Around vehicle trajectory prediction method and system fused with intention perception

Abstract

The invention provides a surrounding vehicle trajectory prediction method and system fused with intention perception. The method comprises the following steps: calling an IA-Informer model; on the basis of an IA-Informer model, intention recognition and intention fusion trajectory prediction are carried out on surrounding vehicles of the vehicle; the IA-Informer model comprises an input matrix module, a target vehicle intention recognition module and an intention fusion trajectory prediction module. According to the invention, an intention prediction module extracts social association according to states of a target vehicle and surrounding vehicles, and outputs an intention probability vector; and the trajectory prediction module introduces the output of the former to further improve the prediction performance.

Description

Technical Field

[0001] The present invention relates to the field of neural network technology, and in particular to a vehicle trajectory prediction method and system integrating intention perception. Background Art

[0002] In current research, when dealing with the problem of vehicle trajectory prediction, a common approach is to transform it into a task of inferring the prior probability distribution of future coordinates based on historical coordinates, and a large number of studies have shown that incorporating the interaction between the target vehicle and surrounding vehicles in the prediction can greatly improve the prediction accuracy. However, there are still two research gaps in existing research. First, the existing research often solves the problems of lane change intention and trajectory prediction independently, that is, corresponding models are established for lane change intention and trajectory prediction respectively. This causes the model to ignore the interaction between intention prediction and trajectory prediction, thereby sacrificing the accuracy of the prediction to a certain extent. Secondly, when considering the interaction between the target vehicle and the surrounding vehicles, many researchers simply use the state of the target vehicle and the state of the surrounding vehicles as a feature vector input of the prediction model, and the analysis of the coupling state is not in-depth enough. Summary of the invention

[0003] One of the purposes of the present invention is to provide a method for predicting surrounding vehicle trajectories that integrates intention perception, wherein the intention prediction module extracts social associations based on the states of the target vehicle and surrounding vehicles and outputs an intention probability vector; the trajectory prediction module introduces the output of the former to further improve the prediction performance; by simultaneously performing intention recognition and trajectory prediction and introducing the output of intention recognition in the trajectory prediction module, a joint prediction framework is formed; and a more sophisticated interactive coding method is designed.

[0004] An embodiment of the present invention provides a method for predicting vehicle trajectories integrating intention perception, including:

[0005] Retrieve the IA-Informer model;

[0006] Based on the IA-Informer model, the vehicle's surrounding intentions are recognized and the trajectory of the fused intentions is predicted.

[0007] Preferably, the IA-Informer model includes: an input matrix module, a target vehicle intention recognition module and a trajectory prediction module of fusion intention.

[0008] Preferably, the step of constructing the input matrix module includes:

[0009] Define the target vehicle as V p0 , the other vehicles in the scene where the target vehicle is located are defined as V si(i∈1,2,…,N). The vehicle’s forward direction is the positive direction of the y-axis, and the direction perpendicular to the y-axis is the x-axis; where N is the total number of other vehicles;

[0010] The encoder and decoder inputs of the IA-Informer model are denoted as X en and X de ; The encoder input is the historical trajectory P of the target vehicle 0 (t) and the interaction information S K (t) is composed of the mask trajectory of the target vehicle at the decoder input

[0011] Define the target vehicle V p0 The trajectory coordinates at time t are:

[0012]

[0013] In the observation domain with domain length mt, the historical trajectory of the target vehicle is:

[0014]

[0015] The interaction relationship of the target vehicle at time t is:

[0016]

[0017] Where K represents the target vehicle number, and Indicates the relative horizontal and vertical distances between the target vehicle and vehicle K. and is the relative lateral and longitudinal speeds between the target vehicle and vehicle K;

[0018] In the observation domain with domain length mt, the interaction information of the target vehicle is:

[0019]

[0020] The encoder input is:

[0021] X en =Concat(P 0 (t),S K (t));

[0022] The input sequence of the decoder at time t is for:

[0023]

[0024] Preferably, the steps of constructing the target vehicle intention recognition module include:

[0025] The intention recognition module predicts the driving intention at the next moment based on the observed target vehicle trajectory information and its interaction information with surrounding vehicles. The ultimate goal is to make the network learn a function:

[0026]

[0027] Among them, ω(t) is the intention probability vector output by the target vehicle intention recognition module;

[0028] in, This is the final trained intention prediction model, ω∈N 1×3 To predict the result of the intention, that is, the intention probability vector defined in this application, the formula is as follows:

[0029] ω=[a1,a2,a3]

[0030] a1+a2+a3=100%;

[0031] Among them, a1, a2, and a3 are the probabilities of the three strategies that the target vehicle may adopt; the three strategies include: lane keeping, left lane change, and right lane change;

[0032] The context vector is calculated layer by layer from bottom to top:

[0033] H1=X en W1+PE temp

[0034] A1=norm(prob(H1,H1,H1)+H1)

[0035] H2=norm(relu(A1W2)W2'+A1)

[0036] H'2 = MaxPool(ELU(Conv1d(H2))

[0037] A2=norm(prob(H'2,H'2,H'2)+H'2)

[0038] F = norm(relu(A2W3)W′3+A2)

[0039] ω=FW4;

[0040] Among them, W1, W2, W2', W3, W3', and W4 are the weights of the fully connected layer, relu is the activation function of the feedforward layer, norm, prob, MaxPool, ELU, and Conv1d are normalization, probabilistic sparse self-attention layer, maximum pooling layer, exponential linear unit, and one-dimensional convolution layer, respectively.

[0041] Preferably, the steps of constructing the trajectory prediction module of the fusion intention include:

[0042] The trajectory prediction module uses the historical trajectory information and intention probability vector of the target vehicle as input and outputs the expected trajectory for the next time period nT, which is as follows:

[0043]

[0044] The decoder encodes the result of intent prediction and inputs it into the multi-head attention module. At the same time, it receives the context vector generated by the encoder and applies it to the second self-attention layer. The process is as follows:

[0045] H3=X de W5+PE temp +ωW6

[0046] A3=norm(prob(H3,H3,H3)+H3)

[0047] A4=norm(mha(A3,F,F)+A3)

[0048] H4=norm(relu(A3W6)W′6+A4)

[0049] H'4 = MaxPool(ELU(Conv1d(H4))

[0050]

[0051] Among them, W6, W6', and W7 are the weights of the fully connected layer; mha is the multi-head attention layer.

[0052] An embodiment of the present invention provides a vehicle trajectory prediction system integrating intention perception, including:

[0053] A calling unit, used to call the IA-Informer model;

[0054] The prediction unit is used to perform intention recognition and trajectory prediction of the vehicle's surrounding vehicles based on the IA-Informer model. Other features and advantages of the present invention will be described in the subsequent description, and partly become obvious from the description, or be understood by implementing the present invention. The objects and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the drawings.

[0055] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0057] Figure 1 A schematic diagram of a method for predicting vehicle trajectories integrating intention perception in an embodiment of the present invention;

[0058] Figure 2 It is a structural diagram of the IA-Informer model in an embodiment of the present invention;

[0059] Figure 3 A schematic diagram of a typical trajectory prediction scenario in an embodiment of the present invention;

[0060] Figure 4 This is a highD data visualization effect diagram in an embodiment of the present invention;

[0061] Figure 5 This is a schematic diagram of lane change data in an embodiment of the present invention;

[0062] Figure 6 This is a diagram showing verification results of different input sequence lengths in an embodiment of the present invention;

[0063] Figure 7 This is a prediction result diagram for a lane keeping scenario in an embodiment of the present invention;

[0064] Figure 8 This is a diagram of the prediction result at the initial moment in a lane-changing scenario in an embodiment of the present invention;

[0065] Fig. 9 This is a diagram showing the final time prediction result in a lane change scenario in an embodiment of the present invention;

[0066] Fig.10 Graph showing the trajectory errors of each model in the embodiment of the present invention. DETAILED DESCRIPTION

[0067] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0068] The embodiment of the present invention provides a method for predicting vehicle trajectory by integrating intention perception. Figure 1 As shown, including:

[0069] S1, call the IA-Informer model;

[0070] S2. Based on the IA-Informer model, the vehicle's surrounding intentions are recognized and the trajectory of the fusion intentions is predicted.

[0071] The IA-Informer model includes: an input matrix module, a target vehicle intention recognition module and a trajectory prediction module of fusion intention.

[0072] The steps of constructing the input matrix module include:

[0073] Define the target vehicle as V p0 , the other vehicles in the scene where the target vehicle is located are defined as V si (i∈1,2,…,N). The vehicle’s forward direction is the positive direction of the y-axis, and the direction perpendicular to the y-axis is the x-axis; where N is the total number of other vehicles;

[0074] The encoder and decoder inputs of the IA-Informer model are denoted as X en and X de ; The encoder input is the historical trajectory P of the target vehicle 0 (t) and the interaction information S K (t) is composed of the mask trajectory of the target vehicle at the decoder input

[0075] Define the target vehicle V p0 The trajectory coordinates at time t are:

[0076]

[0077] In the observation domain with domain length mt, the historical trajectory of the target vehicle is:

[0078]

[0079] The interaction relationship of the target vehicle at time t is:

[0080]

[0081] Where K represents the target vehicle number, and Indicates the relative horizontal and vertical distances between the target vehicle and vehicle K. and is the relative lateral and longitudinal speeds between the target vehicle and vehicle K;

[0082] In the observation domain with domain length mt, the interaction information of the target vehicle is:

[0083]

[0084] The encoder input is:

[0085] X en =Concat(P 0 (t),S K (t));

[0086] The input sequence of the decoder at time t is for:

[0087]

[0088] The steps of constructing the target vehicle intention recognition module include:

[0089] The intention recognition module predicts the driving intention at the next moment based on the observed target vehicle trajectory information and its interaction information with surrounding vehicles. The ultimate goal is to make the network learn a function:

[0090]

[0091] in, This is the final trained intention prediction model, ω∈N 1×3 To predict the result of the intention, that is, the intention probability vector defined in this application, the formula is as follows:

[0092] ω=[a1,a2,a3]

[0093] a1+a2+a3=100%;

[0094] Among them, a1, a2, and a3 are the probabilities of the three strategies that the target vehicle may adopt; the three strategies include: lane keeping, left lane change, and right lane change;

[0095] The context vector is calculated layer by layer from bottom to top:

[0096] H1=X en W1+PE temp

[0097] A1=norm(prob(H1,H1,H1)+H1)

[0098] H2=norm(relu(A1W2)W′2+A1)

[0099] H'2 = MaxPool(ELU(Conv1d(H2))

[0100] A2=norm(prob(H'2,H'2,H'2)+H'2)

[0101] F = norm(relu(A2W3)W′3+A2)

[0102] ω=FW4;

[0103] Among them, W1, W2, W2', W3, W3', and W4 are the weights of the fully connected layer, relu is the activation function of the feedforward layer, norm, prob, MaxPool, ELU, and Conv1d are normalization, probabilistic sparse self-attention layer, maximum pooling layer, exponential linear unit, and one-dimensional convolution layer, respectively.

[0104] The steps to build the trajectory prediction module of fusion intent include:

[0105] The trajectory prediction module uses the historical trajectory information and intention probability vector of the target vehicle as input and outputs the expected trajectory for the next time period nT, which is as follows:

[0106]

[0107] Among them, ω(t) is the intention probability vector output by the target vehicle intention recognition module;

[0108] The decoder encodes the result of intent prediction and inputs it into the multi-head attention module. At the same time, it receives the context vector generated by the encoder and applies it to the second self-attention layer. The process is as follows:

[0109] H3=X de W5+PE temp +ωW6

[0110] A3=norm(prob(H3,H3,H3)+H3)

[0111] A4=norm(mha(A3,F,F)+A3)

[0112] H4=norm(relu(A3W6)W′6+A4)

[0113] H'4 = MaxPool(ELU(Conv1d(H4))

[0114]

[0115] Among them, W6, W6', and W7 are the weights of the fully connected layer; mha is the multi-head attention layer.

[0116] The working principle and beneficial effects of the above technical solution are:

[0117] The main feature of the Informer model is that it adopts a multi-scale temporal encoder and decoder structure and incorporates an adaptive length attention mechanism, which is very suitable for predicting long sequences and multi-label outputs. Vehicle status and trajectory data are data with periodic and continuous characteristics. When predicting trajectories, the interactive features between data are of great value. Therefore, this application will complete the vehicle intent recognition and trajectory prediction tasks based on the framework of the Informer model.

[0118] To solve the above problems, this section proposes an IA-Informer method that integrates intent perception to predict the target vehicle’s intention and trajectory. The overall framework is as follows: Figure 2As shown in the figure; IA stands for Intention Aware. The dashed box on the left side of the figure is the encoder structure, and the one on the right side is the decoder structure.

[0119] For the encoder part, the historical position information of the target vehicle and the interaction information with other vehicles are first input into the embedding module in parallel, so that the model can fully extract the correlation between the vehicle trajectory information and the interaction information. In addition, considering the relationship between the time series and the relative position of the vehicle, the historical trajectory of the vehicle is embedded in time and then input into the encoder together. After that, after calculations by each layer of the encoder, the input information is converted into a context vector. On the one hand, this vector can be input into the decoder as an intermediate variable for extracting information, and on the other hand, the result of intention prediction can be output through a fully connected layer. This application defines an intention recognition probability vector that expresses the vehicle's intention in the form of its probability value for keeping and changing lanes.

[0120] For the decoder, the decoder tries to discover key features from the intermediate representation generated by the encoder so that it can extract features that are highly relevant to the prediction. First, the result of intent recognition - the intent probability vector and the masked trajectory to be predicted are embedded end-to-end in the embedding layer and put into the decoder as a whole. After calculations in each layer of the decoder, the trajectory generator is finally used to generate the position of the future target vehicle.

[0121] The ultimate goal of this application is to achieve the task of vehicle trajectory prediction on structured roads. Therefore, referring to the highD dataset used later, it is necessary to define a general three-lane scenario, such as Figure 3 The yellow vehicle is the ego vehicle (i.e. the autonomous driving vehicle equipped with the algorithm), the blue vehicle is the target vehicle whose trajectory is to be predicted, and the remaining gray vehicles and the ego vehicle can be regarded as surrounding vehicles that may affect the behavior of the target vehicle.

[0122] The encoder and decoder inputs of the IA-Informer model are denoted as X en and X de The encoder input consists of the historical trajectory P of the target vehicle. 0 (t) and the interaction information S K (t) is composed of the mask trajectory of the target vehicle at the decoder input

[0123] Define the target vehicle V p0 The trajectory coordinates at time t are:

[0124]

[0125] Then, in the observation domain with domain length mt, the historical trajectory of the target vehicle can be expressed as:

[0126]

[0127] Similarly, the interaction relationship of the target car at time t can be expressed by the following formula:

[0128]

[0129] Among them, K represents the number of the target vehicle, No. 1 and No. 6 are located in front of and behind the left lane of the target vehicle, No. 2 and No. 5 are located in front of and behind the current lane of the target vehicle, and No. 3 and No. 4 are located in front of and behind the right lane of the target vehicle. and Indicates the relative horizontal and vertical distances between the target car and car K. and is the relative lateral and longitudinal speed between the target vehicle and vehicle K; wherein the target vehicle is the target vehicle, i.e., the vehicle whose intention and trajectory are to be predicted.

[0130] Then, in the observation domain with domain length mt, the interaction information of the target vehicle can be expressed as:

[0131]

[0132] Finally, the encoder input is:

[0133] X en =Concat(P 0 (t),S K (t))(1.5)

[0134] Similar to the encoder, the decoder’s input sequence at time t is It can be expressed as:

[0135]

[0136] The only difference is that this application does not use a dedicated flag as the start marker, but chooses a slice of the input sequence as the marker. Specifically, a segment of length L is added to the input end. token The historical trajectory of is used as a known label to guide the decoder to better calculate the prediction amount, and the historical trajectory is input into the decoder together with the nT segment target trajectory to be predicted. token Take 3s and nT range 5s, that is, use the 3s historical trajectory to predict the 5s future trajectory (the trajectory to be predicted uses 0 as a placeholder).

[0137] The intention recognition module predicts the driving intention at the next moment based on the observed target vehicle trajectory information and its interaction information with surrounding vehicles. The ultimate goal is to make the network learn a function:

[0138]

[0139] in This is the final trained intention prediction model, ω∈N 1×3 To predict the result of the intention, that is, the intention probability vector defined in this application, the formula is as follows:

[0140] ω=[a1,a2,a3](1.8)

[0141] a1+a2+a3=100%(1.9)

[0142] Among them, a1, a2, and a3 are the probabilities of the three strategies that the target vehicle may adopt (lane keeping, left lane change, and right lane change).

[0143] like Figure 2 As shown on the left side of the figure, the overall structure of the model is built based on the Informer encoder. The context vector is calculated layer by layer from bottom to top:

[0144] H1=X en W1+PE temp (1.10)

[0145] A1=norm(prob(H1,H1,H1)+H1)(1.11)

[0146] H2=norm(relu(A1W2)W′2+A1) (1.12)

[0147] H'2=MaxPool(ELU(Conv1d(H2)) (1.13)

[0148] A2=norm(prob(H'2,H'2,H'2)+H'2) (1.14)

[0149] F=norm(relu(A2W3)W′3+A2) (1.15)

[0150] ω=FW4(1.16)

[0151] Among them, all W are the weights of the fully connected layers, relu is the activation function of the feedforward layer, norm, prob, MaxPool, ELU and Conv1d are normalization, probabilistic sparse self-attention layer, maximum pooling layer, exponential linear unit and one-dimensional convolution layer, respectively, which will be explained in detail below.

[0152] Formula (1.10) shows the calculation process of the embedding layer. The input of the embedding layer is the historical trajectory P of the target vehicle. 0 (t) and the interaction information S K In addition to (t), the trajectory sine and cosine space-time coding PE is also addedtemp , and finally combined into the sequence by addition. In order to match the dimensionality range of the two embedded features, a fully connected layer W1 is added to the input sequence to expand its dimension to the same as the embedded feature. Therefore, the fused sequence H1 extracts the information of the original sequence while also highlighting its spatiotemporal characteristics, improving the network's ability to identify and utilize the associations between the nodes of the input sequence.

[0153] Formula (1.11) represents the multi-head probabilistic sparse self-attention layer. In addition, this layer also adds residual connections and layer normalization. Residual connections are additive connections that connect "input" and "output", which improve the gradient dispersion problem and ensure that the network will at least not learn in the wrong direction, making the network converge faster. The purpose of layer normalization is to improve the generalization performance of the network.

[0154] Formula (1.12) is a feedforward layer, which has the same structure as the Transformer and is used to improve the nonlinear ability of the model. It uses the activation function relu to perform two linear projections, and performs residual connections and layer normalization.

[0155] The superposition of embedding layers may result in a very long input sequence, which makes the information contained in the feature vector not very concentrated. Formula (1.13) performs a data distillation. The stride of the maximum pooling layer is set to 2, which means that after downsampling, the dimension of H'2 is half of the original. This layer ensures that the dominant features are extracted and the key context representation vector is generated in the next layer without changing the overall feature distribution, thereby achieving efficient memory usage.

[0156] Formula (1.14) and Formula (1.15) are repeated operations of the network structure of Formula (1.11) and Formula (1.12). By continuously stacking network layers, data can be mined more deeply to deal with more complex problems. In this application, partial extraction of input sequence features has been performed outside the encoder layer, so there is no need to build more layers in the encoder.

[0157] Finally, as shown in formula (1.16), the context vector can obtain the intent probability vector through a fully connected layer.

[0158] The trajectory prediction module uses the historical trajectory information and intention probability vector of the target vehicle as input and outputs the expected trajectory for the next time period nT. Therefore, the problem can be expressed as follows:

[0159]

[0160] like Figure 2 As shown on the right side of , the decoder encodes the result of intent prediction and inputs it into the multi-head attention module. At the same time, it receives the context vector generated by the encoder and applies it to the second self-attention layer. The overall calculation process is as follows:

[0161] H3=X de W5+PE temp +ωW6(1.18)

[0162] A3=norm(prob(H3,H3,H3)+H3) (1.19)

[0163] A4=norm(mha(A3,F,F)+A3)(1.20)

[0164] H4=norm(relu(A3W6)W′6+A4) (1.21)

[0165] H'4=MaxPool(ELU(Conv1d(H4)) (1.22)

[0166]

[0167] Among them, all W are the weights of the fully connected layer. In addition to the sparse probabilistic self-attention layer in formula (1.19), the decoder also has a multi-head attention layer (i.e., mha in formula (1.20)), whose input comes partly from the context vector F extracted from the historical trajectory in the encoder, and partly from the feature A3 extracted from the decoder input. The multi-head attention layer improves the accuracy of trajectory prediction by taking into account the contextual representation of the true trajectory and the historical trajectory. Formulas (1.21) and (1.22) are the same as the encoder calculation, and both perform a feedforward layer and data distillation. Finally, the predicted trajectory is generated by the fully connected layer in formula (1.23).

[0168] The model performs intent recognition and trajectory prediction simultaneously, and introduces the output of intent recognition into the trajectory prediction module to form a joint prediction framework. Specifically:

[0169] The intention recognition module of the model first extracts social associations based on the status of the target vehicle and surrounding vehicles, and outputs an intention probability vector (e.g., the probability of keeping lane, changing left lane, and changing right lane).

[0170] The trajectory prediction module uses the intention probability vector as additional information and combines it with the historical trajectory of the target vehicle to generate a more accurate future trajectory prediction. This approach can make the trajectory prediction more consistent with the actual intention of the vehicle and achieve a synergistic improvement of the two tasks.

[0171] This model designs a more sophisticated interaction encoding method, including the following points:

[0172] Social correlation feature extraction: The model performs social correlation analysis on the status of the target vehicle and surrounding vehicles. It not only inputs simple information such as position and speed, but also considers more descriptive interaction features such as the relative position and relative speed of the target vehicle and surrounding vehicles, capturing the detailed interaction relationship between the front and rear vehicles on each lane.

[0173] Introduction of self-attention mechanism: The model uses the self-attention mechanism (such as sparse self-attention in Informer), which allows the model to automatically identify important interaction features and assign different weights to different interaction relationships, thereby more comprehensively modeling the mutual influence between vehicles.

[0174] This method can more deeply explore the coupling status between the target vehicle and surrounding vehicles. Instead of simply treating them as feature inputs, it fully analyzes their relationships and improves the model's ability to adapt to complex traffic environments.

[0175] In one embodiment, the method for predicting surrounding vehicle trajectories integrating intention perception further includes a test and verification step of the IA-Informer model, including:

[0176] Preprocess the dataset;

[0177] Based on the preprocessing results, the IA-Informer model is trained and simulated for verification.

[0178] The preprocessing of the data set includes:

[0179] The dataset is divided into training set, validation set and test set in a ratio of 7:1:2;

[0180] After the data set is divided, the time series needs to be input into the model in a certain format and length. h The sliding window slides in the input sequence one by one to obtain samples:

[0181]

[0182] in, is the characteristic sequence, Y t is the corresponding label. For the trajectory prediction module, the label can be directly obtained from the horizontal and vertical coordinates of the real trajectory, that is:

[0183] Y t =(x t ,y t );

[0184] For the intention recognition module, the label is the vehicle's driving intention:

[0185] Y t ∈{0,1,2};

[0186] Among them, 0, 1, and 2 represent the vehicle's intention to keep, change lanes left, and change lanes right, respectively;

[0187] The annotation rules of the intent recognition module are as follows:

[0188] For vehicles going straight, all the extracted sequences are marked as 0;

[0189] The steps for marking a left lane-changing vehicle are as follows:

[0190] Calculate the yaw angle for each point on the trajectory:

[0191]

[0192] Among them, x t and t represents the horizontal and vertical coordinates of the vehicle at time t; x t+1 and t+1 Represents the horizontal and vertical coordinates of the vehicle at time t+1;

[0193] Define the heading angle threshold θ b , find the first point B whose yaw angle is less than the threshold as the lane change starting point;

[0194] Considering that the intention is generated before the actual lane change action, a fixed time of 1 second is calculated forward from point B to find point C as the moment when the intention is generated;

[0195] Define the sequence from C to A as 1 and the rest as 0.

[0196] Based on the preprocessing results, the IA-Informer model is trained, including:

[0197] The vehicle intention recognition module uses a multi-classification cross entropy function to calculate the loss:

[0198]

[0199] Among them, m is the number of samples, k represents the intent type, and p ik represents the independent encoding of the i-th sample label, q ik is the probability that the i-th sample belongs to the k-th intent;

[0200] For the trajectory prediction model, the likelihood function that minimizes the negative logarithm is used to train and optimize the parameters in the model:

[0201] L=-log(∑ k Pθ(G|C k ,X obs )P(C k |X obs ));

[0202] Among them, X obs is the historical trajectory sequence of the target vehicle, C k is the intention predicted by the driving recognition module, G is the Gaussian distribution of the future trajectory output by trajectory prediction;

[0203] At each time step, the model calculates the loss and updates the weights of all fully connected layers and attention layers through the back-propagation algorithm to minimize the loss.

[0204] Based on the preprocessing results, the IA-Informer model is simulated and verified, including:

[0205] For the target vehicle intention recognition module, the first indicator and the first comparison model are used to perform simulation verification of the IA-Informer model according to the preprocessing result; wherein the first indicator and the first comparison model include: average accuracy ACC, F1 score, SVM model, HMM model, LSTM model, Transformer model;

[0206] For the trajectory prediction module of fusion intention, the second indicator and the second comparison model are used to perform simulation verification of the IA-Informer model according to the preprocessing results; among them, the second indicator and the second comparison model include: root mean square error RMSE, average displacement error ADE, final displacement error FDE, single LSTM model, dual LSTM model, bidirectional LSTM model, social LSTM and dual Transformer model.

[0207] The working principle and beneficial effects of the above technical solution are:

[0208] In order to verify the effectiveness of the model proposed in this application in predicting the behavior of surrounding vehicles, a public dataset will be used to train and verify the model, and an in-depth comparative analysis will be conducted with the currently widely used network models.

[0209] This application uses the public dataset highD released by the Institute of Automotive Research at RWTH Aachen University in Germany for model training and verification. highD is a large-scale natural vehicle trajectory dataset on German highways, containing 110,000 vehicle driving data at 6 locations for a total of 16.5 hours, including more than 5,000 complete lane change trajectories.

[0210] The dataset contains 60 sub-datasets in total, each of which contains aerial photos of the collected road section, collection point data, and vehicle trajectory data. The collection point data gives information such as location ID, time, collection duration, cumulative vehicle distance, and number of vehicles. The vehicle trajectory data mainly includes vehicle ID, vehicle coordinates, speed, acceleration, and other information. The effect of data visualization in Matlab is as follows: Figure 4Table 1 shows some parameters used to extract and train the model in the vehicle trajectory data.

[0211] Table 1 Description of some parameters of vehicle trajectory data

[0212]

[0213]

[0214] The highD dataset contains a document with statistics of each vehicle, which includes a label called the number of lane changes. Therefore, we select the vehicle IDs whose number of lane changes is not 0 as the target vehicles for lane changes. After that, we can quickly extract a set of lane change sequences based on the frame ID, vehicle ID, road ID, and surrounding vehicle IDs. As for obtaining the straight-ahead data, we can directly select the vehicles whose number of lane changes is 0.

[0215] Figure 5 A schematic diagram of the lane change data is shown. In order to ensure the integrity of the lane change trajectory, this application defines the intersection of the lane change trajectory and the lane line as the lane change critical point A. Only data that exceeds 5s before and after point A can be used, so that the model can fully extract lane change features. Finally, after filtering, this application extracts a total of 28,571 left lane change data, 32,934 right lane change data, and selects 30,000 lane keeping data for model training. Finally, this application divides the entire data set into training set, validation set, and test set in a ratio of 7:1:2, so that historical information will not leak into the test set.

[0216] After the data set is divided, the time series needs to be input into the model in a certain format and length to ensure that the model can save computing resources and learn well. h The sliding window (8s in this application) slides sequentially on the input sequence to obtain samples:

[0217]

[0218] in, is the characteristic sequence, Y t is the corresponding label. For the trajectory prediction module, the label can be directly obtained from the horizontal and vertical coordinates of the real trajectory, that is:

[0219] Y t =(x t ,y t )(1.25)

[0220] For the intention recognition module, the label is the vehicle's driving intention:

[0221] Y t∈{0,1,2}(1.26)

[0222] Among them, 0, 1, and 2 represent the vehicle's intention to keep, change left, and change right lanes, respectively. However, the intention recognition module does not have accurate truth values ​​available, and the annotation method of this application is as follows:

[0223] ① For vehicles going straight, all extracted sequences are marked as 0 (lane keeping).

[0224] ② The markings for vehicles changing lanes left are as follows Figure 5 As shown, the steps are as follows:

[0225] Ⅰ. Calculate the yaw angle of each point on the trajectory:

[0226]

[0227] Among them, x t and t Represents the horizontal and vertical coordinates of the vehicle at time t.

[0228] II. Define the heading angle threshold θ b , find the first point B whose yaw angle is less than the threshold as the lane change starting point.

[0229] III. Considering that the intention is generated before the actual lane-changing action, point C is found by calculating a fixed time of 1 second from point B as the moment when the intention is generated.

[0230] IV. Define the sequence from C to A as 1 (left lane change) and the rest as 0 (lane keeping).

[0231] ③The marking method for vehicles changing lanes right is similar to that for vehicles changing lanes left.

[0232] The model was trained and predicted on a single GeForceRTX3070GPU using the Pytorch-1.8.0 framework in CUDA11.0, cuDNN8.0, Python3.7, and Ubuntu18.04 environments.

[0233] The network parameters of Informer are shown in Table 2. The network was trained for 500 epochs using the Adam optimizer.

[0234] Table 2 Network parameters

[0235]

[0236] The model is trained with the training set, and the actual effect is evaluated based on the test set. Considering that the model designed in this application has two modules: intent recognition and trajectory prediction, the corresponding loss functions are designed respectively.

[0237] The vehicle intention recognition module uses a multi-classification cross entropy function to calculate the loss:

[0238]

[0239] Among them, m is the number of samples, k represents the intent type, and p ik represents the independent encoding of the i-th sample label, q ik is the probability that the i-th sample belongs to the k-th intent.

[0240] For the trajectory prediction model, the likelihood function that minimizes the negative logarithm is used to train and optimize the parameters in the model:

[0241] L=-log(∑ k Pθ(G|C k ,X obs )P(C k |X obs ))(1.29)

[0242] Among them, X obs is the historical trajectory sequence of the target vehicle, C k is the intention predicted by the driving recognition module, and G is the Gaussian distribution of the future trajectory output by trajectory prediction.

[0243] At each time step, the model calculates the loss and updates the weights of all fully connected layers and attention layers through the back-propagation algorithm to minimize the loss.

[0244] In order to verify the effectiveness of the proposed model, this application uses the same verification set, designs corresponding evaluation indicators for the intention recognition module and trajectory prediction module respectively, and selects several basic and common models in the current research field for comparative analysis.

[0245] For the intent recognition model, the evaluation indicators include the following aspects:

[0246] ① Average accuracy ACC: refers to the proportion of correctly predicted samples in the test set to the total test set, and its expression is:

[0247]

[0248] Among them, TP L Indicates that both the predicted result and the true label of the sample are left lane change, TP C Indicates that the predicted result and the true label of the sample are both lane keeping, TP R Indicates that both the predicted result and the true label of the sample are right lane change, N test is the number of samples in the test set.

[0249] Through the above formula, we can not only get the corresponding accuracy for each intent separately, but also calculate the comprehensive accuracy of the entire model, that is, the accuracy of multi-classification problems.

[0250] ②F1 score: An indicator used in statistics to measure the accuracy of a binary classification model, taking into account both the precision P and recall R of the classification model. The expression is as follows:

[0251]

[0252] Among them, TP represents the number of samples whose true label is positive and whose predicted label is positive, FP represents the number of samples whose true label is negative but whose predicted label is positive, and FN represents the number of samples whose true label is positive and whose predicted label is negative.

[0253] Since this application has three driving intentions and is not a binary classification problem, the macro F1 is used as the comprehensive score:

[0254]

[0255] Where K is the number of intentions (there are 3 in this application, K=3), F 1k Indicates the F1 value corresponding to the k-th intent.

[0256] Next, this application selected the following comparison models:

[0257] SVM model: directly input the state of the target vehicle at time t into the SVM to obtain the lane-changing intention at that moment.

[0258] HMM model: The state transition probability and observation probability matrix are obtained through supervised learning. The target vehicle state at the current moment is input into the matrix to calculate the vehicle lane change probability.

[0259] LSTM model: Receive the historical feature sequence of the target vehicle and input it into the LSTM network. The features include the distance between the target vehicle and the lane line, the lateral longitudinal speed and the longitudinal acceleration. The lane change probability is obtained through the Softmax function.

[0260] Transformer model: The historical trajectory of the target vehicle and the status of other vehicles are input into the encoder and decoder of the Transformer respectively, and then the Softmax function is applied to the decoder output to obtain the lane change intention probability vector.

[0261] Different lengths of historical trajectory sequences will inevitably affect the prediction results. If the sequence is too short, the model cannot fully and accurately learn the features, and if the sequence is too long, the inference time will increase significantly. h , this application first takes different T in the test set h={1s,2s,3s,4s,5s}, the model was compared and analyzed. The results are as follows Figure 6 As shown in the figure, it can be seen that when T h When the time is >3s, the accuracy of intent recognition slows down, while the calculation time still increases rapidly. Therefore, after weighing the growth rates of the two, this application selects T h = 3s as the input length of the historical trajectory sequence.

[0262] Table 3 Vehicle intention recognition results

[0263]

[0264] Based on T h =3s, this application trains the proposed model and the comparison model respectively, and finally uses precision, recall and F1 score as evaluation indicators, as shown in Table 3. As can be seen from the table, compared with the three traditional models of SVM, HMM and LSTM, the Transformer and Informer models with self-attention mechanism have significantly improved the effect. This is because the self-attention mechanism can better capture the global context and long-distance dependencies, and better mine the internal features of the data than the traditional model. In addition, since this application encodes the target vehicle itself and the interactive information more finely, rather than just inputting the vehicle information, the model effect has also been improved to a certain extent compared with the previous Transformer.

[0265] Overall, the vehicle intention recognition performance indicators have reached a good level, and the accuracy P of vehicle driving intention prediction has reached more than 93%; the recall rates of left lane change, right lane change and lane keeping have also reached 91%, 86% and 92% respectively; the F1 score is a comprehensive evaluation method of precision and recall, and the scores of the three intentions are also at a good level, 93%, 90% and 95% respectively; finally, the overall accuracy of each model is statistically analyzed, which comprehensively reflects the effectiveness of each method. The accuracy of the method proposed in this application has reached more than 93%, indicating that the module can well identify the driving intention of the target vehicle and meet the pre-requisite requirements of subsequent trajectory prediction.

[0266] First, this application visualizes the prediction trajectory output by the IA-Informer model. The prediction results are as follows: Figures 7 to 9 The target vehicle and surrounding vehicles are marked in blue and black respectively. The black trajectory represents the historical trajectory, the red trajectory represents the actual trajectory of the target vehicle, and the blue trajectory represents the predicted trajectory of the target vehicle.

[0267] Figure 7The trajectory prediction results in the lane keeping scenario are shown. Based on the current driving environment and the historical status of surrounding vehicles, that is, the vehicle in front of the lane has a tendency to change lanes to the left, and the left lane is also occupied, the model believes that the target vehicle will intend to keep the lane, so there will be no lane change trajectory. The model predicts the future trajectory of the target vehicle within 5 seconds in total. From the partial enlarged image, it can be seen that the lateral error is kept within a small range. Even at the end of 5 seconds, the lateral error is only about 0.2m, which has good accuracy.

[0268] Figure 8 and Fig. 9 The trajectory prediction results of the target vehicle when changing lanes are shown, with an interval of 2 seconds. As can be seen from the figure, the lane where the target vehicle is currently located is relatively crowded, with a slow car in front and a vehicle following behind, and there are also vehicles driving in the adjacent lane. Figure 8 It shows that, considering the current environmental characteristics and the historical information of the surrounding vehicles, although the target vehicle has not yet made an obvious lane change behavior, the IA-Informer model analyzes the potential lane change of the ego vehicle based on the invasion of the surrounding vehicles on the future driving space of the ego vehicle, and thus predicts the trajectory of the left lane change. After 2s, if Fig. 9 As shown in the figure, the target vehicle officially starts changing lanes, and the model also predicts the next lane-changing trajectory of the vehicle, which is completely consistent with the actual trajectory in the future. In the overall process, although the predicted trajectory has a certain error when the time domain exceeds 4s, the model has a good control of the overall target vehicle intention and the trajectory prediction error within 3s.

[0269] In addition, due to the different difficulty of the prediction tasks, the model's prediction in the lane keeping scenario is more accurate and closer to the actual trajectory than the lane changing scenario. In general, the trajectory predicted by the IA-Informer model proposed in this application is highly consistent with the actual trajectory.

[0270] In addition to model visualization, it is also necessary to quantitatively analyze the pros and cons of the model. Since the trajectory prediction task is a regression problem, this application uses root mean square error (RMSE), average displacement error (ADE) and final displacement error (FDE) as evaluation indicators:

[0271]

[0272] Among them, N test is the number of test samples, and They represent the predicted sequence and the corresponding true value sequence of the kth test sample at t timestamps, respectively. pre is the predicted length.

[0273] In order to test the effectiveness of the proposed model, this application selected several popular networks that have attracted much attention in current research for comparative analysis:

[0274] Single LSTM model: A simple LSTM architecture is used, which only inputs the vehicle's historical trajectory for prediction.

[0275] Dual LSTM model: Add another LSTM network on the basis of the single LSTM to predict the vehicle's lane change intention and guide the prediction of the trajectory based on the intention.

[0276] Bidirectional LSTM model: Two bidirectional LSTMs are used to encode the historical trajectory of the ego vehicle and the predicted lane change intention of the vehicle in parallel, and a third bidirectional LSTM is used as a decoder to output the predicted trajectory.

[0277] Social LSTM: LSTM modeling is performed on each vehicle, and a social pooling layer is used to extract interaction information in each iteration. In addition to historical trajectory information, the input information also includes information about surrounding participants.

[0278] Dual Transformer Model: Two Transformer modules are used to implement the tasks of intent recognition and trajectory prediction respectively. The intent prediction model extracts social associations based on the input target vehicle state and interaction information and outputs an intent probability vector. The trajectory prediction model adds the intent probability vector to the trajectory to be predicted to achieve more accurate trajectory prediction.

[0279] Table 4 Prediction root mean square error of different models

[0280]

[0281] The RMSE of different models on the test set is shown in Table 4. The results show that in most cases, the model proposed in this application can obtain a smaller prediction error. Table 4 also shows the prediction errors in five different time domains. It can be observed that the prediction error of the IA-Informer model for highD in different time domains is reduced by 19.4%, 14.4%, 3.9%, 10% and 9.5% even compared with the relatively new DualTransformer, which has a greater advantage than other models, thus proving that the model has good results in both short-term and long-term prediction tasks.

[0282] In addition to RMSE, this application also compares the ADE and FDE of the dataset as another evaluation method. Fig.10The distribution box plots of the average displacement error and final displacement error of the method proposed in this application and other reference methods in the prediction time domain of 1-5s are given. It can be observed that as the prediction time increases, the error growth of the model proposed in this application is relatively small. At the same time, from a global perspective, IA-Informer has the smallest error mean and distribution range, indicating its stable performance in trajectory prediction.

[0283] Among them, the intention prediction module extracts social associations according to the status of the target vehicle and surrounding vehicles and outputs the intention probability vector; the trajectory prediction module introduces the output of the former to further improve the prediction performance. Through verification and evaluation of the highD dataset, the accuracy of this model in identifying the driving intention of the target vehicle is as high as 93%, and the root mean square error of trajectory prediction within 1-5s is only 0.29, 0.65, 1.46, 1.98 and 3.04. Compared with several other basic models, they all have better accuracy.

[0284] In one embodiment, the method for predicting vehicle trajectories integrating intention perception further includes:

[0285] S3. Establish an intention sharing mechanism;

[0286] S4. Optimize the IA-Informer model based on the intent sharing mechanism. In S3 and S4, the IA-Informer model is optimized based on the intent sharing mechanism, which can greatly save the intent recognition resources of the IA-Informer model and improve its work efficiency.

[0287] The intention sharing mechanism includes:

[0288] The vehicles that have fallen into or will fall into the sharing request range within the preset time before and after the current time are taken as sharing request targets; wherein the sharing request range is determined based on the historical vehicle trajectory prediction results, the current position of the vehicle and the driving environment map; the preset time can be, for example, 100 seconds; and "will fall into" means that the route planned by the navigation of the vehicle indicates that it will enter the sharing request range within the next 20 seconds; the sharing request vehicle can be determined based on the sharing request range;

[0289] Based on the intention sharing request template, an intention sharing request is made to the sharing request target to obtain the request feedback result; the intention sharing request template is a template for inquiring the sharing request target, for example: requesting it to share the information of the vehicle's own driver's driving behavior prediction / the vehicle's own future driving plan (i.e., the request feedback result), so that the system can directly determine the driving intention of the sharing request target;

[0290] Among them, based on the intention sharing mechanism, the IA-Informer model is optimized, including:

[0291] Input the request feedback result into the IA-Informer model, so that the IA-Informer model reduces the intention resources; when optimizing, directly input the request feedback result into the IA-Informer model;

[0292] The steps for determining the scope of the sharing request based on the historical vehicle trajectory prediction results, the current location of the vehicle, and the driving environment map are as follows:

[0293] Marking the current position of the vehicle on a driving environment map; the driving environment map is a road environment map of the city;

[0294] On the driving environment map, a circular range is drawn with the current position of the marked vehicle as the center and a preset radius length as the radius; the radius length can be 2 meters;

[0295] Divide the circular range into N sector ranges; N is a positive integer; N can be set in advance by the technician;

[0296] Traverse each sector range in turn;

[0297] Each time the traversal is performed, information analysis conditions are generated based on the multimodal features of the sector range; the multimodal features include at least: the distance to the next intersection, etc.; the generated information analysis conditions are conditions for analyzing the historical weekly vehicle trajectory prediction results to determine the value of the sharing intention of the relevant vehicles within the sector range. For example, if the multimodal feature is 20 meters away from the next intersection, the total number of vehicles that have changed lanes by turning left or right in the history reflected in the historical weekly vehicle trajectory prediction results is analyzed, and the analysis value is quantified based on a pre-set quantization table (for example, the quantization table has analysis values ​​corresponding to different types of numerical values) into an analysis value;

[0298] Based on the information analysis conditions, the historical vehicle trajectory prediction results are analyzed to obtain the analysis value; the historical vehicle trajectory prediction results are the results of the IA-Informer model working on the vehicle's vehicle trajectory prediction for the intention recognition and fusion intention within 5 minutes before the current time;

[0299] Associate the analysis value with the traversed fan range;

[0300] After traversing the fan-shaped range sequence, a target adjacent fan-shaped range set is determined from the N fan-shaped ranges; wherein the fan-shaped ranges in the target adjacent fan-shaped range set are adjacent to each other, and the average value of the analysis values ​​associated with each other exceeds the threshold value of the average value; the threshold value of the average value may be 5; the larger the analysis value, the greater the value of the relevant vehicle sharing intention within the fan-shaped range, based on which the target adjacent fan-shaped range set is reasonably determined;

[0301] Based on the target adjacent sector range set, a shared request range is obtained; wherein the shared request range is the concatenated sum of the sector ranges in the target adjacent sector range set.

[0302] The scope of sharing requests is determined in a targeted and reasonable manner to avoid excessive and unnecessary sharing requests that may burden system resources, greatly improving the applicability of the system.

[0303] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A vehicle trajectory prediction method integrating intention perception, characterized in that: include: Retrieve the IA-Informer model; Based on the IA-Informer model, the vehicle's surrounding intentions are recognized and the trajectory of the fused intentions is predicted.

2. The method for predicting vehicle trajectory based on fusion intention perception as claimed in claim 1, characterized in that: The IA-Informer model includes: an input matrix module, a target vehicle intention recognition module and a trajectory prediction module of fusion intention.

3. The method for predicting vehicle trajectory based on fusion intention perception as claimed in claim 2, characterized in that: The steps of constructing the input matrix module include: Define the target vehicle as V p0 , the other vehicles in the scene where the target vehicle is located are defined as V si (i∈1,2,…,N); the vehicle’s forward direction is the positive direction of the y-axis, and the direction perpendicular to the y-axis is the x-axis; where N is the total number of other vehicles; The encoder and decoder inputs of the IA-Informer model are denoted as X en and X de ; The encoder input is the historical trajectory P of the target vehicle 0 (t) and the interaction information S K (t) is composed of the mask trajectory of the target vehicle at the decoder input Define the target vehicle V p0 The trajectory coordinates at time t are: In the observation domain with domain length mt, the historical trajectory of the target vehicle is: The interaction relationship of the target vehicle at time t is: Where K represents the target vehicle number, and Indicates the relative horizontal and vertical distances between the target vehicle and vehicle K. and is the relative lateral and longitudinal speed between the target vehicle and vehicle K; In the observation domain with domain length mt, the interaction information of the target vehicle is: The encoder input is: X en =Concat(P 0 (t),S K (t)); The input sequence of the decoder at time t is for:

4. The method for predicting vehicle trajectory based on fusion intention perception as claimed in claim 2, characterized in that: The steps of constructing the target vehicle intention recognition module include: The intention recognition module predicts the driving intention at the next moment based on the observed target vehicle trajectory information and its interaction information with surrounding vehicles. The ultimate goal is to make the network learn a function: in, This is the final trained intention prediction model, ω∈N 1×3 To predict the result of the intention, that is, the intention probability vector, the formula is as follows: ω=[a1,a2,a3] a1+a2+a3=100%; Among them, a1, a2, and a3 are the probabilities of the three strategies that the target vehicle may adopt; the three strategies include: lane keeping, left lane change, and right lane change; The context vector is calculated layer by layer from bottom to top: H1=X en W1+PE temp A1=norm(prob(H1,H1,H1)+H1) H2=norm(relu(A1W2)W′2+A1) H′2=MaxPool(ELU(Conv1d(H2)) A2=norm(prob(H′2,H′2,H′2)+H′2) F = norm(relu(A2W3)W′3+A2) ω=FW4; Among them, W1, W2, W2', W3, W3', and W4 are the weights of the fully connected layer, relu is the activation function of the feedforward layer, norm, prob, MaxPool, ELU, and Conv1d are normalization, probabilistic sparse self-attention layer, maximum pooling layer, exponential linear unit, and one-dimensional convolution layer, respectively.

5. The method for predicting vehicle trajectory based on fusion intention perception as claimed in claim 2, characterized in that: The steps to build the trajectory prediction module of fusion intent include: The trajectory prediction module uses the historical trajectory information and intention probability vector of the target vehicle as input and outputs the expected trajectory for the next time period nT, which is as follows: Among them, ω(t) is the intention probability vector output by the target vehicle intention recognition module; The decoder encodes the result of intent prediction and inputs it into the multi-head attention module. At the same time, it receives the context vector generated by the encoder and applies it to the second self-attention layer. The process is as follows: H3=X de W5+PE temp +ωW6 A3=norm(prob(H3,H3,H3)+H3) A4=norm(mha(A3,F,F)+A3) H4=norm(relu(A3W6)W′6+A4) H′4=MaxPool(ELU(Conv1d(H4)) Among them, W6, W6', and W7 are the weights of the fully connected layer; mha is the multi-head attention layer.

6. The method for predicting vehicle trajectory based on fusion intention perception as claimed in claim 1, characterized in that: It also includes the testing and validation steps of the IA-Informer model, including: Preprocess the dataset; Based on the preprocessing results, the IA-Informer model is trained and simulated for verification.

7. The method for predicting vehicle trajectory based on fusion intention perception as claimed in claim 6, characterized in that: The preprocessing of the data set includes: The dataset is divided into training set, validation set and test set in a ratio of 7:1:2; After the data set is divided, the time series needs to be input into the model in a certain format and length. h The sliding window slides in the input sequence one by one to obtain samples: in, is the characteristic sequence, Y t is the corresponding label; For the trajectory prediction module, the label can be obtained directly from the horizontal and vertical coordinates of the real trajectory, that is: AND t (x) t ,and t ); For the intention recognition module, the label is the vehicle's driving intention: Y t ∈{0,1,2}； Among them, 0, 1, and 2 represent the vehicle's intention to keep, change lanes left, and change lanes right, respectively; The annotation rules of the intent recognition module are as follows: For vehicles going straight, all the extracted sequences are marked as 0; The steps for marking a left lane-changing vehicle are as follows: Calculate the yaw angle for each point on the trajectory: Among them, x t and t represents the horizontal and vertical coordinates of the vehicle at time t; x t+1 and t+1 Represents the horizontal and vertical coordinates of the vehicle at time t+1; Define the heading angle threshold θ b , find the first point B whose yaw angle is less than the threshold as the lane change starting point; Considering that the intention is generated before the actual lane change action, a fixed time of 1 second is calculated forward from point B to find point C as the moment when the intention is generated; Define the sequence from C to A as 1 and the rest as 0.

8. The method for predicting vehicle trajectory based on fusion intention perception as claimed in claim 7, characterized in that: Based on the preprocessing results, the IA-Informer model is trained, including: The vehicle intention recognition module uses a multi-classification cross entropy function to calculate the loss: Among them, m is the number of samples, k represents the intent type, and p ik represents the independent encoding of the i-th sample label, q ik is the probability that the i-th sample belongs to the k-th intent; For the trajectory prediction model, the likelihood function that minimizes the negative logarithm is used to train and optimize the parameters in the model: L=-log(∑ k P θ (G|C k ,X obs )P(C k |X obs )); Among them, X obs is the historical trajectory sequence of the target vehicle, C k is the intention predicted by the driving recognition module, G is the Gaussian distribution of the future trajectory output by trajectory prediction; P θ (G|C k , X obs ) is the Gaussian distribution probability of the future trajectory G under the given driving intention Ck and historical trajectory sequence Xobs; P(C k |X obs ) is the probability of the intention being Ck given the historical trajectory and Xobs; At each time step, the model calculates the loss and updates the weights of all fully connected layers and attention layers through the back-propagation algorithm to minimize the loss.

9. The method for predicting vehicle trajectory based on fusion intention perception as claimed in claim 7, characterized in that: Based on the preprocessing results, the IA-Informer model is simulated and verified, including: For the target vehicle intention recognition module, the first indicator and the first comparison model are used to perform simulation verification of the IA-Informer model according to the preprocessing result; wherein the first indicator and the first comparison model include: average accuracy ACC, F1 score, SVM model, HMM model, LSTM model, Transformer model; For the trajectory prediction module of fusion intention, the second indicator and the second comparison model are used to perform simulation verification of the IA-Informer model according to the preprocessing results; among them, the second indicator and the second comparison model include: root mean square error RMSE, average displacement error ADE, final displacement error FDE, single LSTM model, dual LSTM model, bidirectional LSTM model, social LSTM and dual Transformer model.

10. A vehicle trajectory prediction system integrating intention perception, characterized in that: include: A calling unit, used to call the IA-Informer model; The prediction unit is used to identify the vehicle's surrounding intentions and predict the trajectory of the fused intentions based on the IA-Informer model.

Images