A vehicle lane-changing trajectory prediction method based on physical information deep learning

Abstract

The application relates to a vehicle lane-changing trajectory prediction method based on physical information deep learning, which comprises the following steps: S1: constructing a data set and dividing the data set into a training set, a verification set and a test set, extracting observation states and matching states I j from the training set, and extracting observation states S2: constructing a physical prediction model, combining the physical prediction model with an Attention-TCN model to form a physical information deep learning model, and training the physical information deep learning model, calculating a loss function and a loss on the verification set, and comparing the loss on the verification set with an optimal loss value and the size of the loss function to selectively update parameters of the Attention-TCN model and parameters of the physical prediction model; S3: using the trained physical information deep learning model to perform a test experiment, and comparing the prediction performance of different lane-changing trajectory prediction models. By combining the constructed physical prediction model with the Attention-TCN model to form the physical information deep learning model, the explainability and descriptiveness of the lane-changing trajectory prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle lane-changing trajectory prediction, and in particular to a vehicle lane-changing trajectory prediction method based on physical information deep learning. BACKGROUND

[0002] With the wide development of sensing technology, control theory and artificial intelligence, autonomous driving has been considered as an effective solution to reduce the incidence of road traffic accidents. Autonomous vehicles need to have the ability to predict the future state of their own vehicles and surrounding vehicles in real time, like human drivers, in order to identify potential dangerous conditions in advance and plan a safe driving path, thereby improving driving safety.

[0003] As a basic driving behavior of vehicles, following and lane-changing behavior is an important research direction for trajectory prediction of autonomous vehicles. In recent years, following trajectory prediction has been widely studied, and many models and methods have been developed. Due to the complexity of lane-changing behavior, the research on trajectory prediction for lane-changing is less than that for following. However, vehicle lane-changing trajectory prediction is crucial for driving safety and needs to be further studied.

[0004] Physical driving method and data-driven method are two main research directions of existing lane-changing trajectory prediction. The physical driving lane-changing trajectory prediction method usually uses kinematic model or microscopic traffic flow model for trajectory prediction, such as quintic polynomial model, sine model, Gipps model, cellular automaton model, etc. The data-driven lane-changing trajectory prediction method refers to using advanced machine learning methods to mine human historical driving behavior information and imitate human driving behavior, such as reinforcement learning, LSTM, Transformer, TCN, etc.

[0005] Physical driving method and data-driven method each have its advantages and disadvantages. On the one hand, the physical driving method which integrates physical knowledge has strong interpretability for traffic phenomena, however, the physical driving model is constructed as a mathematical function with limited parameters, which limits its ability to describe vehicle lane-changing behavior in complex driving environments. On the other hand, although the data-driven method based on deep learning can mine human driving behavior information in complex traffic environment, due to the characteristics of black box model, the data-driven method has poor interpretability and reliability for trajectory prediction. SUMMARY

[0006] Therefore, it is necessary to provide a vehicle lane-changing trajectory prediction method based on physical information deep learning in view of the problems of poor interpretability of data-driven method for trajectory prediction and poor description of physical driving method for trajectory prediction.

[0007] The application provides a vehicle lane-changing trajectory prediction method based on physical information deep learning, which comprises the following steps: S1: constructing a data set by using a public trajectory data set, dividing the data set into a training set, a validation set and a test set, extracting observation states from the training set and collocation state I j from the validation set S2: constructing a physical prediction model, and combining the physical prediction model with an Attention-TCN model to form a physical information deep learning model, and training the physical information deep learning model by using the constructed data set to obtain a trained physical information deep learning model, and the specific training steps are as follows: S21: initializing parameters α, N epoch and best_Loss; let epoch=N epoch ; wherein, is a parameter of the Attention-TCN model, is a parameter of the physical prediction model, α is a weight parameter, N epoch is the number of training times, and best_Loss is the optimal loss value; S22: let epoch=epoch-1; S23: inputting the observation state into the Attention-TCN model to obtain corresponding predicted trajectory coordinates inputting the collocation state I j into the Attention-TCN model and the physical prediction model respectively to obtain corresponding predicted trajectory coordinates and S24: calculating a loss function S25: inputting the observation state in the validation set into the Attention-TCN model to obtain corresponding predicted trajectory coordinate values, and extracting observation trajectory coordinates corresponding to the observation state from the validation set calculating the mean square error between the predicted trajectory coordinate values and the observation trajectory coordinates val to obtain the loss Loss on the validation set val ; S26: judging the size between the loss Loss val on the validation set and the loss function and the size between Loss val and best_Loss;

[0008] If: and Loss val < best_Loss, the parameters and are saved, and otherwise, the parameters and updating; S27: if epoch>0, go to step S22; if epoch=0, stop the loop; S3: use the trained physical information deep learning model to conduct a test experiment, and compare the prediction performance of different lane change trajectory prediction models.

[0009] As a further improvement of the application, the specific steps for training the physical information deep learning model are as follows: the loss function The expression is:

[0010]

[0011] In the above formula, MSE dat is the root mean square error of the data error between the observed trajectory coordinates P i and the predicted trajectory coordinates MSE phy is the root mean square error of the data error between the predicted trajectory coordinates and the predicted trajectory coordinates P i is the observed trajectory coordinates corresponding to the observed state in the training set, and α is the weight of balancing MSE dat and MSE phy , and are the values of P i , and at time t.

[0012] As a further improvement of the application, in S26, the parameters and are updated by using optimizers with learning rates lr dat and lr phy respectively.

[0013] As a further improvement of the application, the specific steps for constructing the data set in S1 are as follows: S11: extract the input state I and label P of the vehicle from the public trajectory data set, and form a data pair by combining the input state I and the label P; wherein the input state I is composed of the historical feature vectors of the target vehicle and surrounding vehicles in the past K time steps, and the label P is composed of the lane change trajectory coordinates of the target vehicle in the future L time steps; S12: divide the input state I into the observed state I O and the matching state I C according to the proportion of N i : N j , extract the observed trajectory coordinates P i corresponding to the observed state I i from the data pair, and use the observed trajectory coordinates Pi and observation state I i pairing the observation data set and the observation data set into a training subset, a validation set and a test set; wherein, the observation state I i including the observation state in the training set the observation state in the validation set and the observation state in the test set observation trajectory coordinates P i is the label corresponding to the observation state I i in P; N O and N C are the number of observation data and collocation data respectively; S13: extracting a sub-collocation state from the collocation state I j inputting the sub-collocation state into the physical prediction model to obtain collocation trajectory coordinates of the target vehicle pairing the collocation state I j and the collocation trajectory coordinates of the target vehicle into a collocation data set; wherein, the sub-collocation state is composed of the trajectory horizontal coordinate, the trajectory vertical coordinate, the lateral velocity and the longitudinal velocity of the target vehicle at time p; S14: combining the collocation data set and the training subset in the observation data set to form a complete training set.

[0014] As a further improvement of the present application, the time p is the time when the target vehicle generates a lane change intention; the historical feature vector is composed of the trajectory horizontal coordinate, the trajectory vertical coordinate, the lateral velocity, the longitudinal velocity, the lateral acceleration and the longitudinal acceleration from time step p-K+1 to p; and the lane change trajectory coordinates are composed of the trajectory horizontal coordinate and the trajectory vertical coordinate from time step p+1 to p+L.

[0015] As a further improvement of the present application, the input state I and the label P of the vehicle are extracted from the public data set in S11, and the specific steps are as follows: S111: screening the trajectories of passenger cars that only perform one lane change operation and have complete lane change trajectories from the public data set; S112: extracting the screened trajectories according to the lane change start time and the end time to obtain the lane change trajectory.

[0016] As a further improvement of the present application, the Attention-TCN model uses TCN as the extractor of the lane change trajectory feature, and captures the dynamic interaction information of vehicles at different times and spaces by means of the time attention mechanism and the spatial attention mechanism.

[0017] As a further improvement of the present application, a quintic polynomial is used to model the physical prediction model, and the lateral trajectory coordinates in the obtained physical prediction model are parameterized by the maximum lateral acceleration ​​ The expression is:

[0018]

[0019] In the above formula, W represents the lane width. The maximum lateral acceleration is given by t, where t is time. This represents the lateral trajectory coordinates of the target vehicle at the start of the lane change in the image coordinate system.

[0020] As a further improvement of the present invention, the trained physical information deep learning model is tested on a test set to compare the performance of different lane change trajectory prediction models. The different lane change trajectory prediction models include: data-driven Attention-TCN model, physical-driven physical prediction model, LSTM model and Transformer model.

[0021] As a further improvement of the present invention, the performance of the above-mentioned different lane change trajectory prediction models is evaluated using the average displacement error, the final displacement error, and the relative displacement error.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. In this invention, a physical prediction model is constructed and combined with the Attention-TCN model to form a physical information deep learning model. Compared with the physical prediction model and the Attention-TCN model alone, this physical information deep learning model is rooted in physical principles and improves the interpretability and understandability of lane change trajectory prediction.

[0024] 2. The Attention-TCN model proposed in this invention captures vehicle dynamic interaction information under different time and space conditions by using temporal and spatial attention mechanisms, thereby improving TCN's ability to extract vehicle features in complex and highly interactive driving environments.

[0025] 3. This invention can predict long-term lane-changing trajectories with high prediction accuracy and high practical applicability. It can help autonomous vehicles detect potential dangerous situations in advance and use them for trajectory planning and decision-making, thereby improving driving safety. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the framework of the lane change trajectory prediction method in an embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram illustrating the construction of the dataset in an embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram of the vehicle lane-changing trajectory according to an embodiment of the present invention;

[0029] Figure 4 A schematic diagram of the data-driven Attention-TCN model in the embodiment of the present application;

[0030] Figure 5 A basic structure schematic diagram of the TCN in the embodiment of the present application. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0032] As shown in the formula (1), the embodiment of the present application proposes a vehicle lane-changing trajectory prediction method based on physical information deep learning, which includes the following steps: Figures 1-5 S1: constructing a data set, wherein the step of constructing the data set is:

[0033] As shown in the formula (2), S11: extracting the input state I and the label P of the vehicle from the public trajectory data set, and forming a data pair with the input state I and the label P, the public trajectory data set can be the HighD data set.

[0034] Figure 2 As shown in the formula (3), the input state I is composed of the historical feature vectors of the target vehicle and the surrounding vehicles in the past K time steps, and the historical feature vector is composed of the trajectory abscissa, the trajectory ordinate, the lateral velocity, the longitudinal velocity, the lateral acceleration and the longitudinal acceleration from time step p-K+1 to p, and p is the time when the target vehicle generates the lane-changing intention.

[0035] The label P is composed of the lane-changing trajectory coordinates of the target vehicle in the future L time steps, and the lane-changing trajectory coordinates are composed of the trajectory abscissa and the trajectory ordinate from time step p+1 to p+L. In this embodiment, K=4s, L=2s.

[0036] As shown in the formula (4), in addition, the input state I and the label P of the vehicle are extracted from the HighD data set in S11, and the specific steps are:

[0037] As shown in the formula (5), S111: track screening, screening the track of the passenger car which only performs one lane-changing operation and has a complete lane-changing trajectory from the HighD data set. Figure 3

[0038] S111: track screening, screening the track of the passenger car which only performs one lane-changing operation and has a complete lane-changing trajectory from the HighD data set.

[0039] ​​S112: Lane change trajectory extraction. The selected trajectory is extracted based on the start and end times of the lane change to obtain the lane change trajectory.

[0040] It should be noted that the definitions of lane-changing start and end times are based on existing technology and are specifically defined according to the reference (Thiemann C, Treiber M, etc. Estimating acceleration and lane-changing dynamics from next generation simulation trajectory data[J]. Transportation Research Record, 2008, 2088(1):90-101.).

[0041] In this embodiment, lane-changing trajectories of 2556 vehicles were selected from the HighD dataset, with 2556 input states I and 2556 labels P.

[0042] S12: Arrange the input state I according to N O :N C The observation state is divided into two categories with a ratio of 3:1. i and matching state I j Extract the observed state I from the data pair i The corresponding observation trajectory coordinates P i The coordinates of the observed trajectory P i and observation state I i The observation dataset is paired up and divided into a training subset, a validation set, and a test set in a 1:1:1 ratio.

[0043] It should be noted that the observation state refers to the input state that is input into the Attention-TCN model, while the collocation state refers to the input state that is input into the physical prediction model.

[0044] Among them, observation state I i Including the observation states in the training set Observational states in the validation set and the observed states in the test set Observation trajectory coordinates P i It is in P and the observed state I i The corresponding label; the N O and N C These refer to the quantity of observational data and the quantity of paired data, respectively.

[0045] S13: From the matching state I j Extracting sub-matching states Pairing the children Input into the physical prediction model to obtain the target vehicle's matching trajectory coordinates The matching state I j And the matching trajectory coordinates of the target vehicle The matching data set is composed of the matching state and the matching trajectory coordinates of the target vehicle.

[0046] Wherein, the sub-matching state is composed of the trajectory horizontal coordinate, the trajectory vertical coordinate, the horizontal velocity and the vertical velocity of the target vehicle at time p.

[0047] S14: Combine the matching data set and the training subset in the observation data set to form a complete training set. In this way, the training set, the validation set in the observation data set and the test set in the observation data together form a data set required for training the physical information deep learning model.

[0048] S2: Build a physical prediction model, and combine it with the Attention-TCN model to form a physical information deep learning model, and train the physical information deep learning model using the built data set to obtain the trained physical information deep learning model. The specific training steps are as follows:

[0049] As Figure 1 shown, S21: initialize parameters α=0.6, N epoch =30 and best_Loss=0.001; let epoch=N epoch ;

[0050] Wherein, is the parameter of the Attention-TCN model, is the parameter of the physical prediction model, and α is the weight parameter, N epoch is the number of training times, and best_Loss is the optimal loss value.

[0051] S22: let epoch=epoch-1.

[0052] S23: input the observation state into the Attention-TCN model to obtain the corresponding predicted trajectory coordinates The matching state I j is input into the Attention-TCN model and the physical prediction model respectively to obtain the corresponding predicted trajectory coordinates And

[0053] S24: calculate the loss function

[0054] Wherein, the loss function is composed of data error and physical error with different weight proportions, The expression is:

[0055]

[0056] In the above formula, MSE dat is the root mean square error of data error between the observed trajectory coordinates P i and the predicted trajectory coordinates ;

[0057] MSE phy is the root mean square error of data error between the predicted trajectory coordinates and the predicted trajectory coordinates ;

[0058] P i is the observed trajectory coordinates corresponding to the observed state in the training set, and α is the weight of balancing MSE dat and MSE phy , and are the values of P i , and at time t respectively.

[0059] S25: input the observed state in the validation set into the Attention-TCN model to obtain the corresponding predicted trajectory coordinate value, and extract the observed trajectory coordinates corresponding to the observed state from the validation set; calculate the mean square error of the predicted trajectory coordinate value and the observed trajectory coordinates , and obtain the loss Loss val on the validation set.

[0060] S26: judge the size of the loss Loss val on the validation set and the loss function and the size of the loss Loss val and best_Loss.

[0061] If: and Loss val < best_Loss, save the parameters and and let

[0062] Otherwise, update the parameters and .

[0063] It should be noted that the updating manner can be selected to update the parameters dat and phy using Adam optimizers with learning rates lr SV = 0.0005 and lr e = 0.0005, respectively. and .

[0064] S27: If epoch > 0, go to step S22. If epoch = 0, it indicates that the training is completed, and the loop is stopped.

[0065] As shown in FIG. 6, it is an Attention-TCN model diagram in the embodiment. The Attention-TCN uses TCN as a lane-changing trajectory feature extractor, and captures vehicle dynamic interaction information at different times and spaces by means of time attention mechanism and space attention mechanism, so as to improve the feature extraction capability thereof. Figure 4 As shown in FIG. 7, it is a basic structure diagram of TCN. Two basic operations of TCN are causal convolution and dilated convolution. The causal convolution is a one-way structure from cause to effect which is time-constrained, and the dilated convolution is to expand the sampling interval of the convolution kernel on the basis of the causal convolution, and the number of neurons in each layer is exponentially reduced by setting the value of the dilated rate d.

[0066] Figure 5 The specific modeling process of the Attention-TCN model is as follows: as shown in FIG. 8,

[0067] Figures 4-5

[0068] First step: input the historical feature vector E of the target vehicle into the TCN encoding layer to obtain the hidden layer state vector H of the target vehicle, and the time attention weight E of the target vehicle is calculated as follows:

[0069] E SV = softmax(tanh(A e H SV )) (1)

[0070] In the formula, A e is a learnable parameter. In the embodiment, α = 0.6, N epoch = 30, and best_Loss = 0.001.

[0071] Second step: weight E SV to H SV to obtain the weighted intermediate state vector Q SV , and the calculation formula of Q SV is as follows:

[0072]

[0073] Similarly, the intermediate state vector of the surrounding vehicles FV T , LV T , FV C , and LV C is obtained and The intermediate state vector of all vehicles is defined as Q.

[0074] Step 3: Set the spatial attention weight of all vehicles as F Then F can be calculated by the following formula:

[0075] F = softmax(tanh(A f Q)) (3)

[0076] In the above formula, A f is a learnable parameter; f SV is the spatial attention weight of the target vehicle, and is the spatial attention weight of the surrounding vehicles. The number of surrounding vehicles selected in this embodiment is four.

[0077] Step 4: Weight Q with F to obtain the weighted intermediate state vector V, and the calculation formula of V is:

[0078]

[0079] In the above formula: w = 1 represents the target vehicle SV, w = 2, 3, 4 and 5 respectively represent the surrounding vehicles FV T , LV T , FV C , and LV C .

[0080] Step 5: Input V into the feedforward neural network layer to predict the lane change trajectory coordinates of the target vehicle in the future L time steps

[0081] It should be noted that the feedforward neural network is part of the structure of the Attention-TCN model.

[0082] The specific steps of constructing the physical prediction model using a quintic polynomial are as follows:

[0083] Step 1: Assume that the longitudinal speed of the target vehicle during the lane change process is constant, then the lane change longitudinal trajectory coordinates of the target vehicle are represented as:

[0084] ​​

[0085] Lane-changing lateral trajectory coordinates is expressed as:

[0086]

[0087] In the above formula, is the trajectory coordinate of the target vehicle at the beginning of lane-changing in the image coordinate system, a1, a2, a3, a4 and a5 are the coefficients of the polynomial.

[0088] Second step: the lateral displacement The derivative can obtain the lateral velocity of the target vehicle in the lane-changing process and the lateral acceleration

[0089] Wherein, the expression of the lateral velocity

[0090] The expression of the lateral acceleration

[0091]

[0092]

[0093] Third step: assuming that the target vehicle satisfies the following constraints at the beginning and end positions of lane-changing:

[0094]

[0095] In the above formula, t e is the total lane-changing time of the target vehicle, y e is the total longitudinal displacement of the target vehicle lane-changing, and W is the lane width. Then the expression of the lateral trajectory coordinate of the target vehicle is:

[0096]

[0097] Further simplified as:

[0098]

[0099] In the above formula, t∈[0, t e ].

[0100] Fourth step: derivative of formula (10) to obtain the lateral velocity and the lateral acceleration

[0101] Wherein, the expression of the lateral velocity

[0102] ​​​

[0103] lateral acceleration The expression of the lateral acceleration is:

[0104]

[0105] Step 5: Derive the formula (12) to get the maximum lateral acceleration of the target vehicle in the lane changing process The expression of the maximum lateral acceleration is:

[0106]

[0107] Step 6: The total lane changing time t e of the target vehicle The expression of the total lane changing time is:

[0108]

[0109] Step 7: Substitute the formula (14) into the formula (10) to get the lane changing lateral trajectory coordinates of the target vehicle with the maximum lateral acceleration as the parameter The expression of the lane changing lateral trajectory coordinates is:

[0110]

[0111] Step 8: When the lane changing ends, i.e., t > t e , the target vehicle continues to travel on the target lane with the longitudinal speed , at this time

[0112] S3: Test the trained physical information deep learning model to compare the prediction performance of different lane changing trajectory prediction models.

[0113] As shown in Figure 1 , specifically, the trained physical information deep learning model is tested on the test set to compare the performance of different lane changing trajectory prediction models, including: data-driven Attention-TCN model, physics-driven physical prediction model, LSTM model and Transformer model. Among them, the historical time step K = 4s and the prediction time step L = 2s are down-sampled every 0.5s, i.e. predicting the lane changing trajectory coordinates of the target vehicle p at 0.5s, 1s, 1.5s and 2s after the time.

[0114] In addition, the performance of the above different lane changing trajectory prediction models is evaluated by using the average displacement error (ADE), the final displacement error (FDE) and the relative displacement error (RDE). ​

[0115] The expression of average displacement error (ADE) is:

[0116]

[0117] The expression of final displacement error (FDE) is:

[0118]

[0119] The expression of relative displacement error (RDE) is:

[0120]

[0121] In the above formula, N is the number of vehicles in the test set, and represents the real trajectory horizontal coordinate and vertical coordinate of the i-th vehicle at time t, and represents the predicted trajectory horizontal coordinate and vertical coordinate of the i-th vehicle at time t, and represents the real trajectory horizontal coordinate and vertical coordinate of the i-th vehicle at time p+L, and represents the predicted trajectory horizontal coordinate and vertical coordinate of the i-th vehicle at time p+L, and represents the real trajectory horizontal coordinate and vertical coordinate of the i-th vehicle at time p.

[0122] The comparison results of the physical information deep learning model with the data-driven Attention-TCN model, the physics-driven physical prediction model, the LSTM model and the Transformer model are shown in Table 1,

[0123] Table 1. Comparison of prediction results of different models

[0124]

[0125]

[0126] As can be seen from Table 1, compared with the data-driven Attention-TCN and the physics-driven physical prediction model, the prediction accuracy of the present application is higher, especially when the prediction time step is increased. Compared with the popular methods such as the LSTM model and the Transformer model, the present application also has better prediction performance. This shows that the present application can fully utilize the advantages of data-driven models and physics-driven models, and realize more accurate and interpretable vehicle lane change trajectory prediction for long time series.

[0127] It should be understood that when an element or layer is referred to as being "on" another element or substrate, it can be directly on the other element or substrate or intervening elements can also be present. In contrast, when an element is referred to as being "directly on" another element or substrate, there are no intervening elements present. It will be further understood that the terms "comprises" and / or "comprising" used in the specification are open-ended and are intended to mean including but not limited to. Similarly, the term "comprising" is used in the claims to mean "including but not limited to".

[0128] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.

[0129] Various features of the above-described embodiments can be combined in any combination, and the description is not limited to the embodiments described above. The above-described embodiments are merely exemplary and do not limit the scope of the application.

[0130] The above-described embodiments are merely exemplary and do not limit the scope of the application. It should be understood that, for the person of ordinary skill in the art, without departing from the concept of the application, a number of variations and modifications can be made, and these all belong to the protection scope of the application. Therefore, the protection scope of the application should be subject to the appended claims.

Claims

1. A vehicle lane-changing trajectory prediction method based on physical information deep learning, characterized in that, It comprises the following steps: S1: Construct a dataset using a public trajectory dataset, and divide the dataset into a training set, a validation set, and a test set, extract observation states from the training set and collocation states , and extract observation states from the validation set ; the observation states refer to input states input into an Attention-TCN model, and the collocation states refer to input states input into a physical prediction model; S2: Construct a physical prediction model, and compose a physical information deep learning model with an Attention-TCN model, and train the physical information deep learning model with the constructed data set to obtain a trained physical information deep learning model, and the specific training steps are as follows: S21: initialize parameters , , α , N epoch and best_Loss ; let epoch = N epoch ; wherein, are parameters of the Attention-TCN model, are parameters of the physical prediction model, α are weight parameters, N epoch is the number of training times, best_Loss is the optimal loss value; S22: Let epoch = epoch - 1; S23: input the observation state into the Attention-TCN model to obtain the corresponding predicted trajectory coordinates , the collocation state into the Attention-TCN model and the physical prediction model respectively, and the corresponding predicted trajectory coordinates are obtained respectively and ; S24: compute the loss function ; S25: the observation state in the verification set is verified The corresponding predicted trajectory coordinate value is obtained by inputting into the Attention-TCN model, and the observation state in the verification set is extracted The corresponding observation trajectory coordinate The mean square error of the predicted trajectory coordinate value and the observation trajectory coordinate is calculated, and the loss on the verification set is obtained ; S26: judging the loss on the verification set with the loss function and with best_Loss the size between; If and then save parameters and and let ; Otherwise, update the parameters and ; S27: If epoch > 0, go to step S22; if epoch = 0, stop the loop; S3: Test the trained physical information deep learning model, and compare the prediction performance of different lane-changing trajectory prediction models; wherein the loss function is expressed as: In the above formulae: MSE dat is the root mean square error of data error between the observed trajectory coordinates and the predicted trajectory coordinates ; MSE phy is the root mean square error of data error between the predicted trajectory coordinates and the predicted trajectory coordinates ; is the observed trajectory coordinates corresponding to the observed state in the training set, α is the weight parameter balance MSE dat and MSE phy the weight of , , and respectively , , and the value at the time t ; N O and N C are the number of observation data and collocation data respectively; L is the prediction time step; p is the prediction target vehicle time; The Attention-TCN model uses TCN as a lane-changing trajectory feature extractor, and captures vehicle dynamic interaction information at different times and spaces through time attention mechanism and space attention mechanism. 2.The vehicle trajectory prediction method based on physical information deep learning according to claim 1, wherein, In S26, the parameters and are updated using an optimizer. 3.The vehicle trajectory prediction method based on physical information deep learning according to claim 1, wherein, The specific steps for constructing the data set in S1 are as follows: S11: extracting an input state of a vehicle from a published trajectory dataset I and a label P and the input state I and a label P into a data pair; wherein the input state I consists of past K time step feature vectors of the target vehicle and surrounding vehicles, and a label P consists of future L time step lane change trajectory coordinates of the target vehicle S12: divide the input state I According to N O : N C The observation state and the matching state , extract the observation trajectory coordinates corresponding to the observation state from the data pair , pair the observation trajectory coordinates and the observation state to form an observation data set, and divide the observation data set into a training subset, a validation set and a test set; wherein the observation state includes observation states in the training set , observation states in the validation set , and observation states in the test set , the observation trajectory coordinates are P labels corresponding to the observation state in the observation trajectory S13: extracting a sub-planning state from the planning state S14: inputting the sub-planning state into a physical prediction model to obtain planning trajectory coordinates of the target vehicle S15: inputting the planning trajectory coordinates of the target vehicle into a planning model to obtain a planning trajectory of the target vehicle S16: inputting the planning trajectory of the target vehicle into a physical prediction model to obtain a planning trajectory coordinate of the target vehicle S17: inputting the planning trajectory coordinate of the target vehicle into a planning model to obtain a planning trajectory of the target vehicle S18: inputting the planning trajectory of the target vehicle into a physical prediction model to obtain a planning trajectory coordinate of the target vehicle S19: inputting the planning trajectory coordinate of the target vehicle into a planning model to obtain a planning trajectory of Wherein, the sub-coordination state is composed of the trajectory abscissa, the trajectory ordinate, the lateral velocity and the longitudinal velocity of the target vehicle at p the moment; S14: Combine the training sub-set in the collocation data set and the observation data set to form a complete training set.

4. The vehicle lane-changing trajectory prediction method based on physical information deep learning according to claim 3, characterized in that, The p The time is the time when the target vehicle generates the lane-changing intention. The history feature vector consists of trajectory cross coordinate, trajectory longitudinal coordinate, cross velocity, longitudinal velocity, cross acceleration, and longitudinal acceleration from time step p-K+1 to p ​ The lane change trajectory coordinates consist of a trajectory lateral coordinate and a trajectory longitudinal coordinate from a time step p+1 to p+L .

5. The vehicle lane-changing trajectory prediction method based on physical information deep learning according to claim 3, characterized in that, In S11, the input state of the vehicle is extracted from the public data set I and the tag P , and the specific steps are: S111: Screen out the trajectories of passenger cars that only perform one lane-changing operation and have complete lane-changing trajectories from the public data set; S112: Extract the screened trajectories according to the lane-changing start time and end time to obtain the lane-changing trajectories.

6. The vehicle trajectory prediction method based on physical information deep learning according to claim 1, wherein, The physical prediction model is modeled using a quintic polynomial, and the expression for the lateral trajectory coordinate with the maximum lateral acceleration as a parameter is: ​ In the above formula, W is the lane width, is the maximum lateral acceleration, t is the time, is the lateral trajectory coordinate of the target vehicle at the beginning of the lane change in the image coordinate system.

7. The vehicle trajectory prediction method based on physical information deep learning according to claim 1, characterized in that, Test the trained physical information deep learning model on the test set, compare the performance of different lane-changing trajectory prediction models, and different lane-changing trajectory prediction models include: data-driven Attention-TCN model, physics-driven physical prediction model, LSTM model and Transformer model. 8.The vehicle trajectory prediction method based on physical information deep learning according to claim 7, wherein, The performance of the above different lane-changing trajectory prediction models is evaluated by using the average displacement error, the final displacement error and the relative displacement error.

Images