A high-definition map construction method based on positioning point query and attention mechanism

Abstract

The present application relates to a kind of high-precision map construction method based on positioning point query and attention mechanism, belong to the field of vehicle-mounted laser point cloud detection.Based on vehicle-mounted laser point cloud data, through vehicle driving trajectory line extraction segmented road surface point cloud, combined with feature pyramid network structure to construct road surface point cloud feature map.For different types of lane elements (long solid line type, short solid line type and non-solid line type), different positioning methods are constructed, the positioning loss of lane elements at the point level and object level is obtained and optimized, the extraction of lane elements is realized, and a high-precision map is generated.The present application constructs a kind of end-to-end generation high-precision map depth model, realizes the automatic positioning, identification and vectorization of multiple types and multiple scales of lane elements in a unified framework, overcomes the influence of different stages on target extraction in traditional multilevel extraction, and improves the precision and efficiency of high-precision map construction.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of vehicle-mounted laser point cloud detection, and particularly relates to a high-precision map construction method based on positioning point query and attention mechanism. BACKGROUND

[0002] A high-precision map is an important basis for realizing unmanned driving, contains rich lane element semantic information and spatial position information, and guarantees safe and orderly traffic behaviors of vehicles and pedestrians. A traditional high-definition map offline construction method needs a large amount of manual labeling and updating cost. Therefore, an automatic generation mode of a high-precision map has become a research mainstream at present, but a map construction still needs to be performed through a multi-stage or multi-task process at present stage, so that the precision of the generated high-precision map is impaired and the efficiency is low.

[0003] At present, some researches mainly perform a high-precision map construction task through image data, trajectory data or point cloud data, perform lane element extraction and expression through spatial view transformation of the image and combination of image processing technology or a deep learning method, but the high-precision map expression is relatively single due to the limitation of the image quality and the element type. A center line and a boundary line for expressing lane traffic are fitted through clustering analysis of the trajectory data, but the trajectory data is affected by a large amount of noise and needs to be analyzed in combination with road prior information in different regional scenes. A scheme based on point cloud data extracts lane elements from road scene point cloud and realizes high-precision map expression in combination with post-processing, but there is a dependency between stage tasks, and error accumulation is easy to occur.

[0004] Vehicle-mounted laser point cloud data can better reflect three-dimensional spatial information and surface reflection intensity information of lane elements, and is more suitable for a high-precision map construction task than image and trajectory data. However, due to the influence of factors such as point cloud data occlusion, wear and tear and missing, the form and texture characteristics of lane elements are impaired, and the difference in the expression form of different lane elements is a difficult problem of high-precision map construction based on vehicle-mounted laser point cloud data at present stage.

[0005] A patent with patent number CN117671143A and the invention name of "a three-dimensional map element extraction method, system, device and medium" encodes and extracts depth features through spatial view conversion of input image data, and realizes vectorization expression of linear map elements through combination of a Transformer attention mechanism model, but the method is limited by the form difference of element objects, and only large-scale lane lines are constructed in lane element of the map.

[0006] The patent with the patent number CN116452852A and the title "A high-precision vector map automatic generation method" realizes the extraction of lane element pixel points by filtering, geometric edge detection, threshold segmentation and other operations on the point cloud image obtained by projection, and then maps it back to the point cloud for lane element segmentation. Combined with post-processing, the vectorized lane element expression is realized, which guarantees the morphological structure of the lane element. However, this method realizes the extraction of lane elements, the prediction of categories and the output of vectorization through multiple tasks, which accumulates errors and is limited by the setting of artificial threshold.

[0007] The patent with the patent number CN114580574A and the title "A method and device for constructing a high-precision map" clusters the trajectory points based on the collected trajectory data to obtain trajectory point clusters. According to the cluster centers of the trajectory point clusters, the initial center line is determined and is smoothed by a filtering algorithm. By combining the preset road width, the boundary lines on both sides of the road center line are generated, thereby constructing a high-precision map. This method realizes the automatic generation of a high-precision map through trajectory data, but the attribute information of different road sections needs to be preset, and the lane lines constructed by the trajectory have a certain position deviation.

[0008] The patent with the patent number CN113920217A and the title "Method, device, equipment and product for generating high-precision map lane line" divides the road surface point cloud into road surface grids, generates the corresponding grid equation by performing plane fitting constraint, edge constraint and plane smoothing constraint on the point cloud in the grid, and generates the lane line based on the grid equation corresponding to each road surface grid. This method improves the accuracy of the generated high-precision map lane line. However, the presence of worn and missing lane elements will result in omission or poor extraction effect, and the continuity between lane lines is poor.

[0009] The patent with the patent number CN118031986A and the title "A method, device, equipment and storage medium for automatically generating high-precision map lane marking" extracts lane element object point cloud according to the intensity difference of road surface point cloud, generates lane feature line through spatial polynomial fitting, and generates lane marking attribute information according to the geometric relationship of the lane feature line. This method generates the geometric and attribute information of the high-precision map lane element, but the multi-stage process leads to the accumulation of errors in the map construction process, and the extraction accuracy of the lane element in the point cloud is relatively high. SUMMARY

[0010] The purpose of the present application is to provide a high-precision map construction method based on positioning point query and attention mechanism, which realizes the end-to-end extraction of vehicle-mounted laser point cloud lane elements and the construction of high-precision map.

[0011] To achieve the above object, the technical scheme of the present application is: a high-precision map construction method based on positioning point query and attention mechanism, comprising the following steps:

[0012] Step A, based on the vehicle driving track point, the vehicle-mounted laser point cloud data is segmented and processed, and the road surface point cloud is extracted;

[0013] Step B, constructing voxelized road surface point cloud, generating road surface point cloud feature map;

[0014] Step C, combining point-level query and instance-level query, constructing positioning methods of different types of lane elements;

[0015] Step D, information interaction of positioning point query is carried out through attention mechanism, and lane element positioning information is output based on road surface features;

[0016] Step E, based on the lane element positioning information, extracting solid line type and non-solid line type lane elements, and constructing high-precision map.

[0017] In an embodiment of the present application, the lane elements include long solid line type, short solid line type and non-solid line type.

[0018] In an embodiment of the present application, the step A specifically comprises the following steps:

[0019] Step A1, based on the coordinate information of the vehicle driving track point, the road point cloud is segmented;

[0020] Step A2, based on the elevation information of the driving track point in the segmented area, the non-road surface background points above the average elevation of the track points are removed.

[0021] In an embodiment of the present application, the step B specifically comprises the following steps:

[0022] Step B1, based on the voxel with length, width and height of w, h and d 3D grid is generated for constructing voxelized road surface point cloud;

[0023] Step B2, the coordinates of the centroid points are calculated voxel by voxel, and the coordinates of the point set in the voxel are updated to the relative coordinates based on the centroid points;

[0024] Step B3, constructing voxelized point cloud features through voxel feature encoding module;

[0025] Step B4, based on the feature pyramid network structure, the multi-scale voxelized point cloud features are fused to obtain the point cloud road surface feature map F fusion .

[0026] In an embodiment of the present application, the step C specifically comprises the following steps:

[0027] Step C1, different positioning methods are constructed for different types of lane elements, wherein the set of positioning points p j is expressed as N p is the number of positioning points of the corresponding type of lane element;

[0028] Step C2, define the instance-level positioning query of the lane element is expressed as where N ins is the number of positioning points of the corresponding type of lane element, and the set of point-level positioning queries shared within the instance is defined as Combining the instance-level positioning query and the point-level positioning query, the set of hierarchical positioning queries of each lane element is expressed as where i is the index of the current positioning lane element.

[0029] In an embodiment of the present application, the step D specifically comprises the following steps:

[0030] Step D1, within the same type of lane element positioning point query method, the lane element positioning point query where N=N ins ·N p is the total number of query points of the corresponding type of lane element, and D is the feature dimension; the self-attention interaction of the positioning point query at the point level and the instance level is constructed, and the updated positioning point query Q' is output:

[0031] Q' = Self_Attention(Q, M)

[0032] In the formula, Self_Attention(·) is a self-attention interaction module, and M is an attention mask of the positioning point query under the same type;

[0033] Step D2, construct the self-attention interaction of the instance-level query between different types of lane element positioning point query methods, and output the updated positioning point query Q'':

[0034] Q'' = Self_Attention((Q' solid + Q' dashed + Q' arrow ), M')

[0035] In the formula, M' is an attention mask of the positioning point query under different types, and Q' solid , Q' dashed , Q' arrow ​​Reference to the updated long solid line type, short solid line type and non-solid line type three types of lane element positioning point query;

[0036] Step D3, extract the initial positioning point coordinates P based on the hierarchical positioning query constructed in step C2 init , combined with the updated positioning point query Q ” Cross attention with road features to realize multiple offset correction of positioning point coordinates:

[0037]

[0038] In the formula, Cross_Attention(·) is the cross attention module, F fusion is the road point cloud feature, P i is the positioning point coordinates of the current attention layer, P0=P init , is the query update of the current attention layer, f(·) is a nonlinear output function;

[0039] Step D4, find the optimal instance-level label assignment between the predicted lane element and the real lane element{y i} with the lowest instance-level matching cost:

[0040]

[0041] In the formula, is the pair matching cost between the predicted lane element and the real lane element y i , considering the class label of the lane element and the position matching cost of the positioning point set, is the predicted lane element index assigned to the real lane element, and arg min represents the index corresponding to the minimum cost value;

[0042] Step D5, calculate the lane element positioning loss based on the optimal matching result for optimization:

[0043]

[0044] In the formula, is the class prediction loss, is the positioning point prediction loss, is the positioning point edge direction loss, is the instance segmentation loss, α a ,α b ,α c ,α d is the weight for balancing different loss terms.

[0045] In an embodiment of the present application, the step E specifically comprises the following steps:

[0046] Step E1, for non-solid line type lane elements in the scene, select samples with less wear and tear and occlusion as templates, and make corresponding vectorized representations;

[0047] Step E2, based on the prediction results of the lane element positioning point query, the vectorized representation of the solid line type lane element is directly obtained by connecting the positioning points under the same instance; the vectorized representation of the non-solid line type lane element is obtained by matching the prediction results of the positioning points and the template, and finally the high-precision map representation under the scene is generated.

[0048] In an embodiment of the present application, step E2 is specifically implemented as follows:

[0049] Step E2-1, considering that the template used for matching and the predicted lane element positioning point may come from different laser scanning systems, which will produce differences in the coordinate system, therefore, the coordinate normalization is performed with the driving vehicle as the center:

[0050] p=P-init(P)

[0051] m=M-init(M)

[0052] In the formula, p and m are points on the matching template P and the element to be matched M respectively, init(·) represents the coordinate point of the corresponding area driving vehicle;

[0053] Step E2-2, calculate the rotation matrix and the translation matrix, first select the corresponding template based on the predicted non-solid line type lane element to be matched, extract the frame points and corner points of the element to be matched and the template element respectively, when the distance of all corresponding point pairs is minimized after matrix transformation, that is, the constraint error function is minimized, then the transformation matrix is considered to be the optimal matrix; define the constraint error function as:

[0054]

[0055] In the formula, R represents the rotation matrix and T represents the translation matrix, p and m are points on the matching template P and the element to be matched M respectively, R and T are the rotation matrix and the translation matrix respectively, and ||·|| is the Euclidean distance, and n is the number of matching points. 2

[0056] The present application also provides a high-precision map construction system based on positioning point query and attention mechanism, comprising a memory, a processor and computer program instructions stored in the memory and capable of being executed by the processor, when the processor executes the computer program instructions, the method steps as described above can be realized.

[0057] ​The application further provides a computer readable storage medium, which stores computer program instructions capable of being run by a processor, and when the processor runs the computer program instructions, the method steps described above can be realized.

[0058] Compared with the prior art, the application has the following beneficial effects: the application constructs a unified extraction framework for various types and scales of lane elements, only differentiates in specific element positioning methods, and improves the class incompatibility of traditional methods for lane element extraction. Combined with the attention mechanism, selective information interaction is performed between element positioning point queries, adaptive learning of geometric structure expression of various lane elements and spatial position relationship of lane elements in the scene is performed, an end-to-end lane element extraction and high-definition map construction scheme is constructed to replace the stage generation method, and error accumulation is avoided. Therefore, the application has strong practicability and broad application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 is a whole method flow schematic diagram of the embodiment of the application;

[0060] Figure 2 is a lane element positioning schematic diagram based on positioning point query and attention mechanism of the embodiment of the application;

[0061] Figure 3 is a voxel feature multi-scale fusion module schematic diagram of the embodiment of the application;

[0062] Figure 4 is a lane element positioning method construction schematic diagram of the embodiment of the application;

[0063] Figure 5 is a positioning point query information interaction schematic diagram based on the attention mechanism of the embodiment of the application;

[0064] Figure 6 is a lane element feature query and positioning point update schematic diagram of the embodiment of the application;

[0065] Figure 7 is a lane element extraction and map construction schematic diagram of the embodiment of the application;

[0066] Figure 8 is a typical scene high-definition map construction result schematic diagram of the embodiment of the application. DETAILED DESCRIPTION

[0067] The technical solutions of the application will be specifically described below with reference to the drawings.

[0068] The application provides a high-precision map construction method based on positioning point query and attention mechanism.

[0069] Please refer to Figure 1 The high-precision map construction method based on positioning point query and attention mechanism specifically comprises the following steps (A-E):

[0070] Step A, based on driving trajectory points, segment processing is performed on vehicle-mounted laser point cloud data, and road surface point clouds are extracted. Specifically, the following steps are included:

[0071] Step A1, based on vehicle driving trajectory points {x1, x2, … x n}, the offset angle of the vehicle driving direction is calculated. For the starting point x a of the current region, when the vehicle drives to x b , if (x b -x a ) is less than 40m, but the direction of the vector formed by (x a+1 -x a ) is greater than 45° (where the direction of the vector formed by (x a+1 -x a ) represents the driving direction of the vehicle at x a ), it is considered that the vehicle has arrived at a corner, and x is taken as the region center point, and (x b -x a ) is the main direction distance for region segmentation. If (x b -x a ) is greater than 40m, then x is directly taken as the region center point, the direction of the vector formed by (x b -x a ) is the main direction, and (x b -x a ) is the main direction distance for region segmentation. Then, x b is taken as the starting point of the next region for region segmentation until the end.

[0072] Step A2, based on the elevation information {h1, h2, … h n} of the driving trajectory points in the segmented region, the average elevation Take this elevation as the near-ground elevation in this small area, and remove the point cloud non-road surface background points with elevation values greater than .

[0073] Please refer to Figure 2 , the present application is based on the positioning point query and attention mechanism lane element positioning process, specifically including the following steps (B-D):

[0074] Step B, construct a voxelized road surface point cloud to generate a road surface point cloud feature map. Specifically, it includes the following steps:

[0075] Step B1, set the road surface point cloud segmentation area to include a three-dimensional space along the Z, X, Y axes with a range of D, W, H respectively. Based on the voxel with a length, width and height of w, h, d Generate a 3D grid to construct a voxelized road surface point cloud. Since the elevation value range of the road surface point cloud is small, in order to reduce the burden of point cloud feature coding, set the voxel height d = H, and based on the voxel point cloud, perform data downsampling preprocessing. Randomly sample the number of points in a single voxel to 10 points if the number of points is greater than 10, and directly retain if the number of points is less than 10.

[0076] Step B2, define the set of points p in a single voxel as V' = {p c ,p i =(x i ,y i ,z i ,r i )} i＝1…k , where p c =(v x ,v y ,v z ) is the centroid of the voxel, x i ,y i ,z i ,r i is the three-dimensional coordinates and surface reflectivity of the point cloud, and k is the number of points in a single voxel. Update the voxel point set description V based on the centroid point coordinates

[0077] Step B3, construct a point cloud voxel feature map for the road surface point cloud by voxel feature coding module As shown in Figure 3 , this step is implemented through the following steps:

[0078] Step B3-1, input the point-level description information in a single voxel into the feature space through a fully connected network to obtain point-level features

[0079] f i =f(V′)

[0080] where f(·) is a linear fully connected layer.

[0081] Step B3-2, aggregate the point-level features within the voxel by a max-pooling function to obtain local aggregated features

[0082] g=maxpool(f i )

[0083] where maxpool is a max-pooling operation.

[0084] Step B3-3, update the point features by concatenating the point-level features and the local aggregated features, and then obtain the final single-voxel feature by a linear network and a pooling function

[0085] f v =maxpool(f([f i ||g]))

[0086] where f(·) is a linear fully connected layer, maxpool is a max-pooling operation, and || represents a vector concatenation operation.

[0087] Step B3-4, arrange the single-voxel features based on the point cloud grid positions divided in step B1 to finally generate a road point cloud voxel feature map

[0088] Step B4, based on the feature pyramid network structure, perform twice convolution down-sampling on the point cloud voxel feature map, where the first convolution down-sampling is F′ v =f conv (F v1 ), and the second convolution down-sampling is F″ v =f conv (F′ v ), respectively sampling the width and height dimensions of the feature map to and of the original feature map dimensions, and then adjusting the width and height dimensions of the feature map to be consistent with F v by interpolation up-sampling. Then, concatenating them and passing them through a convolution layer to obtain the point cloud road feature

[0089] F fusion =f conv (F v ||f up (F′ v )||f up (F″ v ))

[0090] where || represents vector concatenation, and fconv represents the convolutional layer, f up Represents an upsampling layer.

[0091] Step C: Combine point-level query and instance-level query to construct positioning methods for different types of lane features (long solid line, short solid line, and non-solid line). Figure 4 As shown, the specific steps include:

[0092] Step C1: Three types of positioning methods should be set for different types of lane elements. This step is specifically implemented through the following steps:

[0093] Step C1-1: Position the long-distance solid lane feature and the short-distance curved lane feature using 10 equally spaced points:

[0094]

[0095] Where p j Indicates the positioning point, N p Indicates the number of positioning points.

[0096] Step C1-2: Position the short-distance solid lane features and partially truncated straight lane lines using the starting and ending points:

[0097]

[0098] Where p j Indicates the positioning point, N p Indicates the number of positioning points.

[0099] Step C1-3: Non-solid lane features are located using the minimum bounding box points and the selected corner points:

[0100]

[0101] Where p j Indicates the positioning point, N p Indicates the number of positioning points.

[0102] Step C2: Construct a set of instance-level location queries through the vector embedding function Embedding(·) Collection Point-level location query shared within the instance Collection Each lane feature corresponds to a set of hierarchical positioning queries Collection The hierarchical positioning query of the jth point of the i-th map element is expressed as:

[0103]

[0104] where N ins is the number of lane element localization points of the scene, N p is the number of localization points of the corresponding type of lane element.

[0105] Step D, information interaction of the localization point query is carried out through the attention mechanism, and then lane element localization information is output based on the road surface features. Specifically, the following steps are included:

[0106] Step D1, compared with the self-attention interaction of querying all localization points of different types of lane elements, selective masking can make the query pay more attention to information interaction in a certain aspect, and reduce the burden of the network. As shown in Figure 5 , in the same type of lane element localization point query mode, the lane element localization point query where N = N ins ·N p is the number of such query points, and D is the feature dimension. The self-attention interaction of the localization point query is constructed at the point level and the instance level, and the updated localization point query Q' is output:

[0107]

[0108] wherein is a linear projection matrix, Self_Attention(·) is a self-attention interaction module, Softmax(·) is a normalized exponential function, is an attention mask, in the point-level self-attention interaction, N = M inter when the instance-to-instance query is masked, in the instance-level self-attention interaction, M = M intra is the Hadamard product, D k is the feature dimension of the query embedding;

[0109] whereby setting the binary mask M, the model selects the interaction object to calculate the corresponding attention mapping. Define the query embedding of the instance I i as i. For example, query embedding indexes 1, 2,..., N v belong to the first instance, then query embedding indexes N v +1, N v +2,..., 2·N v belong to the second instance, therefore then the binary mask M is constructed as:

[0110]

[0111] Step D2, after attention update based on the same type of lane element positioning point query, Q' solid , dashed , arrow Reference to the updated three types of lane element positioning point query, set the positioning point query Where S is the total number of lane element positioning query points, and D is the feature dimension. Between different types of lane element positioning point query, construct instance-level query self-attention interaction, as shown in Figure 5 The output updated positioning point query Q" is shown as follows:

[0112]

[0113] Where, is a linear projection matrix, Self_Attention(·) is a self-attention interaction module, Softmax(·) is a normalized exponential function, is an attention mask, which masks the positioning point query under the same scale, is Hadamard product, D k is the feature dimension of the query embedding;

[0114] Define the query embedding class C to which the instance I belongs i is denoted as i. As the query embedding index 1,2,…,N v belongs to the first type of query, then The query embedding index N v +1,N v +2,…,2·N v belongs to the second type of query, so Then construct the binary mask M as:

[0115]

[0116] Step D3, as shown in Figure 6 , the hierarchical positioning query constructed in step C2 is extracted through a nonlinear function to obtain the initial positioning point coordinates P init , combined with the updated positioning point query Q" and the road feature to cross attention focus, realize the multiple offset correction of the positioning point coordinates, where the network is set to 6 layers of attention interaction, and the output of each layer will be used for subsequent loss optimization, the purpose is to accelerate the convergence speed of the network:

[0117]

[0118] P init = f(Q hierarchical )

[0119]

[0120] where Cross_Attention(·) is a cross-attention module, F fusion is the road point cloud feature, P i is the positioning point coordinate of the current attention layer (P0=P init ), is the query update of the current attention layer m is the attention head index, k is the positioning point index, K is the total positioning point index number (K<<HW), W m is a linear projection matrix, Δp mqk , A mqk respectively represent the positioning offset and attention weight of the mth attention head and the kth positioning point, both of which are obtained by linear projection, and f(·) is a nonlinear output function.

[0121] Step D4, for the layer-by-layer query output of the cross-attention module, a plurality of predicted results required by multiple tasks are obtained through the corresponding task head. The optimal instance-level label assignment between the predicted lane elements and the real lane elements {y i} is found by using the Hungarian algorithm. is the assignment index of N lane elements with the lowest instance-level matching cost:

[0122]

[0123] wherein, is the pair-wise matching cost between the predicted and the real y i , which takes into account the category label of the lane element and the position of the positioning point set:

[0124]

[0125] wherein, is the category matching cost term, which is the focal loss between the predicted classification score and the target category label c i . is the position matching cost term, which is the distance loss between the predicted point set and the real point set V i . For solid line type markings, the instance segmentation cost is added as an auxiliary cost, which is the binary cross-entropy loss of the predicted instance mask and the real instance mask m;

[0126] Step D5, based on the index of the optimal matching result the lane element positioning loss is calculated for optimization: ​

[0127] Classification loss: For each predicted lane element, assign a class label with the ground truth lane element class to conduct class loss calculation:

[0128]

[0129] where, is the focal loss function, is the predicted classification score, c i is the target class label;

[0130] Point-level distance loss: For each ground truth instance, assign the ground truth position to each predicted position to conduct distance loss calculation:

[0131]

[0132] where, D Manhattan is the Manhattan distance between point sets, is the predicted position of instance i, v i,j is the ground truth position of instance i;

[0133] Edge direction loss: Add edge direction loss to supervise the geometric shape composed of the connection between points, which is obtained by calculating the cosine similarity between the predicted edge of the position and the ground truth edge:

[0134]

[0135] where, cosine_similarity(·) is used to calculate the cosine similarity between vectors, is the edge vector of the jth point in the predicted instance i, e i,j is the edge vector of the jth point in the ground truth instance i;

[0136] Instance segmentation loss: For solid line type lane elements, additionally add instance-level segmentation results as auxiliary loss for optimization, which is obtained by calculating the binary cross-entropy loss between the predicted mask and the ground truth binary mask:

[0137]

[0138] where, φ Seg is the predicted segmentation head, is the layer-wise query output by the cross-attention module, F fusion is the point cloud road feature map, is the cross-entropy loss function, M Gt is the ground truth binary mask;

[0139] In summary, the training loss of the overall network is:

[0140]

[0141] wherein a a , a b , a c , a d is the weight balancing different loss terms.

[0142] Step E, based on the lane element positioning information, extracting solid and non-solid lane elements, and constructing a high-precision map. Specifically, the following steps are included:

[0143] Step E1, for non-solid lane elements in the scene, select samples with less wear and tear and occlusion as templates, and make corresponding vectorized representations.

[0144] Step E2, based on the prediction results of the lane element positioning point query, the vectorized representation of the solid lane element can be directly obtained by connecting the positioning points under the same instance; the vectorized representation of the non-solid lane element is obtained by matching the prediction results of the positioning points and the template. As shown in Figure 7 , the specific matching process is implemented in the following steps:

[0145] Step E2-1, considering that the template used for matching and the predicted lane element positioning points may come from different laser scanning systems, which will produce differences in the coordinate system, therefore, the coordinate normalization is performed with the driving vehicle as the center:

[0146] p = P-init(P)

[0147] m = M-init(M)

[0148] wherein p and m are points on the matching template P and the element to be matched M respectively, and init(·) represents the coordinate point of the driving vehicle in this area.

[0149] Step E2-2, calculate the rotation matrix and translation matrix, first select the corresponding template based on the predicted non-solid lane element category to be matched, extract the frame points and corner points of the element to be matched and the template element respectively, when the distance of all corresponding point pairs is minimized after matrix transformation, that is, the constraint error function is minimized, then the transformation matrix is considered to be the optimal matrix. Define the constraint error function as:

[0150]

[0151] wherein R represents the rotation matrix and T represents the translation matrix, p and m are points on the matching template P and the element to be matched M respectively, R and T are the rotation matrix and the translation matrix respectively, and ‖·‖ 2 is the Euclidean distance, and n is the number of matching points.

[0152] As shown in Figure 8As shown in a high-precision map construction result of a typical scene, the high-precision map construction scheme provided by the application is suitable for lane element extraction under different scales and types. For solid line type lane elements, lane boundary lines, separation lines, sidewalk lines, stop lines and dashed lines with large differences in extraction scales can be compatible, and long distance lane lines with interruptions and omissions can also be described by complete continuous positioning points. At the same time, the positioning output of non-solid line type markings in a unified framework is realized, and the integrity of the geometric structure of the non-solid line type elements is ensured by combining the production of the template library.

[0153] The method directly faces the three-dimensional laser point cloud object, and constructs a positioning point query to construct a high-precision map, overcomes the differences in scales and types of solid line type and non-solid line type lane elements, realizes end-to-end high-precision map generation in a unified framework, selectively interacts information based on the attention mechanism, fully utilizes the geometric structure and spatial position relationship information of the lane element, reduces the interference of background information, improves the precision of lane element extraction in the scene, and provides a new research method for high-precision map construction based on vehicle-mounted laser point cloud.

[0154] The application further provides a high-precision map construction system based on positioning point query and attention mechanism, comprising a memory, a processor and computer program instructions stored in the memory and capable of being executed by the processor, when the processor executes the computer program instructions, the method steps as described above can be realized.

[0155] The application further provides a computer readable storage medium, which stores computer program instructions capable of being executed by a processor, when the processor executes the computer program instructions, the method steps as described above can be realized.

[0156] Those skilled in the art should understand that the embodiments of the application can be provided as a method, a system or a computer program product. Therefore, the application can be in the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the application can be in the form of a computer program product implemented on one or more computer usable storage media containing computer usable program code (including but not limited to disk storage, CD-ROM, optical storage, etc.).

[0157] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0158] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0159] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0160] The above description is only preferred embodiments of the present application, and is not intended to limit the present application to other forms described above. Any person skilled in the art can make modifications or improvements to the above-mentioned disclosed technical content without departing from the technical scope of the present application. However, any simple modification, equivalent change and modification of the above-mentioned embodiments without departing from the technical content of the present application, according to the technical essence of the present application, still belongs to the protection scope of the present application.

Claims

1. A high-definition map construction method based on a positioning point query and an attention mechanism, characterized in that, Comprise the following steps: Step A, segmenting the vehicle-mounted laser point cloud data based on the vehicle driving track points, and extracting the road surface point cloud; Step B, constructing a voxelized road surface point cloud to generate a road surface point cloud feature map; Step C, combining point-level query and instance-level query to construct positioning methods for different types of lane elements; Step D, information interaction of positioning point query is performed through attention mechanism, and lane element positioning information is output based on road surface features; Step E, based on the lane element positioning information, extract solid and non-solid lane elements, and construct a high-precision map; The step C specifically comprises the following steps: Step C1, different positioning methods are constructed for different types of lane elements, wherein the set of positioning points p j is expressed as N p is the number of positioning points corresponding to the type of lane element; Step C2, define lane element instance level positioning query is expressed as a set of where N ins is the number of positioning of the corresponding type of lane element, define the instance-shared point level positioning query is expressed as a set of Combining the instance level positioning query and the point level positioning query, the hierarchical positioning query of each lane element is expressed as a set of where i is the index of the current positioning lane element; The step D specifically comprises the following steps: Step D1, in the same type of lane element positioning point query mode, lane element positioning point query where N = N ins ·N p is the number of total query points of the corresponding class of lane elements, D is the feature dimension; the self-attention interaction of the positioning point query is constructed at the point level and the instance level, and an updated positioning point query Q' is output: Q=Self_Attention(Q,M) In the formula, Self_Attention(·) is a self-attention interaction module, and M is an attention mask of the same type of positioning point query; Step D2, construct a self-attention interaction of instance-level query between different types of lane element positioning point query methods, and output updated positioning point query Q": Q" = Self_Attention((Q solid + Q dashed + Q aarrow ), M') In the formula, M' is the attention mask of the query of the positioning points of different types, and Q' is the query of the positioning points of the long solid line type, the short solid line type, and the non-solid line type. solid Q' dashed Q' arrow denotes the query of the positioning points of the updated long solid line type, the short solid line type, and the non-solid line type. Step D3, the initial positioning point coordinate P is extracted based on the hierarchical positioning query constructed in step C2 init Combined with the cross-attention attention of the updated positioning point query Q" and the road surface features, the multiple offset correction of the positioning point coordinate is realized: In the formula, Cross_Attention(·) is a cross-attention module, F fusion is a road point cloud feature, P i is the positioning point coordinates of the current attention layer, P0=P init , is the query update of the current attention layer, f(·) is a nonlinear output function; Step D4, finding the optimal instance-level label assignment between the predicted lane elements and the real lane elements {y i with the lowest instance-level matching cost: wherein is the predicted lane element and the true lane element y i between the predicted lane element and the true lane element y is the predicted lane element index assigned to the true lane element, and arg min denotes the index corresponding to the minimum cost value. Step D5, calculate the lane element positioning loss based on the optimal matching result for optimization: wherein, is the class prediction loss, is the landmark prediction loss, is the landmark edge direction loss, is the instance segmentation loss, and a ,α b ,α c ,α d are weights balancing different loss terms.

2. The high-definition map construction method based on positioning point query and attention mechanism according to claim 1, characterized in that, The lane elements include long solid line type, short solid line type and non-solid line type.

3. The high-definition map construction method based on positioning point query and attention mechanism according to claim 1, characterized in that, The step A specifically comprises the following steps: Step A1, based on the coordinate information of the vehicle driving track points, the road point cloud is segmented by region; Step A2, based on the elevation information of the driving track points in the segmented area, the non-road surface background points above the average elevation of the track points are removed.

4. The high-definition map construction method based on positioning point query and attention mechanism according to claim 1, characterized in that, The step B specifically comprises the following steps: Step B1, based on a voxel with length, width and height w, h, d generate a 3D grid for constructing a voxelized road point cloud; Step B2, calculate the coordinates of the centroid points voxel by voxel, and update the coordinates of the point set in the voxel to the relative coordinates based on the centroid points; Step B3, construct a voxelized point cloud feature through a voxel feature encoding module; Step B4, based on the feature pyramid network structure, fuse the multi-scale voxelized point cloud features to obtain the point cloud road feature map F fusion .

5. The high-definition map construction method based on positioning point query and attention mechanism according to claim 1, characterized in that, The step E specifically comprises the following steps: Step E1, for non-solid line type lane elements in the scene, select samples with less wear and obstruction as templates, and make corresponding vectorized representations; Step E2, based on the prediction results of the lane element positioning point query, the vectorized representation of the solid line type lane element is directly obtained by connecting the positioning points under the same instance; the vectorized representation of the non-solid line type lane element is obtained by matching the prediction results of the positioning points with the template, and finally a high-precision map representation of the scene is generated.

6. The high-definition map construction method based on positioning point query and attention mechanism according to claim 5, characterized in that, Step E2 is specifically implemented as follows: Step E2-1, considering that the template used for matching and the predicted lane element positioning points may come from different laser scanning systems, which will produce differences in the coordinate system, therefore, the coordinates are normalized with the driving vehicle as the center: p=P-init(P) m=M-init(M) In the formula, p and m are points on the matching template P and the element to be matched M respectively, and init(·) represents the coordinate point of the driving vehicle in the corresponding area; Step E2-2, calculate the rotation matrix and translation matrix, first select the corresponding template based on the predicted non-solid line type lane element to be matched, extract the frame points and corner points of the to-be-matched element and the template element respectively, when the distance of all corresponding point pairs is minimized after matrix transformation, that is, the constraint error function is minimized, then the transformation matrix is considered as the optimal matrix; define the constraint error function as: where R represents a rotation matrix and T represents a translation matrix, p and m are points on the matching template P and the element to be matched M, respectively, R and T are a rotation matrix and a translation matrix, respectively, and ||·|| is the Euclidean distance. 2 is the Euclidean distance, and n is the number of matching points.

7. A high-definition map construction system based on a positioning point query and an attention mechanism, characterized in that, A computer program product comprising a computer readable storage medium having stored thereon computer program instructions, the computer program instructions, when executed by a processor, cause performance of the steps of the method according to any of claims 1-6.

8. A computer readable storage medium having stored thereon computer program instructions, the computer program instructions, when executed by a processor, cause performance of the steps of the method according to any of claims 1-6.

Images