Goal-oriented trajectory prediction system based on graph neural networks

Abstract

The application belongs to the technical field of automatic driving simulation test, and particularly relates to a target-oriented trajectory prediction system based on a graph neural network.The trajectory prediction system comprises a vector representation module, a local information fusion module, a global information interaction module and a decoding module; the vector representation module is connected with the local information fusion module; the local information fusion module is connected with the global information interaction module; and the decoding module is connected with the global information interaction module and the local information fusion module.The application can be used for planning and predicting the driving behaviors of a self vehicle and other vehicles in automatic driving, and can provide the function of inputting target points, so that different prediction results can be obtained by setting different target points, so that the model predicted trajectory is more controllable, has stronger explainability, and has stronger anti-interference for inaccurate target point input.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of automatic driving simulation test, and particularly relates to a target-oriented trajectory prediction system based on a graph neural network method. BACKGROUND

[0002] The automatic driving technology is still in continuous development. Before this technology is widely used, a large amount of testing still needs to be carried out to verify its usability and safety. A commonly used testing method is to carry out field road testing on a closed road. This method can obtain relatively real data, but its efficiency is low, and it is not conducive to the reproduction of the scene. In contrast, it is more efficient and safer to simulate the driving system in a virtual environment. However, a big problem in the virtual environment is to simulate the virtual vehicles in the environment. In the simulation environment, as the driving strategy of the host vehicle changes, the other virtual vehicles should respond as much as possible with real driving behavior, and should be able to generate a variety of reasonable driving trajectories. SUMMARY

[0003] The purpose of the present application is to provide a target-oriented trajectory prediction system based on a graph neural network, which has strong anti-interference ability and high accuracy in target-oriented trajectory prediction.

[0004] The target-oriented trajectory prediction system based on a graph neural network provided by the present application specifically sets a target point for each vehicle. The target point is not the specific position that the vehicle needs to reach, but serves as a reference for the driving intention direction. Therefore, the present application can predict a reasonable trajectory for each vehicle with the target point direction as the driving intention. Specifically, the target-oriented trajectory prediction system based on a graph neural network comprises a vector representation module, a local information fusion module, a global information interaction module, and a decoding module. The vector representation module is connected to the local information fusion module. The local information fusion module is connected to the global information interaction module. The decoding module is connected to the global information interaction module and the local information fusion module. Wherein:

[0005] The vector representation module is used to pre-process the historical trajectory of the vehicle and the map information, vectorize the position and movement distance of the vehicle trajectory at each time, and extract the lane geographical position information in the map for vectorization, so as to process the vehicle trajectory vector representation and the map information vector representation of the local area of each vehicle.

[0006] The local information fusion module is used to obtain the vectorization of the vehicle trajectory and the lane position from the vector representation module, process the vectorization information of each vehicle and other vehicles in its local area, process the map vectorization information of each vehicle and its local area, and fuse the vehicle and map information in the local area of each vehicle to obtain the local information fusion representation of each vehicle.

[0007] The global information interaction module is configured to obtain the local information fusion representation of each vehicle from the local information fusion module, and use a graph neural network to model and fuse the local information fusion representations of different vehicles to obtain a global information interaction representation of each vehicle.

[0008] The decoding module is configured to obtain the global information interaction representation of each vehicle from the global information interaction module, and accept a coordinate input of a target point, and then decode a plurality of target-oriented predicted trajectories of each vehicle according to the global information interaction representation of each vehicle and the target point.

[0009] Further, the vector representation module specifically comprises:

[0010] The vehicle information vector representation unit is configured to obtain and process information of each vehicle to obtain a vehicle coordinate vector, a vehicle orientation angle, and a vehicle displacement vector relative to a previous time stamp corresponding to each time stamp, wherein the time stamp size Δt is 0.1 s, and the vehicle history trajectory vector contains vehicle coordinates and orientation angles at 20 continuous time points.

[0011] The map information vector representation unit is configured to obtain a vector representation of a map lane center line from a high-definition map, including a number of each lane center line and coordinates of a point set contained in the lane center line.

[0012] The local information screening and processing unit is configured to obtain vector representations from the vehicle information vector representation unit and the map information vector representation unit, screen out vehicle trajectory information vectors and map information vectors in a local area range of each vehicle, obtain vector representations of local information of each vehicle, and rotate and translate the vector representations of all local information so that the coordinates and orientation angles of the vehicle itself are (0, 0) and 0° respectively, and the transformed coordinates are:

[0013]

[0014] The transformed orientation angle θ' = θ - θ 0 ,

[0015] wherein the coordinates of the vehicle at the current time are (x 0 ,y 0 ), the orientation angle at the current time is θ 0 , the coordinates of the local information vector to be transformed are (x, y), the orientation angle is θ, the coordinates of the transformed local information vector are (x', y'), and the orientation angle at the current time is θ'.

[0016] Further, the local information fusion module specifically comprises:

[0017] An inter-vehicle interaction unit is configured to obtain a vector representation of local vehicle information of each vehicle of the vector representation module, integrate the information using a Transformer model (Vaswani, A. "Attention is all you need." Advances in Neural Information Processing Systems (2017).) to obtain a local inter-vehicle information fusion representation of each vehicle;

[0018] A vehicle-road interaction unit is configured to obtain a vector representation of lane information near each vehicle of the vector representation module, integrate the information using a time series model to obtain a local vehicle-lane information fusion representation of each vehicle;

[0019] The calculation formula is as follows:

[0020] z t =GNN AA ((x' 1,t ,y' 1,t ),...,(x' n,t ,y' n,t )), (2)

[0021] a 0 =Transformer(z1,z2,...,z t ), (3)

[0022] h 0 =GNN AL (a 0 ,(x' n+1,t ,y' n+1,t ),...,(x' n+m,t ,y' n+m,t )), (4)

[0023] wherein z t is the inter-vehicle information representation vector at time t, GNN AA is a graph neural network algorithm (F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner and G. Monfardini, "The Graph Neural Network Model," in IEEE Transactions on Neural Networks, vol. 20, no. 1, pp. 61-80, Jan. 2009) that integrates inter-vehicle information, (x' 1,t ,y' 1,t ),……, (x' n,t ,y' n,t) is the rotated and translated coordinate of the n vehicles at time t, (x n+1,t ,y' n+1,t ) is the rotated and translated coordinate of the n vehicles at time t, (x n+m,t ,y' n+m,t ) is the rotated and translated coordinate of the n vehicles at time t, (x 0 is the local vehicle information fusion representation of the current vehicle, Transformer is a transformer algorithm, h 0 is the local vehicle lane information fusion representation of the current vehicle, GNN AL is a graph neural network algorithm that fuses vehicle lane information.

[0024] Further, the global information interaction module specifically comprises:

[0025] a graph neural network modeling unit, configured to establish a corresponding node for each vehicle, establish an interaction relationship between vehicles as an edge between nodes corresponding to the vehicles, and perform graph modeling;

[0026] an information interaction unit, configured to acquire a graph of the interaction between vehicles from the graph neural network modeling unit, and acquire a local information fusion representation of the vehicle from the local information fusion module, and then fuse global information between vehicles by using a graph neural network to obtain a global information interaction representation of each vehicle, the calculation formula being as follows:

[0027]

[0028] wherein, is the global information interaction representation of the vehicle numbered k, GNN GI is a graph neural network that fuses global information between vehicles (F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner and G. Monfardini, "The Graph Neural Network Model," in IEEE Transactions on Neural Networks, vol. 20, no. 1, pp. 61-80, Jan. 2009).

[0029] Further, the decoding module specifically comprises:

[0030] a decoding unit, configured to acquire a local information fusion representation of a vehicle from the local information fusion module, acquire a global information interaction representation of the vehicle from the global information interaction module, and acquire a plurality of target point coordinates of each vehicle from the input, and then generate a corresponding predicted trajectory for each target point of each vehicle based on the acquired information, the calculation formula being as follows:

[0031]

[0032] wherein, is the global information interaction representation of the k-th vehicle fusing the i-th target point coordinate, is the i-th predicted trajectory decoded by the k-th vehicle, is the coordinate of the i-th target point, MLP P is the multi-layer perceptron algorithm decoding the target point information, MLP T is the multi-layer perceptron algorithm decoding the predicted trajectory (Murtagh, Fionn. "Multilayer perceptrons for classification and regression." Neurocomputing 2.5-6 (1991): 183-197.).

[0033] The application also provides a target-oriented trajectory prediction system based on a graph neural network, which further comprises a vehicle-mounted laser radar and a GPS positioning system, which are used to detect and identify the positions of the ego vehicle, adjacent vehicles and obstacles in real time, and process the coordinates, length, width and orientation data of the vehicles and obstacles, specifically including the identified positions of the ego vehicle at the current time T (x 1,T ,y 1,T ), the orientation angle is θ 1,T , the identified positions of the adjacent vehicles and obstacles at the current time T are (x 2,T ,y 2,T ),..., (x n,T ,y n,T ), the orientation angles are θ 2,T ,..., θ n,T .

[0034] Further, by using high-precision map data, lane data near the ego vehicle is obtained and processed into lane vector representation, and the high-precision map data vector representation identified at a time stamp t before the current time is (x n+1,T-t ,y n+1,T-t ),..., (x n+m,T-t ,y n+m,T-t ).

[0035] Further, a historical data storage unit is used to store the positions of the ego vehicle, adjacent vehicles and obstacles identified in the past period of time in the storage medium, and the historical position of the ego vehicle at a time stamp t before the current time is (x 1,T-t ,y 1,T-t , the historical positions of the adjacent vehicles and obstacles are (x 2,T-t ,y 2,T-t ),..., (x n,T-t ,y n,T-t ).

[0036] In the present application, based on vehicle data, obstacle data and high-precision map data, the target-oriented trajectory prediction system based on graph neural network is trained using the gradient descent method, 60 consecutive timestamp segments are intercepted from the historical data storage unit, the data of the first 20 timestamps are input as the vehicle data, obstacle data and high-precision map data of the target-oriented trajectory prediction system based on graph neural network, the positions of the vehicles at the uniformly randomly selected timestamps in the 40th to 60th timestamps are input as the target point data of the target-oriented trajectory prediction system based on graph neural network, and the predicted trajectory T of each vehicle is obtained k The loss function L of the target-oriented trajectory prediction system based on graph neural network is calculated according to the following formula:

[0037] L = MSE (T k ,<(x k,21 ,y k,21 ),(x k,22 ,y k,22 ),...,(x k,50 ,y k,50 )>), (8)

[0038] Wherein, MSE is the root mean square error;

[0039] The target-oriented trajectory prediction system based on graph neural network is used to input vehicle data, obstacle data, high-precision map data and multiple target point data, and multiple trajectory predictions of the vehicle are obtained.

[0040] Compared with the prior art, the present application has the following advantages:

[0041] The target-oriented trajectory prediction system based on graph neural network can be used for planning and prediction of driving behavior of the ego vehicle and other vehicles in the automatic driving algorithm, and provides the function of inputting target points, so that different prediction results can be obtained by setting different target points, so that the model prediction trajectory is more controllable, and the prediction trajectory has stronger explainability; meanwhile, the target point is uniformly and randomly selected in a certain timestamp during model training, so that the target-oriented trajectory prediction system based on graph neural network has stronger anti-interference performance for inaccurate target point input. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 The figure is a structure diagram of the target-oriented trajectory prediction system based on graph neural network of the present application.

[0043] Figure 2 The figure is a detail diagram of the implementation scenario of the target-oriented trajectory prediction system based on graph neural network of the present application.

[0044] Figure 3 The figure is a simulation experiment diagram of the present application. DETAILED DESCRIPTION

[0045] The application will be further described below with reference to specific examples and accompanying drawings.

[0046] Figure 1 A structure diagram of a target-oriented trajectory prediction system based on a graph neural network provided by the application. It includes a vector representation module, a local information fusion module, a global information interaction module, and a decoding module; the vector representation module is connected to the local information fusion module; the local information fusion module is connected to the global information interaction module; the decoding module is connected to the global information interaction module and the local information fusion module.

[0047] Figure 2 A schematic diagram of an implementation scenario of a target-oriented trajectory prediction system based on a graph neural network provided by the application.

[0048] There are three vehicles in this embodiment, and the black circular points in 201 represent the historical trajectory information of the three vehicles detected and recognized in real time by the vehicle-mounted laser radar and GPS positioning system. The historical trajectory information is represented by the trajectory coordinates of 20 consecutive time points, and the timestamp size is Δt=0.1s. There are 16 lanes in this embodiment, and there are 16 lane center lines near the ego vehicle obtained by using high-precision map data. The black diamond points in 202 represent the center line of one of the lanes, which is represented by a sequence of point coordinates with an interval of 2 meters. In this embodiment, the vehicle in the middle of the three vehicles is the ego vehicle, and the white triangular points and black triangular points in 204 represent the guidance information of two future points of the given ego vehicle. The target-oriented trajectory prediction method based on a graph neural network provided by the application uses the above information, and through the processing of the vector representation module, the local information fusion module, the global information interaction module, and the decoding module, the predicted trajectory of the ego vehicle is obtained. The white square points and black square points in 203 represent the two future trajectories predicted by the system and method provided by the application for the ego vehicle. Among them, the white square predicted trajectory is guided by the white triangular point to form a left turn trajectory, and the black square predicted trajectory is guided by the black triangular point to form a straight trajectory.

[0049] The vector representation module includes a vehicle information vector representation unit, a map information vector representation unit, and a local information screening processing unit. The vehicle information vector representation unit is used to obtain and process the information of each vehicle to obtain the vehicle coordinate vector corresponding to each timestamp, the vehicle orientation angle, and the displacement vector of the vehicle relative to the previous timestamp, with a timestamp size of Δt=0.1s. The vehicle historical trajectory vector includes the vehicle coordinates and orientation angle of 20 consecutive time points. The coordinates of the ith vehicle at the tth timestamp are (x i,t ,y i,t ), and the orientation angle is θ i,t, the displacement vector relative to the previous timestamp is (x i,t -x i,t-1 ,y i,t -y i,t-1 ); the map information vector representation unit is configured to obtain a vector representation of a map lane center line from a high-definition map, including a number of each lane center line and coordinates of a point set contained by the lane center line, assuming that there are n vehicles in total and m points are contained by the lane center line, then the coordinates of the jth point are (x n+j ,y n+j ); the local information screening processing unit is configured to obtain the vector representation from the vehicle information vector representation unit and the map information vector representation unit, and screen out the vehicle trajectory information vector and the map information vector within the local area range of each vehicle to obtain the vector representation of the local information of each vehicle, in the embodiment, the vehicle trajectory information vector and the map information vector within a range of not more than 20 meters from the vehicle are selected; then, in the embodiment, taking the vehicle with the number 1 as an example, the vector representations of all local information of the vehicle are rotated and translated so that the coordinates and the heading angle of the vehicle with the number 1 at the current time T are (0, 0) and 0° respectively, specifically, for the coordinates (x, y) in any vector, the transformed coordinates are:

[0050]

[0051] for any heading angle θ, the transformed heading angle is θ' = θ - θ 1,T

[0052] where the coordinates of the vehicle with the number 1 at the current time T are (x 1,T ,y 1,T ), and the heading angle at the current time is θ1 T .

[0053] Further, the local information fusion module includes an inter-vehicle interaction unit and a vehicle-road interaction unit; the inter-vehicle interaction unit is configured to obtain the vector representation of the local vehicle information of each vehicle of the vector representation module, integrate the information using a time series model to obtain a local inter-vehicle information fusion representation of each vehicle; the vehicle-road interaction unit is configured to obtain the vector representation of the lane information near each vehicle of the vector representation module, integrate the information using a time series model to obtain a local vehicle-lane information fusion representation of each vehicle; in the embodiment, taking the vehicle with the number 1 as an example, the calculation formula is as follows:

[0054] z t =GNN AA ((x' 1,t ,y' 1,t ),...,(x' n,t ,y' n,t ))

[0055] a 1 =Transformer(z1,z2,...,z t )

[0056] h 1 =GNN AL (a 1 ,(x' n+1,t ,y' n+1,t ),...,(x' n+m,t ,y' n+m,t ))

[0057] wherein z t is a vector of length 32, is the inter-vehicle information representation vector of the t-th moment, GNN AA is a graph neural network algorithm for fusing inter-vehicle information, (x' 1,t , y' 1,t ) to (x' n,t , y' n,t ) are the coordinates of n vehicles respectively at the t-th moment after rotation and translation, (x' n+1,t , y' n+1,t ) to (x' n+m,t , y' n+m,t ) are the map information vector representations after rotation and translation, a 1 is a vector of length 32, is the local inter-vehicle information fusion representation of the vehicle numbered 1, Transformer is a transformer algorithm, h 1 is a vector of length 32, is the local vehicle lane information fusion representation of the vehicle numbered 1, GNN AL is a graph neural network algorithm for fusing vehicle lane information.

[0058] Further, the global information interaction module comprises a graph neural network modeling unit and an information interaction unit; the graph neural network modeling unit is configured to establish a corresponding node for each vehicle, establish the interaction relationship between vehicles as an edge between the nodes corresponding to the vehicles, and perform graph modeling; the information interaction unit is configured to obtain the graph of the interaction between vehicles from the neural network modeling unit, and obtain the local information fusion representation of the vehicle from the local information fusion module, and then fuse the global information between vehicles by using the graph neural network to obtain the global information interaction representation of each vehicle, and the calculation formula is as follows:

[0059]

[0060] wherein, is a vector of length 32, is the global information interaction representation of the vehicle numbered k, GNN GI is a graph neural network for fusing global information between vehicles.

[0061] Further, the decoding module comprises a decoding unit; the decoding unit is configured to obtain the local information fusion representation of the vehicle from the local information fusion module, obtain the global information interaction representation of the vehicle from the global information interaction module, and obtain the plurality of target point coordinates of each vehicle from the input, and then generate a corresponding predicted trajectory for each target point of each vehicle based on the obtained information, and the calculation formula is as follows:

[0062]

[0063] wherein, is a vector with a length of 32, is the global information interaction representation of the k-th vehicle fusion i-th target point coordinate, is a matrix with a size of 30x2, representing the i-th predicted trajectory decoded by the k-th vehicle, which contains the predicted coordinates of 30 future time stamps, is the coordinate of the i-th target point for guiding the future prediction direction, and MLP P is a multi-layer perceptron algorithm for decoding target point information, MLP R is a multi-layer perceptron algorithm for decoding predicted trajectory.

[0064] In the simulation environment, the system is tested, and the specific steps are as follows:

[0065] (1) Play back a section of real sampled intersection data, with a time stamp size of 0.1s, and play the historical real data of the first 2 seconds for each vehicle.

[0066] (2) Run the target-oriented trajectory prediction system based on the graph neural network, and continuously predict the driving trajectory of each vehicle, and obtain the predicted data of the vehicle position at the next time each time.

[0067] (3) Continue to perform step 2, and each execution will obtain the position of all vehicles at the next time, until the simulation ends. The results show that the vehicle runs in a trajectory close to the real and reasonable trajectory, as shown in Figure 3 . The simulation process has a complete video.

Claims

1. A goal-oriented trajectory prediction system based on graph neural networks, characterized in that, A target point is set for each vehicle, which is not a specific position to be reached by the vehicle, but a reference for the driving intention direction, so that a reasonable trajectory with the target point direction as the driving intention is predicted for each vehicle; Specifically, it comprises a vector representation module, a local information fusion module, a global information interaction module and a decoding module; the vector representation module is connected with the local information fusion module; the local information fusion module is connected with the global information interaction module; the decoding module is connected with the global information interaction module and the local information fusion module; wherein: The vector representation module is used for preprocessing the historical trajectory of the vehicle and the map information, vectorizing the position and moving distance of the vehicle trajectory at each time, and extracting the lane geographical position information in the map for vectorization, so as to process the vehicle trajectory vector representation and the map information vector representation of the local area for each vehicle; The local information fusion module is used for obtaining the vectorization of the vehicle trajectory and the lane position from the vector representation module, processing the vectorization information of each vehicle and other vehicles in its local area, processing the vectorization information of each vehicle and the map in its local area, and fusing the vehicle and map information in the local area of each vehicle to obtain the local information fusion representation of each vehicle; The global information interaction module is used for obtaining the local information fusion representation of each vehicle from the local information fusion module, using a graph neural network model to fuse the local information fusion representations of different vehicles to obtain the global information interaction representation of each vehicle; The decoding module is used for obtaining the global information interaction representation of each vehicle from the global information interaction module, accepting the coordinate input of a target point, and then decoding a plurality of predicted trajectories with target guidance for each vehicle according to the global information interaction representation of each vehicle and the target point; The vector representation module specifically comprises: A vehicle information vector representation unit is used for obtaining and processing the information of each vehicle to obtain the vehicle coordinate vector, the vehicle orientation angle and the displacement vector of the vehicle relative to the last time stamp corresponding to each time stamp, the time stamp size Δt=0.1s, and the vehicle historical trajectory vector contains the vehicle coordinates and orientation angle of 20 consecutive time points; A map information vector representation unit is used for obtaining the vector representation of the map lane center line from the high-precision map, including the number of each lane center line and the coordinates of the point set contained therein; A local information screening processing unit is used for obtaining the vector representation from the vehicle information vector representation unit and the map information vector representation unit, screening out the vehicle trajectory information vector and the map information vector in the local area range of each vehicle, obtaining the vector representation of the local information of each vehicle, and rotating and translating all the vector representations of the local information so that the coordinates and orientation angle of the vehicle itself are (0, 0) and 0° respectively, and the transformed coordinates are: ， （1） Transformed orientation angle θ' = θ - θ 0 , Wherein, the vehicle's current coordinates are (x... 0 ,y 0 The orientation angle at the current moment is θ. 0 The coordinates of the local information vector to be transformed are (x, y), and the orientation angle is θ. The coordinates of the transformed local information vector are (x', y'), and the orientation angle at the current moment is θ'.

2. The graph neural network-based goal-directed trajectory prediction system of claim 1, wherein, The local information fusion module specifically comprises: An inter-vehicle interaction unit is used for obtaining the vector representation of the local vehicle information of each vehicle of the vector representation module, integrating the information using a time series model to obtain the local inter-vehicle information fusion representation of each vehicle; A vehicle-road interaction unit is configured to obtain a vector representation of lane information near each vehicle of the vector representation module, integrate the information using a time series model, and obtain a local vehicle-lane information fusion representation of each vehicle. The calculation formula is as follows: ， Among them, z t for t Vehicle-to-vehicle information representation vector at time step, GNN AA A graph neural network algorithm for fusing information between vehicles. for n The vehicles were respectively in t The coordinates of the time after rotation and translation. Let a be the vector representation of the rotated and translated map information. 0 This represents the local inter-vehicle information fusion representation of the current vehicle, and Transformer is the transform algorithm. h 0 GNN is a fusion representation of local vehicle lane information for the current vehicle. AL A graph neural network algorithm for fusing vehicle lane information.

3. The graph neural network-based goal-directed trajectory prediction system of claim 2, wherein, The global information interaction module specifically includes: A graph neural network modeling unit is configured to establish a corresponding node for each vehicle, establish an interaction relationship between vehicles as an edge between nodes corresponding to the vehicles, and perform graph modeling. An information interaction unit is configured to obtain an interaction graph between vehicles from the neural network modeling unit, obtain a local information fusion representation of the vehicles from the local information fusion module, and then fuse global information between the vehicles using a graph neural network to obtain a global information interaction representation of each vehicle. ， （5） wherein, is a global information interaction representation, GNN k for the vehicle numbered GI is a graph neural network that fuses global inter-vehicle information.

4. The graph neural network-based goal-directed trajectory prediction system of claim 3, wherein, The decoding module specifically includes: A decoding unit is configured to obtain a local information fusion representation of the vehicles from the local information fusion module, obtain a global information interaction representation of the vehicles from the global information interaction module, and obtain a plurality of target point coordinates of each vehicle from the input, and then generate a corresponding predicted trajectory for each target point of each vehicle based on the obtained information. ， （6） ， （7） wherein, is a global information interaction representation of the vehicle numbered k fusing the first i target point coordinate, represents the first k predicted trajectory decoded by the vehicle numbered i , is the coordinate of the first i target point, MLP P is a multi-layer perceptron algorithm decoding the target point information, MLP T is a multi-layer perceptron algorithm decoding the predicted trajectory.

5. The graph neural network-based goal-directed trajectory prediction system of claim 4, wherein, Also included are a vehicle-mounted laser radar and a GPS positioning system for real-time detection and identification of the positions of the ego vehicle, adjacent vehicles and obstacles, and processing into coordinate, length, width and orientation data of the vehicles and obstacles, specifically including the identified positions of the ego vehicle at the current time T (x 1,T ,y 1,T ), the orientation angle θ 1,T , and the identified positions of the current adjacent vehicles and obstacles (x 2,T ,y 2,T ),…, (x n,T ,y n,T ), the orientation angles θ 2,T ,…,θ n,T ; It also uses high-precision map data to obtain lane data near the vehicle and processes it into lane vector representations, representing the distance from the current time. t The high-precision map data vector identified by the timestamp is represented as (x n+1,T-t ,y n+1,T-t ),…,(x n+m,T-t ,y n+m,T-t ); It also uses historical data storage units to store the positions of the vehicle, neighboring vehicles, and obstacles identified over a past period of time in the storage medium, up to the current moment. t The timestamp's location in the vehicle's history is (x 1,T-t ,y 1,T-t The historical positions of adjacent vehicles and obstacles are (x 2,T-t ,y 2,T-t ),…,(x n,T-t ,y n,T-t ).

6. The graph neural network-based goal-directed trajectory prediction system of claim 5, wherein, Based on the vehicle data, the obstacle data and the high-precision map data, the target-oriented trajectory prediction system based on the graph neural network is trained using a gradient descent method, 60 continuous timestamp segments are intercepted from a historical data storage unit, data of the first 20 timestamps are taken as vehicle data, obstacle data and high-precision map data of the target-oriented trajectory prediction system based on the graph neural network, positions of vehicles at randomly selected timestamps in the 40th to 60th timestamps are taken as target point data of the target-oriented trajectory prediction system based on the graph neural network, and a predicted trajectory T of each vehicle is obtained K The loss function of the target-oriented trajectory prediction system based on the graph neural network L The calculation formula is as follows: ， （8） Wherein, MSE is the root mean square error.

7. The graph neural network-based goal-directed trajectory prediction system of claim 6, wherein, The trained target-oriented trajectory prediction system based on the graph neural network is used to input vehicle data, obstacle data, high-precision map data and a plurality of target point data to obtain a plurality of trajectory predictions of the vehicle.

Images