Multi-vehicle joint trajectory prediction method based on intent learning of surrounding vehicles

By constructing a multi-vehicle joint trajectory prediction model, the problem of the inability to integrate multi-scale vehicle motion characteristics and spatial interaction relationships in existing technologies is solved, enabling accurate prediction of vehicle intentions in complex road environments and improving the safety and decision-making efficiency of autonomous vehicles.

CN118968441APending Publication Date: 2024-11-15JILIN UNIVERSITY +1

Patent Information

Application Number
CN202410940828.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-15
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

Existing technologies cannot effectively integrate multi-scale vehicle motion characteristics, cannot accurately predict the various potential intentions of vehicles in complex road environments, and fail to consider the spatial interaction between vehicles and the impact of road information on trajectory prediction.

Method used

A multi-vehicle joint trajectory prediction model is constructed, including a proxy feature encoding layer, a map feature encoding layer, a global interaction relationship encoding layer, an intent self-attention layer, and an interaction attention decoding layer. Through a feature pyramid network, a convolutional neural network, a Transformer encoder, and a multilayer perceptron, the spatial interaction relationship between vehicle and road information is captured, and multi-vehicle joint multimodal predicted trajectories are generated through intent queries.

Benefits of technology

It improves the accuracy and safety of trajectory prediction for autonomous vehicles in complex road environments, can better integrate multi-scale vehicle motion characteristics, capture multiple potential intentions of vehicles, and output multi-vehicle joint multimodal predicted trajectories, thereby improving the efficiency and safety of autonomous driving decision-making and planning.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118968441A_ABST
    Figure CN118968441A_ABST
Patent Text Reader

Abstract

The invention discloses a multi-vehicle joint trajectory prediction method based on surrounding vehicle intention learning, and the method comprises the steps: inputting a surrounding vehicle historical trajectory set into a proxy feature coding layer, and obtaining a surrounding vehicle historical trajectory feature embedding vector; inputting the lane center line position point set into a map feature coding layer to obtain a lane center line position point embedding vector; connecting the peripheral vehicle historical track feature embedded vector and the lane center line position point embedded vector, adding the connected vectors with the position codes, and inputting a global interaction relation coding layer to obtain a global interaction feature vector; inputting the historical trajectory feature embedded vector and the intention query vector of the surrounding vehicles into an intention self-attention layer to obtain intention coding features of the surrounding vehicles; inputting the surrounding vehicle intention coding features and the global interaction feature vector into an interaction attention decoding layer to obtain a multi-vehicle joint decoding feature vector; and according to the multi-vehicle joint decoding feature vector, obtaining a prediction track corresponding to the intentions of the surrounding vehicles and the confidence coefficient of the prediction track.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving, and in particular relates to a multi-vehicle joint trajectory prediction method based on surrounding vehicle intention learning. Background Art

[0002] The autonomous driving system mainly includes three parts: perception, decision-making planning and control. The perception system mainly obtains the vehicle's motion status and surrounding environment information through cameras, lidar, millimeter-wave radar and GPS, and passes the acquired information to the recognition module that processes this information. In general, the autonomous driving system consists of an adaptive detection and recognition framework, a control system, a vehicle deviation warning system and an unknown obstacle recognition system. These processed information are integrated and passed to the decision-making and planning stage. The decision-making planning system uses the data collected during the perception process and combines it with high-precision map information to predict the movement of traffic participants around the vehicle, including trajectory prediction, behavior prediction and risk prediction, and then combines perception, map information and prediction results to plan the vehicle's movement. Finally, based on the results of the decision-making planning module, the control system uses the steering wheel and accelerator pedal to control the lateral and longitudinal movement of the vehicle.

[0003] When a vehicle is driving on the road, it will frequently interact with surrounding traffic participants, and their future motion trajectories will directly affect the decision-making of the autonomous vehicle. Without the ability to predict trajectories, the operating efficiency and safety of the autonomous vehicle will be greatly reduced. Therefore, in the implementation of autonomous driving, trajectory prediction is a crucial task. It involves how to accurately predict the motion trajectories of intelligent agents such as vehicles and pedestrians in the future. It is very important for the autonomous driving system to effectively plan paths, make decisions, and avoid potential dangers. The challenge of trajectory prediction is to take into account multiple factors, such as interactions between vehicles, road environment, and traffic control information. In addition, trajectory prediction also needs to take into account uncertain factors, such as sensor noise and incomplete input information. With the continuous advancement of sensor technology, vehicles can obtain rich and diverse environmental information, such as cameras, lidar, radar, and ultrasonic sensors. These sensors can provide data about roads, traffic signs, pedestrians, other vehicles, etc., providing a basis for trajectory prediction.

[0004] Patent CN 202310702890 discloses a vehicle trajectory prediction method for self-learning of vehicle-to-vehicle interactive motion relations. This technical solution simply divides vehicle intentions into three categories: left lane change, right lane change, and straight driving, which cannot accurately express the vehicle's current intentions. Moreover, the output is a single trajectory of a single vehicle, which cannot take into account the uncertainty involved in predicting future trajectories.

[0005] Patent CN 202110746125 discloses a vehicle trajectory multimodal prediction method. This technical solution only considers the temporal interaction of trajectories, but does not consider the spatial interaction between different vehicle trajectories. In addition, the LSTM-based method cannot integrate multi-scale vehicle motion features. In addition, this technical solution does not consider the impact of road information such as lane lines on trajectory prediction. Summary of the invention

[0006] The purpose of the present invention is to provide a multi-vehicle joint trajectory prediction method based on surrounding vehicle intention learning, which overcomes the deficiency of the existing methods that they cannot integrate multi-scale vehicle motion features, and designs a learnable intention query that can capture multiple potential intentions of vehicles in complex road environments, and finally output multi-vehicle joint multimodal prediction trajectories.

[0007] The technical solution provided by the present invention is:

[0008] A multi-vehicle joint trajectory prediction method based on surrounding vehicle intention learning includes the following steps:

[0009] Step 1: Build a multi-vehicle joint trajectory prediction model, which includes: agent feature encoding layer, map feature encoding layer, global interaction relationship encoding layer, intention self-attention layer and interaction attention decoding layer;

[0010] Step 2: Obtain multiple frames of historical trajectories of vehicles around the main vehicle to form a set of historical trajectories of surrounding vehicles;

[0011] Inputting the surrounding vehicle historical trajectory set into the proxy feature encoding layer for feature extraction to obtain a surrounding vehicle historical trajectory feature embedding vector;

[0012] Acquire multiple frames of lane centerline position point data corresponding to the vehicle historical trajectory time to form a lane centerline position point set;

[0013] Inputting the lane centerline position point set into the map feature coding layer for feature extraction to obtain a lane centerline position point embedding vector;

[0014] Step 3: Concatenate the surrounding vehicle historical trajectory feature embedding vector and the lane centerline position point embedding vector, add the vector to the position code, and input the vector to the global interaction relationship coding layer to obtain a global interaction feature vector.

[0015] The position code is a set of distances between the center point of all surrounding vehicles and lane centerline and the main vehicle in the current frame;

[0016] Step 4: embedding the surrounding vehicle historical trajectory feature vector and the intention query vector into the intention self-attention layer to obtain the surrounding vehicle intention encoding feature;

[0017] Input the surrounding vehicle intention encoding features and the global interaction feature vector into the interactive attention decoding layer, use the global interaction feature vector as the key vector and value vector of the interactive attention decoding layer, and obtain a multi-vehicle joint decoding feature vector;

[0018] The method for obtaining the intention query vector is as follows: clustering the feasible terminal positions of the surrounding vehicles, taking each class in the clustering result as a surrounding vehicle intention, and obtaining the surrounding vehicle initial intention set; performing dimension conversion on the surrounding vehicle initial intention set to obtain the intention query vector;

[0019] Step 5: Pass the multi-vehicle joint decoding feature vector through two multi-layer perceptrons to obtain the predicted trajectory corresponding to the intention of each surrounding vehicle and the confidence of each predicted trajectory.

[0020] Preferably, the proxy feature encoding layer adopts a feature pyramid network.

[0021] Preferably, the map feature encoding layer is composed of a convolutional neural network and a multi-layer perceptron.

[0022] Preferably, the global interaction relationship encoding layer adopts a Transformer encoder.

[0023] Preferably, before the step 2, the method further includes using data from the Argoverse2 motion prediction dataset to train, verify and test the constructed multi-vehicle joint trajectory prediction model.

[0024] Preferably, the positions of surrounding vehicles corresponding to the predicted endpoint frame in the data of the Argoverse2 motion prediction data set are used as the feasible endpoint positions of the surrounding vehicles.

[0025] Preferably, the loss function used when training the multi-vehicle joint trajectory prediction model is:

[0026] L=L reg +L cls ;

[0027] Among them, L reg = HuberLoss(traj,y); L cls =CrossEntropyLoss(pi,mode);

[0028] Where, L reg is the trajectory regression loss, L clsIt is divided into class loss; y represents the future true trajectory of the vehicle; HuberLoss represents smooth L1 loss; CrossEntropyLoss represents cross entropy loss; mode represents the index of the trajectory with the minimum mean absolute error corresponding to each surrounding vehicle; traj represents the multi-vehicle joint multimodal trajectory finally output by the prediction model; pi represents the confidence level corresponding to each predicted trajectory.

[0029] Preferably, AdamW is used as the optimizer during the training of the multi-vehicle joint trajectory prediction model.

[0030] The beneficial effects of the present invention are:

[0031] The multi-vehicle joint trajectory prediction method based on weekly vehicle intention learning provided by the present invention overcomes the deficiency of the existing methods that they cannot integrate multi-scale vehicle motion characteristics, and designs a learnable intention query, which can capture multiple potential intentions of vehicles in complex road environments, and finally output multi-vehicle joint multimodal prediction trajectories, providing more efficient and reasonable input information for the autonomous driving decision-making and planning system, thereby improving the safety of autonomous driving vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flow chart of the multi-vehicle joint trajectory prediction method based on surrounding vehicle intention learning described in the present invention.

[0033] Figure 2 This is a framework diagram of the multi-vehicle joint trajectory prediction model described in the present invention.

[0034] Figure 3 This is a clustering point diagram obtained by clustering the feasible terminal positions of surrounding vehicles in the present invention. DETAILED DESCRIPTION

[0035] The present invention is further described in detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0036] like Figure 1-2 As shown, the present invention provides a multi-vehicle joint trajectory prediction method based on surrounding vehicle intention learning, and the specific implementation process is as follows.

[0037] 1. Construct a multi-vehicle joint trajectory prediction model and train the multi-vehicle joint trajectory prediction model

[0038] The multi-vehicle joint trajectory prediction model includes: agent feature encoding layer, map feature encoding layer, global interaction relationship encoding layer, intention self-attention layer and interaction attention decoding layer.

[0039] The multi-vehicle joint trajectory prediction model is trained and evaluated on the Argoverse2 (AV2) motion prediction dataset. The dataset includes 250K non-overlapping urban road scenes, divided into 199908, 24988 and 24984 samples for training, validation and testing respectively. Each sample contains 5 seconds of history and needs to be predicted for 6 seconds in the future, with a sampling frequency of 10Hz.

[0040] Step 1: After obtaining the historical trajectory of the vehicles in the scene and the lane centerline data, initialize the data: convert the historical trajectory data into relative positions, that is, the coordinate position of vehicle i around the main vehicle in the scene at time t Indicates the relative distance between the surrounding vehicles at this moment and the previous moment; in the map information, lane lines are represented in the form of segments, and each lane line segment consists of 20 lane centerline position points. The data p of all lane centerline points in the jth segment are j =(x j ,y j ) to transform the coordinate system to the current position of the main vehicle a obs =(x obs ,y obs ) is the origin and the heading angle is the X-axis.

[0041] Step 2: Put the initialized vehicle history trajectory and lane centerline data into different feature information embedding layers for feature extraction.

[0042] For the set of all surrounding vehicle historical trajectories in frame T, a∈R N×T×2 , extract features from the data through a proxy feature encoding layer consisting of a feature pyramid network (FPN), the formula is as follows:

[0043] A=FPN(a),A∈R N×C

[0044] Where A represents the vehicle historical trajectory feature embedding vector obtained through the pyramid network; N represents the number of vehicles in the scene; and C represents the dimension of the embedding vector.

[0045] For the set of all lane centerline points p∈R M×2 , the data is extracted through a map feature encoding layer composed of a multi-layer perceptron (MLP) and a convolutional neural network (CNN). The formula is as follows:

[0046] P = CNN(MLP(p)), P∈R M×C

[0047] Where P represents the lane centerline feature embedding vector; CNN represents the convolutional neural network operation; and M represents the number of lane centerlines in the scene.

[0048] Step 3: Concatenate the embedding vectors obtained in step 2, and add the distances between the center point of all surrounding vehicles and lane centerline and the main vehicle at the current moment as the position code and the embedding vector, and input them into the global interaction relationship encoding layer composed of the Transformer encoder. The multi-head self-attention MHSA is used to capture the spatial interaction relationship between vehicles and vehicles, and between vehicles and road information, which is recorded as the global interaction feature vector. The formula is as follows:

[0049] S=MHSA(concat(A,P)+PE),S∈R (N+M)×C

[0050] Where S represents the global interaction feature vector; concat represents the concatenation operation of the vector; PE represents the position encoding, and its value is the distance between the current frame of the surrounding vehicles or the middle point of the lane centerline and the current frame of the main vehicle.

[0051] Step 4: Convert each initial intention corresponding to each surrounding vehicle from the local coordinate system centered on the current frame position of the surrounding vehicle to the global coordinate system centered on the current frame position of the main vehicle. Finally, an MLP is used to transform the dimension of the intention into D, as follows:

[0052] Q=Kmeans(I),Q∈R k×2

[0053]

[0054]

[0055] intention=MLP(Q′),intention∈R N×k×D

[0056] In the formula, rot i represents the rotation matrix of the surrounding vehicle i; matmul represents matrix multiplication; Represents the current frame position coordinates of the surrounding vehicle i; intention represents the intention query vector corresponding to each surrounding vehicle.

[0057] The method for determining the initial intention corresponding to each surrounding vehicle is as follows: first, extract the last frame position of the surrounding vehicles in all scenes in the training set, cluster the last frame position I by the Kmeans clustering algorithm, and generate k kinds of initial intentions Q∈R K×D , each class in the clustering result corresponds to an initial intention. In this embodiment, k=64 is set, and the obtained clustering points are as follows Figure 3As shown. These intentions implicitly model the possible future motion states of the surrounding vehicle and refine driving behaviors such as going straight, turning left, turning right, accelerating, decelerating, stopping, and turning around. In this embodiment, the training sample contains 5 seconds of historical records, and a 6-second prediction is required for the future; each second includes 10 frames, and the last frame here is the last frame (frame 110) in the next 6 seconds.

[0058] The intent query vector and the vehicle historical trajectory feature embedding vector A in step 2 are input into an intention self-attention layer composed of a Transformer encoder to obtain the intention encoding feature A′ of the surrounding vehicles. The formula is as follows:

[0059] A′=MHSA(intention+A,intention+A,A),A′∈R N×k×D

[0060] The intention encoding feature A′ and the global interaction feature vector S are input into the interactive attention decoding layer composed of the Tranformer decoder, and the multi-vehicle joint decoding feature vector O is output through the multi-head interactive attention MHCA. The formula is as follows:

[0061] O=MHCA(A′,S,S),O∈R N×k×D ;

[0062] Among them, the global interaction feature vector is used as the key vector and value vector of the interaction attention decoding layer.

[0063] Step 5: The multi-vehicle joint decoding feature vector O output in step 4 is passed through two multi-layer perceptrons to output the predicted trajectory corresponding to each intention of the surrounding vehicles and the confidence of each predicted trajectory. The formula is as follows:

[0064] traj=MLP(O′),traj∈R N×k×step×2

[0065] C=MLP(O′),C∈R N×k×1

[0066] Where traj represents the predicted future trajectory coordinates; C represents the confidence level corresponding to the predicted trajectory; step represents the number of steps of the future trajectory to be predicted.

[0067] In the trajectory prediction training phase, the loss function L consists of two items: trajectory regression loss L reg and classification loss L cls , the formula is as follows:

[0068] L=L reg +L cls

[0069] L reg= HuberLoss(traj,y)

[0070] L cls =CrossEntropyLoss(pi,mode)

[0071] Where y represents the future real trajectory of the vehicle; HuberLoss represents the smooth L1 loss; CrossEntropyLoss represents the cross entropy loss; mode represents the index of the trajectory with the minimum mean absolute error for each surrounding vehicle. The training was conducted for 60 rounds, the optimizer used was AdamW, the learning rate was set to 0.001, and the cosine was decreased to 0, and the weight decay was set to 0.0001. The verification evaluation indicators include avgMinADE, avgMinFDE, actorMR and avgBrierMinFDE.

[0072] Table 1 illustrates the hyperparameters of the network architecture of the present invention.

[0073] Table 1

[0074]

[0075]

[0076] 2. Use the trained multi-vehicle joint trajectory prediction model to predict multi-vehicle joint trajectories

[0077] The prediction process is basically the same as the training process. The specific process is as follows:

[0078] Step 1: After obtaining the historical trajectory of multiple frames of vehicles around the main vehicle (trajectory before the current moment) and lane centerline data, initialize the data: convert the historical trajectory data into relative positions, that is, the coordinate position of vehicle i around the main vehicle at time t represents the relative distance between the surrounding vehicles at this moment and the previous moment; the lane line is represented in the form of segments, and each lane line segment consists of 20 lane centerline position points. The data p of all lane centerline points in the jth segment are j =(x j ,y j ) to transform the coordinate system to the current position of the main vehicle a obs =(x obs ,y obs ) is the origin and the heading angle is the X-axis.

[0079] Step 2: Put the initialized vehicle history trajectory and lane centerline data into different feature information embedding layers for feature extraction.

[0080] For the set of all surrounding vehicle historical trajectories in frame T, a∈RN×T×2 , extract features from the data through a proxy feature encoding layer consisting of a feature pyramid network (FPN), the formula is as follows:

[0081] A=FPN(a),A∈R N×C

[0082] Where A represents the vehicle historical trajectory feature embedding vector obtained through the pyramid network; N represents the number of vehicles in the scene; and C represents the dimension of the embedding vector.

[0083] For the set of all lane centerline points p∈R M×2 , the data is extracted through a map feature encoding layer composed of a multi-layer perceptron (MLP) and a convolutional neural network (CNN). The formula is as follows:

[0084] P = CNN(MLP(p)), P∈R M×C

[0085] Where P represents the lane centerline feature embedding vector; CNN represents the convolutional neural network operation; and M represents the number of lane centerlines in the scene.

[0086] Step 3: Concatenate the embedding vectors obtained in step 2, and add the distances between the center point of all surrounding vehicles and lane centerline and the main vehicle at the current moment as the position code and the embedding vector, and input them into the global interaction relationship encoding layer composed of the Transformer encoder. The multi-head self-attention MHSA is used to capture the spatial interaction relationship between vehicles and vehicles, and between vehicles and road information, which is recorded as the global interaction feature vector. The formula is as follows:

[0087] S=MHSA(concat(A,P)+PE),S∈R (N+M)×C

[0088] Where S represents the global interaction feature vector; concat represents the concatenation operation of the vector; PE represents the position encoding, and its value is the distance between the current frame of the surrounding vehicles or the middle point of the lane centerline and the current frame of the main vehicle.

[0089] Step 4: Transform each initial intention corresponding to each surrounding vehicle (the same as the initial intention used in the training process) from the local coordinate system centered on the current frame position of the surrounding vehicle to the global coordinate system centered on the current frame position of the main vehicle. Finally, an MLP is used to transform the dimension of the intention into D, as follows:

[0090] Q=Kmeans(I),Q∈R k×2

[0091]

[0092]

[0093] intention=MLP(Q′),intention∈R N×k×D

[0094] In the formula, rot i represents the rotation matrix of the surrounding vehicle i; matmul represents matrix multiplication; Represents the current frame position coordinates of the surrounding vehicle i; intention represents the intention query vector corresponding to each surrounding vehicle.

[0095] The intent query vector and the vehicle historical trajectory feature embedding vector A in step 2 are input into an intention self-attention layer composed of a Transformer encoder to obtain the intention encoding feature A′ of the surrounding vehicles. The formula is as follows:

[0096] A′=MHSA(intention+A,intention+A,A),A′∈R N×k×D

[0097] The intention encoding feature A′ and the global interaction feature vector S are input into the interactive attention decoding layer composed of the Tranformer decoder, and the multi-vehicle joint decoding feature vector O is output through the multi-head interactive attention MHCA. The formula is as follows:

[0098] O=MHCA(A′,S,S),O∈R N×k×D ;

[0099] Among them, the global interaction feature vector is used as the key vector and value vector of the interaction attention decoding layer.

[0100] Step 5: The multi-vehicle joint decoding feature vector O output in step 4 is passed through two multi-layer perceptrons to output the predicted trajectory corresponding to each intention of the surrounding vehicles and the confidence of each predicted trajectory. The formula is as follows:

[0101] traj=MLP(O′),traj∈R N×k×step×2

[0102] C=MLP(O′),C∈R N×k×1

[0103] Where traj represents the predicted future trajectory coordinates; C represents the confidence level corresponding to the predicted trajectory; step represents the number of steps of the future trajectory to be predicted.

[0104] The multi-vehicle joint trajectory prediction method based on vehicle intention learning provided by the present invention can perform multi-vehicle joint trajectory prediction; the feature information embedding method can fuse multi-scale vehicle motion features and capture the complex spatial relationships that may exist between vehicles. The learnable intention query adopted by the present invention can capture multiple potential intentions of vehicles in complex road environments; the multi-modal trajectory generation method uses fewer parameters to efficiently generate multi-modal prediction trajectories, which can better adapt to the uncertainty of human driving behavior.

[0105] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation modes, and they can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A multi-vehicle joint trajectory prediction method based on surrounding vehicle intention learning, characterized in that: The steps include: Step 1: Build a multi-vehicle joint trajectory prediction model, which includes: agent feature encoding layer, map feature encoding layer, global interaction relationship encoding layer, intention self-attention layer and interaction attention decoding layer; Step 2: Obtain multiple frames of historical trajectories of vehicles around the main vehicle to form a set of historical trajectories of surrounding vehicles; Inputting the surrounding vehicle historical trajectory set into the proxy feature encoding layer for feature extraction to obtain a surrounding vehicle historical trajectory feature embedding vector; Acquire multiple frames of lane centerline position point data corresponding to the vehicle historical trajectory time to form a lane centerline position point set; Inputting the lane centerline position point set into the map feature coding layer for feature extraction to obtain a lane centerline position point embedding vector; Step 3: Concatenate the surrounding vehicle historical trajectory feature embedding vector and the lane centerline position point embedding vector, add the vector to the position code, and input the vector to the global interaction relationship coding layer to obtain a global interaction feature vector. The position code is a set of distances between the center point of all surrounding vehicles and lane centerline and the main vehicle in the current frame; Step 4: embedding the surrounding vehicle historical trajectory feature vector and the intention query vector into the intention self-attention layer to obtain the surrounding vehicle intention encoding feature; Input the surrounding vehicle intention encoding features and the global interaction feature vector into the interactive attention decoding layer, use the global interaction feature vector as the key vector and value vector of the interactive attention decoding layer, and obtain a multi-vehicle joint decoding feature vector; The method for obtaining the intention query vector is as follows: clustering the feasible terminal positions of the surrounding vehicles, taking each class in the clustering result as a surrounding vehicle intention, and obtaining the surrounding vehicle initial intention set; performing dimension conversion on the surrounding vehicle initial intention set to obtain the intention query vector; Step 5: Pass the multi-vehicle joint decoding feature vector through two multi-layer perceptrons to obtain the predicted trajectory corresponding to the intention of each surrounding vehicle and the confidence of each predicted trajectory.

2. The multi-vehicle joint trajectory prediction method based on surrounding vehicle intention learning according to claim 1 is characterized in that: The proxy feature encoding layer adopts a feature pyramid network.

3. The multi-vehicle joint trajectory prediction method based on surrounding vehicle intention learning according to claim 2 is characterized in that: The map feature encoding layer is composed of a convolutional neural network and a multi-layer perceptron.

4. The multi-vehicle joint trajectory prediction method based on surrounding vehicle intention learning according to claim 3 is characterized in that: The global interaction relationship encoding layer adopts a Transformer encoder.

5. The multi-vehicle joint trajectory prediction method based on surrounding vehicle intention learning according to any one of claims 1 to 4, characterized in that: Before the step 2, the method also includes training, verifying and testing the constructed multi-vehicle joint trajectory prediction model using data from the Argoverse2 motion prediction dataset.

6. The multi-vehicle joint trajectory prediction method based on surrounding vehicle intention learning according to claim 5 is characterized in that: The positions of surrounding vehicles corresponding to the predicted endpoint frame in the Argoverse2 motion prediction dataset are used as the feasible endpoint positions of surrounding vehicles.

7. The multi-vehicle joint trajectory prediction method based on surrounding vehicle intention learning according to claim 6 is characterized in that: The loss function used when training the multi-vehicle joint trajectory prediction model is: L=L reg +L cls ; Among them, L reg = HuberLoss(traj,y); L cls =CrossEntropyLoss(pi,mode); Where, L reg is the trajectory regression loss, L cls It is divided into class loss; y represents the future true trajectory of the vehicle; HuberLoss represents smooth L1 loss; CrossEntropyLoss represents cross entropy loss; mode represents the index of the trajectory with the minimum mean absolute error corresponding to each surrounding vehicle; traj represents the multi-vehicle joint multimodal trajectory finally output by the prediction model; pi represents the confidence level corresponding to each predicted trajectory.

8. The multi-vehicle joint trajectory prediction method based on surrounding vehicle intention learning according to claim 7 is characterized in that: AdamW is used as the optimizer in the process of training the multi-vehicle joint trajectory prediction model.

Citation Information

Patent Citations

  • Vehicle track multi-modal prediction method

    CN113657433A

  • Vehicle trajectory prediction method based on self-learning of interactive motion relation between workshops

    CN116702986A

Cited By

  • Adaptive iteration trajectory prediction method and system based on attention mechanism

    CN119840643A

  • Adaptive Iterative Trajectory Prediction Method and System Based on Attention Mechanism

    CN119840643B

  • Vehicle multi-modal trajectory prediction method taking graph as center

    CN120182937A

  • A graph-centric vehicle multimodal trajectory prediction method

    CN120182937B

  • Transform-based automatic driving long time sequence track prediction method and system

    CN120408105A