Weighted graph attention network-based interpretable vehicle trajectory prediction method and system

Through an interpretable vehicle trajectory prediction method based on a weighted graph attention network, combined with data-driven and knowledge-driven modules, the problem of lack of interpretability of existing models is solved, and high-precision and multimodal trajectory prediction is achieved, suitable for complex traffic Scene.

CN119028130AInactive Publication Date: 2024-11-26UNIV OF SHANGHAI FOR SCI & TECH

Patent Information

Application Number
CN202411051119.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-11-26
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing vehicle trajectory prediction model lacks interpretability and is difficult to apply to actual autonomous driving systems, especially in complex traffic scenarios, factors such as vehicle interaction and road structure cannot be effectively considered.

Method used

Using an interpretable vehicle trajectory prediction method based on a weighted graph attention network, the L-MNL sampler and trajectory generator are constructed, combined with the data drive module and the knowledge drive module, interpretable path candidate areas and predicted trajectories are generated to realize the vehicle. High-precision prediction of trajectory.

Benefits of technology

It realizes high-precision and interpretability of vehicle trajectory prediction, can effectively learn the interaction between the target vehicle and the environment in the graph structure, provide multimodal prediction trajectory, improve prediction accuracy, and show competitiveness on large data sets.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119028130A_ABST
    Figure CN119028130A_ABST
Patent Text Reader

Abstract

The invention discloses an interpretable vehicle trajectory prediction method and system based on a weighted graph attention network, and the method comprises the steps: collecting the historical features of all traffic participants in an unmanned scene; based on the historical features, constructing an interpretable vehicle trajectory prediction model; and using the interpretable vehicle trajectory prediction model to complete vehicle trajectory prediction. According to the method, non-additive fusion is carried out on explainable and unexplainable effects, and high prediction precision and explainability are achieved at the same time. Meanwhile, the weighted graph attention network can more effectively learn the interaction between the target vehicle and the environment in the graph structure, so that the prediction precision is improved. The method can provide interpretability for predicted regional distribution. According to the method, high-precision interpretability is realized on a big data set, and the method can be applied to other fields to research an interpretability mechanism of a deep learning model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of vehicle trajectory prediction, and in particular to an explainable vehicle trajectory prediction method and system based on a weighted graph attention network. Background Art

[0002] With the rapid development of artificial intelligence, autonomous driving is gaining widespread attention and is expected to solve many existing problems, such as traffic congestion, driving safety, and energy efficiency. In mixed traffic scenarios, vehicle trajectory prediction is one of the key technologies for autonomous driving, as autonomous vehicles need to predict the trajectories of surrounding human vehicles to make safer driving decisions. However, many factors affect the prediction of the vehicle's future trajectory, such as maneuver diversity, interaction with the environment, information uncertainty, and timeliness of calculation. In recent years, researchers in the field of autonomous driving have designed various vehicle trajectory prediction models to address these challenges. Existing vehicle trajectory prediction methods can be roughly divided into three categories: physics-based models, driving behavior-based models, and interactive perception models.

[0003] Physics-based models use motion properties such as velocity and acceleration as control inputs to predict the future trajectory of the vehicle through dynamic models, including constant velocity (CV) model, constant acceleration (CA) model, constant speed and acceleration (CTRA) model, etc. Since these models usually rely on idealized physical assumptions such as constant velocity or constant acceleration, they are limited to short-term predictions.

[0004] Driving behavior-based models, such as the driving behavior-based Long Short-Term Memory (LSTM) model and the driving behavior-based CRTA model, assume that maneuvers such as turning or going straight significantly affect the future trajectory of the vehicle. Generally, the prediction accuracy of driving behavior-based models is better than that of physics-based models and is suitable for long-term trajectory prediction. However, maneuver-based models lack consideration of vehicle interactions and road structures, which have a significant impact on the future trajectory of the vehicle. Summary of the invention

[0005] In order to solve the technical problems in the above background, the present invention proposes an interpretable vehicle trajectory prediction model to solve the problem that the existing mainstream models are difficult to apply due to the lack of interpretability.

[0006] In order to achieve the above object, the present invention proposes an interpretable vehicle trajectory prediction method based on a weighted graph attention network, the steps comprising:

[0007] Collect historical characteristics of all traffic participants in the unmanned driving scenario;

[0008] Based on the historical features, construct an interpretable vehicle trajectory prediction model;

[0009] The vehicle trajectory prediction is completed using the explainable vehicle trajectory prediction model.

[0010] Preferably, the historical characteristics include historical characteristics of the target vehicle and historical characteristics of other traffic participants; the historical characteristics of other traffic participants include: coordinates, speed, acceleration and yaw angle.

[0011] Preferably, the constructed interpretable vehicle trajectory prediction model includes: an L-MNL sampler and a trajectory generator;

[0012] The L-MNL sampler is used to generate the total utility of the path candidate region;

[0013] The trajectory generator is used to generate a predicted trajectory of the target vehicle based on the total utility of the path candidate area.

[0014] Preferably, the L-MNL sampler includes: a data-driven module and a knowledge-driven module; the total utility of the path candidate region includes: an explainable utility and an unexplainable utility of the path candidate region;

[0015] The data-driven module is used to generate unexplained utilities of the path candidate regions;

[0016] The knowledge-driven module is used to generate interpretable utilities of path candidate regions.

[0017] Preferably, the method for the data-driven module to calculate the unexplained utility of each candidate region includes:

[0018] z i =linear(concat(a i ,h T , g i )) Among them: linear() is a linear projection; z i is the unexplained utility of the ith candidate region; g i It is obtained by linear mapping the coordinates of the midpoint of the i-th candidate region.

[0019] Preferably, the knowledge-driven module calculates the interpretable utility of each candidate region as follows:

[0020] u i =β dir dir i +β occ occ i +β col col i

[0021] Among them, u i is the interpretable utility of the ith candidate region; dir i , occ i and col i As explanatory variables, they represent the preferences for maintaining direction, avoiding crowding, and avoiding collision in the i-th candidate area; β dir , β occ and β col are the coefficients of the three explanatory variables, which are defined as the explanatory variables for trajectory prediction as follows:

[0022] dir i =|d fd -d i |

[0023]

[0024] Where, d fd is the target vehicle at time point t obs Direction angle; d i is the direction angle of the center of the i-th candidate area; H is the number of surrounding traffic participants; dist hi is the distance between the hth traffic participant and the center of the ith candidate area; and The definition is as follows:

[0025]

[0026] Among them, max l is the radius of the entire sector area; and is the direction angle of the left and right boundaries of the ith candidate area; d h is the direction angle of the hth other traffic participant; D h is the distance between the hth other traffic participant and the target vehicle; |θ h -θ i | is the angle between the moving direction of the hth other traffic participant and the direction of the i-th candidate area.

[0027] Preferably, the method for the trajectory generator to generate the predicted trajectory of the target vehicle includes:

[0028] c j =concat(a j ,h T , g j )

[0029] traj j =MLP(concat(oj ,lv j ))

[0030] Among them: GRU dec () is GRU, used to decode the context vector c j ; is a fully connected layer used to embed c j ;lv j is a latent variable randomly sampled from a standard Gaussian distribution; MLP() generates latent trajectory samples traj using two fully connected layers j .

[0031] The present invention also provides an explainable vehicle trajectory prediction system based on a weighted graph attention network, the system is used to implement the above method, including: a collection module, a construction module and a prediction module;

[0032] The collection module is used to collect historical characteristics of all traffic participants in the unmanned driving scene;

[0033] The construction module is used to construct an explainable vehicle trajectory prediction model based on the historical features;

[0034] The prediction module is used to complete vehicle trajectory prediction using the interpretable vehicle trajectory prediction model.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] The present invention achieves high prediction accuracy and interpretability by non-additive fusion of interpretable and uninterpretable utilities. At the same time, the weighted graph attention network can more effectively learn the interaction between the target vehicle and the environment in the graph structure, thereby improving the prediction accuracy. The present invention can provide interpretability for the predicted regional distribution. This method achieves high-precision interpretability on large data sets and can be applied to other fields to study the interpretability mechanism of deep learning models. Finally, multimodal prediction trajectories can be generated by the L-MNL sampler and the k-means clustering algorithm. Experiments show that the multimodal trajectory prediction accuracy of the present invention is competitive with that of existing models. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0038] Figure 1 A schematic diagram of a method flow chart of an embodiment of the present invention;

[0039] Figure 2 A schematic diagram of a model framework of an embodiment of the present invention;

[0040] Figure 3 Schematic diagram of the structure of an L-MNL sampler according to an embodiment of the present invention;

[0041] Figure 4 A schematic diagram of a graph tensor according to an embodiment of the present invention;

[0042] Figure 5 A weighted graph attention network structure diagram of an embodiment of the present invention;

[0043] Figure 6 It is a fan-shaped schematic diagram of an embodiment of the present invention;

[0044] Figure 7 Schematic diagram of a simulation scenario according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] Embodiment 1

[0048] like Figure 1 As shown, it is a schematic diagram of the method flow of an embodiment, and the steps include:

[0049] S1. Collect historical characteristics of all traffic participants in the unmanned driving scenario.

[0050] In this embodiment, the traffic participants include the target vehicle and other traffic participants. and from t = 1 to t = t obs Historical characteristics of other traffic participants during the time period The characteristics of these other traffic participants include: coordinates speed Acceleration and yaw angle Expressed as Since the future trajectory has inherent uncertainty, the goal of this embodiment is to generate obs +1 to t=t obs+t fut K multimodal prediction trajectories within a time period and its corresponding probability Each predicted trajectory are represented by the distribution of estimated coordinates, and the predicted trajectory should be close to the ground truth y t =(x t ,y t ).

[0051] S2. Based on historical features, an explainable vehicle trajectory prediction model is constructed.

[0052] In order to accurately predict the future trajectory of the target vehicle, this paper proposes an interpretable vehicle trajectory prediction model DCM-GAT-GRU based on weighted graph attention network, which can make predictions based on interactive information and explain the factors affecting the prediction. Figure 2 The framework of the interpretable vehicle trajectory prediction model is presented, which is divided into two parts: L-MNL sampler and trajectory generator; Figure 3 As shown in Figure 1, the L-MNL sampler consists of a graph-based data-driven module and a knowledge-driven module. In the graph-based data-driven module part, GRU is used to encode the historical trajectories of multiple traffic participants in the scene and embed them into a graph tensor containing road information. Subsequently, the interaction relationship between traffic participants is learned through a weighted graph attention network. Finally, the output of the weighted graph attention network is combined with the encoded historical trajectory of the target vehicle and input into the multi-head attention mechanism to learn the impact of environmental information on the target vehicle. A set of fan-shaped regions are constructed to represent the possible target areas of the target vehicle, and the coordinates of each region are encoded into a linear layer, which is then combined with the output of the graph-based data-driven module as the unexplainable utility of the driver's decision. In the knowledge-driven module, the explanatory variables consisting of scene information are input into the CNN to obtain the explainable utility. The explainable utility and the unexplainable utility are added and the distribution of the vehicle's alternative areas is obtained through the softmax function. In the trajectory generation part, the output head of the knowledge-driven module and the encoded region coordinates are sampled according to the distribution, and then these sampled outputs are decoded by the GRU to generate potential trajectories. Finally, the potential trajectories are clustered by the k-means clustering algorithm to generate multimodal trajectories.

[0053] The interpretable vehicle trajectory prediction model includes: an L-MNL sampler and a trajectory generator; the L-MNL sampler is used to generate an interpretable candidate area distribution; the data-driven module is used to generate the uninterpretable utility of the path candidate area; and the knowledge-driven module is used to generate the interpretable utility of the path candidate area.

[0054] Specifically, to achieve accurate vehicle trajectory prediction, it is necessary to describe the interaction between the target vehicle and its surroundings. Therefore, a data-driven module based on a graph structure is constructed as an interaction perception model. This module inputs data into a deep learning model to effectively describe the interaction features.

[0055] GRU is used to encode the historical trajectories of the target vehicle and other traffic participants. The encoding calculation process is as follows:

[0056]

[0057] In the formula, h i represents the encoding feature of the i-th other traffic participant; h T Represents the encoding features of the target vehicle; GRU enc () represents the GRU used to encode the trajectories of traffic participants; Represents a fully connected layer, which is used to embed features into a higher dimension.

[0058] In this embodiment, the high-precision map is represented as a lane graph, where the nodes are the center lines of the lanes; then a weighted graph is constructed based on the distances between the nodes, where there is an edge between any two nodes, and the weight is the inverse of the distance between the two nodes.

[0059] like Figure 4 As shown in Figure 1, a graph tensor is built to embed the encoded features of other traffic participants into nodes corresponding to their positions. The embedding process uses masked attention, which is calculated as:

[0060]

[0061] Where gt0 is the graph tensor; q node 、v node is the projection of the coded features of traffic participants; k node It is the projection of node features, including the midpoint coordinates of the node, the road direction, and whether there are sidewalks and parking lines; d node It is k node The length of the feature dimension; M is the mask matrix, where each element is calculated as:

[0062]

[0063] Then, a weighted graph attention network is used to learn the implicit interactions in the graph tensor. Figure 5 As shown, unlike the basic graph attention network that uses the additive attention mechanism, the weighted graph attention network of this embodiment uses the dot product attention mechanism to improve computational efficiency. In the lth layer of the weighted graph attention network, q l , k l and v lIt is through gt l-1 Then, by l and k l Perform dot product, scaling, and softmax operations to calculate the learnable attention weight matrix A l Finally, output gt l The calculation process is as follows:

[0064] gt l =W pre A l v l

[0065] W pre (i,j)=1 / D(i,j)

[0066] Among them, W pre represents a predefined weight matrix; D represents a distance matrix that records the distances between node centers.

[0067] The multi-head attention mechanism is used to calculate the impact of interactive information on the target vehicle trajectory. The calculation process is as follows:

[0068]

[0069] Among them: a i is the i-th head of multi-head attention; k i 、v i is obtained by linear mapping, gt L is the output of the last layer of weighted graph attention network, representing the interactive information in the environment; q i It is through h T Obtained by linear mapping; It is k i The length of the feature dimension.

[0070] The i heads of attention correspond to i candidate regions to generate uninterpretable utilities. Figure 6 As shown, this embodiment constructs a segmented sector area in front of the target vehicle, representing the candidate area that the target vehicle may reach in time. The angle θ of the sector is evenly divided into P parts. The radius of the sector is equal to And is divided into Q parts equally.

[0071] The calculation process of the unexplained utility of each candidate region is as follows:

[0072] z i =linear(concat(a i ,h T , g i ))

[0073] Among them: linear() is a linear projection; z i is the unexplained utility of the ith candidate region; g i It is obtained by linear mapping the coordinates of the midpoint of the i-th candidate region.

[0074] For safety and efficiency reasons, it is necessary to intuitively understand the degree of influence of each factor on the prediction results. Therefore, this embodiment establishes a knowledge-driven module to explain the influence of some explainable variables on the future driving decision of the vehicle.

[0075] The knowledge-driven module obtains the explainable utility composed of explanatory variables by simulating the utility calculation method of MNL. The explainable utility of each candidate region is calculated as follows:

[0076] u i =β dir dir i +β occ occ i +β col col i

[0077] Among them, u i is the interpretable utility of the ith candidate region; dir i , occ i and col i As explanatory variables, they represent the preferences for maintaining direction, avoiding crowding, and avoiding collision in the i-th candidate area; β dir , β occ and β col are the coefficients of the three explanatory variables, which are defined as the explanatory variables for trajectory prediction as follows:

[0078] dir i =|d fd -d i |

[0079]

[0080] Where, d fd is the target vehicle at time point t obs Direction angle; d i is the direction angle of the center of the i-th candidate area; H is the number of surrounding traffic participants; dist hi is the distance between the hth traffic participant and the center of the ith candidate area; and The definition is as follows:

[0081]

[0082] Among them, max l is the radius of the entire sector area; and is the direction angle of the left and right boundaries of the ith candidate area; d h is the direction angle of the hth other traffic participant; D h is the distance between the hth other traffic participant and the target vehicle; |θ h -θ i | is the angle between the moving direction of the hth other traffic participant and the direction of the i-th candidate area.

[0083] The probability calculation process of the target vehicle arriving at each candidate area in the future is as follows:

[0084] p i =softmax(z i +u i )

[0085] Among them, p i is the probability of selecting the i-th candidate region.

[0086] The L-MNL sampler is distributed according to Output header for knowledge-driven modules and coding regions The coordinates are sampled j times. These samples are used to generate the predicted trajectory in the trajectory generator.

[0087] To generate trajectories for K modalities, J trajectory samples are generated. These trajectory samples are then divided into K categories based on clustering methods, corresponding to the K modalities.

[0088] The trajectory samples are calculated as follows:

[0089] c j =concat(a j ,h T , g j )

[0090]

[0091] traj j =MLP(concat(o j ,lv j ))

[0092] Among them: GRU dec () is GRU, used to decode the context vector cj; is a fully connected layer used to embed c j ;lv jis a latent variable randomly sampled from a standard Gaussian distribution; MLP() generates latent trajectory samples traj using two fully connected layers j .

[0093] Finally, the k-means clustering algorithm is used to divide the potential trajectory samples into k categories. The center of each cluster represents the final predicted trajectory of the kth modality. The weight of the number of samples in each category to the total number of samples is the probability θ of the kth mode k .

[0094] The loss function used to train this model consists of three parts: negative log-likelihood loss L traj And two cross entropy losses L prob and L goal .

[0095] Loss function L traj The goal is to make the error between the predicted trajectory mode closest to the ground truth as small as possible. traj is defined as:

[0096]

[0097] Loss function L prob The goal is to make the probability of the trajectory mode that is closest to the ground truth in the predicted trajectory as large as possible. prob is defined as:

[0098]

[0099] Where k = k * At that time k (k) is equal to 1 otherwise 0; k * is defined as:

[0100]

[0101] Loss function L goal The goal is to maximize the probability that the target vehicle will reach the real area in the future. goal is defined as:

[0102]

[0103] Among them, when the candidate area i is the real arrival area, σ i* (i) is equal to 1 otherwise it is 0.

[0104] The total loss function L is defined as:

[0105] L=L traj +L prob +L goal

[0106] S3. Use the explainable vehicle trajectory prediction model to complete vehicle trajectory prediction.

[0107] After the above steps are completed, the constructed interpretable vehicle trajectory prediction model is used to complete the vehicle trajectory prediction in the unmanned driving scenario.

[0108] Embodiment 2

[0109] The feasibility and advancement of the method provided by the present invention will be verified in conjunction with this embodiment.

[0110] This embodiment uses the autonomous driving open dataset nuScenes for training and evaluation. nuScenes is an open dataset for autonomous driving developed by Motional, an autonomous driving technology company. The data for the nuScenes vehicle trajectory prediction task is collected from more than a thousand driving scenes in cities such as Singapore and Boston, and each scene lasts up to 20 seconds. The coordinates, yaw angle and other information of the target vehicle and surrounding traffic participants are collected through sensors on the vehicle used for data collection. In addition, the surrounding traffic participants are also marked. There are 49,787 samples in the dataset, of which 32,186 are used for training, 8,560 are used for verification, and 9,041 are used for testing.

[0111] The range of the interactive space that the target vehicle can perceive is a rectangular area, 80 meters forward, 20 meters backward, and 50 meters to the left and right of the target vehicle. The fan-shaped arc of the candidate area is 180 degrees, and the radius is 9 times the speed of the target vehicle at the time step. The number of longitudinal divisions of the fan-shaped area is 3, and the number of horizontal divisions is 8, so there are 24 potential targets in total. There are 1024 heads in the multi-head attention output, of which 24 heads correspond to 24 candidate areas and 1000 heads correspond to 1000 generated trajectory samples.

[0112] This embodiment implements the model in the Pytorch environment and runs on a server equipped with an Intel (R) Core (TM) i7-1270 CPU and an NVIDIA (NVIDIA) GeForce RTX 4060 GPU. During the model training process, the batch size is 32 and the number of learning times is 100. The initial learning rate is set to 0.001, and the learning rate is halved after every 10 epochs. The average inference time of the model of the present invention is 9.4 milliseconds, which means that the model can achieve real-time prediction based on sensor data.

[0113] The prediction performance of the vehicle trajectory prediction model is evaluated according to the following indicators:

[0114] 1) Minimum Average Displacement Error (MinimumAverage DisplacementErroroverk): The minimum of the average L2 distances between the top k most probable predicted trajectories and the ground truth trajectories.

[0115] 2) Minimum Final Displacement Error (overk): The minimum absolute error between the final predicted positions of the top k most probable locations and the ground truth positions.

[0116] 3) Miss Rate: The proportion of predicted trajectory points whose minimum deviation from the true trajectory is greater than 2 meters among the top k maximum probabilities.

[0117] In order to verify the accuracy of the proposed model, the model is compared with the most advanced models on the nuScenes dataset. Table 1 shows the evaluation indicators of the existing models on the nuScenes dataset in 6 seconds. Compared with the existing models, the proposed model is very competitive. In the nuScenes trajectory prediction task, minADE5, minADE 10 , minFDE1 and MR 10 It is second only to PGP, which ranks first among mainstream models, but is better than other leading algorithms such as P2T and Trajection++.

[0118] Table 1

[0119]

[0120] The minFDE1 and MR5 of the proposed model reach the state-of-the-art (SOTA), which reflects the excellent prediction accuracy of the proposed model. Although the prediction accuracy of the proposed model is slightly lower than that of PGP overall, it provides additional interpretability, which is not available in PGP. The interpretability evaluation of the proposed model is introduced below:

[0121] In this example, the estimated coefficients of the explanatory variables obtained by training the model are: i , occ i and col i They are -0.4, -0.2 and -0.1 respectively. dir and β col is a negative value, which means that the driver is more inclined to maintain the direction and avoid collision; occ is a negative value, which means that the driver is more willing to avoid driving into a congested area. Compared with the mainstream model, the absolute value of the coefficient of the model of the present invention is much smaller, because the data-driven module of the model of the present invention learns more implicit influencing factors, thereby reducing the weight of the knowledge-driven module.

[0122] This embodiment selects two scenarios to demonstrate the interpretability of the model. According to the first embodiment, the total utility of the candidate area can be divided into unexplained utility and explainable utility, and the explainable utility can be further divided into direction-keeping utility, congestion avoidance utility, and collision avoidance utility. The direction-keeping utility of each area is fixed and is determined by the distance between each area and the front. The degree of congestion in each area determines the utility of avoiding occupancy. Figure 7 As shown in the figure, in scenario 1, the front of the target vehicle is crowded; in scenario 2, the right front is crowded, so the utility of these locations in avoiding crowded parts is shaded. The utility of avoiding collision depends on whether there are vehicles traveling in opposite directions in each area. In scenario 1, there is an oncoming vehicle in the left front, so the left front is dark; in scenario 2, all the moving vehicles turn left, and only the row of cars on the near right side is parked in a different direction from the target car, so the near right side is shaded.

[0123] The results show that the impact of explainable utility on the model as a whole depends on the size of the dataset used for model training. In the model trained on the full dataset containing 32,186 samples, the sum of unexplained utilities in all regions is much higher than the sum of explainable utilities. In scenario 1, unexplained utility accounts for 96.62% of the total utility, while explainable utility accounts for only 3.38%. In scenario 2, unexplained utility accounts for 95.77%, while explainable utility accounts for 4.23%. In contrast, for the model trained on a small sample dataset consisting of 100 samples randomly selected from the full dataset, the sum of explainable utility is relatively close to the sum of unexplained utility. In scenario 1, unexplained utility accounts for 55.59% of the total utility, while explainable utility accounts for 44.41%. In scenario 2, unexplained utility accounts for 52.18% and explainable utility accounts for 47.82%. However, the small dataset brings lower prediction accuracy, with minADE5, minADE 10 , minFDE1 and MR 10 They are 2.97, 2.64, 10.91, 0.86 and 0.80 respectively. The inverse relationship between the dataset size and interpretability shows that when the dataset size is large, the model of the present invention can predict vehicle trajectories with high accuracy and provide additional interpretability. On the contrary, when the dataset is limited and the traditional neural network model is difficult to apply, our model can still be effectively applied and provide higher interpretability.

[0124] Embodiment 3

[0125] This embodiment also provides an interpretable vehicle trajectory prediction system based on a weighted graph attention network, including: an acquisition module, a construction module and a prediction module; the acquisition module is used to collect historical features of all traffic participants in the unmanned driving scenario; the construction module is used to construct an interpretable vehicle trajectory prediction model based on historical features; the prediction module is used to complete vehicle trajectory prediction using the interpretable vehicle trajectory prediction model.

[0126] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. An interpretable vehicle trajectory prediction method based on weighted graph attention network, characterized in that the steps include: Collect historical characteristics of all traffic participants in the unmanned driving scenario; Based on the historical features, construct an interpretable vehicle trajectory prediction model; The vehicle trajectory prediction is completed using the explainable vehicle trajectory prediction model.

2. The interpretable vehicle trajectory prediction method based on weighted graph attention network according to claim 1 is characterized in that: The historical characteristics include the historical characteristics of the target vehicle and the historical characteristics of other traffic participants; the historical characteristics of other traffic participants include: coordinates, speed, acceleration and yaw angle.

3. The explainable vehicle trajectory prediction method based on weighted graph attention network according to claim 1 is characterized in that: The constructed interpretable vehicle trajectory prediction model includes: an L-MNL sampler and a trajectory generator; The L-MNL sampler is used to generate the total utility of the path candidate region; The trajectory generator is used to generate a predicted trajectory of the target vehicle based on the total utility of the path candidate area.

4. The explainable vehicle trajectory prediction method based on weighted graph attention network according to claim 3 is characterized in that: The L-MNL sampler includes: a data-driven module and a knowledge-driven module; the total utility of the path candidate region includes: an explainable utility and an unexplainable utility of the path candidate region; The data-driven module is used to generate unexplained utilities of the path candidate regions; The knowledge-driven module is used to generate interpretable utilities of path candidate regions.

5. The explainable vehicle trajectory prediction method based on weighted graph attention network according to claim 4 is characterized in that: The method for the data-driven module to calculate the unexplained utility of each candidate region includes: z i =linear(concat(α i ,h T ,g i )) Where: linearO is a linear projection; z i is the unexplained utility of the ith candidate region; g i It is obtained by linear mapping the coordinates of the midpoint of the i-th candidate region.

6. The interpretable vehicle trajectory prediction method based on weighted graph attention network according to claim 4 is characterized in that: The knowledge-driven module calculates the interpretable utility of each candidate region as follows: you i =b dir three i +b occ occ i +b col col i Among them, u i is the interpretable utility of the ith candidate region; dir i , occ i and col i As explanatory variables, they represent the preferences for maintaining direction, avoiding crowding, and avoiding collision in the i-th candidate area; β dir , β occ and β col are the coefficients of the three explanatory variables, which are defined as the explanatory variables for trajectory prediction as follows: you i =|d fd -d i | Where, d fd is the target vehicle at time t obs Direction angle; d i is the direction angle of the center of the i-th candidate area; H is the number of surrounding traffic participants; dist hi is the distance between the hth traffic participant and the center of the ith candidate area; and The definition is as follows: Among them, max l is the radius of the entire sector area; and is the direction angle of the left and right boundaries of the ith candidate area; d h is the direction angle of the hth other traffic participant; D h is the distance between the hth other traffic participant and the target vehicle; |θ h -θ i | is the angle between the moving direction of the hth other traffic participant and the direction of the i-th candidate area.

7. The explainable vehicle trajectory prediction method based on weighted graph attention network according to claim 3 is characterized in that: The method for the trajectory generator to generate the predicted trajectory of the target vehicle includes: c j =concat(a j ,h T ,g j ) three j =MLP(concat(o j ,lv j )) Among them: GRU dec O is a GRU, used to decode the context vector c j ; is a fully connected layer used to embed c j ;lv j is a latent variable randomly sampled from a standard Gaussian distribution; MLP() generates latent trajectory samples traj using two fully connected layers j .

8. An interpretable vehicle trajectory prediction system based on a weighted graph attention network, the system being used to implement the method of any one of claims 1-7, characterized in that: include: Acquisition module, construction module and prediction module; The collection module is used to collect historical characteristics of all traffic participants in the unmanned driving scene; The construction module is used to construct an explainable vehicle trajectory prediction model based on the historical features; The prediction module is used to complete vehicle trajectory prediction using the interpretable vehicle trajectory prediction model.

Citation Information

Patent Citations

  • Discrete selection hotel room demand model

    CN116324837A

  • Road network flow prediction method for urban traffic signal control

    CN116994431A

  • Automatic driving track prediction method and device, electronic equipment and storage medium

    CN117104275A

Cited By

  • Automatic driving vehicle behavior prediction method and system based on artificial intelligence

    CN119578658A

  • Tunnel support system collaborative design method and system based on artificial intelligence

    CN119939747A

  • Public traffic scheduling optimization method driven by multi-modal time series data analysis

    CN120355045A

  • A public transportation scheduling optimization method driven by multimodal time series data analysis

    CN120355045B

  • Trajectory prediction method and apparatus, and storage medium

    WO2026157311A1