Intersection center line automatic generation method and system based on vehicle trajectory data

Abstract

The invention discloses an intersection center line automatic generation method based on vehicle trajectory data. The intersection center line automatic generation method comprises the following steps: acquiring own vehicle trajectory data and other vehicle trajectory data of a target intersection area; generating a trajectory density map representing a vehicle driving mode based on the own vehicle trajectory data and the other vehicle trajectory data; inputting the trajectory density map into a trained intersection center line generation model to obtain vector data which is output by the intersection center line generation model and represents a local center line; and converting the vector data of the plurality of local center lines into a unified coordinate system, and splicing the vector data to generate a complete center line of the target intersection. According to the method, a traditional mode depending on manual drawing is thoroughly changed, track data are automatically processed through an algorithm model, a center line vector is output, and drawing personnel are liberated from heavy manual work.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of automatic driving, in particular to a method and system for automatically generating intersection centerlines based on vehicle trajectory data. BACKGROUND

[0002] Lane-level high-precision maps, as a key input for autonomous driving systems, provide indispensable environmental information for vehicle positioning, planning, and control modules. Such maps are typically constructed by professional survey vehicles equipped with laser radars, optical cameras, and other sensors, which acquire point cloud and image data of the road, and automatically identify and extract lane lines, curbs, traffic signs, and other map elements based on these data.

[0003] Among the many elements of lane-level maps, intersection centerlines are crucial for guiding vehicles to safely and legally navigate complex intersections. However, unlike lane lines on road segments, which have clear visual markings, intersection centerlines often lack corresponding physical markings in real-world roads, making them difficult to automatically detect and extract through geometric or texture features of point clouds or images. This characteristic has led to the long-term reliance on manual drawing by cartographers, resulting in high production costs, low update efficiency, and technical bottlenecks that hinder the large-scale production and rapid iteration of high-precision maps.

[0004] To reduce reliance on manual drawing, existing technologies have proposed some automated or semi-automated solutions, with the mainstream method mainly based on mathematical curve fitting of the two endpoints of the centerline. While this method introduces some automation, its practical application still has significant limitations: on the one hand, the parameter model relied upon by curve fitting cannot be generalized to complex intersections of various shapes (such as irregular intersections, roundabouts, and multiple diverging roads), leading to deviations between the generated centerline geometry and real vehicle driving patterns; on the other hand, such methods generally lack the ability to perceive and avoid physical obstacles in intersection areas (such as safety islands, flow guide islands, and greenery flowerpots), and the generated centerline may cross non-drivable areas, thus still requiring extensive manual verification and correction in practical applications. SUMMARY

[0005] To overcome the above technical defects, the present application provides a method and system for automatically generating intersection centerlines based on vehicle trajectory data.

[0006] To solve the above problems, the present application is implemented according to the following technical solutions:

[0007] In a first aspect, the present application provides a method for automatically generating a centerline of an intersection based on vehicle trajectory data, comprising the following steps: obtaining ego vehicle trajectory data and other vehicle trajectory data of a target intersection region; generating a trajectory density map representing a vehicle driving pattern based on the ego vehicle trajectory data and the other vehicle trajectory data; inputting the trajectory density map into a trained intersection centerline generation model to obtain vector data representing a local centerline output by the intersection centerline generation model; converting the vector data of a plurality of local centerlines to a unified coordinate system and splicing them to generate a complete centerline of the target intersection.

[0008] In combination with the first aspect, the present application provides a first specific implementation of the first aspect, in particular, the ego vehicle trajectory data and the other vehicle trajectory data of the target intersection region are obtained by the following steps: collecting positioning information, point cloud data and image data of the ego vehicle through a vehicle-mounted sensor; based on the point cloud data and the image data, a multi-modal 3D target detection algorithm is used to detect and output three-dimensional center point coordinates of other vehicles in the target intersection region to generate other vehicle driving trajectories.

[0009] In combination with the first aspect, the present application provides a second specific implementation of the first aspect, in particular, based on the point cloud data and the image data, a multi-modal 3D target detection algorithm is used to detect and output three-dimensional center point coordinates of other vehicles in the target intersection region to generate other vehicle driving trajectories, which specifically includes the following steps: using a Transformer encoder to extract feature embeddings of the point cloud data and the image data respectively; projecting image features to a bird's eye view space to align them with point cloud feature space; splicing and fusing point cloud features and image features to form a unified multi-modal bird's eye view feature map; based on the fused features, a Transformer decoder outputs three-dimensional center point coordinates of all detected vehicles.

[0010] In combination with the first aspect, the present application provides a third specific implementation of the first aspect, in particular, the trajectory density map representing the vehicle driving pattern is generated by the following steps: sampling the ego vehicle trajectory at a preset time interval to obtain discrete trajectory points; centering on each trajectory point, a region with a specified edge length is demarcated; superimposing the ego vehicle trajectory points and the other vehicle trajectory points in the region and rendering them into a single-channel grayscale image; wherein the grayscale value of each pixel is determined based on the density or existence of trajectory points in the corresponding grid cell.

[0011] In combination with the first aspect, the present application provides a fourth specific implementation of the first aspect. Specifically, the step of splicing the vector data of the plurality of local center lines is a sequential splicing process based on the driving order of the ego vehicle, including: initializing a historical center line set with a local center line corresponding to a first ego vehicle trajectory point; for each subsequent local center line corresponding to an ego vehicle trajectory point, calculating the matching cost between the local center line and each center line in the historical center line set; based on the matching cost, matching the current local center line with the historical center line set; merging the matched center line pairs and updating the historical center line set; iterating the above process until all local center lines are processed.

[0012] In combination with the first aspect, the present application provides a fifth specific implementation of the first aspect. Specifically, the matching cost is Chamfer distance, and the Hungarian algorithm is used for matching; only the center line pairs that are successfully matched and have Chamfer distance lower than a preset threshold are merged.

[0013] In the second aspect, the present application also provides a vehicle trajectory data-based intersection center line automatic generation system, including: a data acquisition module for acquiring ego vehicle trajectory data and other vehicle trajectory data of a target intersection area; a density map generation module for generating a trajectory density map representing vehicle driving patterns based on the ego vehicle trajectory data and the other vehicle trajectory data; a center line inference module for inputting the trajectory density map into a trained intersection center line generation model to obtain vector data representing local center lines output by the intersection center line generation model; and a center line splicing module for converting the vector data of the plurality of local center lines to a unified coordinate system and splicing to generate a complete center line of the target intersection.

[0014] In combination with the second aspect, the present application provides a first specific implementation of the second aspect. Specifically, the data acquisition module includes: a data acquisition unit for acquiring positioning information, point cloud data and image data of the ego vehicle through a vehicle-mounted sensor; and an other vehicle trajectory detection unit for detecting and outputting three-dimensional center point coordinates of other vehicles in the target intersection area based on the point cloud data and the image data using a multi-modal 3D target detection algorithm to generate other vehicle driving trajectories.

[0015] In combination with the second aspect, the present application provides a second specific implementation of the second aspect. Specifically, the other vehicle trajectory detection unit performs the following steps: using a Transformer encoder to extract feature embeddings of the point cloud data and the image data respectively; projecting the image features to a bird's eye view space to align them with the point cloud feature space; splicing and fusing the point cloud features and the image features to form a unified multi-modal bird's eye view feature map; and based on the fused features, a Transformer decoder outputs the three-dimensional center point coordinates of all detected vehicles.

[0016] In combination with the second aspect, the present application provides a third implementation of the second aspect, in particular, the density map generation module is specifically configured to: sample the ego vehicle trajectory at a preset time interval to obtain discrete trajectory points; define a region with a specified edge length centered on each trajectory point; superimpose the ego vehicle trajectory points and the other vehicle trajectory points in the region, and render them into a single-channel grayscale image; wherein the grayscale value of each pixel is determined based on the density or existence of trajectory points in the corresponding grid cell.

[0017] Compared with the prior art, the present application has the following beneficial effects:

[0018] The present application completely changes the traditional mode of relying on manual drawing, automatically processes trajectory data and outputs center line vectors through algorithm model, and liberates the cartographer from heavy manual work. This greatly improves the production and updating efficiency of lane-level maps, and greatly reduces the labor and time cost. The present application is not based on simple geometric assumptions for curve fitting, but learns and extracts driving patterns from massive real vehicle trajectory data. Therefore, the generated center line can accurately reflect the actual passing path of the vehicle in the intersection, avoiding the deviation of the fitted model from the actual situation, and providing more accurate and reliable guidance for autonomous vehicles. The method directly uses widely collected vehicle trajectory data, which is rich in data sources and relatively low in acquisition cost. Through the fusion of "ego vehicle + other vehicle" data, the passing mode of the intersection can be more comprehensively represented, making the generated center line result more objective and reliable, and reducing the error caused by subjective judgment. BRIEF DESCRIPTION OF DRAWINGS

[0019] The specific embodiments of the present application will be further described in detail below in combination with the drawings, in which:

[0020] Figure 1 is a flowchart of a vehicle trajectory data-based intersection center line automatic generation method of the present application. DETAILED DESCRIPTION

[0021] The preferred embodiments of the present application will be described below in combination with the drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and not to limit the present application.

[0022] As shown in Figure 1 , a vehicle trajectory data-based intersection center line automatic generation method and system according to the present application.

[0023] Example 1

[0024] The application provides a kind of intersection center line automatic generation method based on vehicle trajectory data, comprising the following steps: obtaining the trajectory data of self-vehicle and other vehicle trajectory data in the target intersection area;Based on the trajectory data of self-vehicle and other vehicle trajectory data, a trajectory density map representing vehicle driving mode is generated;The trajectory density map is input into the trained intersection center line generation model to obtain the vector data representing the local center line output by the intersection center line generation model;The vector data of a plurality of local center lines is converted to a unified coordinate system and spliced to generate a complete center line of the target intersection.

[0025] Specifically, the embodiment takes a city intersection as an example, obtains the self-vehicle trajectory data within 200 meters around the target intersection from the high-precision map collection team, and obtains the other vehicle trajectory data (GPS data) in the region from the partner within a week. The original data is preprocessed as follows: eliminate abnormal trajectory points (such as positioning drift points) that deviate from the road, convert the coordinates of all trajectory points to WGS-84 coordinate system, and prevent data from being too dense by resampling all trajectories at a time interval to ensure uniformity of trajectory points. The target intersection area (e.g. 200m x 200m) is divided into 0.2m x 0.2m grids. All preprocessed trajectory points are traversed, and the number of trajectory points in each grid is counted. Then, the statistical results are smoothed using a Gaussian kernel function, and the frequency is normalized to the interval of 0-255, and finally a gray trajectory density map is generated.

[0026] The generated gray trajectory density map (resolution adjusted to 512x512) is input into the pre-trained U-Net model. U-Net uses an encoder-decoder structure, and outputs a probability map Y ^ ∈[0,1] 512×512 , where each pixel value represents the probability of belonging to the center line. The model is trained using thousands of pairs of manually annotated "trajectory density map-center line map", and the center line map is a binary image where the center line pixel value is 1 and the others are 0. A combination of binary cross-entropy loss (BCE) and Dice loss is used:

[0027] Binary cross-entropy loss: where y i is the true label, is the predicted probability, and N is the total number of pixels.

[0028] Dice loss:

[0029] Total loss: L seg =L BCE +λL Dice ; where λ is the weight, usually set to 1.

[0030] The probability map is thresholded (e.g., threshold 0.5), then a skeletonization algorithm (e.g., Zhang-Suen algorithm) is used to extract the centerline skeleton, and finally vectorized into a set of vector polylines L = {L1, L2, …, L k} where each L i is a local centerline. All the local centerline vector data output by the model are unified into the UTM coordinate system through a pre-computed transformation matrix. Then, the following strategy is used for splicing: for two local centerline endpoints with a distance less than a threshold (e.g., 1 meter) and a direction angle less than a threshold (e.g., 15 degrees), it is determined that they are breakpoints of the same centerline. A cubic B-spline curve is used to connect these matched endpoints, ensuring smooth transition at the connection. Finally, a complete, smooth, and topologically correct centerline network of the intersection is generated. In a preferred embodiment, the obtaining of the ego vehicle trajectory data and the other vehicle trajectory data in the target intersection region specifically includes the following steps: collecting positioning information, point cloud data, and image data of the ego vehicle through a vehicle-mounted sensor; based on the point cloud data and the image data, a multi-modal 3D target detection algorithm is used to detect and output three-dimensional center point coordinates of other vehicles in the target intersection region, to generate a driving trajectory of the other vehicle.

[0031] Specifically, a professional collection vehicle drives into the target intersection region, and a collection system integrated on the roof of the vehicle works synchronously. GNSS / IMU provides positioning information with a frequency of 100 Hz and an absolute accuracy better than 5 cm; a 64-line laser radar scans the surrounding environment at a frequency of 10 Hz to generate point clouds; multiple high-definition cameras collect 360° surround view images at a frame rate of 30 fps. A multi-modal 3D target detection algorithm runs in real time on a vehicle-mounted industrial computer. The algorithm fuses the vehicle 2D bounding box and semantic information recognized from the camera image with the laser radar point cloud to detect the three-dimensional bounding box of the surrounding vehicles. Then, the center point at the bottom of each bounding box is extracted as the 3D position of the vehicle in the current frame. A multi-target tracking algorithm (such as SORT or DeepSORT) is used to associate the same vehicle in different frames, and finally form a continuous and smooth 3D trajectory of each other vehicle in the intersection region. The ego vehicle trajectory and all the generated other vehicle trajectories are fused into a trajectory dataset. Then, the dataset is cleaned to remove short tracking loss or jump trajectory segments caused by occlusion and other reasons.

[0032] In a preferred embodiment, the step of detecting and outputting the three-dimensional center point coordinates of other vehicles within the target intersection area based on the point cloud data and image data using a multimodal 3D target detection algorithm to generate the driving trajectories of other vehicles specifically includes the following steps: using a Transformer encoder to extract feature embeddings from the point cloud data and image data respectively; projecting the image features onto the bird's-eye view space to align it with the point cloud feature space; stitching and fusing the point cloud features and image features to form a unified multimodal bird's-eye view feature map; and having the Transformer decoder regress and output the three-dimensional center point coordinates of all detected vehicles based on the fused features.

[0033] Specifically, a Transformer encoder architecture is used to extract feature embeddings from point cloud data and image data, respectively. Specifically, the point cloud data is voxelized and input into a Transformer-based point cloud encoder, which outputs point cloud features. Where, N v C is the number of voxels, and C is the feature dimension. Image data is input to a Transformer-based image encoder, which outputs image features. Where H and W are the height and width of the image feature map. The image features contain rich two-dimensional texture and semantic features. The image features are projected into the bird's-eye view space through perspective transformation or a depth estimation-based method, aligning them spatially with the point cloud features natively existing in the bird's-eye view (BEV) space. The projection formula, based on perspective transformation or depth estimation, generates the projection matrix P. The image feature projection formula is: Where X and Y are the dimensions of the BEV mesh. Within the bird's-eye view space, for features F from the point cloud... point and features from the projected image By stitching and merging the data, a unified and information-rich multimodal bird's-eye view feature map is formed: The features are then compressed using a convolutional layer to obtain the fused features. This fusion process fully utilizes the semantic information of the images and the precise geometric information of the point clouds. The Transformer decoder then uses the fused multimodal bird's-eye view feature map F... fusion The regression outputs the 3D bounding boxes of all detected vehicles, and outputs the 3D bounding box parameters of the detected vehicle. Where (x) i ,y i ,z i ) represents the coordinates of the center point, w i ,l i ,h i It is the size, θ iIt is the direction angle; extract the coordinates of the center point at the bottom of the bounding box as the position of the other vehicle.

[0034] The SORT or DeepSORT algorithm is used to associate the same vehicle in different frames, generating continuous trajectories of other vehicles. The training loss function of the multimodal 3D object detection algorithm includes classification loss, regression loss, and orientation loss: L det =L cls +L reg +L dir Classification loss L cls Focal Loss is used to distinguish between vehicles and non-vehicles: Where p i This represents the predicted probability, and α and γ are hyperparameters. The regression loss L... reg Using Smooth L1 Loss for bounding box parameter regression: Where b j It is the actual value. This is the predicted value. Directional loss L dir Using a sinusoidal loss function for the direction angle θ i :

[0035] In a preferred embodiment, generating a trajectory density map representing the vehicle's driving mode specifically includes the following steps: sampling the vehicle's trajectory at preset time intervals to obtain discrete trajectory points; defining an area with a specified side length centered on each trajectory point; superimposing the vehicle's trajectory points within the area with the trajectory points of other vehicles, and rendering it as a single-channel grayscale image; wherein the grayscale value of each pixel is determined based on the density or existence of trajectory points within the corresponding grid cell.

[0036] Specifically, the preprocessed trajectory data (both the vehicle's and other vehicles') is resampled at preset time intervals (e.g., 0.1 seconds) to obtain a set of discrete trajectory points. Divide the target intersection area into a grid (e.g., 0.2m × 0.2m), with a grid size Δ = 0.2m. Initialize an all-zero matrix. (Corresponding mesh, H = W = 200 / 0.2 = 1000). Traverse all preprocessed trajectory points P = {x, y} and find their corresponding mesh element g. ij A local area R(p) is defined centered on this point (corresponding to a 5×5 grid, i.e., 1m×1m). For each grid cell g∈R(p), the count is incremented: C(g)←C(g)+1. A Gaussian kernel function is used to smooth matrix C, generating a density map D. The Gaussian kernel function is G(x,y)= Where σ is the standard deviation (e.g., σ = 0.5), the smoothing operation is achieved through convolution: D = C * C, and D is normalized to the range of 0-255. The final grayscale trajectory density map was obtained, and the resolution was adjusted to 512×512.

[0037] In a preferred embodiment, the step of stitching together the vector data of multiple local centerlines is specifically an ordered stitching process based on the vehicle's driving sequence, including: initializing a historical centerline set with the local centerline corresponding to the first vehicle trajectory point; calculating the matching cost between the local centerline corresponding to each subsequent vehicle trajectory point and each centerline in the historical centerline set; matching the current local centerline with the historical centerline set based on the matching cost; merging the successfully matched centerline pairs and updating the historical centerline set; iterating the above process until all local centerlines have been processed.

[0038] Specifically, an ordered stitching algorithm based on the vehicle's driving sequence is executed: Let the set of historical centerlines S = {L1}, where L1 is the local centerline corresponding to the first vehicle's trajectory point. For each subsequent local centerline L... k (k = 2, 3, ..., K), execute: calculate L using Chamfer distance. k With each centerline L in S s The cost. Chamfer distance formula Among them, |L k | and |L s | represent the center line L respectively k and L s The number of points on the L, where || is the Euclidean distance. Optimal matching is achieved using the Hungarian algorithm: Let L... k The center line in S is considered a bipartite graph. The pairing with the minimum total matching cost is found using the Hungarian algorithm, only if d Chamfer (L k ,L s ) < τ (threshold such as 1 meter) and the included angle of direction is less than θ th A successful match is determined when the angle is 15 degrees. For successfully matched centerline pairs, a cubic B-spline curve is used for smooth connection. Given a control point sequence P0, P1, ..., P... n cubic B-spline curve segment Q i (t) parameterized as: A new centerline L is generated after connection. new Use L new Replace L in S s If L k If no centerline is matched, it is added to S. The process is repeated until all local centerlines are processed, and finally S contains the complete centerline network.

[0039] In a preferred embodiment, the matching cost is Chamfer distance, and the Hungarian algorithm is used for matching; only the center lines with successful matching and Chamfer distance lower than a preset threshold are merged.

[0040] Embodiment 2

[0041] The application further provides a vehicle trajectory data-based intersection center line automatic generation system, comprising: a data acquisition module configured to acquire ego trajectory data and other vehicle trajectory data of a target intersection region; a density map generation module configured to generate a trajectory density map representing vehicle driving patterns based on the ego trajectory data and the other vehicle trajectory data; a center line inference module configured to input the trajectory density map into a trained intersection center line generation model to obtain vector data representing local center lines output by the intersection center line generation model; and a center line splicing module configured to convert the vector data of the plurality of local center lines to a unified coordinate system and splice them to generate a complete center line of the target intersection.

[0042] In a preferred embodiment, the data acquisition module comprises: a data acquisition unit configured to acquire positioning information, point cloud data and image data of an ego vehicle through a vehicle-mounted sensor; and an other vehicle trajectory detection unit configured to detect and output three-dimensional center point coordinates of other vehicles in the target intersection region based on the point cloud data and the image data by using a multi-modal 3D target detection algorithm to generate other vehicle driving trajectories.

[0043] In a preferred embodiment, the other vehicle trajectory detection unit performs the following steps: using a Transformer encoder to extract feature embeddings of the point cloud data and the image data respectively; projecting image features to a bird's eye view space to align them with point cloud feature space; splicing and fusing the point cloud features and the image features to form a unified multi-modal bird's eye view feature map; and using a Transformer decoder to regress and output three-dimensional center point coordinates of all detected vehicles based on the fused features.

[0044] In a preferred embodiment, the density map generation module is specifically configured to: sample the ego trajectory at preset time intervals to obtain discrete trajectory points; define a region with a specified edge length centered on each trajectory point; superimpose the ego trajectory points and the other vehicle trajectory points in the region and render them into a single-channel grayscale image; and determine the grayscale value of each pixel based on the density or existence of trajectory points in the corresponding grid cell.

[0045] The above description is only a preferred embodiment of the present application, and does not limit the present application in any form. Any modification, equivalent change and modification of the above embodiment based on the technical essence of the present application, without departing from the technical solution of the present application, are still within the scope of the present application.

Claims

1. A method for automatically generating intersection centerlines based on vehicle trajectory data, characterized in that, Includes the following steps: Acquire the trajectory data of your own vehicle and other vehicles in the target intersection area; A trajectory density map representing the vehicle's driving mode is generated based on the vehicle's trajectory data and the trajectory data of other vehicles. The trajectory density map is input into the trained intersection centerline generation model to obtain vector data representing the local centerline output by the intersection centerline generation model. The vector data of multiple local centerlines are converted to a unified coordinate system and then stitched together to generate the complete centerline of the target intersection.

2. The method for automatically generating intersection centerlines based on vehicle trajectory data according to claim 1, characterized in that, The acquisition of the vehicle trajectory data and other vehicle trajectory data in the target intersection area specifically includes the following steps: The vehicle's location information, point cloud data, and image data are collected through onboard sensors. Based on the point cloud data and image data, a multimodal 3D target detection algorithm is used to detect and output the three-dimensional center point coordinates of other vehicles in the target intersection area, so as to generate the driving trajectory of other vehicles.

3. The method for automatically generating intersection centerlines based on vehicle trajectory data according to claim 2, characterized in that, Based on the point cloud data and image data, a multimodal 3D target detection algorithm is used to detect and output the three-dimensional center point coordinates of other vehicles within the target intersection area to generate their driving trajectories. This process specifically includes the following steps: The Transformer encoder is used to extract feature embeddings from point cloud data and image data, respectively. Project the image features into the bird's-eye view space to align them with the point cloud feature space; Point cloud features and image features are stitched and fused to form a unified multimodal bird's-eye view feature map; Based on the fused features, the Transformer decoder regresses and outputs the three-dimensional center point coordinates of all detected vehicles.

4. The method for automatically generating intersection centerlines based on vehicle trajectory data according to claim 1, characterized in that, The generation of the trajectory density map representing the vehicle's driving mode specifically includes the following steps: The vehicle trajectory is sampled at preset time intervals to obtain discrete trajectory points; Delineate an area with a specified side length centered on each of the aforementioned trajectory points; The trajectory points of the vehicle within the specified area are superimposed with the trajectory points of other vehicles and rendered as a single-channel grayscale image; wherein, the grayscale value of each pixel is determined based on the density or existence of trajectory points within the corresponding grid cell.

5. The method for automatically generating intersection centerlines based on vehicle trajectory data according to claim 1, characterized in that, The step of stitching together the vector data of multiple local centerlines is specifically an ordered stitching process based on the vehicle's driving sequence, including: Initialize a set of historical centerlines using the local centerline corresponding to the first vehicle trajectory point; For each subsequent local centerline corresponding to a vehicle trajectory point, calculate the matching cost between it and each centerline in the historical centerline set. Based on the matching cost, the current local centerline is matched with the set of historical centerlines; Merge successfully matched centerline pairs and update the historical centerline set; Iterate through the above process until all local centerlines have been processed.

6. The method for automatically generating intersection centerlines based on vehicle trajectory data according to claim 5, characterized in that, The matching cost is the Chamfer distance, and the Hungarian algorithm is used for matching; only centerline pairs that are successfully matched and whose Chamfer distance is less than a preset threshold are merged.

7. An automatic intersection centerline generation system based on vehicle trajectory data, characterized in that, include: The data acquisition module is used to acquire the trajectory data of the vehicle itself and the trajectory data of other vehicles in the target intersection area; The density map generation module is used to generate a trajectory density map representing the vehicle's driving mode based on the vehicle's trajectory data and other vehicle trajectory data. The centerline inference module is used to input the trajectory density map into the trained intersection centerline generation model to obtain vector data representing the local centerline output by the intersection centerline generation model. The centerline stitching module is used to convert the vector data of multiple local centerlines to a unified coordinate system and stitch them together to generate the complete centerline of the target intersection.

8. The automatic intersection centerline generation system based on vehicle trajectory data according to claim 7, characterized in that, The data acquisition module includes: The data acquisition unit is used to collect the vehicle's positioning information, point cloud data, and image data through onboard sensors; The vehicle trajectory detection unit is used to detect and output the three-dimensional center point coordinates of other vehicles in the target intersection area based on the point cloud data and image data, using a multimodal 3D target detection algorithm, so as to generate the driving trajectory of other vehicles.

9. The automatic intersection centerline generation system based on vehicle trajectory data according to claim 8, characterized in that, The vehicle trajectory detection unit performs the following steps: The Transformer encoder is used to extract feature embeddings from point cloud data and image data, respectively. Project the image features into the bird's-eye view space to align them with the point cloud feature space; Point cloud features and image features are stitched and fused to form a unified multimodal bird's-eye view feature map; Based on the fused features, the Transformer decoder regresses and outputs the three-dimensional center point coordinates of all detected vehicles.

10. The automatic intersection centerline generation system based on vehicle trajectory data according to claim 7, characterized in that, The density map generation module is specifically configured as follows: The vehicle trajectory is sampled at preset time intervals to obtain discrete trajectory points; Delineate an area with a specified side length centered on each of the aforementioned trajectory points; The trajectory points of the vehicle within the specified area are superimposed with the trajectory points of other vehicles and rendered as a single-channel grayscale image; wherein, the grayscale value of each pixel is determined based on the density or existence of trajectory points within the corresponding grid cell.

Images