Intersection lane-level topology detection method, device and equipment based on trajectory data

Abstract

The invention relates to the technical field of intelligent traffic information, in particular to an intersection lane-level topology detection method, device and equipment based on trajectory data, and the method comprises the steps: obtaining the trajectory data in an intersection; extracting shape features, direction features and relation features among the tracks in the track data; and constructing a trajectory relationship graph according to the shape features, the direction features and the relationship features among the trajectories, inputting the trajectory relationship graph into a graph neural network architecture, and outputting a lane-level intersection topology by the graph neural network architecture. Therefore, the problems of high data source acquisition cost, high data processing complexity, insufficient algorithm precision and robustness, real-time dynamic update challenge and the like in the prior art are solved.

Description

Technical Field

[0001] The present application relates to the field of intelligent traffic information technology, and in particular to a method, device and equipment for detecting lane-level topology of an intersection based on trajectory data. Background Art

[0002] Lane-level road network data, as a core element of high-precision map construction, plays a vital role in the development of cutting-edge intelligent transportation technologies such as lane-level navigation and autonomous driving. In a complex road network, intersections are the convergence and dispersion points of traffic flow. Their complex structural characteristics and high risk of traffic accidents make it an indispensable task to accurately obtain lane information and intersection topology.

[0003] In terms of data source selection, current technology mainly relies on high-resolution remote sensing images, LiDAR scanning and rich trajectory data. Although high-precision remote sensing images can provide detailed visual information, they are difficult to obtain and are easily affected by environmental factors; LiDAR is known for its high-precision three-dimensional point cloud data, which ensures the accuracy of map construction, but its high cost has become a bottleneck for its widespread application. In contrast, trajectory data has become a data source with great potential due to its advantages of low cost, easy acquisition and high frequency of updates.

[0004] Among the related technologies, the method of combining road centerline information, trajectory tracking and density clustering has successfully achieved effective recognition and extraction of the lane-level structure of the intersection. In addition, by introducing advanced technologies such as multi-feature integration and semi-supervised anti-counterfeiting, the accuracy of intersection position recognition has been further improved, and the complex topological relationship and road section attributes between intersections have been deeply analyzed.

[0005] However, the complexity of the intersection area cannot be ignored. As a transition area where roads in different directions meet, driving behaviors in this area are frequent and changeable. Driving behaviors such as turning, changing lanes, merging, and diverting are intertwined, making the distribution of trajectory data extremely complex. The overlap, intersection, and break between trajectories not only increase the difficulty of data processing, but also put forward higher requirements on the stability and accuracy of the algorithm. Summary of the invention

[0006] The present application provides a method, device and equipment for detecting lane-level topology of an intersection based on trajectory data to solve the problems in the prior art such as high cost of acquiring data sources, high complexity of data processing, insufficient algorithm accuracy and robustness, and challenges of real-time dynamic updating.

[0007] The first aspect of the present application provides a method for detecting lane-level topology of an intersection based on trajectory data, comprising the following steps: acquiring trajectory data inside the intersection; extracting shape features, direction features, and relationship features between trajectories in the trajectory data; constructing a trajectory relationship graph based on the shape features, direction features, and relationship features between trajectories, inputting the trajectory relationship graph into a graph neural network architecture, and the graph neural network architecture outputs a lane-level intersection topology.

[0008] Optionally, constructing a trajectory relationship graph according to shape features, direction features and relationship features between trajectories includes: using shape features and direction features as node features, using relationship features between trajectories as edge features; and generating a trajectory relationship graph according to node features and edge features.

[0009] Optionally, the node features include at least one of trajectory distortion, trajectory curvature, trajectory azimuth change, start point and end point azimuth, and start point to end point orientation, and the edge features include at least one of the Hausdorff distance between trajectories and the distance sequence of sampling points between trajectories, wherein the calculation formula of the Hausdorff distance between trajectories is:

[0010]

[0011] Among them, d(p,τ B ) represents the distance from point p to τ B The minimum distance between any points in A ,q) represents τ A The minimum distance from any point in to point q;

[0012] The distance sequence of sampling points between trajectories is:

[0013] d i =∥p i -q i ∥;

[0014] D sp = <d1,d2,…,d m >

[0015] Among them, m is a fixed number of uniformly sampled samples at equal intervals, and p i is τ A The corresponding sampling point in q i is τ B The corresponding sampling point in d i is the sampling point q i distance.

[0016] Optionally, the graph neural network architecture includes a node embedding component and a node classification component, wherein the node embedding component includes multiple node embedding blocks, each embedding block includes multiple fully connected layers, graph learning layers and normalization layers, each fully connected layer uses ReLU as an activation function, and in each embedding block, the node features are input into multiple fully connected layers to obtain edge features between nodes, the node features and the edge features between nodes are input into the graph learning layer, and the output of the graph learning layer is normalized using the normalization layer; the node classification component includes a classification block consisting of two fully connected layers, the first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the Sigmoid activation function to predict the probability of each node in the graph as a lane trajectory. During the prediction process, if the predicted probability of the node exceeds the preset probability, the node is marked as a lane trajectory.

[0017] Optionally, the graph learning layer includes a graph and a multi-head attention mechanism, wherein the graph is generated according to node features and edge features between nodes, and the d-th head attention of the multi-head attention mechanism for each edge from j to i is:

[0018] q c,i =W c,q x i +b c,q

[0019] e c,ij =W c,e e ij +b c,e

[0020] k c,j =W c,k x j +b c,k

[0021]

[0022] Among them, x i is the node feature vector of the i-th node, x j is the node feature vector of the jth node, q c,i is the query vector of the cth attention head of the ith node, W c,q and b c,q To calculate q c,i The trainable parameters, k c,j is the key vector of the cth attention head of the jth node, W c,k and b c,k To calculate k c,i The trainable parameters, e ij for e c,ij is the edge feature vector between the i-th and j-th nodes, W c,e and b c,eTo calculate e c,ij The trainable parameters of c,ij is the attention score between the i-th node and the j-th node.

[0023] When the number of attention heads is h, the updated node features are:

[0024] v c,j =W c,v x j +b c,v

[0025]

[0026] Among them, ∥ represents the serial operation, v c,j is the value vector of the cth attention head of the jth node, W c,v and b c,v Calculate v c,j The trainable parameters, x i ′ is the updated node feature.

[0027] Optionally, the graph neural network architecture is trained based on a set of road network extraction loss functions.

[0028] Optionally, the bipartite matching step of the set-based road network extraction loss function includes: obtaining a real lane set and a trajectory prediction set; matching the real lane set and the trajectory prediction set, determining the optimal arrangement according to the bipartite matching algorithm, and calculating the minimized matching cost according to the optimal arrangement; decomposing into classification matching cost and position matching cost according to the minimized matching cost, wherein the classification matching cost is calculated for non-empty real lanes, and the position matching cost is normalized by Hausdorff distance; calculating the total matching cost according to the position matching cost and the classification matching cost after Hausdorff distance normalization; and optimizing the prediction model according to the total matching cost.

[0029] Optionally, determining the optimal arrangement according to a bisection matching algorithm, and calculating the minimum matching cost according to the optimal arrangement, including: calculating a set-based road network extraction loss function:

[0030]

[0031] in, is the best allocation, The output of the model is represented by The probability that the indexed trajectory is classified as a lane trajectory.

[0032] The second aspect of the present application provides an intersection lane-level topology detection device based on trajectory data, including: an acquisition module for acquiring trajectory data inside the intersection; an extraction module for extracting shape features, direction features and relationship features between trajectories in the trajectory data; an output module for constructing a trajectory relationship graph based on the shape features, direction features and relationship features between trajectories, inputting the trajectory relationship graph into a graph neural network architecture, and the graph neural network architecture outputs a lane-level intersection topology.

[0033] The third aspect of the present application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a method for detecting intersection lane-level topology based on trajectory data as described in the above embodiment.

[0034] Therefore, this application includes the following beneficial effects:

[0035] The embodiment of the present application achieves high-precision and high-robustness recognition and extraction of lane lines inside the intersection by constructing a trajectory relationship diagram inside the intersection and using graph neural network learning technology. Even when faced with complex and changeable trajectory data, it can still accurately capture representative trajectories as a direct mapping of lane lines, thereby generating a detailed and accurate intersection lane-level road network topology, significantly reducing the cost of building a high-precision intersection lane-level road network. At the same time, it also supports real-time dynamic updates, providing reliable high-precision map underlying data support for intelligent transportation systems such as autonomous driving. As a result, technical problems such as high data source acquisition costs, high data processing complexity, insufficient algorithm accuracy and robustness, and real-time dynamic update challenges in the prior art are solved.

[0036] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0038] Figure 1 A flowchart of a method for detecting intersection lane-level topology based on trajectory data according to an embodiment of the present application;

[0039] Figure 2 A schematic diagram for constructing an internal trajectory relationship diagram of an intersection provided according to an embodiment of the present application;

[0040] Figure 3 A schematic diagram of node characteristics of a trajectory relationship graph provided according to an embodiment of the present application;

[0041] Figure 4A schematic diagram of the Hausdorff distance between trajectories provided according to an embodiment of the present application;

[0042] Figure 5 A schematic diagram of the distance between sampling points of trajectories provided according to an embodiment of the present application;

[0043] Figure 6 A schematic diagram of a graph neural network architecture provided according to an embodiment of the present application;

[0044] Figure 7 A schematic diagram of a lane extraction loss function training based on a set according to an embodiment of the present application;

[0045] Figure 8 A flowchart of a method for detecting intersection lane-level topology based on trajectory data provided by one embodiment of the present application;

[0046] Fig. 9 A schematic diagram of the effect of extracting a lane-level road network at an intersection provided according to an embodiment of the present application;

[0047] Fig.10 This is an example diagram of an intersection lane-level topology detection device based on trajectory data according to an embodiment of the present application;

[0048] Fig.11 Schematic diagram of the structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0049] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0050] In terms of data processing and extraction technology, a variety of innovative methods have been proposed and applied. Among them, in related technology 1, the existing road centerline-level road network map is used to determine the intersection location, set a buffer zone, identify turning change points through trajectory tracking, and then use the density clustering algorithm to cluster the turning change points to construct the intersection range circle. Finally, through trajectory tracking and spatial matching methods, the turning entry and exit points of the intersection are extracted, and matched with the center lines of the lanes of the adjacent road sections to complete the extraction of the lane-level structure of the intersection. This process combines trajectory data analysis with spatial matching technology to achieve the precise construction of the lane-level structure of the intersection;

[0051] In related technology 2, multi-dimensional features are introduced to accurately identify intersections and semi-supervised de-aliasing, and the topological relationship of intersections is identified using Delaunay triangulation, trajectory distribution, and road geometry features, and the morphological results are integrated to optimize the geometry of road sections. Based on this information, taxi trajectories are analyzed to calculate intersection turning rules and one-way and two-way attributes of road sections, providing key data support for intelligent transportation systems.

[0052] However, the intersection area has a complex structure and is also a transition zone between roads in different directions. Driving behaviors such as turning, lane changing, merging and diverging often occur in these areas, and the trajectory distribution is complex, making it a challenging task to extract the lane-level topology of intersections based on trajectory data.

[0053] The following describes the intersection lane-level topology detection method, device and equipment based on trajectory data of the embodiment of the present application with reference to the accompanying drawings. In view of the difficulty in extracting the intersection lane-level topology based on trajectory data mentioned in the above background technology, the present application provides a method for detecting the intersection lane-level topology based on trajectory data. In this method, by constructing a trajectory relationship diagram inside the intersection and using graph neural network learning technology, high-precision and high-robustness recognition and extraction of lane lines inside the intersection are achieved. Even in the face of complex and changeable trajectory data, representative trajectories can still be accurately captured as a direct mapping of lane lines, thereby generating a detailed and accurate intersection lane-level road network topology, significantly reducing the cost of constructing a high-precision intersection lane-level road network, while also supporting real-time dynamic updates, providing reliable high-precision map underlying data support for intelligent transportation systems such as autonomous driving. As a result, the problems of high data source acquisition cost, high data processing complexity, insufficient algorithm accuracy and robustness, and real-time dynamic update challenges in the prior art are solved.

[0054] Specifically, Figure 1 A flowchart of a method for detecting lane-level topology at an intersection based on trajectory data is provided in an embodiment of the present application.

[0055] like Figure 1 As shown, the intersection lane-level topology detection method based on trajectory data includes the following steps:

[0056] In step S101 , the trajectory data inside the intersection is obtained.

[0057] The intersection may be an area where two or more roads intersect in a road network, and the trajectory data may be motion path data left by vehicles, pedestrians or other moving objects in the intersection area.

[0058] It can be understood that the embodiment of the present application can grasp the traffic flow, vehicle speed, turning behavior and other information in real time by analyzing the trajectory data inside the intersection, and provide data support for subsequent lane-level topology detection of the intersection.

[0059] In step S102, shape features, direction features, and relationship features between trajectories in the trajectory data are extracted.

[0060] Among them, the shape feature can be the shape characteristics formed by the trajectory data in space, such as straight line, curve, ring, etc.; the direction feature can be the moving direction represented by the trajectory data, including the direction of entering the intersection, the turning direction in the intersection, and the direction of leaving the intersection; the relationship feature can be the relationship between different trajectories in space and time, such as parallel, intersection, following, etc.

[0061] It can be understood that the embodiments of the present application extract shape features, direction features and relationship features of trajectories in trajectory data, which can simplify complex trajectory data into a form that is easier to process and analyze, thereby improving the efficiency and accuracy of data processing.

[0062] In step S103, a trajectory relationship graph is constructed based on shape features, direction features, and relationship features between trajectories, and the trajectory relationship graph is input into a graph neural network architecture, which outputs a lane-level intersection topology.

[0063] Among them, the trajectory relationship graph can be a data structure that graphically represents the shape features, direction features and relationship features between trajectories in the trajectory data. The graph neural network architecture can include node embedding components and node classification components. The lane-level intersection topology refers to the intersection topology structure obtained after fine division of lanes in the intersection area.

[0064] It can be understood that the embodiment of the present application integrates shape features, direction features and relationship features between trajectories into the trajectory relationship graph, which can comprehensively capture the complexity and dynamics of traffic flow in the intersection, and help to more accurately reflect the actual traffic conditions at the intersection; the trajectory relationship graph is input into the graph neural network architecture, and the graph neural network architecture outputs the lane-level intersection topology, providing more accurate and detailed information for path planning and navigation services.

[0065] In an embodiment of the present application, a trajectory relationship graph is constructed based on shape features, direction features, and relationship features between trajectories, including: using shape features and direction features as node features, and using relationship features between trajectories as edge features; and generating a trajectory relationship graph based on node features and edge features.

[0066] The node features may include at least one of trajectory distortion, trajectory curvature, trajectory azimuth change, start point and end point azimuths, and start point to end point orientation; the edge features may include at least one of inter-trajectory Hausdorff distance and inter-trajectory sampling point distance sequence.

[0067] It can be understood that the embodiment of the present application uses shape features and direction features as node features, which can accurately describe the shape and movement direction of a single trajectory in space; the relationship features between trajectories are used as edge features, which capture the interaction and mutual influence between different trajectories; by combining node features and edge features to generate a trajectory relationship graph, a complex network reflecting the traffic flow situation inside the intersection can be constructed, providing refined data support for traffic management and enhancing the accuracy and efficiency of traffic management.

[0068] Specifically, Figure 2 As shown in the figure, the internal trajectory of the intersection is taken as input, the shape and direction features of the trajectory are used as node features, and the relationship between trajectories is used as edge features to construct a fully connected graph.

[0069] like Figure 3 As shown, the node features include:

[0070] Trajectory distortion: The ratio of trajectory length L to its start-end distance D, quantifies the distortion of the trajectory. This feature describes the degree of tortuosity of the trajectory line and is defined as Among them, L represents the length of the trajectory line, which is the straight-line distance from the starting point to the end point of the trajectory;

[0071] Track curvature: The curvature of the track, calculated as Among them, Δα i is the midpoint of the trajectory i Relative to p i-1 The azimuth change, n is the number of points in the trajectory, and L is the length of the trajectory line;

[0072] Track azimuth change: The direction of the track changes. Considering the characteristics of the azimuth, it is decomposed into sine and cosine values, respectively, by sin(α n -α1) and cos(α n -α1) calculation, where α1 is the azimuth trajectory of the starting point, α n is the azimuth trajectory of the end point;

[0073] Starting and ending azimuth: The starting and ending directions of the trajectory. These features are also decomposed into sine and cosine values, namely sinα1, cosα1, sinα n and cosα n ;

[0074] Start-to-end orientation: The vector direction from the start point to the end point of the trajectory. This feature is decomposed into sine and cosine values, that is, and in is the trajectory from the starting point p1 to the end point p n The direction of the vector.

[0075] In the embodiment of the present application, the calculation formula of the Hausdorff distance between trajectories is:

[0076]

[0077] Among them, d(p,τ B ) represents the distance from point p to τ B The minimum distance between any points in A ,q) represents τ A The minimum distance from any point in to point q;

[0078] The distance sequence of sampling points between trajectories is:

[0079] d i =∥p i -q i ∥;

[0080] D sp = <d1,d2,…,d m >

[0081] Among them, m is a fixed number of uniformly sampled samples at equal intervals, and p i is τ A The corresponding sampling point in q i is τ B The corresponding sampling point in d i is the sampling point q i distance.

[0082] Specifically, the edge features include the Hausdorff distance between trajectories and the distance sequence of sampling points, where the Hausdorff distance is as follows: Figure 4 As shown, given two trajectories, τ A and τ B , the Hausdorff distance is τ A and τ B The Hausdorff distance between the points is the maximum value among the minimum distances between them, d hd The calculation is as follows:

[0083]

[0084] Among them, d(p,τ B ) represents the distance from point p to τ B The minimum distance between any points in A ,q) represents τA The minimum distance from any point in to point q;

[0085] like Figure 5 As shown, the distance sequence calculation of sampling points between trajectories includes the following steps:

[0086] (1) Along two given trajectories τ A and τ B Uniformly sample a fixed number m of points at equal intervals, for example 10 points;

[0087] (2) For τ A Each sampling point p in i , calculate τ B Sampling point q i The distance d i :d i =∥p i -q i ∥;

[0088] (3) Forming the distance sequence of sampling points: D sp = <d1,d2,…,d m >.

[0089] Given that the above distance ranges from 0 to more than 50m, directly inputting these distances into the network may lead to unstable optimization process. Therefore, exponential scaling is used to normalize these distance features. For a distance feature d, the normalized feature is defined as:

[0090]

[0091] Here, β is a scaling parameter.

[0092] The output of the normalization function falls in the range of (0, 1], which is particularly sensitive to value changes at smaller distances, helping the model understand and distinguish nearby trajectories. For longer distances, the output tends to be close to 0.

[0093] In an embodiment of the present application, a node embedding component includes multiple node embedding blocks, each embedding block includes multiple fully connected layers, a graph learning layer and a normalization layer, each fully connected layer uses ReLU as an activation function, and in each embedding block, node features are input into multiple fully connected layers to obtain edge features between nodes, the node features and the edge features between nodes are input into the graph learning layer, and the output of the graph learning layer is normalized using the normalization layer; the node classification component includes a classification block consisting of two fully connected layers, the first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the Sigmoid activation function to predict the probability of each node in the graph as a lane trajectory. During the prediction process, if the predicted probability of the node exceeds the preset probability, the node is marked as a lane trajectory.

[0094] Among them, the node embedding block can be the basic building unit in the node embedding component, the graph learning layer can include a graph and a multi-head attention mechanism, and the preset probability can be specifically calibrated, such as 70%.

[0095] It is understandable that the embodiments of the present application achieve deep mining and integration of node features in trajectory data and efficient lane trajectory recognition through node embedding components and node classification components. The node embedding component uses multiple embedding blocks and core processing layers to enhance the expressiveness of node features and capture the complex relationships between nodes. The node classification component uses two fully connected layers, respectively, using ReLU and Sigmoid activation functions, to achieve further extraction and classification of node representations, and finally accurately identifies lane trajectories through preset probability thresholds, thereby improving the accuracy and efficiency of traffic data analysis.

[0096] Specifically, Figure 6 As shown, the network takes the trajectory relationship graph as input and contains two main components, node embedding and node classification components.

[0097] The node embedding component in the network consists of M node embedding blocks, which can generate an embedding vector for each node based on the features of each node and the edge features between nodes. In the node embedding block, the node features First, it is input into two fully connected layers to learn the relationship between various trajectory attributes and further learn node representation. Each fully connected layer has 64 neurons and uses ReLU as the activation function. Then, the learned node features are and edge features between nodes It is fed to the graph learning layer. After passing through the graph learning layer, the output is normalized using layer normalization to improve training stability. Similarly, ReLU is used to introduce nonlinearity. Each node embedding block shares the same structure. After passing through M node embedding blocks, the node embedding component finally outputs a node embedding vector.

[0098] The node classification component in the network contains a classification block consisting of two fully connected layers. The first fully connected layer has 64 neurons and uses the ReLU activation function. Considering that the model is a binary classification task, the trajectories are divided into lane trajectories and other trajectories. The second fully connected layer maps the 64-dimensional input to a one-dimensional output and uses the Sigmoid activation function to predict the probability of each node in the graph being a lane trajectory. During the prediction process, if the predicted probability of a node exceeds 70%, the model will mark the node as a lane trajectory.

[0099] In the embodiment of the present application, a graph is generated based on the node features and the edge features between the nodes, and the d-th head attention of the multi-head attention mechanism for each edge from j to i is:

[0100] q c,i =W c,q x i +b c,q

[0101] e c,ij =W c,e e ij +b c,e

[0102] k c,j =W c,k x j +b c,k

[0103]

[0104] Among them, x i is the node feature vector of the i-th node, x j is the node feature vector of the jth node, q c,i is the query vector of the cth attention head of the ith node, W c,q and b c,q To calculate q c,i The trainable parameters, k c,j is the key vector of the cth attention head of the jth node, W c,k and b c,k To calculate k c,i The trainable parameters, e ij for e c,ij is the edge feature vector between the i-th and j-th nodes, W c,e and b c,e To calculate e c,ij The trainable parameters of c,ij is the attention score between the i-th node and the j-th node.

[0105] When the number of attention heads is h, the updated node features are:

[0106] v c,j =W c,v x j +b c,v

[0107]

[0108] Among them, ∥ represents the serial operation, v c,j is the value vector of the cth attention head of the jth node, W c,v and b c,v Calculate v c,j The trainable parameters, x i′ is the updated node feature.

[0109] Specifically, the GraphTransformer layer is used as the graph learning layer: given a graph G = (X, E), its node features X = {x1, x2, ..., x n}, x and edge feature E = {e 12 ,e 13 ,…,e n-1,n}, the c-th head attention of the multi-head attention mechanism for each edge from j to i is:

[0110] q c,i =W c,q x i +b c,q

[0111] e c,ij =W c,e e ij +b c,e

[0112] k c,j =W c,k x j +b c,k

[0113]

[0114] Among them, x i is the node feature vector of the i-th node, x j is the node feature vector of the jth node, q c,i is the query vector of the cth attention head of the ith node, W c,q and b c,q To calculate q c,i The trainable parameters, k c,j is the key vector of the cth attention head of the jth node, W c,k and b c,k To calculate k c,i The trainable parameters, e ij for e c,ij is the edge feature vector between the i-th and j-th nodes, W c,e and b c,e To calculate e c,ij The trainable parameters of c,ij is the attention score between the i-th node and the j-th node.

[0115] When the number of attention heads is h, the updated node features are:

[0116] v c,j =W c,v xj +b c,v

[0117]

[0118] Among them, ∥ represents the serial operation, v c,j is the value vector of the cth attention head of the jth node, W c,v and b c,v Calculate v c,j The trainable parameters, x i ′ is the updated node feature.

[0119] Each head generates features with a dimension of 64, and the outputs of h heads are concatenated to form a (64×h)-dimensional vector.

[0120] In an embodiment of the present application, the graph neural network architecture is trained based on a set of road network extraction loss functions.

[0121] It can be understood that the embodiment of the present application achieves accurate extraction of road network data features and optimization of model performance by adopting a set-based road network extraction loss function for training, thereby enhancing the generalization ability of the model and enabling it to adapt to road network structures in different scenarios.

[0122] Specifically, Figure 7 As shown, the set-based road network extraction loss function is: This loss regards intersection lane extraction as a set prediction task, holistically evaluates the entire output of the network at the intersection, and solves potential lane extraction missing and redundancy problems.

[0123] In an embodiment of the present application, the bipartite matching step of the set-based road network extraction loss function includes: obtaining a real lane set and a trajectory prediction set; matching the real lane set and the trajectory prediction set, determining the optimal arrangement according to the bipartite matching algorithm, and calculating the minimized matching cost according to the optimal arrangement; decomposing the minimized matching cost into a classification matching cost and a position matching cost, wherein the classification matching cost is calculated for non-empty real lanes, and the position matching cost is normalized by Hausdorff distance; calculating the total matching cost according to the position matching cost and the classification matching cost after Hausdorff distance normalization; and optimizing the prediction model according to the total matching cost.

[0124] It can be understood that the embodiments of the present application obtain a real lane set and a trajectory prediction set, match the two through a binary matching algorithm and calculate the minimum matching cost; decompose the matching cost into a classification matching cost and a position matching cost, and then calculate the total matching cost; optimize the prediction model according to the total matching cost to improve the consistency between the predicted trajectory and the real lane.

[0125] Specifically, Figure 7 As shown in Figure 2, for an intersection with n trajectories, y represents the ground truth set of lanes in the intersection, and Since the number of trajectories exceeds the number of real lanes in the intersection, we treat y as a set of size n and use To obtain the optimal bipartite matching between these two sets, that is, to find the one that minimizes the cost Arrangement of:

[0126]

[0127] in, Represents the true value y i and the cost of pairwise matching between the trajectories indexed by σ(i).

[0128] The matching cost takes into account the probability of each trajectory being a lane trajectory and the position accuracy of each lane trajectory. The matching cost consists of two parts, the classification matching cost and the similarity matching cost:

[0129]

[0130] Considering that y contains The filling of the existing real lane and No lanes in y are not calculated. Defined as:

[0131]

[0132] where c i Represents y i categories, including "no lane" and "existing lane", represents the probability that the trajectory indexed by σ(i) is classified as a lane trajectory by the model.

[0133] The positional similarity between the predicted trajectory and the ground truth lane is considered, which can be measured by the Hausdorff distance. However, the Hausdorff distance between the intersection trajectory and the ground truth lane can range from less than 1 meter to more than 100 meters. The range is between [-1, 0]. Directly using the Hausdorff distance as the location matching cost may result in and Losing the consistency constraint, this may cause the matching cost to ignore the classification accuracy and affect the performance of the algorithm. Therefore, it is necessary to normalize the Hausdorff distance.

[0134] The Hausdorff distance normalization of the trajectory-lane matching relationship includes the following steps for each intersection:

[0135] (1) Distance calculation: Calculate the trajectory of each lane The Hausdorff distance between

[0136] (2) Trajectory matching: According to the Hausdorff distance, the nearest lane is found as the matching lane for each trajectory. This matching is bidirectional. Since the number of trajectories at an intersection usually exceeds the number of lanes, a lane may match multiple trajectories. For each lane Define the trajectory index set S that matches it i ;

[0137] (3) Distance normalization; for trajectory τ j , ground truth lane and its matching trajectory index set S i , normalized Hausdorff distance Defined as:

[0138]

[0139] Among them, |S i | represents the difference with the i-th ground truth lane The number of matched trajectories. a is a correction term less than 1.

[0140] Specifically, when the lane With the trajectory τ j If it does not match, Set to 1, indicating the maximum matching distance. When a lane matches only one trajectory, the distance between them is set to 0. When a lane is matched with multiple trajectories, the Hausdorff distances of these corresponding lane-trajectory pairs are normalized to the range between 0 and 1. Note that this may result in pairs that are matched but far away. The values ​​are close to 1. To distinguish them from mismatched pairs, the distance is multiplied by a correction term a, where a is less than 1.

[0141] Through the above steps, the normalized Hausdorff distance of any trajectory-lane pair in the intersection can be calculated The range is [0, 1] and can be used as the location matching cost.

[0142] The matching cost is calculated as follows:

[0143]

[0144] in, is the normalized Hausdorff distance between the i-th ground truth lane and the σ(i)-th intersection trajectory.

[0145] In the embodiment of the present application, the optimal arrangement is determined according to the bipartite matching algorithm, and the minimization matching cost is calculated according to the optimal arrangement, including: calculating the set-based road network extraction loss function:

[0146]

[0147] in, is the best allocation, The output of the model is represented by The probability that the indexed trajectory is classified as a “lane trajectory”.

[0148] Specifically, the matching cost is calculated by bipartite matching The best bipartite match between the trajectory of the intersection and the ground truth lane is generated using a bipartite matching algorithm. Referring to the common loss design in target detection, the loss function is defined as a linear combination of the negative log-likelihood of the category prediction and the position accuracy loss:

[0149]

[0150] in, is the best allocation, The output of the model is represented by The probability that the indexed trajectory is classified as a “lane trajectory”.

[0151] It should be noted that in order to solve the imbalance problem between positive and negative samples, When , the log probability term is multiplied by 0.1.

[0152] In summary, if Figure 8 As shown, the trajectory relationship graph construction module constructs a trajectory relationship graph from the intersection trajectory based on the shape and direction characteristics of the trajectory data and the relationship between the trajectories; the intersection topology detection module based on the graph network directly extracts representative trajectories as lanes from the intersection trajectory relationship graph and constructs the topology through the graph neural network.

[0153] The embodiment of the present application constructs an intersection trajectory relationship graph through the shape of the internal trajectory of the intersection, the direction characteristics and the relationship between the trajectories, and directly extracts the representative trajectories of the intersection as lane lines to extract the intersection road network through the graph neural network. The challenge of complex structure and trajectory distribution to intersection lane extraction is addressed through the construction of the trajectory relationship graph and the graph deep learning method, and the internal lane lines of the intersection are directly extracted through the graph neural network. Fig. 9 As shown in the figure, the lane-level road network extraction effect of the intersection only relies on trajectory data to extract the lane-level road network of the intersection, which reduces the construction cost of the lane-level road network of the intersection. It is easy to implement and can ensure the integrity of the extracted road network. It provides underlying data support for high-precision maps for future upper-level autonomous driving applications and has great application potential.

[0154] According to the intersection lane-level topology detection method based on trajectory data proposed in the embodiment of the present application, by constructing a trajectory relationship diagram inside the intersection and using graph neural network learning technology, high-precision and high-robustness recognition and extraction of lane lines inside the intersection are achieved. Even in the face of complex and changeable trajectory data, representative trajectories can still be accurately captured as a direct mapping of lane lines, thereby generating a detailed and accurate intersection lane-level road network topology, significantly reducing the cost of building a high-precision intersection lane-level road network, while also supporting real-time dynamic updates, providing reliable high-precision map underlying data support for intelligent transportation systems such as autonomous driving. As a result, the problems of high data source acquisition costs, high data processing complexity, insufficient algorithm accuracy and robustness, and real-time dynamic update challenges in the prior art are solved.

[0155] Next, a device for detecting intersection lane-level topology based on trajectory data according to an embodiment of the present application will be described with reference to the accompanying drawings.

[0156] Fig.10 It is a block diagram of an intersection lane-level topology detection device based on trajectory data according to an embodiment of the present application.

[0157] like Fig.10 As shown, the intersection lane-level topology detection device 10 based on trajectory data includes: an acquisition module 100, an extraction module 200 and an output module 300.

[0158] Among them, the acquisition module 100 is used to obtain the trajectory data inside the intersection; the extraction module 200 is used to extract the shape features, direction features and relationship features between trajectories in the trajectory data; the output module 300 is used to construct a trajectory relationship graph based on the shape features, direction features and relationship features between trajectories, and input the trajectory relationship graph into the graph neural network architecture, which outputs the lane-level intersection topology.

[0159] It should be noted that the aforementioned explanation of the embodiment of the intersection lane-level topology detection method based on trajectory data is also applicable to the intersection lane-level topology detection device based on trajectory data of this embodiment, and will not be repeated here.

[0160] According to the intersection lane-level topology detection device based on trajectory data proposed in the embodiment of the present application, by constructing a trajectory relationship diagram inside the intersection and using graph neural network learning technology, high-precision and high-robustness recognition and extraction of lane lines inside the intersection are achieved. Even in the face of complex and changeable trajectory data, representative trajectories can still be accurately captured as a direct mapping of lane lines, thereby generating a detailed and accurate intersection lane-level road network topology, significantly reducing the cost of building a high-precision intersection lane-level road network. At the same time, it also supports real-time dynamic updates, providing reliable high-precision map underlying data support for intelligent transportation systems such as autonomous driving. As a result, the problems of high data source acquisition costs, high data processing complexity, insufficient algorithm accuracy and robustness, and real-time dynamic update challenges in the prior art are solved.

[0161] Fig.11 A schematic diagram of an electronic device of a vehicle provided in an embodiment of the present application. The electronic device may include:

[0162] A memory 1101 , a processor 1102 , and a computer program stored in the memory 1101 and executable on the processor 1102 .

[0163] When the processor 1102 executes the program, the intersection lane-level topology detection method based on trajectory data provided in the above embodiment is implemented.

[0164] Furthermore, the electronic device further comprises:

[0165] The communication interface 1103 is used for communication between the memory 1101 and the processor 1102 .

[0166] The memory 1101 is used to store computer programs that can be executed on the processor 1102 .

[0167] The memory 1101 may include a high-speed RAM (Random Access Memory) memory, and may also include a non-volatile memory, such as at least one disk memory.

[0168] If the memory 1101, the processor 1102 and the communication interface 1103 are implemented independently, the communication interface 1103, the memory 1101 and the processor 1102 can be connected to each other through a bus and communicate with each other. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Fig.11 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0169] Optionally, in a specific implementation, if the memory 1101, the processor 1102 and the communication interface 1103 are integrated on a chip, the memory 1101, the processor 1102 and the communication interface 1103 can communicate with each other through an internal interface.

[0170] The processor 1102 may be a CPU (Central Processing Unit), or an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present application.

[0171] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0172] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0173] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or N executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0174] It should be understood that the various parts of the present application can be implemented in hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, the steps or methods can be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array, a field programmable gate array, etc.

[0175] A person of ordinary skill in the art may understand that all or part of the steps carried by the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the above-mentioned program may be stored in a computer-readable storage medium, which, when executed, includes one of the steps of the method embodiment or a combination thereof.

[0176] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A method for detecting lane-level topology at an intersection based on trajectory data, characterized in that: The following steps are involved: Get the trajectory data inside the intersection; Extracting shape features, direction features, and relationship features between trajectories in the trajectory data; A trajectory relationship graph is constructed according to the shape features, the direction features and the relationship features between the trajectories, the trajectory relationship graph is input into a graph neural network architecture, and the graph neural network architecture outputs a lane-level intersection topology.

2. The method for detecting intersection lane-level topology based on trajectory data according to claim 1, characterized in that: The constructing a trajectory relationship graph according to the shape feature, the direction feature and the relationship feature between the trajectories comprises: The shape feature and the direction feature are used as node features, Using the relationship features between the trajectories as edge features; The trajectory relationship graph is generated according to the node features and the edge features.

3. The intersection lane level topology detection method based on trajectory data according to claim 2 is characterized in that: The node features include at least one of trajectory distortion, trajectory curvature, trajectory azimuth change, start point and end point azimuth, and start point to end point azimuth, and the edge features include at least one of the Hausdorff distance between trajectories and the distance sequence of sampling points between trajectories, wherein: The calculation formula of the Hausdorff distance between the trajectories is: Among them, d(p,τ B ) represents the distance from point p to τ B The minimum distance between any points in A ,q) represents τ A The minimum distance from any point in to point q; The distance sequence of sampling points between the trajectories is: d i =∥p i -q i ∥; D sp =<d1,d2,…,d m >; Among them, m is a fixed number of uniformly sampled samples at equal intervals, and p i is τ A The corresponding sampling point in q i is τ B The corresponding sampling point in , d i is the sampling point q i distance.

4. The method for detecting intersection lane-level topology based on trajectory data according to claim 1, characterized in that: The graph neural network architecture includes a node embedding component and a node classification component, wherein: The node embedding component includes a plurality of node embedding blocks, each embedding block includes a plurality of fully connected layers, a graph learning layer and a normalization layer, each fully connected layer uses ReLU as an activation function, in each embedding block, the node features are input into the plurality of fully connected layers to obtain edge features between nodes, the node features and the edge features between the nodes are input into the graph learning layer, and the output of the graph learning layer is normalized using the normalization layer; The node classification component includes a classification block consisting of two fully connected layers. The first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the Sigmoid activation function to predict the probability of each node in the graph as a lane trajectory. During the prediction process, if the predicted probability of the node exceeds the preset probability, the node is marked as a lane trajectory.

5. The method for detecting intersection lane-level topology based on trajectory data according to claim 4, characterized in that: The graph learning layer includes a graph and a multi-head attention mechanism, where: A graph is generated based on the node features and the edge features between the nodes, and the d-th head attention of the multi-head attention mechanism for each edge from j to i is: q c,i =W c,q x i +b c,q have been c,ij =W c,e have been ij +b c,e k c,j =W c,k x j +b c,k Among them, x i is the node feature vector of the i-th node, x j is the node feature vector of the jth node, q c,i is the query vector of the cth attention head of the ith node, W c,q and b c,q To calculate q c,i The trainable parameters, k c,j is the key vector of the cth attention head of the jth node, W c,k and b c,k To calculate k c,i The trainable parameters, e ij for e c,ij is the edge feature vector between the i-th and j-th nodes, W c,e and b c,e To calculate e c,ij The trainable parameters of c,ij is the attention score between the i-th node and the j-th node; When the number of attention heads is h, the updated node features are: v c,j =W c,v x j +b c,v Among them, ∥ represents the serial operation, v c,j is the value vector of the cth attention head of the jth node, W c,v and b c,v Calculate v c,j The trainable parameters, x i ′ is the updated node feature.

6. The method for detecting intersection lane-level topology based on trajectory data according to claim 4, characterized in that: The graph neural network architecture is trained based on a set of road network extraction loss functions.

7. The method for detecting intersection lane-level topology based on trajectory data according to claim 6, characterized in that: The bipartite matching step of the set-based road network extraction loss function includes: Obtain the real lane set and trajectory prediction set; Matching the real lane set with the trajectory prediction set, determining an optimal arrangement according to a bisection matching algorithm, and calculating a minimum matching cost according to the optimal arrangement; Decomposing the minimization matching cost into a classification matching cost and a position matching cost, wherein the classification matching cost is calculated for a non-empty real lane, and the position matching cost is normalized by Hausdorff distance; The total matching cost is calculated based on the position matching cost and classification matching cost after the Hausdorff distance normalization; The prediction model is optimized according to the total matching cost.

8. The method for detecting intersection lane-level topology based on trajectory data according to claim 7, characterized in that: Determine the optimal arrangement according to the binary matching algorithm, and calculate the minimum matching cost according to the optimal arrangement, including: Calculate the set-based road network extraction loss function: in, is the best allocation, The output of the model is represented by The probability that the indexed trajectory is classified as a lane trajectory.

9. An intersection lane-level topology detection device based on trajectory data, characterized in that: include: An acquisition module, used to obtain trajectory data inside the intersection; An extraction module, used to extract shape features, direction features and relationship features between trajectories in the trajectory data; An output module is used to construct a trajectory relationship graph according to the shape features, the direction features and the relationship features between the trajectories, input the trajectory relationship graph into a graph neural network architecture, and the graph neural network architecture outputs a lane-level intersection topology.

10. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the intersection lane-level topology detection method based on trajectory data as described in any one of claims 1 to 8.

Images