A fusion intention-aware surrounding vehicle trajectory prediction method and system

Abstract

The application provides a fusion intention perception surrounding vehicle trajectory prediction method and system, wherein the method comprises the following steps: calling an IA-Informer model; based on the IA-Informer model, performing intention recognition on surrounding vehicles and performing trajectory prediction of the surrounding vehicles in a fusion intention mode; the IA-Informer model comprises an input matrix module, a target vehicle intention recognition module and a trajectory prediction module in a fusion intention mode. In the application, an intention prediction module extracts social correlation according to the states of target vehicles and surrounding vehicles, and outputs an intention probability vector; a trajectory prediction module introduces the output of the former, and further improves the prediction performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of neural networks, and particularly relates to a surrounding vehicle trajectory prediction method and system fusing intention perception. BACKGROUND

[0002] In current research, when dealing with vehicle trajectory prediction problems, the common method is to transform it into a task of inferring the prior probability distribution of future coordinates according to historical coordinates, and a large number of studies have shown that incorporating the interaction between the target vehicle and the surrounding vehicles in prediction can greatly improve the prediction accuracy. However, there are still two research gaps in existing research. First, the existing research on the solution to the lane-changing intention and trajectory prediction problems is often independent, that is, a corresponding model is established for the lane-changing intention and trajectory prediction problems respectively. This results in the model ignoring the interaction between intention prediction and trajectory prediction, thereby sacrificing the accuracy of prediction to some extent. Second, when considering the interaction between the target vehicle and the surrounding vehicles, many researchers simply input the state of the target vehicle and the state of the surrounding vehicles as a feature vector of the prediction model, and the analysis of the coupled states is not deep enough. SUMMARY

[0003] One of the purposes of the present application is to provide a surrounding vehicle trajectory prediction method fusing intention perception, wherein the intention prediction module extracts social correlation according to the state of the target vehicle and the surrounding vehicles, and outputs an intention probability vector; the trajectory prediction module introduces the output of the former, further improving 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; a more refined interaction encoding method is designed.

[0004] The surrounding vehicle trajectory prediction method fusing intention perception provided by the embodiment of the present application comprises:

[0005] The IA-Informer model is called;

[0006] Based on the IA-Informer model, intention recognition and intention-fused trajectory prediction are performed on the surrounding vehicles of the vehicle.

[0007] Preferably, the IA-Informer model comprises an input matrix module, a target vehicle intention recognition module and an intention-fused trajectory prediction module.

[0008] Preferably, the construction step of the input matrix module comprises:

[0009] The target vehicle is defined as V p0 , and other vehicles in the scene where the target vehicle is located are defined as V si(i∈1,2,…,N). The forward direction of the vehicle is the positive direction of the y-axis, and the vertical direction of the y-axis is the x-axis; wherein 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 respectively; wherein the encoder input is composed of the historical trajectory P 0 (t) of the target vehicle and the interaction information S K (t), and the decoder input is the masked trajectory

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

[0012]

[0013] The historical trajectory of the target vehicle under the observation domain with a length of mtis:

[0014]

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

[0016]

[0017] wherein K represents the number of the target vehicle, and represent the horizontal and longitudinal relative distances between the target vehicle and the Kth vehicle, and are the horizontal and longitudinal relative speeds between the target vehicle and the Kth vehicle;

[0018] The interaction information of the target vehicle under the observation domain with a length of mtis:

[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:

[0023]

[0024] Preferably, the construction step of the target vehicle intention recognition module comprises:

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

[0026]

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

[0028] wherein, is the finally trained intention prediction model, ω ∈ N 1×3 is the result of predicting the intention, i.e., the intention probability vector defined in the present application, and the formula is as follows:

[0029] ω=[a1,a2,a3]

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

[0031] wherein a1, a2, and a3 are the probabilities of the three strategies that the target vehicle may appear; the three strategies include lane keeping, left lane changing, and right lane changing;

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

[0041] Preferably, the construction steps of the trajectory prediction module integrating intention include:

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

[0043]

[0044] The decoder encodes the result of intention prediction and inputs it into the multi-head attention module, and applies the context vector generated by the encoder to the second self-attention layer, 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] Wherein, W6, W6', W7 are the weight values of the full connection layer; and mha is a multi-head attention layer.

[0052] The intention perception-fused surrounding vehicle trajectory prediction system provided by the embodiment of the present application comprises:

[0053] The calling unit is configured to call the IA-Informer model.

[0054] The prediction unit is configured to perform intention recognition on the surrounding vehicle of the vehicle and perform intention-fused trajectory prediction on the surrounding vehicle of the vehicle based on the IA-Informer model.

[0055] The technical solutions of the present application will be further described in detail below with the aid of the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

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

[0057] Figure 1 A schematic diagram of a fusion intention perception surrounding vehicle trajectory prediction method in an embodiment of the present application;

[0058] Figure 2 A structure diagram of an IA-Informer model in an embodiment of the present application;

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

[0060] Figure 4 A highD data visualization effect diagram in an embodiment of the present application;

[0061] Figure 5 A lane change data schematic diagram in an embodiment of the present application;

[0062] Figure 6 A different input sequence length verification result diagram in an embodiment of the present application;

[0063] Figure 7 A prediction result diagram in a lane keeping scene in an embodiment of the present application;

[0064] Figure 8 A prediction result diagram at an initial time in a lane change scene in an embodiment of the present application;

[0065] Figure 9 A prediction result diagram at a final time in a lane change scene in an embodiment of the present application;

[0066] Figure 10 A trajectory error distribution diagram of each model in an embodiment of the present application. DETAILED DESCRIPTION

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

[0068] An embodiment of the present application provides a surrounding vehicle trajectory prediction method fusing intention perception, as shown in Figure 1 , comprising:

[0069] S1, calling an IA-Informer model;

[0070] S2, based on the IA-Informer model, performing intention recognition on the surrounding vehicles of the vehicle and trajectory prediction fusing intention.

[0071] The IA-Informer model comprises an input matrix module, a target vehicle intention recognition module, and a trajectory prediction module fusing intention.

[0072] The construction step of the input matrix module includes:

[0073] Define the target vehicle V p0 , and other vehicles in the scene where the target vehicle is located as V si (i∈1,2,…,N). The forward direction of the vehicle is the positive direction of the y-axis, and the vertical direction of the y-axis is the x-axis; wherein 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 respectively; wherein the encoder input is composed of the historical trajectory P 0 (t) of the target vehicle and the interaction information S K (t); the decoder input is the mask trajectory

[0075] Define the trajectory coordinates of the target vehicle V p0 at time t as:

[0076]

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

[0078]

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

[0080]

[0081] wherein K represents the number of the target vehicle, and represent the horizontal and longitudinal relative distances between the target vehicle and the Kth vehicle, and are the horizontal and longitudinal relative speeds between the target vehicle and the Kth vehicle;

[0082] Under the observation domain with a length of 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:

[0087] ​

[0088] The construction step of the target vehicle intention recognition module includes:

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

[0090]

[0091] Wherein, is the finally trained intention prediction model, ω∈N 1×3 To predict the intention result, i.e. the intention probability vector defined in the application, the formula is as follows:

[0092] ω=[a1,a2,a3]

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

[0094] Wherein, a1, a2, a3 are the probabilities of the three strategies that the target vehicle may appear; the three strategies include: lane keeping, left lane changing and right lane changing;

[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] Wherein, W1, W2, W2', W3, W3', W4 are the weight values of the full connection layer, relu is the activation function of the feedforward layer, norm, prob, MaxPool, ELU and Conv1d are normalization, probability sparse self-attention layer, maximum pooling layer, exponential linear unit and one-dimensional convolution layer respectively.

[0104] The construction step of the trajectory prediction module fusing the intention includes:

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

[0106]

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

[0108] The decoder encodes the result of intention prediction and inputs it into the multi-head attention module, and simultaneously receives the context vector generated by the encoder and applies it to the second self-attention layer, and 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] Wherein, W6, W6', and W7 are the weight values of the full connection layer; and mha is the multi-head attention layer.

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

[0117] The main feature of the Informer model is to adopt a multi-scale time encoder and decoder structure, and to integrate an adaptive length attention mechanism, which is very suitable for predicting long sequences and multi-label outputs. Vehicle state and trajectory data are a kind of data with periodicity and continuity, and the interaction features between the data have important value when predicting the trajectory. Therefore, the vehicle intention recognition and trajectory prediction tasks are completed based on the framework of the Informer model.

[0118] To solve the above problems, this section proposes an IA-Informer method fusing intention perception to realize the prediction of the target vehicle intention and trajectory, and the overall framework is as follows: Figure 2IA is Intention Aware. The left dashed box in the figure is the encoder structure, and the right 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 time series and vehicle relative position, the historical trajectory of the ego vehicle is embedded in time and then input into the encoder together. After the calculation of each layer of the encoder, the input information is converted into a context vector. This vector can be used as an intermediate variable to extract information and input into the decoder, and on the other hand, the result of intention prediction can be output through a fully connected layer. The present application defines an intention recognition probability vector, which expresses the intention of the vehicle in the form of probability values of keeping and lane changing.

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

[0121] The ultimate goal of the present application is to realize the vehicle trajectory prediction task on a structured road, so a general three-lane scene needs to be defined for subsequent use of the highD dataset, as shown in Figure 3 . The yellow vehicle is the ego vehicle (i.e. the autonomous vehicle carrying the algorithm), and the blue vehicle is the target vehicle whose trajectory is to be predicted. 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 , respectively. The encoder input is composed of the historical trajectory P 0 (t) of the target vehicle and the interaction information S K (t), and the decoder input is the masked trajectory

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

[0124]

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

[0126]

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

[0128]

[0129] Where K represents the target vehicle's number, vehicles 1 and 6 are located in front of and behind the target vehicle in the left lane, vehicles 2 and 5 are located in front of and behind the target vehicle in its current lane, and vehicles 3 and 4 are located in front of and behind the target vehicle in the right lane. and This indicates the relative distance between the target vehicle and vehicle K in both the horizontal and vertical directions. and It is the lateral and longitudinal relative speed between the target vehicle and vehicle K; where the target vehicle is the vehicle whose intention and trajectory are to be predicted.

[0130] Therefore, in an observation domain of 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 receives the input sequence at time t. It can be represented as:

[0135]

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

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

[0138]

[0139] wherein is the final trained intention prediction model, ω ∈ N 1×3 is the result of predicting the intention, i.e., the intention probability vector defined in the present application, the formula is as follows:

[0140] ω = [a1, a2, a3] (1.8)

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

[0142] wherein a1, a2, a3 are the probabilities of the three strategies (lane keeping, left lane changing, and right lane changing) that the target vehicle is likely to appear.

[0143] As shown on the left side of Figure 2 , the overall structure of the model is constructed 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] wherein all W 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, probability sparse self-attention layer, maximum pooling layer, exponential linear unit, and one-dimensional convolution layer, respectively, which will be described in detail below.

[0152] Formula (1.10) shows the calculation process of the embedding layer, in addition to the historical trajectory P 0 (t) of the target vehicle and the interaction information S K (t), the sine-cosine space-time encoding PE of the trajectory is also addedtemp , and finally combined into the sequence by addition. In order to match the dimension 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 features. Therefore, the fusion sequence H1 extracts the information of the original sequence while highlighting its spatiotemporal features, improving the network's ability to recognize and utilize the correlation between the nodes of the input sequence.

[0153] Equation (1.11) represents the multi-head probability sparse self-attention layer. In addition to this, a residual connection and layer normalization are added to this layer. The residual connection is an additive connection that connects the "input" and the "output", which improves the problem of gradient dispersion, ensures that the network at least does not learn in the wrong direction, makes the network converge faster, and the purpose of layer normalization is to improve the generalization performance of the network.

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

[0155] The superposition of the embedding layer may cause the input sequence to be very long, which makes the feature vector contain not very concentrated information. Equation (1.13) performs a data distillation. The stride of the max 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 without changing the overall feature distribution, and a key context representation vector is generated in the next layer, thereby achieving efficient memory usage.

[0156] Equations (1.14) and (1.15) are repeated operations of the network structure of equations (1.11) and (1.12). By continuously stacking network layers, data can be more deeply mined to deal with more complex problems. In this application, partial extraction of the 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 equation (1.16), the context vector can obtain the intention probability vector through a fully connected layer.

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

[0159]

[0160] As shown on the right side of Figure 2 , the decoder encodes the result of intention prediction and inputs it into the multi-head attention module, while receiving the context vector generated by the encoder and applying 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] where all Ws are the weights of the fully connected layers. Except for the sparse probability self-attention layer in equation (1.19), the decoder also has a multi-head attention layer (i.e., mha in equation (1.20)) whose inputs are partly from the context vector F extracted from the historical trajectory in the encoder and partly from the features A3 extracted from the decoder input. Since the context representation of the real trajectory and the historical trajectory are considered together, the multi-head attention layer improves the accuracy of trajectory prediction. Equations (1.21) and (1.22) and the encoder calculation are the same, both of which are a feedforward layer and data distillation. Finally, the predicted trajectory is generated by the fully connected layer in equation (1.23).

[0168] This 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 intent recognition module of the model first extracts social connections according to the states of the target vehicle and the surrounding vehicles and outputs an intent probability vector (e.g., the probabilities of keeping the lane, left lane changing, and right lane changing).

[0170] The trajectory prediction module then takes the intent probability vector as additional information and combines the historical trajectory of the target vehicle to generate a more accurate future trajectory prediction. This way can make the trajectory prediction more consistent with the actual intent of the vehicle and achieve a collaborative improvement of the two tasks.

[0171] The model designs a more refined interactive encoding method, which includes the following points:

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

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

[0174] This approach can more deeply explore the coupling state between the target vehicle and the surrounding vehicles, no longer simply treating them as feature inputs, but fully analyzing their mutual relationship to improve the model's adaptability to complex traffic environments.

[0175] In one embodiment, the intention-aware surrounding vehicle trajectory prediction method further includes testing and verifying steps of the IA-Informer model, including:

[0176] Preprocessing the data set;

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

[0178] The preprocessing of the data set includes:

[0179] According to the proportions of 7:1:2, the data set is divided into a training set, a validation set and a test set;

[0180] After completing the division of the data set, the time series needs to be input into the model in a certain format and length. By using a sliding window with a length of t h , the input sequence is sequentially slid to obtain samples:

[0181]

[0182] wherein, is the feature 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 true trajectory, i.e.

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

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

[0185] Y t ∈ {0, 1, 2}.

[0186] where 0, 1, 2 represent the keep, left-lane change and right-lane change intentions of the vehicle respectively;

[0187] The labeling rules of the intention recognition module are as follows:

[0188] For the vehicle that is straight driving, all the extracted sequences are marked as 0;

[0189] The labeling steps for the left-lane change vehicle are as follows:

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

[0191]

[0192] where x t and y t represent the horizontal and vertical coordinates of the vehicle at time t; x t+1 and y t+1 represent 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 occurs before the actual lane change action, find point C by pushing back 1s from point B as the intention generation time;

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

[0196] Based on the preprocessing results, the training of the IA-Informer model is carried out, including:

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

[0198]

[0199] where m is the number of samples, k represents the intention type, 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 intention;

[0200] For the trajectory prediction model, the negative log likelihood function 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] where X obs is the history trajectory sequence of the target vehicle, C k is the predicted intention of the driving recognition module, and G is the Gaussian distribution of the future trajectory predicted by the trajectory prediction output.

[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, simulation verification of the IA-Informer model is carried out, including:

[0205] For the target vehicle intention recognition module, the first index and the first contrast model are used to carry out simulation verification of the IA-Informer model according to the preprocessing results; wherein the first index and the first contrast model include: average accuracy ACC, F1 score, SVM model, HMM model, LSTM model, Transformer model.

[0206] For the trajectory prediction module of the fusion intention, the second index and the second contrast model are used to carry out simulation verification of the IA-Informer model according to the preprocessing results; wherein the second index and the second contrast model include: root mean square error RMSE, average displacement error ADE, final displacement error FDE, single LSTM model, double LSTM model, bidirectional LSTM model, social LSTM and double Transformer model.

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

[0208] In order to verify the effectiveness of the model proposed in the present application in predicting the behavior of surrounding vehicles, the model will be trained and verified using public data sets, and a deep comparative analysis will be carried out with the current widely used network model.

[0209] The present application selects the public data set highD released by the Institute of Automotive Engineering of Aachen University of Technology in Germany for training and verification of the model. highD is a large-scale natural vehicle trajectory data set of German highways, which contains 110,000 vehicle driving data of 6 locations, with a total duration of 16.5 hours, and more than 5,000 complete lane-changing trajectories.

[0210] The data set contains a total of 60 sub-data sets, each of which contains aerial photographs, collection point data, and vehicle trajectory data of the collection section. The collection point data provides information such as location ID, time, collection duration, vehicle cumulative travel distance, and vehicle number. The vehicle trajectory data mainly includes vehicle ID, vehicle coordinates, speed, acceleration, and other information. The effect of the data after visualization in Matlab is as follows: Figure 4The part of vehicle trajectory data used for model extraction and training is shown in Table 1.

[0211] Table 1. Part of vehicle trajectory data parameter description

[0212]

[0213]

[0214] There is a document in the highD dataset that contains the statistics of each vehicle, and there is a label for the number of lane changes. Therefore, the vehicle ID with a non-zero number of lane changes is selected as the target vehicle for lane changing. Then, according to the frame ID, vehicle ID, road ID, and surrounding vehicle ID, a set of lane changing sequences can be quickly extracted. As for the straight data, those vehicles with a lane change count of 0 can be directly selected.

[0215] Figure 5 A schematic diagram of lane changing data is shown. In order to ensure the integrity of the lane changing trajectory, the intersection of the lane changing trajectory and the lane line is defined as the lane changing critical point A in this application. Only data that exceeds 5s before and after point A can be used, so that the model can fully extract the lane changing features. Finally, after filtering, this application extracts a total of 28571 left lane changing data, 32934 right lane changing data, and selects 30000 lane keeping data for model training. Finally, this application divides the entire dataset into training set, validation set and test set according to the ratio of 7:1:2, so that the historical information will not be leaked into the test set.

[0216] After completing the division of the dataset, 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. This application uses a sliding window with a length of t h (8s in this application) to slide the input sequence to obtain samples:

[0217]

[0218] wherein, is the feature sequence, and 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 driving intention of the vehicle:

[0221] Y t∈ {0, 1, 2} (1.26)

[0222] where 0, 1, 2 represent the vehicle's keep, left-lane change and right-lane change intention respectively. However, the intention recognition module does not have accurate true values available, and the labeling method of the present application is as follows:

[0223] ① For vehicles that are straight ahead, all the extracted sequences are labeled as 0 (lane keeping).

[0224] ② The left-lane change vehicle labeling is shown as Figure 5 , and the steps are as follows:

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

[0226]

[0227] where x t and y t represent 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 start point of lane change.

[0229] III. Considering that the intention occurs before the actual lane change action, find point C by pushing back 1s from point B as the intention generation time.

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

[0231] ③ The right-lane change vehicle labeling method is similar to the left-lane change.

[0232] The model was trained and predicted on a single GeForce RTX 3070 GPU using the Pytorch-1.8.0 framework under the CUDA11.0, cuDNN8.0, Python3.7 and Ubuntu18.04 environment.

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

[0234] Table 2 Network parameters

[0235]

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

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

[0238]

[0239] where m is the number of samples, k represents the intention type, 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 intention.

[0240] For the trajectory prediction model, the parameters in the optimization model are trained by minimizing the negative log-likelihood function:

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

[0242] where 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 the 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 backpropagation algorithm to minimize the loss.

[0244] To verify the effectiveness of the proposed model, the same validation set is used to design corresponding evaluation indicators for the intention recognition module and the trajectory prediction module, and several basic and common models in the current research field are selected for comparative analysis.

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

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

[0247]

[0248] where TP L represents that both the prediction result and the true label of the sample are left lane changing, TP C represents that both the prediction result and the true label of the sample are lane keeping, TP R represents that both the prediction result and the true label of the sample are right lane changing, and N test is the number of test set samples.

[0249] Through the above formula, the corresponding accuracy rate can be obtained for each intention, and the comprehensive accuracy rate of the entire model, i.e., the accuracy rate of the multi-classification problem, can also be calculated.

[0250] ②F1 score: an index used in statistics to measure the accuracy of a binary classification model, which takes into account both the precision P and recall R of the classification model, and the expression is as follows:

[0251]

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

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

[0254]

[0255] Among them, K is the number of intentions (the present application has 3, K = 3), F 1k represents the F1 value corresponding to the kth intention.

[0256] Next, the following comparative models are selected:

[0257] SVM model: directly input the state of the target vehicle t into the SVM to obtain the lane change intention at this moment.

[0258] HMM model: obtain the state transition probability and observation probability matrix through supervised learning, input the state of the target vehicle at the current moment into the matrix, and calculate the lane change probability of the vehicle.

[0259] LSTM model: receives the historical time feature sequence of the target vehicle input into the LSTM network, the features include the distance from the target vehicle to the lane line, the lateral longitudinal speed and the longitudinal acceleration, and obtains the lane change probability through the Softmax function.

[0260] Transformer model: input the historical trajectory of the target vehicle and the state of the other vehicle into the encoder and decoder of the Transformer, respectively, and then use the Softmax function on the output of the decoder 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 learn the features accurately and comprehensively, and if the sequence is too long, the inference time will be greatly increased. In order to obtain a suitable historical sequence length T h , the present application first takes different T h= {1s, 2s, 3s, 4s, 5s} are compared and analyzed. The results are shown in Figure 6 As can be seen from the figure, when T h > 3s, the accuracy of intention recognition grows slowly, while the calculation time still rises rapidly. Therefore, after weighing the growth rate of both, the present application selects T h = 3s as the input length of the historical trajectory sequence.

[0262] Table 3 Vehicle intention recognition results

[0263]

[0264] On the premise of T h = 3s, the proposed model and the comparative model are trained, and finally the precision, recall and F1 score are used as evaluation indexes, 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 a very significant effect improvement, because the self-attention mechanism can better capture the global context and long-distance dependency, and better mine the internal features of the data than the traditional models. In addition, since the target vehicle itself and the interaction information are encoded more finely, not just the vehicle information is input, the model effect has also improved compared to the previous Transformer.

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

[0266] Firstly, the present application visualizes the prediction trajectory output by the IA-Informer model, and the prediction result is shown in Figures 7 to 9 The target vehicle and the surrounding vehicles are marked in blue and black respectively, the black trajectory represents the historical trajectory, the red trajectory represents the true trajectory of the target vehicle, and the blue trajectory represents the predicted trajectory of the target vehicle.

[0267] Figure 7The trajectory prediction result in the lane keeping scenario is shown. Based on the current driving environment and the historical state of the surrounding vehicles, that is, the vehicle in front of the current lane has a left lane change tendency, and the left lane is also occupied, the model considers that the target vehicle will have a lane keeping intention, so there is no lane change trajectory. The model predicts the future trajectory of the target vehicle within 5s, and from the local enlarged view, it can be seen that the lateral error is kept in a small range. Even at the final time of 5s, the lateral error is only about 0.2m, with good accuracy.

[0268] Figure 8 and Figure 9 The trajectory prediction result when the target vehicle changes lanes is shown, and the interval between the two times is 2s. As can be seen from the figure, the current lane of the target vehicle is relatively crowded, there are slow vehicles in front, there are also vehicles following behind, and there are also vehicles driving in the adjacent lane. Figure 8 It is shown that considering the current environmental characteristics and the historical information of the surrounding vehicles, although the target vehicle has not yet shown obvious lane changing behavior, the IA-Informer model analyzes the potential lane changing possibility of the target vehicle according to the occupation of the future driving space of the target vehicle by the surrounding vehicles, and thus predicts a left lane changing trajectory. As shown in Figure 9 , the target vehicle formally starts the lane changing behavior, and the model also predicts the next lane changing trajectory of the vehicle, which is completely consistent with the future real trajectory. In the whole process, although there is a certain error in the prediction trajectory in the time domain exceeding 4s or more, the model controls the overall target vehicle intention and the trajectory prediction error within 3s very well.

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

[0270] In addition to model visualization, quantitative analysis of the pros and cons of the model is also needed. Since the trajectory prediction task belongs to the regression problem, the root mean square error (RMSE), the average displacement error (ADE) and the final displacement error (FDE) are used as evaluation indexes:

[0271]

[0272] where N test is the number of test samples, and respectively represent the prediction sequence and the corresponding true value sequence of the kth test sample at t time stamps, and T pre is the prediction length.

[0273] To verify the effectiveness of the proposed model, the present application selects several popular networks that are highly regarded in current research for comparative analysis:

[0274] Single LSTM model: A simple LSTM architecture is used, and only the historical trajectory of the vehicle is input for prediction.

[0275] Dual LSTM model: Another LSTM network is added to the single LSTM to predict the vehicle's lane change intention, which guides the prediction of the trajectory.

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

[0277] Social LSTM: Each vehicle is modeled using an LSTM, and a social pooling layer is used to extract interaction information at 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 intention recognition and trajectory prediction tasks. The intention prediction model extracts social relationships based on the input target vehicle state and interaction information, and outputs an intention probability vector. The trajectory prediction model adds the intention 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 proposed model can obtain smaller prediction error. Table 4 also shows the prediction error in five different time domains. It can be observed that the IA-Informer model reduces the prediction error of highD by 19.4%, 14.4%, 3.9%, 10%, and 9.5% in different time domains compared with the relatively novel DualTransformer, and has a greater advantage compared with other models, which proves that the model has good effect in short-term and long-term prediction tasks.

[0282] In addition to RMSE, the present application also compares the ADE and FDE of the dataset as another evaluation method. Figure 10The average displacement error and the distribution box plot of the final displacement error of the method proposed in the application and other reference methods in the prediction time domain of 1-5s are given. It is observed from the box plot that the model error of the method proposed in the application grows relatively small with the increase of the prediction time. At the same time, from the global point of view, IA-Informer has the smallest error average and distribution range, indicating its stable performance in trajectory prediction.

[0283] Among them, the intention prediction module extracts social correlation according to the state of the target vehicle and the surrounding vehicles, and outputs an intention probability vector; the trajectory prediction module introduces the output of the former, which further improves the prediction performance. Through the verification and evaluation of the highD data set, the accuracy of the 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, it has better accuracy.

[0284] In one embodiment, the intention-aware surrounding vehicle trajectory prediction method further comprises:

[0285] S3, establishing an intention sharing mechanism;

[0286] S4, optimizing the IA-Informer model based on the intention sharing mechanism; in S3 and S4, optimizing the IA-Informer model based on the intention sharing mechanism can greatly save the intention recognition resources of the IA-Informer model and improve its work efficiency;

[0287] Among them, the intention sharing mechanism comprises:

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

[0289] Based on the intention sharing request template, the intention sharing request is performed on the sharing request target to obtain a 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 driver driving behavior prediction of the vehicle itself / future driving planning of the vehicle itself (i.e. the request feedback result), so as to facilitate the system to directly determine the driving intention of the sharing request target;

[0290] Among them, optimizing the IA-Informer model based on the intention sharing mechanism comprises:

[0291] The request feedback result is input into the IA-Informer model, so that the IA-Informer model reduces the intention resource; in optimization, the request feedback result is directly input into the IA-Informer model;

[0292] The step of determining the sharing request range based on the historical weekly vehicle trajectory prediction result, the current position of the vehicle and the driving environment map is as follows:

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

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

[0295] Dividing the circular range into N sector ranges; wherein N is a positive integer; N can be set in advance by technical personnel;

[0296] Traversing each sector range in turn;

[0297] Each time of traversal, generating an information analysis condition based on the multi-modal features of the sector range; the multi-modal features at least include: distance to the next intersection, etc.; the generated information analysis condition is a condition for analyzing the historical weekly vehicle trajectory prediction result to determine the degree of value of the sharing intention of the related vehicles in the sector range, for example: if the multi-modal feature is the distance to the next intersection of 20 meters, then the total number of vehicles that have made left turns, right turns and lane changes in the historical weekly vehicle trajectory prediction result is quantified into an analysis value based on a pre-set quantification table (for example: the quantification table has different types of numerical values corresponding to analysis values);

[0298] Based on the information analysis condition, analyzing the historical weekly vehicle trajectory prediction result to obtain an analysis value; the historical weekly vehicle trajectory prediction result is the result of the intention recognition and trajectory prediction of the weekly vehicle of the vehicle by the IA-Informer model within 5 minutes before the current time;

[0299] Associating the analysis value with the traversed sector range;

[0300] After traversing the sequence of sector ranges, determining a target adjacent sector range set from the N sector ranges; wherein each sector range in the target adjacent sector range set is adjacent to each other, and the average value of the analysis values associated with each sector range exceeds the threshold value of the average value; the threshold value of the average value can be 5; the larger the analysis value, the greater the degree of value of the sharing intention of the related vehicles in the sector range, based on which the target adjacent sector range set is reasonably determined;

[0301] Based on the target adjacent sector range set, a sharing request range is obtained; wherein the sharing request range is the splicing sum of each sector range in the target adjacent sector range set.

[0302] The sharing request range is determined in a targeted and reasonable manner, unnecessary sharing requests are avoided, and the system resource burden is reduced, thereby greatly improving the applicability of the system.

[0303] Obviously, various modifications and changes can be made to the present application by those skilled in the art without departing from the spirit and scope of the present application. Thus, if these modifications and changes of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and changes.

Claims

1. A fusion intention-aware surrounding vehicle trajectory prediction method, characterized in that, The method comprises the following steps: calling an IA-Informer model; based on the IA-Informer model, performing intention recognition on the surrounding vehicles of the target vehicle and performing trajectory prediction based on the fused intention; the IA-Informer model comprises an input matrix module, a target vehicle intention recognition module, and a trajectory prediction module based on fused intention; the construction steps of the input matrix module comprise: Defining a target vehicle , other vehicles in a scene where the target vehicle is located are defined as ; the forward direction of the vehicle is the positive direction of the y-axis, and the vertical direction of the y-axis is the x-axis; wherein N is the total number of other vehicles; The encoder and decoder inputs of the IA-Informer model are denoted as and respectively; where the encoder input consists of the target vehicle's historical trajectory and interaction information , and the decoder input is the target vehicle's masked trajectory ; Defining the target vehicle The trajectory coordinates at time t are: ； under an observation domain with a length of mT, the historical trajectory of the target vehicle is: ； at time t, the interaction relationship of the target vehicle is: ； wherein K denotes the number of the target vehicle, and denotes the lateral and longitudinal relative distance of the target vehicle to the Kth vehicle, and is the lateral and longitudinal relative speed of the target vehicle to the Kth vehicle. under an observation domain with a length of mT, the interaction information of the target vehicle is: ； the input of the encoder is: ； Input sequence to decoder at time t is: ; wherein, is the length of the historical trajectory; the construction steps of the target vehicle intention recognition module comprise: the intention recognition module is based on the observed trajectory information of the target vehicle and the interaction information with the surrounding vehicles to predict the driving intention at the next time, and the ultimate goal is to make the network learn a function: ； wherein, is the final trained intent prediction model, to predict the result of the intent, i.e. the intent probability vector, the formula is as follows: ； wherein, respectively, are the probabilities of three strategies that the target vehicle can take; the three strategies include: lane keeping, left lane changing, and right lane changing; the context vector F is calculated layer by layer from bottom to top: ； Wherein, W1, W2, W2', W3, W3', W4 are the weights of the full connection layer, relu is the activation function of the feedforward layer, norm, prob, MaxPool, ELU and Conv1d are normalization, probability sparse self-attention layer, maximum pooling layer, exponential linear unit and one-dimensional convolution layer respectively, is the cosine-sine spatiotemporal encoding of the trajectory; the construction steps of the trajectory prediction module based on fused intention comprise: the trajectory prediction module takes the historical trajectory information of the target vehicle and the intention probability vector as input, and outputs the expected trajectory in the next time period nT, which is as follows: ; wherein is an intention probability vector output by the target vehicle intention recognition module; the decoder encodes the intention prediction result and inputs it into the multi-head attention module, and receives the context vector F generated by the encoder and applies it to the second self-attention layer, the process is as follows: wherein, W6, W6', W7 are the weights of the full connection layer; and mha is the multi-head attention layer. 2.The fusion intent-aware peripheral vehicle trajectory prediction method of claim 1, wherein, Further comprising the testing and verifying steps of the IA-Informer model, comprising: preprocessing the data set; based on the preprocessing result, training and simulation verification of the IA-Informer model are performed respectively. 3.The fusion intent-aware peripheral vehicle trajectory prediction method of claim 2, wherein, The preprocessing of the data set comprises: dividing the data set into training set, validation set and test set according to the ratio of 7:1:2; After the completion of the division of the dataset, the time series needs to be input into the model in a certain format and length, by using a sliding window of length to slide the input sequence in turn to obtain samples: ； wherein is a feature sequence, is a 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: ； for the intention recognition module, the label is the driving intention of the vehicle: ； wherein, 0, 1 and 2 respectively represent the keep, left lane change and right lane change intentions of the vehicle; the labeling rule of the intention recognition module is as follows: for the vehicle that is straight driving, all the extracted sequences are marked as 0; the labeling steps of the left lane change vehicle are as follows: calculate the yaw angle of each point on the trajectory: ； wherein, and represent the lateral and longitudinal coordinates of the vehicle at time t; and represent the lateral and longitudinal coordinates of the vehicle at time t+1; Defining a threshold for the heading angle finds a first point B where the yaw angle is less than the threshold as the lane-changing starting point considering that the intention is generated before the actual lane change action, a fixed time length of 1s is calculated from point B to find point C as the intention generation time; the sequence from C to A is defined as 1, and the rest is 0; wherein, A is the lane change critical point.

4. The fusion intent-aware peripheral vehicle trajectory prediction method of claim 3, wherein, Based on the preprocessing result, the IA-Informer model is trained, comprising: the vehicle intention recognition module adopts a multi-classification cross-entropy function to calculate the loss: ； wherein m is the number of samples, k represents the intent type, denotes an individual encoding of the i-th sample label, is the probability that the i-th sample belongs to the k-th intent. for the trajectory prediction model, a negative log likelihood function is used to train and optimize the parameters in the model: ； wherein, is a sequence of historical trajectories of the target vehicle, is an intention predicted by the target vehicle intention recognition module, is a Gaussian distribution of future trajectories predicted by the trajectory prediction output; is a Gaussian distribution probability of a future trajectory G given a driving intention and a sequence of historical trajectories ; is a probability that the intention is given a historical trajectory ; at each time step, the model calculates the loss and updates the weights of all full connection layers and attention layers through the back propagation algorithm to minimize the loss.

5. The fusion intent-aware peripheral vehicle trajectory prediction method of claim 3, wherein, Based on the preprocessing result, the simulation verification of the IA-Informer model is performed, comprising: For the target vehicle intention recognition module, the simulation verification of the IA-Informer model is performed according to the preprocessing result using a first index and a first contrast model; wherein the first index includes: average accuracy , score; the first contrast model includes: an SVM model, an HMM model, an LSTM model, and a Transformer model; For the trajectory prediction module of the fusion intention, the second index and the second contrast model are used to perform simulation verification of the IA-Informer model according to the preprocessing result; wherein the second index includes: root mean square error RMSE, average displacement error ADE, final displacement error FDE; the second contrast model includes: single LSTM model, double LSTM model, bidirectional LSTM model, social LSTM and double Transformer model.

6. A fusion intent-aware circum-vehicle trajectory prediction system, comprising: Comprise: The calling unit is used for calling the IA-Informer model; The prediction unit is used for identifying the intention and predicting the trajectory of the fusion intention of the vehicle based on the IA-Informer model; The IA-Informer model comprises: an input matrix module, a target vehicle intention identification module and a trajectory prediction module of the fusion intention; The construction steps of the input matrix module comprise: Defining a target vehicle , other vehicles in a scene where the target vehicle is located are defined as ; the forward direction of the vehicle is the positive direction of the y-axis, and the vertical direction of the y-axis is the x-axis; wherein N is the total number of other vehicles; The encoder and decoder inputs of the IA-Informer model are denoted as and respectively; wherein the encoder input consists of the target vehicle’s historical trajectory and interaction information , and the decoder input is the target vehicle’s masked trajectory ; Defining the target vehicle The trajectory coordinates at time t are: ； Under the observation domain with the domain length of mT, the historical trajectory of the target vehicle is: ； The interaction relationship of the target vehicle at t time is: ； where K denotes the number of the target vehicle, and denotes the lateral and longitudinal relative distance of the target vehicle to the Kth vehicle, and is the lateral and longitudinal relative speed of the target vehicle to the Kth vehicle; Under the observation domain with the domain length of mT, the interaction information of the target vehicle is: ； The encoder input is: ； Input sequence to decoder at time t is: ; wherein, is the length of the historical trajectory; The construction steps of the target vehicle intention identification module comprise: The intention identification module is based on the observed target vehicle trajectory information and the interaction information with the surrounding vehicles to predict the driving intention at the next time, and the final target is to make the network learn a function: ； wherein, is the final trained intent prediction model, to predict the result of the intent, i.e. the intent probability vector, the formula is as follows: ； wherein, are the probabilities of the three strategies that the target vehicle can take; the three strategies include: lane keeping, left lane changing, and right lane changing; The context vector F is calculated from bottom to top layer by layer: ； Wherein, W1, W2, W2', W3, W3', W4 are the weights of the full connection layer, relu is the activation function of the feedforward layer, norm, prob, MaxPool, ELU and Conv1d are normalization, probability sparse self-attention layer, maximum pooling layer, exponential linear unit and one-dimensional convolution layer respectively, is the cosine-sine spatiotemporal encoding of the trajectory; The construction steps of the trajectory prediction module of the fusion intention comprise: The trajectory prediction module uses the historical trajectory information and the intention probability vector of the target vehicle as input, and outputs the expected trajectory of the next time period nT, which is as follows: ; wherein is an intention probability vector output by the target vehicle intention recognition module; The decoder encodes the intention prediction result and inputs it into the multi-head attention module, and receives the context vector F generated by the encoder and applies it to the second self-attention layer, and the process is as follows: Wherein, W6, W6', W7 are the weight values of the full connection layer; mha is the multi-head attention layer.

Images