Method and system for predicting vehicle track in area without lane line

Abstract

The invention discloses a vehicle trajectory prediction method and system for a lane-free area, and the method comprises the steps: obtaining the trajectory data of a target vehicle in the lane-free area in real time, carrying out the data preprocessing, and obtaining a derivative information set of the trajectory data at each moment; inputting all the derivative information sets into a hidden Markov model to carry out vehicle transverse motion intention recognition training so as to predict a transverse motion state and add the transverse motion state into the corresponding derivative information sets; classifying variables in a derivative information set containing a transverse motion state prediction result, and inputting the variables into a time sequence fusion Transformer model to carry out vehicle track prediction training so as to predict a vehicle position; and obtaining new trajectory data of the target vehicle in real time, and inputting the new trajectory data into the trained hidden Markov model and the time sequence fusion Transform model in sequence to obtain a trajectory prediction result. According to the method, the accuracy of predicting the running track of the vehicle in the non-lineation area by the time sequence fusion Transform model is improved.

Description

Technical Field

[0001] The present invention relates to traffic safety technology, and in particular to a vehicle trajectory prediction method and system in a lane-free area. Background Art

[0002] Accurate trajectory prediction is an important basis for autonomous vehicles to make correct driving decisions and avoid driving hazards. Previous trajectory prediction research methods are mainly applicable to conventional roads, and can achieve accurate vehicle trajectory prediction in environments with lane markings. However, on complex roads without lane markings, such as road intersections, diverging and merging areas of toll plazas, and diverging and merging areas of ramps on highways and expressways, vehicle driving is not restricted by lane rules, but only by its target driving position. Compared with ordinary roads, the traffic environment in unmarked areas is more complex, so vehicle driving often presents disordered characteristics, such as sudden turns or lane changes, resulting in traffic conflicts at various angles, which brings huge challenges to vehicle trajectory prediction.

[0003] With the continuous development of machine learning technology, more and more deep learning methods are used to predict time series data. One of the advanced representative methods is the Temporal Fusion Transformer (TFT) model. This model can flexibly input variables with the highest correlation with the predicted time series data and can capture the short-term and long-term dependencies of the time series data. However, this algorithm is not used to predict time series data such as vehicle trajectories, especially in areas without lane lines. The starting and ending points of the vehicle's lateral movement are not certain, and the direction and angle of movement are more variable. In complex traffic environments and where vehicles are prone to sudden and disordered lateral movement, the trajectory prediction accuracy of the TFT model in this area is insufficient. Summary of the invention

[0004] The technical problem to be solved by the present invention is: in response to the above-mentioned problems of the prior art, a method and system for predicting vehicle trajectory in an area without lane lines is provided, which improves the accuracy of the time series fusion Transformer model in predicting the vehicle driving trajectory in an area without lane lines by adding two measures: early prediction of the vehicle's lateral disordered movement intention and classification of the model's input variables.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A method for predicting vehicle trajectories in a lane-free area, comprising the following steps:

[0007] Step 1: Obtain the trajectory data of the target vehicle in the area without lane lines in real time, and perform data preprocessing to obtain the derived information set of the trajectory data at each moment;

[0008] Step 2: All derived information sets are input into the hidden Markov model for vehicle lateral motion intention recognition training, so that the trained hidden Markov model predicts the lateral motion state of the target vehicle at the next moment based on the derived information set of the target vehicle's historical trajectory data, and adds the lateral motion state prediction result at each moment to the corresponding derived information set;

[0009] Step 3: Classify the variables in the derived information set containing the lateral motion state prediction results and input them into the time series fusion Transformer model for vehicle trajectory prediction training, so that the trained time series fusion Transformer model can predict the vehicle position at the next moment based on the derived information set containing the lateral motion state prediction results at the historical moment of the target vehicle;

[0010] Step 4: After acquiring the new trajectory data of the target vehicle in real time and performing data preprocessing, the corresponding derivative information set is input into the trained hidden Markov model to obtain a derivative information set containing the lateral motion state prediction result. The derivative information set containing the lateral motion state prediction result is then classified and input into the trained time series fusion Transformer model to obtain the trajectory prediction result.

[0011] Furthermore, the steps of data preprocessing in step 1 and step 4 both include:

[0012] The target vehicle’s trajectory data is added to the corresponding trajectory data set to obtain the target vehicle’s vehicle type T. Based on the coordinates of the trajectory data at the initial and final moments in the trajectory data set, the target vehicle’s initial lane L is obtained. a and the final lane L leaving the lane-free area b , the vehicle type T of the target vehicle, the initial lane L a and final lane L b Add the derivative information set at the corresponding moment;

[0013] According to the coordinates of the trajectory data of the target vehicle at adjacent moments in the trajectory data set, the heading angle θ of the target vehicle at each moment and the longitudinal speed v of the target vehicle at each moment are calculated. x , lateral velocity v y , the longitudinal speed difference between the target vehicle and other vehicles in the affected area Δv1, Δv2, Δv3, Δv4, and then add the target vehicle heading angle θ, longitudinal speed v x , lateral velocity v y , longitudinal speed differences Δv1, Δv2, Δv3, Δv4 are added to the derivative information set at the corresponding time.

[0014] Furthermore, when calculating the longitudinal speed differences Δv1, Δv2, Δv3, Δv4 between the target vehicle and other vehicles in the affected area based on the vehicle trajectory data set, it includes:

[0015] The geometric center of the target vehicle is taken as the origin of the affected area, and the affected area is determined according to the specified length and width. The affected area includes four areas: the left front, right front, left rear, and right rear of the target vehicle. The longitudinal speed differences Δv1, Δv2, Δv3, and Δv4 between the target vehicle and other vehicles in each area are calculated respectively. If there is no other vehicle in a certain area, the corresponding longitudinal speed difference is 0.

[0016] Furthermore, in step 1, the step of performing data preprocessing also includes: normalizing the variables in the derived information set, and mapping all variable values ​​in the derived information set to the interval [0,1].

[0017] Furthermore, in step 2, when the derived information set is input into the hidden Markov model for vehicle lateral motion intention recognition training, it includes:

[0018] Divide the derived information set at each moment into training set, validation set, and test set according to the specified ratio;

[0019] First, the training set is input into the hidden Markov model for training to determine the model parameters, then the validation set is input into the trained hidden Markov model to adjust the model parameters, and finally the test set is input into the adjusted hidden Markov model to verify the model prediction effect. This step is repeated until the loss function corresponding to the hidden Markov model meets the convergence condition.

[0020] Furthermore, in step 2, when the trained hidden Markov model predicts the lateral motion state of the target vehicle at the next moment based on the derived information set of the target vehicle's historical moment trajectory data, the trained hidden Markov model obtains the derived information set of each historical moment trajectory data, and outputs the recognition result with the maximum probability as the lateral motion state variable D of the target vehicle at the next moment. y Furthermore, in step 3, the variables in the derived information set containing the lateral motion state prediction result are classified and input into the time series fusion Transformer model for vehicle trajectory prediction training, including:

[0021] Dividing the derived information set into a training set, a validation set, and a test set according to a specified ratio, and classifying the variables in the derived information sets of the training set, the validation set, and the test set respectively;

[0022] First, the classified variables of the training set are input into the time series fusion Transformer model for training to determine the model parameters. Then, the classified variables of the validation set are input into the trained time series fusion Transformer model to adjust the model parameters. Finally, the classified variables of the test set are input into the adjusted time series fusion Transformer model to verify the model prediction effect. This step is repeated until the loss function corresponding to the time series fusion Transformer model meets the convergence conditions.

[0023] Further, the variables in the derived information sets of the training set, the validation set, and the test set are classified respectively, and the derived information set containing the lateral motion state prediction result is classified and input into the trained time series fusion Transformer model in step 4, both of which include: the trajectory data coordinates {x i ,y i} is used as the historical data of the predicted target in the time series fusion Transformer model, and the target vehicle heading angle θ and the target vehicle longitudinal speed v in the derived information set are x , lateral velocity v y The longitudinal speed differences Δv1, Δv2, Δv3, and Δv4 between the target vehicle and other vehicles in the affected area are used as observable input variables in the temporal fusion Transformer model, and the lateral motion state D in the derived information set is converted into y The prediction results are used as known variables in the time series fusion Transformer model, and the vehicle type T of the target vehicle in the derived information set, the initial lane L of the target vehicle entering the lane-free area, and the lane type L of the target vehicle in the lane-free area. a and the final lane L leaving the lane-free area b As a static covariate in the time series fusion Transformer model.

[0024] Furthermore, in step 3, when the trained time series fusion Transformer model predicts the vehicle position at the next moment based on the derived information set containing the lateral motion state prediction results at the target vehicle's historical moments, it includes: the trained time series fusion Transformer model obtains the derived information set containing the lateral motion state prediction results at each historical moment, and outputs the prediction result with the largest weight as the vehicle position at the next moment.

[0025] The present invention also proposes a vehicle trajectory prediction system for a lane-free area, comprising a storage medium of interconnected microprocessors, wherein the microprocessors are programmed or configured to execute any of the vehicle trajectory prediction methods for a lane-free area.

[0026] Compared with the prior art, the advantages of the present invention are:

[0027] Before using the time series fusion Transformer model to predict the trajectory, the present invention first uses a hidden Markov model to predict the vehicle's lateral movement intention; when using the time series fusion Transformer model to predict the trajectory, the variables in the information set are classified and input into the model, which effectively improves the prediction accuracy of the time series fusion Transformer model and can accurately predict the driving trajectory of the target vehicle in the area without lane lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of a method according to an embodiment of the present invention.

[0029] Figure 2 Schematic diagram of the affected area of ​​the vehicle in the embodiment of the present invention. DETAILED DESCRIPTION

[0030] The present invention is further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present invention is not limited thereby.

[0031] Embodiment 1

[0032] In order to improve the accuracy of the TFT algorithm in predicting the vehicle's driving trajectory in the unlined area, this embodiment proposes a vehicle trajectory prediction method in the laneless area, which improves the prediction accuracy of the TFT model by increasing the early prediction of the vehicle's lateral disordered movement intention and classifying the input variables of the model. Figure 1 As shown, the specific steps include:

[0033] Step 1: Obtain the trajectory data of the target vehicle in the area without lane lines in real time, and perform data preprocessing to obtain a derivative information set of the trajectory data at each moment. The steps of data preprocessing include:

[0034] Step 1.1 Number each recorded vehicle, record it as vehicle ID, and sort it from small to large; at the same time, number the time of the entire data collection process, record it as frame ID, the interval between two adjacent collection moments of each vehicle's trajectory data is 0.1s, and the frame ID is sorted from small to large. Add the trajectory data of the target vehicle to the corresponding trajectory data set, and the trajectory data set can be expressed as P = {P1, P2, P3, ..., P i}, where i represents the current position of the vehicle at the i-th time point, including the horizontal and vertical coordinates (x i ,y i ).

[0035] Step 1.2: Based on the trajectory data set of each target vehicle, the following data can be directly obtained and added to the derived information set at the corresponding time:

[0036] ① Vehicle type T (according to the scenario type, the vehicle type can be divided into cars, trucks, and buses, and can also be divided into ETC vehicles and MTC vehicles according to the highway toll collection method, etc.);

[0037] ②According to the coordinates of the trajectory data at the initial and final moments in the trajectory data set, the initial lane L of the vehicle entering the lane-free area is obtained. a and the final lane L leaving the lane-free area b .

[0038] Step 1.3 Based on the trajectory data set of each target vehicle, the following data can be calculated and added to the derived information set at the corresponding time:

[0039] ① The vehicle heading angle at each moment

[0040] ②The longitudinal speed of the vehicle at each moment Lateral speed

[0041] ③ The longitudinal speed difference Δv1, Δv2, Δv3, Δv4 between the vehicle and other vehicles in the affected area. When the longitudinal speed difference Δv1, Δv2, Δv3, Δv4 between the target vehicle and other vehicles in the affected area is calculated based on the vehicle trajectory data set, it includes:

[0042] Taking the geometric center of the target vehicle as the origin of the affected area, the area around the target vehicle is divided into four areas: left front, right front, left rear, and right rear according to the range of the first length before and after the origin and the second length on the left and right sides. Other vehicles in any area will affect the driving path of the target vehicle. The longitudinal speed difference between the target vehicle and other vehicles in each area is calculated respectively, where the longitudinal speed difference between the target vehicle and other vehicles in the left front area is Δv1, the longitudinal speed difference between the target vehicle and other vehicles in the right front area is Δv2, the longitudinal speed difference between the target vehicle and other vehicles in the left rear area is Δv3, and the longitudinal speed difference between the target vehicle and other vehicles in the right rear area is Δv4. In this embodiment, the first length is 30m and the second length is 5m. Figure 2 As shown in the figure, the four areas of the target vehicle's left front, right front, left rear and right rear are regarded as its affected areas. Vehicles entering these four areas will affect the driving path of the current vehicle. At each moment, the longitudinal speed difference between the current vehicle and other vehicles in the four surrounding areas is calculated, which are recorded as Δv1, Δv2, Δv3 and Δv4 respectively. If there is no other vehicle in a certain area, the speed difference is recorded as 0.

[0043] In summary, in step 1, the derived information set contained in the trajectory data of the target vehicle at the i-th time point is Q i ={frame ID,vehicle ID,xi ,y i ,v x ,v y ,θ,T,L a ,L b ,Δv1,Δv2,Δv3,Δv4}. At the same time, in order to eliminate the influence of inconsistent variable size range on the prediction results, the derived information set Q i The variables in are normalized and the derived information set Q i All variable values ​​in are mapped to the interval [0,1].

[0044] Step 2: All derived information sets are input into the hidden Markov model for vehicle lateral motion intention recognition training, so that the trained hidden Markov model predicts the lateral motion state of the target vehicle at the next moment based on the derived information set of the target vehicle's historical trajectory data, and adds the lateral motion state prediction result at each moment to the corresponding derived information set;

[0045] The expressions of hidden Markov models include:

[0046] λ={A,B,π} (1)

[0047] A=[α ij ] N×N (2)

[0048] B=[b j (k)] N×M (3)

[0049] Among them, A is an N×N dimensional state transition probability matrix, α ij is the probability of state transition from i to j; α ij Satisfy the conditions as well as B is an N×M observable probability distribution matrix that satisfies π is the initial state probability.

[0050] In this embodiment, the derived information set of each vehicle obtained in step 1 is

[0051] {frame ID, vehicle ID, x i ,y i ,v x ,v y ,θ,T,L a ,L b,Δv1,Δv2,Δv3,Δv4} are added as vehicle features to the Hidden Markov Model (HMM) for vehicle lateral motion intention recognition training. The goal is to train the model to predict the lateral motion state D of the vehicle at the i-th moment based on the vehicle's trajectory data from time 0 to the i-1th moment. y . D y It is a binary variable, 1 indicates that the vehicle is in a disordered lateral motion state, and 0 indicates that the vehicle is not in this state. The hidden Markov model uses the maximum likelihood estimation method to calculate the result with the highest probability, which is the vehicle lateral motion state recognition result.

[0052] In this embodiment, when the derived information set is input into the hidden Markov model for vehicle lateral motion intention recognition training in step 2, it includes:

[0053] The derived information set {frame ID, vehicle ID, x i ,y i ,v x ,v y ,θ,T,L a ,L b ,Δv1,Δv2,Δv3,Δv4} are divided into training set, validation set and test set according to the specified ratio. In this embodiment, the specified ratio is 6:2:2;

[0054] First, the training set is input into the hidden Markov model for training to determine the model parameters, then the validation set is input into the trained hidden Markov model to adjust the model parameters, and finally the test set is input into the adjusted hidden Markov model to verify the model prediction effect. This step is repeated until the loss function corresponding to the hidden Markov model meets the convergence condition, and a trained hidden Markov model is obtained.

[0055] Correspondingly, in step 2, the trained hidden Markov model predicts the lateral motion state of the target vehicle at the next moment based on the derived information set of the target vehicle's historical trajectory data, including: the trained hidden Markov model obtains the derived information set {frame ID, vehicle ID, x i ,y i ,v x ,v y ,θ,T,L a ,L b ,Δv1,Δv2,Δv3,Δv4}, and output the recognition result with the maximum probability as the lateral motion state variable D of the target vehicle at the next moment. y , and added to the derived information set at the corresponding time, so the derived information set contained in the target vehicle at the i-th time point is updated to Q i={frame ID,vehicle ID,x i ,y i , v x ,v y ,θ,T,L a ,L b ,D y , Δv1, Δv2, Δv3, Δv4}.

[0056] Step 3: Classify the variables in the derived information set containing the lateral motion state prediction results and input them into the time series fusion Transformer model for vehicle trajectory prediction training, so that the trained time series fusion Transformer model can predict the vehicle position at the next moment based on the derived information set containing the lateral motion state prediction results at the historical moment of the target vehicle;

[0057] In this embodiment, the information set Q containing the prediction results of the hidden Markov model in step 2 is i The goal is to train the model to predict the position of the vehicle at time i (x i ,y i ).

[0058] The construction of the time series fusion Transformer model includes the following steps:

[0059] Step 3.1: Build a Gated Residual Network (GRN):

[0060] GRN ω (α,c)=LayerNorm(α+GLU ω (η1))

[0061] Among them, α represents the primary information vector; c is the background information vector; LayerNorm is the standard normalization layer; GLU ω (η1) is a gated linear unit, which is used to control the entry of non-essential information into the model; ω is a weight sharing parameter.

[0062] Step 3.2: Construct variable selection network:

[0063] The variable selection network is used to identify the most important relevant variables for the prediction task and eliminate irrelevant variables. The selection of variables depends on the calculated variable weights. The calculation method is:

[0064]

[0065] Among them, tis the flattened vector representation of all variables at time t; c s is the background vector obtained from the static covariate; softmax() is the normalized exponential function;

[0066] Step 3.3: Static covariate encoding:

[0067]

[0068] Among them, ζ static variable selects the network output.

[0069] Step 3.4: Introduce an interpretable multi-head attention mechanism:

[0070]

[0071]

[0072] Among them, Q, K, V are the input matrices of dot product attention, are the weights of Q, K, and V respectively, is the multi-head attention mechanism parameter, H h W H A linear combination of .

[0073] Step 3.5: Build the temporal fusion decoder:

[0074] ①Establish gated skip connections

[0075]

[0076] in, To process the feature vector, n is the location parameter, and n∈[-k,τ max ].

[0077] ②Establish a static enrichment layer θ(t,n):

[0078]

[0079] Among them, c e For time characteristics.

[0080] ③ Establishing the temporal self-attention layer:

[0081] Collect all time features into a matrix Θ(t) = [θ(t, -k), ...θ(t, τ)] T In , for each prediction time step, the interpretable multi-head attention mechanism can be expressed as:

[0082] B(t)=Interpretable MultiHead(Θ(t),Θ(t),Θ(t))

[0083] At the same time, an additional gating layer v(t,n) is established to facilitate model training:

[0084] δ(t,n)=layerNorm(θ(t,n)+GLU δ (β(t,n)))

[0085] ④ Additional nonlinear processing of the self-attention layer:

[0086] ψ(t,n)=GRN ψ (δ(t,n))

[0087] In step S3 of this embodiment, when the variables in the derived information set including the lateral motion state prediction result are classified and input into the time series fusion Transformer model for vehicle trajectory prediction training, it includes:

[0088] The derived information set is divided into a training set, a validation set, and a test set according to a specified ratio. In this embodiment, the specified ratio is 6:2:2, and the variables in the derived information sets of the training set, the validation set, and the test set are classified respectively;

[0089] First, the variables classified in the training set are input into the time series fusion Transformer model for training to determine the model parameters, then the variables classified in the validation set are input into the trained time series fusion Transformer model to adjust the model parameters, and finally the variables classified in the test set are input into the adjusted time series fusion Transformer model to verify the model prediction effect, and this step is repeated until the loss function corresponding to the time series fusion Transformer model meets the convergence condition. When training the model, the accuracy of the trajectory prediction result is expressed by the mean square error:

[0090]

[0091] The time series fusion Transformer model divides the input variables into three categories and inputs them into the model: static variables, known variables, and unknown variables. Static variables will not change with the movement of the vehicle. The known variables are the lateral movement intention predicted by the hidden Markov model. The unknown variables are the movement states of the surrounding vehicles that change in real time. These variables are of different types and play different roles. The variables are classified and input into the model through the above steps 3.1 and 3.2. During the training process, the model will judge the importance of each variable at each moment, thereby generating self-judgment ability during the training process. Specifically, in this embodiment, the time series fusion Transformer model is represented as:

[0092]

[0093] in, represents the prediction target, τ is the prediction point, y i,t-k:t Represents the historical data of the predicted target, o i,t-k:t is the observable model input variable, corresponding to the variable {x i ,y i ,v x ,v y ,θ,Δv1,Δv2,Δv3,Δv4},k i,t-k:t is a known variable, corresponding to the lateral motion state prediction result D in the derived information set y ,s i is a static covariate, corresponding to the variable {T,L a ,L b}.

[0094] Therefore, in this embodiment, when the variables in the derived information sets of the training set, the validation set, and the test set are classified, they all include: classifying the trajectory data coordinates {x i ,y i} is used as the historical data of the predicted target in the time series fusion Transformer model, and the target vehicle heading angle θ and the target vehicle longitudinal speed v in the derived information set are x , lateral velocity v y The longitudinal speed differences Δv1, Δv2, Δv3, and Δv4 between the target vehicle and other vehicles in the affected area are used as observable input variables in the temporal fusion Transformer model, and the lateral motion state D in the derived information set is converted into y The prediction results are used as known variables in the time series fusion Transformer model, and the vehicle type T of the target vehicle in the derived information set, the initial lane L of the target vehicle entering the lane-free area, and the lane type L of the target vehicle in the lane-free area. a and the final lane L leaving the lane-free area b As a static covariate in the time series fusion Transformer model.

[0095] Correspondingly, in step 3, the trained time series fusion Transformer model predicts the vehicle position at the next moment based on the derived information set containing the lateral motion state prediction results of the target vehicle at the historical moment, including: the trained time series fusion Transformer model obtains the derived information set {frame ID, vehicle ID, x i ,y i ,v x , v y ,θ,T,L a , L b , D y,Δv1,Δv2,Δv3,Δv4}, where the trajectory data coordinates {x i ,y i} As the historical data of the predicted target, the target vehicle heading angle θ, the target vehicle longitudinal speed v x , lateral velocity v y , the longitudinal velocity differences Δv1, Δv2, Δv3, Δv4 between the target vehicle and other vehicles in the affected area are used as observable input variables, and the lateral motion state D y The prediction results are used as known variables, including the vehicle type T of the target vehicle, the initial lane L of the target vehicle entering the lane-free area, and the target vehicle’s initial lane L. a and the final lane L leaving the lane-free area b As a static covariate, the prediction result with the largest weight is output as the vehicle position at the next moment.

[0096] After the above steps, the trained hidden Markov model can predict the lateral motion state of the target vehicle at the next moment based on the derived information set of the target vehicle's historical trajectory data. At the same time, the trained time series fusion Transformer model can predict the vehicle position at the next moment based on the derived information set containing the lateral motion state prediction results of the target vehicle's historical moments. Next, the two trained models can be used to predict the vehicle trajectory. The method of this embodiment also includes:

[0097] Step 4: After acquiring the new trajectory data of the target vehicle in real time and performing data preprocessing, the corresponding derivative information set is input into the trained hidden Markov model to obtain a derivative information set containing the lateral motion state prediction result. The derivative information set containing the lateral motion state prediction result is then classified and input into the trained time series fusion Transformer model to obtain the trajectory prediction result.

[0098] Specifically, when obtaining the new trajectory data of the target vehicle in real time and performing data preprocessing, the horizontal and vertical coordinates of the vehicle trajectory are substituted, and the same steps as steps 1.1 to 1.3 are performed to obtain the derivative information set {frame ID, vehicle ID, x i ,y i ,v x ,v y ,θ,T,L a ,L b , Δv1, Δv2, Δv3, Δv4}. All variable values ​​are normalized and mapped to the [0, 1] interval.

[0099] When the corresponding derivative information set is input into the trained hidden Markov model to obtain the derivative information set containing the lateral motion state prediction result, the trained hidden Markov model obtains the derivative information set {frame ID, vehicle ID, x i ,y i , v x , v y ,θ,T,L a , L b , Δv1, Δv2, Δv3, Δv4}, and output the recognition result with the maximum probability as the lateral motion state variable D of the target vehicle at the next moment y , get the derivative information set {frame ID, vehicle ID, x i ,y i , v x , v y ,θ,T,L a , L b , D y , Δv1, Δv2, Δv3, Δv4}.

[0100] The derived information set containing the lateral motion state prediction result is classified and input into the trained time series fusion Transformer model. When the trajectory prediction result is obtained, the trajectory data coordinates {x i ,y i} is used as the historical data of the predicted target in the time series fusion Transformer model, and the target vehicle heading angle θ and the target vehicle longitudinal speed v in the derived information set are x , lateral velocity v y The longitudinal speed differences Δv1, Δv2, Δv3, and Δv4 between the target vehicle and other vehicles in the affected area are used as observable input variables in the temporal fusion Transformer model, and the lateral motion state D in the derived information set is converted into y The prediction results are used as known variables in the time series fusion Transformer model, and the vehicle type T of the target vehicle in the derived information set, the initial lane L of the target vehicle entering the lane-free area, and the lane type L of the target vehicle in the lane-free area. a and the final lane L leaving the lane-free area b As a static covariate in the time series fusion Transformer model; the trained time series fusion Transformer model obtains a derivative information set containing the lateral motion state prediction results from time 0 to time i-1

[0101] {frame ID, vehicle ID, x i ,y i, v x , v y ,θ,T,L a , L b , D y ,Δv1,Δv2,Δv3,Δv4}, where the trajectory data coordinates {x i ,y i} As the historical data of the predicted target, the target vehicle heading angle θ, the target vehicle longitudinal speed v x , lateral velocity v y , the longitudinal velocity differences Δv1, Δv2, Δv3, Δv4 between the target vehicle and other vehicles in the affected area are used as observable input variables, and the lateral motion state D y The prediction results are used as known variables, including the vehicle type T of the target vehicle, the initial lane L of the target vehicle entering the lane-free area, and the target vehicle’s initial lane L. a and the final lane L leaving the lane-free area b As a static covariate, the prediction result with the largest weight is output as the vehicle position at the next moment.

[0102] Embodiment 2

[0103] This embodiment provides a vehicle trajectory prediction system for a lane-free area, including a storage medium of interconnected microprocessors, wherein the microprocessors are programmed or configured to execute the vehicle trajectory prediction method for a lane-free area described in the first embodiment.

[0104] In summary, before using the time series fusion Transformer model to predict the trajectory, the present invention first uses the hidden Markov model to predict the vehicle's lateral movement intention; when using the time series fusion Transformer model to predict the trajectory, the variables in the information set are classified and input into the model, which overcomes the problem of insufficient trajectory prediction accuracy of the TFT model in areas without lane lines, complex traffic environments and vehicles prone to sudden and disordered lateral movements, effectively improves the prediction accuracy of the time series fusion Transformer model, and can accurately predict the driving trajectory of the target vehicle in areas without lane lines.

[0105] The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment, it is not intended to limit the present invention. Therefore, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for predicting vehicle trajectories in a laneless area, characterized in that the steps include: Step 1: Obtain the trajectory data of the target vehicle in the area without lane lines in real time, and perform data preprocessing to obtain the derived information set of the trajectory data at each moment; Step 2: All derived information sets are input into the hidden Markov model for vehicle lateral motion intention recognition training, so that the trained hidden Markov model predicts the lateral motion state of the target vehicle at the next moment based on the derived information set of the target vehicle's historical trajectory data, and adds the lateral motion state prediction result at each moment to the corresponding derived information set; Step 3: Classify the variables in the derived information set containing the lateral motion state prediction results and input them into the time series fusion Transformer model for vehicle trajectory prediction training, so that the trained time series fusion Transformer model can predict the vehicle position at the next moment based on the derived information set containing the lateral motion state prediction results at the historical moment of the target vehicle; Step 4: After acquiring the new trajectory data of the target vehicle in real time and performing data preprocessing, the corresponding derivative information set is input into the trained hidden Markov model to obtain a derivative information set containing the lateral motion state prediction result. The derivative information set containing the lateral motion state prediction result is then classified and input into the trained time series fusion Transformer model to obtain the trajectory prediction result.

2. The vehicle trajectory prediction method in a lane-free area according to claim 1, characterized in that: The steps of data preprocessing in step 1 and step 4 both include: The target vehicle’s trajectory data is added to the corresponding trajectory data set to obtain the target vehicle’s vehicle type T. Based on the coordinates of the trajectory data at the initial and final moments in the trajectory data set, the target vehicle’s initial lane L is obtained. a and the final lane L leaving the lane-free area b , the vehicle type T of the target vehicle, the initial lane L a and final lane L b Add the derivative information set at the corresponding moment; According to the coordinates of the trajectory data of the target vehicle at adjacent moments in the trajectory data set, the heading angle θ of the target vehicle at each moment and the longitudinal speed v of the target vehicle at each moment are calculated. x , lateral velocity v y , the longitudinal speed difference between the target vehicle and other vehicles in the affected area Δv1, Δv2, Δv3, Δv4, and then add the target vehicle heading angle θ, longitudinal speed v x , lateral velocity v y , longitudinal speed differences Δv1, Δv2, Δv3, Δv4 are added to the derivative information set at the corresponding time.

3. The vehicle trajectory prediction method in a lane-free area according to claim 2, characterized in that: When calculating the longitudinal speed differences Δv1, Δv2, Δv3, Δv4 between the target vehicle and other vehicles in the affected area based on the vehicle trajectory data set, it includes: The geometric center of the target vehicle is taken as the origin of the affected area, and the affected area is determined according to the specified length and width. The affected area includes four areas: the left front, right front, left rear, and right rear of the target vehicle. The longitudinal speed differences Δv1, Δv2, Δv3, and Δv4 between the target vehicle and other vehicles in each area are calculated respectively. If there is no other vehicle in a certain area, the corresponding longitudinal speed difference is 0.

4. The vehicle trajectory prediction method in a lane-free area according to claim 1, characterized in that: In step 1, the step of performing data preprocessing also includes: normalizing the variables in the derived information set, and mapping all variable values ​​in the derived information set to the interval [0, 1].

5. The vehicle trajectory prediction method in a lane-free area according to claim 1, characterized in that: In step 2, when the derived information set is input into the hidden Markov model for vehicle lateral motion intention recognition training, it includes: Divide the derived information set at each moment into training set, validation set, and test set according to the specified ratio; First, the training set is input into the hidden Markov model for training to determine the model parameters, then the validation set is input into the trained hidden Markov model to adjust the model parameters, and finally the test set is input into the adjusted hidden Markov model to verify the model prediction effect. This step is repeated until the loss function corresponding to the hidden Markov model meets the convergence condition.

6. The vehicle trajectory prediction method in a lane-free area according to claim 1, characterized in that: In step 2, when the trained hidden Markov model predicts the lateral motion state of the target vehicle at the next moment based on the derived information set of the target vehicle's historical moment trajectory data, the following steps are performed: the trained hidden Markov model obtains the derived information set of each historical moment trajectory data, and outputs the recognition result of the maximum probability as the lateral motion state variable D of the target vehicle at the next moment. y .

7. The vehicle trajectory prediction method in a lane-free area according to claim 1, characterized in that: In step 3, when the variables in the derived information set containing the lateral motion state prediction result are classified and input into the time series fusion Transformer model for vehicle trajectory prediction training, it includes: Dividing the derived information set into a training set, a validation set, and a test set according to a specified ratio, and classifying the variables in the derived information sets of the training set, the validation set, and the test set respectively; First, the classified variables of the training set are input into the time series fusion Transformer model for training to determine the model parameters. Then, the classified variables of the validation set are input into the trained time series fusion Transformer model to adjust the model parameters. Finally, the classified variables of the test set are input into the adjusted time series fusion Transformer model to verify the model prediction effect. This step is repeated until the loss function corresponding to the time series fusion Transformer model meets the convergence conditions.

8. The vehicle trajectory prediction method in a lane-free area according to claim 7, characterized in that: The variables in the derived information sets of the training set, the validation set, and the test set are classified respectively, and the derived information set containing the lateral motion state prediction result is classified and input into the trained time series fusion Transformer model in step 4, both of which include: the trajectory data coordinates {x i ,y i } is used as the historical data of the predicted target in the time series fusion Transformer model, and the target vehicle heading angle θ and the target vehicle longitudinal speed v in the derived information set are x , lateral velocity v y The longitudinal speed differences Δv1, Δv2, Δv3, and Δv4 between the target vehicle and other vehicles in the affected area are used as observable input variables in the temporal fusion Transformer model, and the lateral motion state D in the derived information set is converted into y The prediction results are used as known variables in the time series fusion Transformer model, and the vehicle type T of the target vehicle in the derived information set, the initial lane L of the target vehicle entering the lane-free area, and the lane type L of the target vehicle in the lane-free area. a and the final lane L leaving the lane-free area b As a static covariate in the time series fusion Transformer model.

9. The vehicle trajectory prediction method in a lane-free area according to claim 1, characterized in that: In step 3, when the trained time series fusion Transformer model predicts the vehicle position at the next moment based on the derived information set containing the lateral motion state prediction results at the historical moments of the target vehicle, it includes: the trained time series fusion Transformer model obtains the derived information set containing the lateral motion state prediction results at each historical moment, and outputs the prediction result with the largest weight as the vehicle position at the next moment.

10. A vehicle trajectory prediction system in a lane-free area, characterized in that: A storage medium comprising interconnected microprocessors, wherein the microprocessors are programmed or configured to execute the vehicle trajectory prediction method for a lane-free area according to any one of claims 1 to 9.

Images