-processing to obtain map features. This approach is prone to issues such as broken linear features and large fitting errors for irregularly shaped features, leading to insufficient completeness and accuracy in map feature extraction. Summary of the Invention

[0004] The map feature extraction method and apparatus provided in this application are intended to improve the efficiency, accuracy and completeness of map feature extraction.

[0005] In a first aspect, embodiments of this application provide a method for extracting map features, including:

[0006] Obtain a top-down view of the point cloud of the target area;

[0007] Based on a preset multi-task network model, feature extraction is performed on the top view of the point cloud to obtain an image feature map, and semantic segmentation processing is performed on the image feature map to obtain a semantic segmentation result.

[0008] Based on a preset multi-task network model, vector feature extraction processing is performed on the image feature map in the row direction and column direction respectively to obtain the row vector point order of the row vector target and the column vector point order of the column vector target;

[0009] Based on the semantic segmentation results, the row vector point sequence of the row vector target and the column vector point sequence of the column vector target are fused to obtain the vector point sequence of map features in the target area.

[0010] In one possible implementation, the step of performing vector feature extraction processing on the image feature map in the row and column directions based on a preset multi-task network model to obtain the row vector point order of the row vector target and the column vector point order of the column vector target includes:

[0011] Based on a preset multi-task network model, the image feature map is cropped and stacked according to preset row anchor boxes to obtain a row vector feature map; and the image feature map is cropped and stacked according to preset column anchor boxes to obtain a column vector feature map.

[0012] The row vector feature map is processed by vector feature extraction to obtain the row vector point order of the row vector target; and the column vector feature map is processed by vector feature extraction to obtain the column vector point order of the column vector target.

[0013] In one possible implementation, the step of performing vector feature extraction processing on the row vector feature map to obtain the row vector point order of the row vector target, and the step of performing vector feature extraction processing on the column vector feature map to obtain the column vector point order of the column vector target, include:

[0014] The vector feature map is classified using a pre-defined classification head network to determine the confidence level; wherein, the confidence level represents the probability that a vector target exists in each anchor frame; the vector feature map is a row vector feature map or a column vector feature map, and the vector target is a row vector target or a column vector target;

[0015] The vector feature map is processed by a preset recognition head network to determine the initial vector point order of the vector target within each anchor frame; wherein, the initial vector point order is either the initial row vector point order or the initial column vector point order.

[0016] Based on the confidence level and the initial vector point order of the vector target, the vector point order of the vector target is obtained; wherein, the vector point order of the vector target is the row vector point order of the row vector target or the column vector point order of the column vector target.

[0017] In one possible implementation, the step of performing recognition processing on the vector feature map through a preset recognition head network to determine the initial vector point order of the vector target within each anchor frame includes:

[0018] The recognition head network performs recognition processing on the vector feature map to obtain the score of each preset candidate position within each preset interval of each anchor frame in the vector feature map;

[0019] For each preset interval, the coordinate information of the vector target within the preset interval is determined based on the position index corresponding to the maximum score among the preset candidate positions within that preset interval.

[0020] The coordinate information of vector targets within each preset interval is sequentially spliced ​​to obtain the initial vector point sequence of vector targets within each anchor frame.

[0021] In one possible implementation, obtaining the vector point order of the vector target based on the confidence level and the initial vector point order of the vector target includes:

[0022] The initial vector point sequence of the vector target corresponding to the anchor frame with a confidence level greater than a preset threshold is sequentially spliced ​​to obtain the vector point sequence of the vector target.

[0023] In one possible implementation, the step of fusing the row vector point order of the row vector target and the column vector point order of the column vector target based on the semantic segmentation result to obtain the vector point order of map features in the target area includes:

[0024] Based on the semantic segmentation results and the row vector point order of the row vector target, the feature attributes of the row vector target are determined; and based on the semantic segmentation results and the column vector point order of the column vector target, the feature attributes of the column vector target are determined.

[0025] Based on the feature attributes of the row vector targets and the feature attributes of the column vector targets, the row vector point sequence of each row vector target and the column vector point sequence of each column vector target are fused to obtain the vector point sequence of map features in the target area.

[0026] In one possible implementation, the step of fusing the row vector point sequence of each row vector target and the column vector point sequence of each column vector target based on the feature attributes of the row vector targets and the feature attributes of the column vector targets to obtain the vector point sequence of map features in the target area includes:

[0027] For a pair of row vector targets and column vector targets, if it is determined that the feature attributes of the row vector target and the feature attributes of the column vector target are consistent, and the row vector point sequence of the row vector target and the column vector point sequence of the column vector target meet the preset conditions, then the row vector point sequence and the column vector point sequence are concatenated to obtain the vector point sequence of the map feature corresponding to the feature attribute.

[0028] In one possible implementation, the method further includes:

[0029] Based on the heading angle and coordinate transformation matrix of the point cloud top view, the vector point sequence of each map element is transformed to obtain the transformed vector point sequence of the map element.

[0030] A vector map is generated based on the vector point sequence of the converted map features; wherein the vector map is used to update the map.

[0031] Secondly, embodiments of this application provide a map feature extraction device, comprising:

[0032] The acquisition module is used to acquire a top-down view of the point cloud of the target area;

[0033] The first processing module is used to extract features from the point cloud top view based on a preset multi-task network model to obtain an image feature map, and to perform semantic segmentation processing on the image feature map to obtain a semantic segmentation result.

[0034] The second processing module is used to perform vector feature extraction processing on the image feature map in the row direction and column direction respectively based on a preset multi-task network model to obtain the row vector point order of the row vector target and the column vector point order of the column vector target.

[0035] The fusion module is used to fuse the row vector point sequence of the row vector target and the column vector point sequence of the column vector target according to the semantic segmentation result, so as to obtain the vector point sequence of map elements in the target area.

[0036] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0037] The memory stores computer-executed instructions;

[0038] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0039] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0040] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0041] The map feature extraction method and apparatus provided in this application first acquire a point cloud top view of the target area, then extract features from the point cloud top view using a preset multi-task network model to obtain an image feature map, and perform semantic segmentation processing on the image feature map to obtain a semantic segmentation result. Next, perform vector feature extraction processing on the image feature map in the row and column directions using the multi-task network model to obtain the row vector point sequence of the row vector target and the column vector point sequence of the column vector target. Finally, combine the semantic segmentation result to perform fusion processing on the row vector point sequence and the column vector point sequence to obtain the vector point sequence of map features in the target area. This method relies on a multi-task network model to achieve end-to-end processing from point cloud top view to map feature attributes and vector point sequence, eliminating the traditional manual delineation and annotation, semantic segmentation and geometric fitting steps. This effectively reduces the manual cost and processing cycle of map feature extraction and improves processing efficiency. At the same time, through the extraction and fusion of vector features in both row and column directions, it avoids the problems of large fitting errors of irregular map features and linear feature breakage, significantly improving the completeness, accuracy and automation of map feature extraction. It can support the direct generation of linear and polygonal features, better meeting the real-time and high-precision requirements of high-precision maps in scenarios such as autonomous driving, vehicle-road cooperation and digital twin cities. Attached Figure Description

[0042] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0043] Figure 1 A flowchart illustrating the map feature extraction method provided in this application. Figure 1 ;

[0044] Figure 2 A flowchart illustrating the map feature extraction method provided in this application. Figure 2 ;

[0045] Figure 3 A schematic diagram illustrating one application scenario provided in this application;

[0046] Figure 4 A schematic diagram of the structure of the map feature extraction device provided in this application;

[0047] Figure 5 A schematic diagram of the structure of the electronic device provided in this application.

[0048] The accompanying drawings have illustrated specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0049] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0050] Applications such as autonomous driving, vehicle-road cooperation, and digital twin cities place stringent requirements on the real-time performance and accuracy of high-precision maps, exceeding 99%. However, traditional high-precision map production still heavily relies on manual processes. Data collection vehicles equipped with high-end LiDAR and inertial navigation can acquire more than 1TB of point cloud data daily. Subsequently, office staff need to use specialized software to draw map elements such as lane lines, stop lines, zebra crossings, and curbs frame by frame and assign corresponding attributes. A single person can only complete the vectorization of elements for about 5 kilometers of urban roads per day, with labor costs accounting for more than 55%, which has become a key bottleneck restricting the rapid updating of map data.

[0051] In recent years, although semi-automatic extraction schemes based on semantic segmentation combined with post-processing have emerged, for non-standard shape elements such as curved lane lines and irregular pedestrian crossings, the edge detection and fitting method will produce large errors due to the fragmented process of segmentation followed by fitting. Furthermore, the fitting algorithm relies on manually designed rules, which is difficult to adapt to diverse element shapes. At the same time, polygonal elements such as complex intersections and irregular green belts cannot be directly generated and require manual verification to complete the point sequence or adjust the fitting parameters, resulting in interruptions in the end-to-end processing flow and long mapping cycles. Mapping a single intersection often takes several hours, further increasing production costs. The overall extraction efficiency and automation level cannot meet the needs of practical applications.

[0052] In summary, the efficiency and accuracy of map feature extraction methods in related technologies are relatively low. Therefore, to address these issues, the inventors propose a solution using a multi-task end-to-end deep learning network with a dual-branch structure. By setting semantic segmentation and vector feature extraction branches, the deep learning network directly and simultaneously completes feature category discrimination and vector point order generation from a point cloud intensity top-down view, skipping the edge detection and geometric fitting steps in the traditional extraction process, and achieving direct output from input data to attributed vector point orders. However, the global coordinate regression task carries the risk of error accumulation. To address this, the inventors further introduce an anchor box processing mechanism, dividing the image feature map into row and column anchor box regions, decomposing the global task into local classification tasks in the row and column directions, reducing prediction errors, and thereby improving the efficiency, completeness, and accuracy of map feature extraction.

[0053] The execution subject of this application embodiment can be an electronic device with processing capabilities, such as a computer, server, etc., and this application embodiment is not limited thereto.

[0054] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0055] Figure 1 A flowchart illustrating the map feature extraction method provided in this application. Figure 1 ,like Figure 1 As shown, the method includes:

[0056] S101. Obtain a top-down view of the point cloud of the target area.

[0057] For example, the target area refers to the geographic spatial area from which map features are to be extracted. This application embodiment does not limit the scope of the target area.

[0058] A point cloud top view is a two-dimensional top view image formed by projecting the collected point cloud data of the target area onto a ground coordinate system.

[0059] In one example, a target area can be scanned using LiDAR to output point cloud data. This point cloud data can include 3D coordinates, reflection intensity, elevation information, and distance information. Then, the point cloud data is projected onto a ground coordinate system to generate a point cloud intensity top-down view. The vertical axis represents the vehicle's travel direction, the horizontal axis is perpendicular to the travel direction, the pixel value is the point cloud reflection intensity, and the heading angle and the coordinate transformation matrix between pixels and latitude / longitude are recorded. The heading angle of the point cloud top-down view characterizes the geographic orientation offset of the view; it is the angle Tm between the travel direction and the latitude-decreasing direction. The travel direction refers to the actual direction the onboard device collecting the point cloud data is traveling within the target area, and the latitude-decreasing direction refers to the direction in which the latitude value gradually decreases in geographic space. The coordinate transformation matrix Hi between pixels and latitude / longitude is used to convert the pixel coordinates of the point cloud top-down view into geographic latitude / longitude coordinates; it can also be called the scaling factor.

[0060] S102. Based on the preset multi-task network model, feature extraction is performed on the top view of the point cloud to obtain an image feature map, and semantic segmentation processing is performed on the image feature map to obtain the semantic segmentation result.

[0061] For example, a multi-task network model refers to a deep learning network model that can perform multiple data processing tasks simultaneously. This application does not limit the type of multi-task network model, but it can be a convolutional neural network model.

[0062] Feature extraction refers to the processing operation of extracting spatial texture and geometric structure information from a point cloud top-view. An image feature map is data containing deep features such as spatial texture and geometric structure information from the point cloud top-view, obtained after feature extraction.

[0063] Semantic segmentation refers to the pixel-level classification of image feature maps. The semantic segmentation result is pixel-level classification data labeled with map feature category attributes.

[0064] In one example, the multi-task network model may include a feature extraction network and a semantic segmentation head network. It should be noted that this application does not limit the types of the feature extraction network and the semantic segmentation head network. For example, the feature extraction network can be a Residual Network-50 (ResNet-50) or a Vision Transformer (ViT), while the semantic segmentation head network can be, for example, a Fully Convolutional Network (FCN), a Short-Term Dense Concatenation Network (STDC), or a Convolutional Network for Biomedical Image Segmentation (U-Net), etc., and can be specifically configured according to actual needs.

[0065] For example, for feature extraction networks, when the computing power of electronic devices is limited, Mobile Neural Network version 3 (MobileNetV3) can be chosen instead of ResNet-50; if higher accuracy is required, the ViT network can be chosen. For semantic segmentation head networks, if faster inference speed is needed, a Deep Convolutional Encoder-Decoder Architecture for Image Segmentation (SegNet) network can be chosen instead of U-Net; if higher segmentation accuracy is required, the DeepLabv3+ network can be chosen.

[0066] Furthermore, the point cloud top view of the target area can be input into the feature extraction network of the preset multi-task network model to extract features from the point cloud top view and obtain the corresponding image feature map. Then, the semantic segmentation head network of the multi-task network model performs pixel-level semantic recognition on the image feature map to obtain the semantic segmentation result labeled with map element category attributes.

[0067] S103. Based on the preset multi-task network model, vector feature extraction processing is performed on the image feature map in the row direction and column direction respectively to obtain the row vector point order of the row vector target and the column vector point order of the column vector target.

[0068] For example, the row direction refers to the horizontal extension direction of the image feature map, and the column direction refers to the vertical extension direction of the image feature map. Vector feature extraction processing refers to the process of extracting an ordered set of coordinate points from the image feature map.

[0069] Vector targets refer to candidate map features with continuous geometric shapes, capable of representing spatial location and point sequence structure, but do not contain feature category attributes. Row vector targets are existing vector targets extracted along the row direction; the row vector point sequence refers to the ordered set of coordinate points corresponding to the row vector target. Column vector targets are existing vector targets extracted along the column direction; the column vector point sequence refers to the ordered set of coordinate points corresponding to the column vector target.

[0070] In one example, the multi-task network model can also include two parallel vector feature extraction networks. These networks can extract the vector point sequence of vector targets in the image feature map along the row and column directions. For example, these vector feature extraction networks can be fully convolutional networks. Furthermore, the image feature map can be input into the two vector feature extraction networks for vector feature extraction processing, yielding the row vector point sequence of row vector targets and the column vector point sequence of column vector targets.

[0071] In another example, the multi-task network model can also include two parallel vector feature extraction networks. Taking the row-direction vector feature extraction network as an example, this network can divide the image feature map into anchor boxes along the row direction to obtain row vector feature maps. A classification head network determines the confidence level of the presence of vector targets in the row vector feature maps, and a recognition head network obtains the initial vector point order of the vector targets. Based on the confidence level, the initial vector point order is filtered and concatenated in sequence to obtain the row vector point order of the vector targets. The column-direction vector feature extraction network is similar and will not be described further here.

[0072] This approach employs an architecture combining an end-to-end deep learning network (multi-task network model) with semantic segmentation and vector feature point coordinate prediction and classification. It can directly learn the mapping relationship between the top view of point cloud intensity and the vector point sequence and attributes of arbitrary shapes / polygonal elements, skipping the edge detection and fitting steps of traditional processes, and achieving high-precision mapping without human intervention.

[0073] S104. Based on the semantic segmentation results, the row vector point sequence of the row vector target and the column vector point sequence of the column vector target are fused to obtain the vector point sequence of map features in the target area.

[0074] For example, fusion processing refers to the processing operations of attribute matching, spatial splicing and deduplication filtering of row vector point order and column vector point order.

[0075] Map elements refer to geographic entities in high-definition maps, such as lane lines, stop lines, zebra crossings, and curbs. The vector point sequence of map elements refers to the complete and ordered set of coordinate points corresponding to various map elements within the target area.

[0076] In one example, as mentioned above, the multi-task network model includes a two-branch structure of semantic segmentation and vector feature extraction. Semantic features provide type attribute information, and vector feature extraction provides point order information. It outputs the semantic segmentation result, the row vector point order of the row vector target, and the column vector point order of the column vector target. Then, the electronic device can assign corresponding feature attributes to the row vector point order and column vector point order according to the semantic segmentation result, and stitch together similar features with the same feature attributes to obtain the vector point order of each map feature in the target area.

[0077] Understandably, traditional methods, which use manually designed rules to fit missing segments, struggle to match the true geometric shapes of irregular features. They can only use fixed patterns for approximate fitting, inevitably leading to significant fitting errors. However, single-direction vector feature extraction can only extract stable point sequences for map features that tend towards that direction. For linear features perpendicular to that direction or tilted at a large angle, as well as irregular map features such as curves and irregular polygons, discontinuous feature capture and missing key points can easily occur, directly resulting in broken linear features. By employing a parallel extraction method in both row and column directions, row-direction vector feature extraction can reliably capture the complete point sequence of horizontally extending and gently curving map elements, while column-direction vector feature extraction can accurately capture the complete point sequence of vertically extending and steeply curving map elements. The two methods complement each other in terms of geometric features, comprehensively covering map element features of any angle and shape, thus avoiding the feature loss problem caused by extraction in a single direction from the source. On this basis, through fusion processing, attribute matching and spatial stitching are performed on the point sequences extracted in both directions, which can complete the same map element fragments extracted in different directions into a continuous and complete vector point sequence. The entire process does not require relying on manually designed fitting rules to simulate and correct the shape of the elements, and can directly restore the true geometric contour of irregular map elements. Therefore, it can effectively avoid the problems of large fitting errors for irregular map elements and the breakage of linear elements.

[0078] Optionally, the vector point sequence of map elements can be applied to real-time updates of high-precision maps for autonomous vehicles, digital construction of urban road networks, and digitalization of intersection and lane elements in Intelligent Transportation Systems (ITS). In urban road network digitalization scenarios, it can efficiently support the generation of road network vector maps for digital twin cities. In intelligent transportation systems, it can realize rapid digital acquisition and modeling of intersection and lane elements. At the same time, it can provide real-time, accurate, and stable vector map data for autonomous vehicles, meeting the practical needs of dynamic updates and continuous iteration of high-precision maps.

[0079] The map feature extraction method provided in this application first obtains a point cloud top view of the target area, then extracts features from the point cloud top view using a preset multi-task network model to obtain an image feature map, and performs semantic segmentation processing on the image feature map to obtain a semantic segmentation result. Next, the multi-task network model performs vector feature extraction processing on the image feature map in the row and column directions respectively to obtain the row vector point sequence of the row vector target and the column vector point sequence of the column vector target. Finally, the row vector point sequence and the column vector point sequence are fused together with the semantic segmentation result to obtain the vector point sequence of map features in the target area. This method relies on a multi-task network model to achieve end-to-end processing from point cloud top view to map feature attributes and vector point sequence, eliminating the traditional manual delineation and annotation, semantic segmentation and geometric fitting steps. This effectively reduces the manual cost and processing cycle of map feature extraction and improves processing efficiency. At the same time, through the extraction and fusion of vector features in both row and column directions, it avoids the problems of large fitting errors of irregular map features and linear feature breakage, significantly improving the completeness, accuracy and automation of map feature extraction. It can support the direct generation of linear and polygonal features, better meeting the real-time and high-precision requirements of high-precision maps in scenarios such as autonomous driving, vehicle-road cooperation and digital twin cities.

[0080] Figure 2 A flowchart illustrating the map feature extraction method provided in this application. Figure 2 , Figure 3 A schematic diagram illustrating an application scenario provided in this application, such as... Figure 2 and Figure 3 As shown, in this embodiment... Figure 1 Based on the examples, the method for extracting map features is described in detail, which includes:

[0081] S201. Obtain a top-down view of the point cloud of the target area.

[0082] It should be noted that this step is similar to the aforementioned step S101, and will not be repeated here.

[0083] S202. Based on the preset multi-task network model, feature extraction is performed on the top view of the point cloud to obtain an image feature map, and semantic segmentation processing is performed on the image feature map to obtain the semantic segmentation result.

[0084] For example, the ViT network in the multi-task network model is used as a feature extraction network to process the image and output an image feature map Fi (imgFeatMap). Then, the STDC network in the multi-task network model is used to process the image feature map and output a semantic segmentation result, which can be represented as a segmentation label mask Rslm (segLabelMask).

[0085] S203. Based on the preset multi-task network model, the image feature map is cropped and stacked according to the preset row anchor boxes to obtain the row vector feature map; and the image feature map is cropped and stacked according to the preset column anchor boxes to obtain the column vector feature map.

[0086] For example, a row anchor box refers to a reference box set along the horizontal direction of the image feature map for extracting vector features in the row direction. Cropping refers to the operation of cropping the image feature map according to the anchor box range, and stacking refers to the operation of stacking multiple cropped local feature maps dimensionally to form a new feature map. A row vector feature map refers to the feature data obtained after cropping and stacking row anchor boxes for extracting vector features in the row direction. A column anchor box refers to a reference box set along the width direction of the image feature map for extracting vector features in the column direction; a column vector feature map refers to the feature data obtained after cropping and stacking column anchor boxes for extracting vector features in the column direction.

[0087] In one example, each vector feature extraction network in the multi-task network model includes an Anchor network. This network can set multiple vertically arranged row anchor boxes according to preset sizes and strides. Taking 16 row anchor boxes as an example, each row anchor box extracts a local feature map with a height of N / 8 and a width of M. These 16 extracted local feature maps are then stacked to form a row vector feature map with dimensions M×(N / 8)×16×C. Similarly, multiple horizontally arranged column anchor boxes can be set to extract and stack image feature maps in the width direction to obtain column vector feature maps. It should be noted that the number of Anchors is not limited in this embodiment; it can be set according to actual needs. For example, it can be set to 16. If the detail requirements of the geographic features are high (such as narrow lane lines), the number of Anchors can be increased or doubled, for example, changing 16 anchors to 32.

[0088] For example, for an image feature map with width M, height N, and number of channels C, 16 row anchor boxes are preset. 16 row labels are selected in the height dimension with a step size of N / 16. Starting from each row label, a local region of N / 8 is cropped in the height dimension, resulting in 16 local feature maps. If the height of the cropped region is less than N / 8, zeros are added to the end of the height direction to reach N / 8. The 16 local feature maps are then stacked according to the dimensions of the row anchor boxes to obtain a row vector feature map Frv (rowVectorFeatMap) with width M, height N, number of anchor boxes 16, and number of channels C, respectively.

[0089] Similarly, for an image feature map with width M, height N, and number of channels C, 16 column anchor boxes are preset. 16 column labels are obtained in the width dimension with a step size of M / 16. Starting from each column label, a local region of M / 8 is truncated in the width dimension to obtain 16 local feature maps. For the part of the truncated region that is less than M / 8, zeros are added to the end. The 16 local feature maps are stacked according to the dimensions of the column anchor boxes to obtain a column vector feature map Fcv (colVectorFeatMap) with width, height, number of anchor boxes, and number of channels of M / 8, N, 16, and C, respectively.

[0090] S204. Perform vector feature extraction processing on the row vector feature map to obtain the row vector point order of the row vector target; and perform vector feature extraction processing on the column vector feature map to obtain the column vector point order of the column vector target.

[0091] For example, vector feature extraction processing refers to the operation of extracting an ordered set of coordinate points from a vector feature map.

[0092] In one example, each vector feature extraction network in the multi-task network model also includes a parallel classification head (clsHead) and a recognition head (recHead). The classification head network detects each row and column vector feature map in parallel to obtain the confidence score of the presence of a vector target. At the same time, the recognition head network outputs the initial vector point sequence of the corresponding vector target in parallel. The initial vector point sequences with confidence scores greater than a preset threshold are concatenated in order to obtain the row vector point sequence of the row vector target and the column vector point sequence of the column vector target, respectively.

[0093] Specifically, step S204 may include the following steps:

[0094] S2041. The vector feature map is classified using a pre-defined classification head network to determine the confidence level.

[0095] Here, the vector feature map is either a row vector feature map or a column vector feature map, and the vector target is either a row vector target or a column vector target. In other words, steps S2041 to S2043 can be performed on the row vector feature map to obtain the row vector point sequence of the row vector target; or, the column vector feature map can be performed to obtain the column vector point sequence of the column vector target. The principle of both is the same.

[0096] For example, a classification head network refers to a neural network module in a multi-task network model used to perform binary classification tasks. Classification processing refers to the operation performed by the classification head network on the local features corresponding to each anchor box in the vector feature map, including feature mapping and probability calculation. Confidence represents the probability that a vector target exists within each anchor box in the vector feature map; that is, it is a numerical value that quantifies the probability of a vector target existing in each anchor box. Its value typically ranges from 0 to 1, with a value closer to 1 indicating a higher probability of the presence of a vector target.

[0097] For example, taking row vector feature maps as an example, the row vector feature maps are input into a pre-defined classification head network. The classification head network uses convolutional layers, fully connected layers, and activation functions to independently extract features and calculate probabilities for the local features corresponding to each anchor box dimension of the vector feature map. It outputs a confidence value for each anchor box, thereby determining whether a row vector target exists within each anchor box of the row vector feature map. Taking 16 row anchor boxes as an example, the classification head network can output a 1×16 feature vector Frac(anchorClsRes). This feature vector represents whether a vector element exists within the 16 anchor box intervals, and each value in the feature vector represents the confidence score of each anchor box.

[0098] S2042. The vector feature map is processed by a preset recognition head network to determine the initial vector point sequence of the vector target in each anchor frame.

[0099] The initial vector point order is either the initial row vector point order or the initial column vector point order.

[0100] For example, the recognition head network refers to the neural network module in a multi-task network model used to extract the spatial location information of vector targets. Recognition processing refers to the operations performed by the recognition head network on each anchor box feature in the vector feature map, including interval-by-interval analysis, score calculation, and position index determination. The initial vector point sequence refers to the sequence of coordinate points of the vector targets directly output by the recognition head network, without undergoing confidence level filtering.

[0101] For example, taking a row vector feature map as an example, the row vector feature map is input into a preset recognition head network. The recognition head network extracts the local features of each row anchor box in the row vector feature map and outputs the initial row vector point sequence of the vector target in each row anchor box.

[0102] Specifically, the recognition head network is used to process the row vector feature map to obtain the score of each preset candidate position in each preset interval of each row anchor box of the row vector feature map; for each preset interval, the coordinate information of the row vector target in the preset interval is determined according to the position index corresponding to the maximum score of the preset candidate position in the preset interval; the coordinate information of the row vector targets in each preset interval is sequentially concatenated to obtain the initial row vector point sequence of the row vector targets in each row anchor box.

[0103] Here, the preset interval refers to the equidistant pixel intervals pre-divided along the width of each row anchor frame, with each preset interval corresponding to a single pixel column or a group of consecutive pixel columns in the horizontal direction of the row anchor frame. Preset candidate positions refer to the pixel positions in the height direction of each preset interval where a row vector target may exist. The score is the quantified probability value of the recognition head network for the existence of a row vector target at each preset candidate position within each preset interval. The position index is the sequence number of the preset candidate position within the preset interval. The coordinate information refers to the pixel coordinates of the row vector target in the image coordinate system, obtained based on the position index transformation. The initial row vector point sequence is the sequence of coordinate points of the row vector target formed by splicing the coordinate information of all preset intervals under each row anchor frame in spatial order.

[0104] For example, when a row vector feature map of dimension M×N / 8×16×C is input into the recognition head network, the network first processes each of the 16 row anchor boxes independently: for the feature slice (of dimension M×N / 8×C) corresponding to a single row anchor box, it is divided into M equal-width preset intervals along the width dimension (each preset interval corresponds to 1 pixel column in the width direction and covers the entire N / 8 in the height direction), and each preset interval contains N / 8 preset candidate positions arranged along the height dimension; the recognition head network processes each preset interval through convolutional layers and fully connected layers. The scores of the N / 8 preset candidate positions within the single row anchor box are calculated to obtain the score matrix of M preset intervals × N / 8 preset candidate positions. Then, the matrix is ​​integrated and output as a three-dimensional matrix with dimensions M×(N / 8)×16, where dimension M corresponds to each preset interval, dimension N / 8 corresponds to each preset candidate position within each preset interval, and dimension 16 corresponds to the 16 preset row anchor boxes. Each feature value in this three-dimensional score matrix is ​​the score of the row vector target in the preset interval and corresponding preset candidate position of each row anchor box predicted by the recognition head network.

[0105] For each preset interval, the maximum value index (argmax) operation is performed along the N / 8 dimensions of the preset candidate position based on the three-dimensional matrix. The maximum value is selected from the scores of the N / 8 preset candidate positions, and the position index of the preset candidate position corresponding to the maximum value is determined, resulting in an output matrix Frap (anchorPtsRes) with a dimension of M×16. The element value at coordinate (i,j) in the output matrix Frap represents the position index of the row vector target in the i-th preset interval and the j-th row anchor box, that is, the relative coordinate value of the row vector point sequence in the current anchor box.

[0106] Based on the resolution of the image feature map (Fi), and combined with the step size of N / 16 in the height dimension of the row anchor box, the relative coordinate value is converted into the absolute coordinate information of the row vector target in the image feature map. The absolute coordinate is (i, Frap(i,j)+j×(N / 16)), which is the coordinate information of the row vector target in the i-th preset interval. The coordinate information of the row vector targets corresponding to all preset intervals is sequentially spliced ​​according to the spatial arrangement order of the preset intervals to finally obtain the initial row vector point sequence of the complete row vector target.

[0107] Combining the pixel horizontal coordinate i (i ranges from 0 to M-1) in the width dimension of the preset interval, the position index idx, and the starting row index j×(N / 16) of the row anchor frame in the height dimension of the image feature map (j is the index of the row anchor frame, ranging from 0 to 15), the coordinate information of the row vector target within the preset interval is converted to (i, idx+j×(N / 16)). Finally, according to the spatial order from left to right of the preset interval (horizontal coordinate i ranges from 0 to M-1), the coordinate information of the M preset intervals under the row anchor frame is sequentially concatenated to form the initial row vector point sequence of the row vector target within the row anchor frame. Repeat the above operation to complete the recognition processing of 16 row anchor frames, and finally obtain the initial row vector point sequence corresponding to each row anchor frame.

[0108] For the column vector feature map, the recognition head network is used to recognize and process the column vector feature map to obtain the score of each preset candidate position in each preset interval of each column anchor frame of the column vector feature map; for each preset interval, the coordinate information of the column vector target in the preset interval is determined according to the position index corresponding to the maximum score of the preset candidate position in the preset interval; the coordinate information of the column vector targets in each preset interval is sequentially concatenated to obtain the initial column vector point sequence of the column vector targets in each column anchor frame.

[0109] For example, a recognition head network is used to process the column vector feature map Fcv. The recognition head network outputs a three-dimensional score matrix with dimensions (M / 8)×N×16, where dimension M / 8 corresponds to each preset candidate position within a preset interval, dimension N corresponds to each preset interval, and dimension 16 corresponds to 16 preset column anchor boxes. Each feature value in this three-dimensional score matrix is ​​the score predicted by the recognition head network for the presence of a column vector target within the corresponding preset interval and at the corresponding preset candidate position. For this three-dimensional score matrix, the maximum value index (argmax) operation is performed along the M / 8 dimension where the preset candidate positions are located to filter out the position index corresponding to the preset candidate position with the highest score under each preset interval and each column anchor box. Based on this, an output matrix Fcap (an) with dimension N×16 is obtained. The element value at coordinate (i,j) in the output matrix Fcap represents the position index of the column vector target within the i-th preset interval and the j-th column anchor box, i.e., the relative coordinate value of the column vector point sequence within the current anchor box. Based on the resolution of the image feature map (Fi) and combined with the step size M / 16 in the width dimension of the column anchor box, the relative coordinate value is converted into the absolute coordinate information of the column vector target in the image feature map. The absolute coordinate is (Fcap(i,j)+j×(M / 16),i), which is the coordinate information of the column vector target within the i-th preset interval. The coordinate information of the column vector targets corresponding to all preset intervals is sequentially concatenated according to the spatial arrangement order of the preset intervals to finally obtain the initial column vector point sequence of the column vector targets within each column anchor box.

[0110] This method processes image features by row and column, breaking them down into smaller parts. It transforms the global coordinate regression task into a local coordinate classification task through anchor boxes. Furthermore, processing anchor boxes in both directions separately allows for more comprehensive acquisition of feature information of linear elements in the horizontal and vertical directions, which helps improve the positional accuracy of vector points.

[0111] S2043. Based on the confidence level and the initial vector point sequence of the vector target, obtain the vector point sequence of the vector target.

[0112] Among them, the vector point order of the vector target is the row vector point order of the row vector target or the column vector point order of the column vector target.

[0113] For example, the core of this step is to filter the initial vector point sequence based on confidence level. The vector point sequence of the vector target refers to the row vector point sequence or column vector point sequence that has been filtered and is considered valid.

[0114] Specifically, the initial vector point sequence of the vector targets corresponding to anchor boxes with confidence scores greater than a preset threshold is sequentially concatenated to obtain the vector point sequence of the vector targets. For example, a confidence score threshold thr is preset, and the confidence score output by the classification head network is compared with this threshold. Only the initial vector point sequence corresponding to anchor boxes with confidence scores greater than the threshold is retained. Then, according to the spatial arrangement order of the row anchor boxes or column anchor boxes, the retained initial row vector point sequence or initial column vector point sequence is sequentially concatenated to finally obtain the row vector point sequence of the row vector targets and the column vector point sequence of the column vector targets.

[0115] It should be noted that the reference Figure 3 As shown, for any vector feature extraction network, taking row vector feature extraction as an example, the row vector feature map can be input into the classification head network and the recognition head network in parallel for processing. At the same time, column vector feature extraction and row vector feature extraction are also processed in parallel.

[0116] S205. Based on the semantic segmentation results and the row vector point sequence of the row vector target, determine the feature attributes of the row vector target; and based on the semantic segmentation results and the column vector point sequence of the column vector target, determine the feature attributes of the column vector target.

[0117] For example, feature attributes refer to the map feature category identifier corresponding to the vector target, which may include category information such as lane lines, stop lines, zebra crossings, and curb strips.

[0118] In one example, based on the coordinate positions of the row and column vector points, the pixel category at the corresponding location is queried in the semantic segmentation results. The retrieved category information is then used as the feature attributes of the corresponding row and column vector targets. Furthermore, the feature attributes and row vector point order of the row vector targets are integrated to obtain the row vector information Rrvi (rowVectorInfo); correspondingly, the feature attributes and column vector point order of the column vector targets are integrated to obtain the column vector information Rcvi (colVectorInfo).

[0119] S206. Based on the feature attributes of the row vector targets and the feature attributes of the column vector targets, the row vector point sequence of each row vector target and the column vector point sequence of each column vector target are fused to obtain the vector point sequence of map features in the target area.

[0120] For example, fusion processing refers to the operation of attribute matching, spatial splicing, and deduplication filtering of row vector point order and column vector point order.

[0121] In one example, row and column vector point sequences with consistent feature attributes and spatial locations that meet preset proximity or continuity conditions are selected. The selected point sequences are then spliced ​​together and duplicate fragments are removed to obtain the vector point sequence of each map feature in the target area, which is the vectorized result of the polygon map features.

[0122] Specifically, for a pair of row vector targets and column vector targets, if it is determined that the feature attributes of the row vector target and the feature attributes of the column vector target are consistent, and the row vector point sequence of the row vector target and the column vector point sequence of the column vector target meet the preset conditions, then the row vector point sequence and the column vector point sequence are spliced ​​to obtain the vector point sequence of the map feature corresponding to the feature attribute.

[0123] In this context, a pair of row and column vector targets refers to row and column vector targets that correspond spatially and are located within the same geographical area of ​​the target region. Consistent element attributes mean that the map element category identifiers corresponding to the row and column vector targets are exactly the same. Stitching processing refers to the operation of integrating coordinates, removing redundant points, and spatially reordering the two sets of vector point sequences. Preset conditions refer to any one of the following: the line spacing between the row and column vector point sequences in the image coordinate system is less than a preset threshold; the coordinate points of the row and column vector point sequences are continuous; or there is local overlap between the row and column vector point sequences.

[0124] For example, by traversing all row and column vector targets, a set of row and column vector targets corresponding to the spatial location is selected, and the element attributes of the two are read. If the two are the same map element category such as lane lines or zebra crossings, then the coordinate positions of the row and column vector points are further judged. If the coordinate positions meet the preset conditions, the two vector targets are considered to be the same map element. Then the coordinate points in the row and column vector points are merged, the duplicate and redundant coordinate points are removed, and the points are reordered according to the actual arrangement order in the image space. After the stitching process is completed, the complete and continuous vector point sequence of the map element corresponding to the element attribute can be obtained.

[0125] Optionally, if two element attributes are identical (e.g., both are lane lines), the row and column vector point sequences are compared using coordinates to calculate the linear similarity and line spacing between the two point sequences. If the linear similarity is greater than a preset similarity threshold (determined as linear overlap), or the line spacing is less than a preset line spacing threshold, then these two lane lines are determined to be the same actual lane line within the target area. Subsequently, the scores corresponding to these two lane lines (i.e., the confidence scores output by the classification head network from their respective vector feature maps) are extracted. The two scores are compared with a preset confidence threshold, filtering out low-confidence lane lines with scores below the threshold, and retaining only the vector point sequences of high-confidence lane lines with scores higher than or equal to the threshold as the valid vector point sequences of the actual lane line, thus avoiding lane line vector point sequence splicing deviations caused by low-confidence data.

[0126] S207. Based on the heading angle and coordinate transformation matrix of the point cloud top view, perform coordinate transformation on the vector point sequence of each map element to obtain the transformed vector point sequence of the map elements.

[0127] For example, the heading angle refers to the orientation angle of the acquisition device corresponding to the top view of the point cloud, the coordinate transformation matrix refers to the parameter matrix used to realize the mutual conversion between image coordinates and geospatial coordinates, and the coordinate transformation processing refers to the operation of converting the vector point sequence from the image coordinate system to the geographic coordinate system.

[0128] In one example, the vector point sequence of map features in the image coordinate system is calculated by combining the heading angle of the point cloud top view and the preset coordinate transformation matrix, and then converted into a vector point sequence in the geospatial coordinate system.

[0129] S208. Generate a vector map based on the vector point sequence of the converted map elements.

[0130] Vector graphics are used to update maps.

[0131] For example, a vector map refers to map feature graphic data composed of vector point sequences under geographic spatial coordinates. Map updating refers to the operation of synchronizing newly generated vector maps to a high-definition map to replace or supplement the original map data.

[0132] In one example, the corresponding vector graphics are constructed according to the converted map feature vector point order, all vector graphics are integrated to generate a vector image, and the vector image is written into the high-precision map database to complete the map data update.

[0133] The map feature extraction method provided in this application first obtains a point cloud top view of the target area, then extracts features from the point cloud top view using a preset multi-task network model to obtain an image feature map and simultaneously performs semantic segmentation processing to obtain a semantic segmentation result. Next, the multi-task network model uses row and column anchor boxes to truncate and stack the image feature map to obtain row vector feature maps and column vector feature maps. Then, vector feature extraction processing is performed on the two types of vector feature maps to obtain the row vector point sequence of the row vector target and the column vector point sequence of the column vector target. Subsequently, the semantic segmentation result is combined to determine the corresponding feature attributes for the row vector target and the column vector target. Then, the row vector point sequence and the column vector point sequence are fused according to the feature attributes to obtain the vector point sequence of the map feature. After that, the coordinate transformation of the vector point sequence is performed using the heading angle and coordinate transformation matrix. Finally, a vector map is generated based on the transformed vector point sequence and the map is updated. This method, relying on a multi-task network model, achieves fully automated processing from point cloud top-down views to geospatial vector maps. By extracting and fusing vector features in both row and column directions, it achieves complementary geometric features, avoiding the linear feature breakage problem caused by single-direction extraction from the source. The entire process eliminates the need for manual sketching and rule-based fitting operations, effectively eliminating fitting errors of irregular map features. At the same time, it significantly reduces the manual cost and processing cycle of high-precision map production, improves the completeness, accuracy, and automation level of map feature extraction, and can better adapt to the real-time and high-precision usage requirements of high-precision maps in scenarios such as autonomous driving, vehicle-road cooperation, and digital twin cities.

[0134] In summary, for reference Figure 3 As shown, this embodiment of the application first generates a top-down view of the point cloud using point cloud data as initial input. This top-down view is then input into the Image Feature Extraction Network (VIT) of the multi-task network model for feature extraction, resulting in an image feature map (imgFeatMap). Based on this image feature map, three parallel processing steps are performed simultaneously:

[0135] First, semantic segmentation is performed through the Semantic Segmentation Head Network (STDC), and the semantic segmentation result seglabelMask is output.

[0136] Secondly, the anchor box processing in the row direction generates a row vector feature map rowVectorFeatMap. Then, the anchor box classification result anchorClsRes (confidence) is output by the classification head network clsHead, and the anchor box point order result anchorPtsRes (initial vector point order of the row vector target) is output by the recognition head network recHead. Finally, the row vector point order and attributes are merged by combining the semantic segmentation result seglabelMask to obtain the row vector information rowVectorInfo.

[0137] Third, the anchor box processing in the column direction generates the column vector feature map colVectorFeatMap. Similarly, the anchor box classification result anchorClsRes (confidence) and the anchor box point order result anchorPtsRes (initial vector point order of the column vector target) are output by the classification head network clsHead and the recognition head network recHead, respectively. Then, the column vector point order and attributes are merged by combining the semantic segmentation result seglabelMask to obtain the column vector information colVectorInfo.

[0138] Then, a merge operation is performed on the row vector information (rowVectorInfo) and column vector information (colVectorInfo) to filter out row and column vector point sequences with consistent feature attributes and spatial locations that meet preset proximity or continuity conditions. The filtered point sequences are then spliced ​​and duplicate fragments are removed to obtain the vector point sequence (vectorInfo) of each map feature in the target area. Then, combined with the heading angle and coordinate transformation matrix, the coordinate transformation of the vector point sequence (vectorInfo) of each map feature is performed. Finally, a vector map is output based on the transformed vector point sequence of the map features for map updating.

[0139] Figure 4 A schematic diagram of the structure of the map feature extraction device provided in this application is shown below. Figure 4 As shown, the map feature extraction device 300 provided in this embodiment includes:

[0140] Module 301 is used to acquire a top-down view of the point cloud of the target area;

[0141] The first processing module 302 is used to extract features from the top view of the point cloud based on a preset multi-task network model to obtain an image feature map, and to perform semantic segmentation processing on the image feature map to obtain a semantic segmentation result.

[0142] The second processing module 303 is used to perform vector feature extraction processing on the image feature map in the row direction and column direction respectively based on a preset multi-task network model to obtain the row vector point order of the row vector target and the column vector point order of the column vector target.

[0143] The fusion module 304 is used to fuse the row vector point sequence of the row vector target and the column vector point sequence of the column vector target according to the semantic segmentation result, so as to obtain the vector point sequence of map features in the target area.

[0144] In one possible implementation, the second processing module 303 is configured to:

[0145] Based on a pre-defined multi-task network model, the image feature map is cropped and stacked according to the pre-defined row anchor boxes to obtain a row vector feature map; and the image feature map is cropped and stacked according to the pre-defined column anchor boxes to obtain a column vector feature map.

[0146] Vector feature extraction is performed on the row vector feature map to obtain the row vector point order of the row vector target; and vector feature extraction is performed on the column vector feature map to obtain the column vector point order of the column vector target.

[0147] In one possible implementation, the second processing module 303 is configured to:

[0148] The vector feature map is classified using a pre-defined classification head network to determine the confidence level; the confidence level represents the probability that a vector target exists in each anchor box; the vector feature map is either a row vector feature map or a column vector feature map, and the vector target is either a row vector target or a column vector target;

[0149] The vector feature map is processed by a pre-set recognition head network to determine the initial vector point order of the vector target within each anchor frame; wherein the initial vector point order is either the initial row vector point order or the initial column vector point order.

[0150] Based on the confidence level and the initial vector point order of the vector target, the vector point order of the vector target is obtained; wherein, the vector point order of the vector target is either the row vector point order of the row vector target or the column vector point order of the column vector target.

[0151] In one possible implementation, the second processing module 303 is configured to:

[0152] The vector feature map is processed by the recognition head network to obtain the score of each preset candidate position in each preset interval of each anchor box in the vector feature map;

[0153] For each preset interval, the coordinate information of the vector target within the preset interval is determined based on the position index corresponding to the maximum score among the preset candidate positions within that preset interval.

[0154] The coordinate information of vector targets within each preset interval is sequentially spliced ​​to obtain the initial vector point sequence of vector targets within each anchor frame.

[0155] In one possible implementation, the second processing module 303 is configured to:

[0156] The initial vector point sequence of the vector targets corresponding to the anchor frames with confidence levels greater than a preset threshold is sequentially concatenated to obtain the vector point sequence of the vector targets.

[0157] In one possible implementation, the fusion module 304 is used for:

[0158] Based on the semantic segmentation results and the row vector point order of the row vector targets, determine the feature attributes of the row vector targets; and based on the semantic segmentation results and the column vector point order of the column vector targets, determine the feature attributes of the column vector targets.

[0159] Based on the feature attributes of row vector targets and column vector targets, the row vector point sequence of each row vector target and the column vector point sequence of each column vector target are fused to obtain the vector point sequence of map features in the target area.

[0160] In one possible implementation, the fusion module 304 is used for:

[0161] For a pair of row vector targets and column vector targets, if it is determined that the feature attributes of the row vector target and the feature attributes of the column vector target are consistent, and the row vector point sequence of the row vector target and the column vector point sequence of the column vector target meet the preset conditions, then the row vector point sequence and the column vector point sequence are concatenated to obtain the vector point sequence of the map feature corresponding to the feature attribute.

[0162] In one possible implementation, the device further includes a generation module for:

[0163] Based on the heading angle and coordinate transformation matrix of the point cloud top view, the vector point sequence of each map element is transformed to obtain the transformed vector point sequence of the map element.

[0164] A vector map is generated based on the vector point sequence of the converted map features; the vector map is used to update the map.

[0165] The map feature extraction device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0166] Figure 5 A schematic diagram of the structure of the electronic device provided in this application. Figure 5 As shown, the electronic device 400 provided in this embodiment includes at least one processor 401 and a memory 402. Optionally, the electronic device 400 further includes a communication component 403. The processor 401, memory 402, and communication component 403 are connected via a bus.

[0167] In a specific implementation, at least one processor 401 executes computer execution instructions stored in memory 402, causing at least one processor 401 to perform the above-described method.

[0168] The specific implementation process of processor 401 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0169] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0170] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0171] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0172] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0173] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0174] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0175] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0176] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0177] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0178] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0179] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0180] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0181] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for extracting map features, characterized in that, include: Obtain a top-down view of the point cloud of the target area; Based on a preset multi-task network model, feature extraction is performed on the top view of the point cloud to obtain an image feature map, and semantic segmentation processing is performed on the image feature map to obtain a semantic segmentation result. Based on a preset multi-task network model, vector feature extraction processing is performed on the image feature map in the row direction and column direction respectively to obtain the row vector point order of the row vector target and the column vector point order of the column vector target; Based on the semantic segmentation results, the row vector point sequence of the row vector target and the column vector point sequence of the column vector target are fused to obtain the vector point sequence of map features in the target area.

2. The method according to claim 1, characterized in that, The preset multi-task network model performs vector feature extraction processing on the image feature map in the row and column directions respectively, to obtain the row vector point order of the row vector target and the column vector point order of the column vector target, including: Based on a preset multi-task network model, the image feature map is cropped and stacked according to preset row anchor boxes to obtain a row vector feature map; and the image feature map is cropped and stacked according to preset column anchor boxes to obtain a column vector feature map. The row vector feature map is processed by vector feature extraction to obtain the row vector point order of the row vector target; and the column vector feature map is processed by vector feature extraction to obtain the column vector point order of the column vector target.

3. The method according to claim 2, characterized in that, The process involves performing vector feature extraction on the row vector feature map to obtain the row vector point order of the row vector target, and performing vector feature extraction on the column vector feature map to obtain the column vector point order of the column vector target, including: The vector feature map is classified using a pre-defined classification head network to determine the confidence level; wherein, the confidence level represents the probability that a vector target exists in each anchor frame; the vector feature map is a row vector feature map or a column vector feature map, and the vector target is a row vector target or a column vector target; The vector feature map is processed by a preset recognition head network to determine the initial vector point order of the vector target within each anchor frame; wherein, the initial vector point order is either the initial row vector point order or the initial column vector point order. Based on the confidence level and the initial vector point order of the vector target, the vector point order of the vector target is obtained; wherein, the vector point order of the vector target is the row vector point order of the row vector target or the column vector point order of the column vector target.

4. The method according to claim 3, characterized in that, The step of performing recognition processing on the vector feature map through a preset recognition head network to determine the initial vector point order of the vector target within each anchor frame includes: The recognition head network performs recognition processing on the vector feature map to obtain the score of each preset candidate position within each preset interval of each anchor frame in the vector feature map; For each preset interval, the coordinate information of the vector target within the preset interval is determined based on the position index corresponding to the maximum score among the preset candidate positions within that preset interval. The coordinate information of vector targets within each preset interval is sequentially spliced ​​to obtain the initial vector point sequence of vector targets within each anchor frame.

5. The method according to claim 3, characterized in that, The step of obtaining the vector point order of the vector target based on the confidence level and the initial vector point order of the vector target includes: The initial vector point sequence of the vector target corresponding to the anchor frame with a confidence level greater than a preset threshold is sequentially spliced ​​to obtain the vector point sequence of the vector target.

6. The method according to claim 1, characterized in that, The step of fusing the row vector point order of the row vector target and the column vector point order of the column vector target based on the semantic segmentation result to obtain the vector point order of map features in the target area includes: Based on the semantic segmentation results and the row vector point order of the row vector target, the feature attributes of the row vector target are determined; and based on the semantic segmentation results and the column vector point order of the column vector target, the feature attributes of the column vector target are determined. Based on the feature attributes of the row vector targets and the feature attributes of the column vector targets, the row vector point sequence of each row vector target and the column vector point sequence of each column vector target are fused to obtain the vector point sequence of map features in the target area.

7. The method according to claim 6, characterized in that, The step of fusing the row vector point sequence of each row vector target and the column vector point sequence of each column vector target based on the feature attributes of the row vector targets and the feature attributes of the column vector targets to obtain the vector point sequence of map features in the target area includes: For a pair of row vector targets and column vector targets, if it is determined that the feature attributes of the row vector target and the feature attributes of the column vector target are consistent, and the row vector point sequence of the row vector target and the column vector point sequence of the column vector target meet the preset conditions, then the row vector point sequence and the column vector point sequence are concatenated to obtain the vector point sequence of the map feature corresponding to the feature attribute.

8. The method according to any one of claims 1-7, characterized in that, The method further includes: Based on the heading angle and coordinate transformation matrix of the point cloud top view, the vector point sequence of each map element is transformed to obtain the transformed vector point sequence of the map element. A vector map is generated based on the vector point sequence of the converted map features; wherein the vector map is used to update the map.

9. A device for extracting map elements, characterized in that, include: The acquisition module is used to acquire a top-down view of the point cloud of the target area; The first processing module is used to extract features from the point cloud top view based on a preset multi-task network model to obtain an image feature map, and to perform semantic segmentation processing on the image feature map to obtain a semantic segmentation result. The second processing module is used to perform vector feature extraction processing on the image feature map in the row direction and column direction respectively based on a preset multi-task network model to obtain the row vector point order of the row vector target and the column vector point order of the column vector target. The fusion module is used to fuse the row vector point sequence of the row vector target and the column vector point sequence of the column vector target according to the semantic segmentation result, so as to obtain the vector point sequence of map elements in the target area.

10. An electronic device / computer-readable storage medium / computer program product, characterized in that, The electronic device includes: a memory and a processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-8; The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-8; The computer program product includes a computer program that, when executed by a processor, is used to implement the method as described in any one of claims 1-8.

Images