A vehicle trajectory prediction method based on vehicle trajectory interaction dynamics

Abstract

The present application relates to a kind of based on vehicle trajectory interaction dynamics of the week vehicle trajectory prediction method, in the vehicle trajectory prediction process, to the time-dependent feature, space-time-dependent feature, space-time interaction feature that can represent the hidden features such as driver driving intent, driving style, driving strategy these are extracted and fused into the final prediction result, so as to more comprehensive comprehensive consideration the interaction between various factors that affect the week vehicle trajectory, make the predicted trajectory more accurate.In addition, since the modular feature extraction method is used, the vehicle predicted trajectory is improved, which is more convenient for the debugging and optimization of the trajectory prediction model, and people can prevent and avoid the vehicle trajectory prediction model from predicting the trajectory of the vehicle in a specific situation, thereby reducing the potential collision hazard of the vehicle on the road.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of trajectory prediction, in particular to the field of surrounding vehicle trajectory prediction, and more particularly to a surrounding vehicle trajectory prediction method based on vehicle trajectory interaction dynamics. BACKGROUND

[0002] In the complex and changeable traffic environment of modern society, as a key component of the automatic driving system and the advanced auxiliary driving system, the surrounding vehicle trajectory prediction can accurately predict the motion trajectory of the surrounding vehicles centered on the ego vehicle on the road, so that the automatic driving system can timely foresee potential dangers and avoid potential collision risks, thereby improving the safety of vehicle driving. Therefore, it is of great significance to accurately predict the surrounding vehicle trajectory on the road.

[0003] The existing surrounding vehicle trajectory prediction methods can be roughly divided into two types, namely the method based on physics and the method based on deep learning algorithm model. The traditional trajectory prediction method based on physics models through dynamics or kinematics formula, so as to realize the prediction of the surrounding vehicle trajectory, but this method is difficult to comprehensively consider various complex factors affecting the surrounding vehicle trajectory in the road traffic environment, and thus the prediction accuracy of the surrounding vehicle trajectory is poor.

[0004] The method based on deep learning algorithm model can more comprehensively consider the interaction between various factors affecting the surrounding vehicle trajectory under complex road conditions, and can learn by itself, so that the prediction result is more accurate compared with the traditional trajectory prediction method.

[0005] However, due to the characteristics of the deep learning algorithm model, people cannot fully understand the data changes in the algorithm model during the prediction process, which leads to the problem of weak explainability of the final output prediction result. This problem undoubtedly increases the difficulty of debugging and optimizing the algorithm model itself, thereby affecting the accuracy of the surrounding vehicle trajectory prediction. In addition, this problem also makes it difficult for users to accurately judge the prediction result output by the algorithm model, so that people lack confidence in the accuracy of the prediction result output by the algorithm model. At the same time, due to the inability to fully understand the data changes in the algorithm model during the prediction process, it is difficult for users to fully predict and prevent the algorithm model from making wrong predictions of the surrounding vehicle trajectory in certain situations, thereby increasing the potential safety hazards in road traffic. SUMMARY

[0006] The present application provides a surrounding vehicle trajectory prediction method based on vehicle trajectory interaction dynamics to overcome the shortcomings of the prior art.

[0007] The application is realized by the technical scheme, and provides a surrounding vehicle trajectory prediction method based on vehicle trajectory interaction dynamics.

[0008] Step one: collect state information of all surrounding vehicles by using a sensor, and pre-process the collected information to obtain original time feature sets of surrounding vehicles f and g from t-T time to t time and original space interaction feature sets of surrounding vehicles f and g from t-T time to t time

[0009] Step two: establish a time-dependent feature extraction module and a space-time-dependent feature extraction module, input the original time feature sets into the time-dependent feature extraction module to obtain time-dependent feature sets input the original space interaction feature sets into the space-time-dependent feature extraction module to obtain space-time-dependent feature sets

[0010] Step three: establish a space-time interaction feature extraction module, input the original time feature sets original space interaction feature sets time-dependent feature sets and space-time-dependent feature sets into the space-time interaction feature extraction module to obtain space-time interaction feature sets

[0011] Step four: splice the space-time interaction feature sets time-dependent feature sets space-time-dependent feature sets to obtain fusion features;

[0012] Step five: input the fusion features into a trajectory decoder to obtain a predicted trajectory of the surrounding vehicle f;

[0013] Step six: repeat steps one to five to obtain predicted trajectories of all surrounding vehicles.

[0014] 10、Further, in the original time feature sets the original time feature of the surrounding vehicle f at t time is the original time feature is composed of a longitudinal coordinate x, a horizontal coordinate y, a speed v, an acceleration a and a heading angle γ of the surrounding vehicle f at t time;

[0015] in the original space interaction feature sets the original space interaction feature of the surrounding vehicles f and g at t time is The original space interaction feature Composed of the relative distance d, relative speed Δv, relative acceleration Δa of the surrounding vehicles f and g at time t.

[0016] Compared with the prior art, the present application has the following beneficial effects:

[0017] Since the predicted trajectory of the vehicle is based on the original time feature, the original space interaction feature, and the time-dependent feature, the space-time-dependent feature, and the space-time interaction feature, and the time-dependent feature, the space-time-dependent feature, and the space-time interaction feature are extracted separately by the time-dependent feature extraction module, the space-time-dependent feature extraction module, and the space-time interaction feature extraction module, therefore, through this modular design, it is beneficial for people to master the data change in the feature extraction process, thereby improving the explainability of the vehicle trajectory prediction result;

[0018] At the same time, based on the understanding of the vehicle trajectory interaction dynamics, the influencing factors of the surrounding vehicle trajectory prediction not only include the original time features of the vehicle coordinates, speed, acceleration, and heading angle of the surrounding vehicle f, and the original space interaction features of the relative distance, relative speed, and relative acceleration between the surrounding vehicles f and g, but also include higher-dimensional time-dependent characteristics, space-time-dependent characteristics, and space-time interaction characteristics, wherein the time-dependent feature can represent the driver's driving intention, driving style, and driving strategy of the surrounding vehicle f, the space-time-dependent characteristic can represent the change of the interaction strategy between the surrounding vehicles f and g with time, and the space-time interaction characteristic can represent the motion state of the surrounding vehicle f under the comprehensive influence of the driver's driving intention, driving style, and driving strategy and the interaction strategy of the surrounding vehicle g, and simultaneously coupling the interaction strategy between the surrounding vehicle f and other surrounding vehicles other than the surrounding vehicles f and g.

[0019] And the present application can extract representative time-dependent features, space-time-dependent features, and space-time interaction features from the time-dependent characteristics, space-time-dependent characteristics, and space-time interaction characteristics of the surrounding vehicle trajectory in three dimensions based on the original time features and original space interaction features, and fuse the three features to obtain the final predicted trajectory of the surrounding vehicle, thereby greatly improving the final trajectory prediction accuracy.

[0020] Further, the method for obtaining the time-dependent feature set by the time-dependent feature extraction module is:

[0021] ​An LSTM network layer is arranged in the time-dependent feature extraction module;

[0022] The original time feature set is input into the LSTM network layer in the time-dependent feature extraction module, and based on a first formula, the hidden features of the original time feature set are obtained.

[0023] The hidden features of the original time feature set are obtained. Based on a second formula, a time-dependent feature set is obtained.

[0024] The first formula is:

[0025]

[0026] Where c0 is the initial state of the LSTM network layer, h0 is the initial state of the LSTM network layer hidden state, and c0 and h0 are randomly set.

[0027] The second formula is:

[0028]

[0029] Where LeakyReLU(*) is a nonlinear activation function, MaxPool(*) is a maximum pooling function, and MultiheadAtt(*) is an attention mechanism function.

[0030] Further, the method for obtaining the space-time dependent features through the space-time dependent feature extraction module is:

[0031] An LSTM network layer is arranged in the space-time dependent feature extraction module.

[0032] The original space interaction feature set is input into the LSTM network layer in the space-time dependent feature extraction module, and based on a third formula, the hidden features of the original space interaction feature set are obtained.

[0033] The hidden features of the original space interaction feature set are obtained. Based on a fourth formula, a space-time dependent feature set is obtained.

[0034] The third formula is:

[0035]

[0036] Wherein c0 is the initial state of the LSTM network layer, h0 is the initial state of the LSTM network layer hidden state, c0, h0 are randomly set;

[0037] The fourth formula is:

[0038]

[0039] Wherein, LeakyReLU(*) is a nonlinear activation function, MaxPool(*) is a maximum pooling function, and MultiheadAtt(*) is an attention mechanism function.

[0040] The beneficial effects of the above step are: since the original time features and the original space interaction features of the surrounding vehicle f are time-varying sequences, that is, these features change with time, and the LSTM network layer is a typical time sequence feature extraction network, therefore, the LSTM network layers respectively arranged in the time-dependent feature extraction module and the space-time-dependent feature extraction module can effectively capture the time-dependent features and the space-time-dependent features from the original time features and the original space interaction features. Meanwhile, as a variant of recurrent neural network, although the LSTM network layer has improved in dealing with long-term dependencies, it still cannot completely solve the problem of long-term dependencies and gradient instability, therefore, by adding the attention mechanism, the extraction ability and accuracy of the time-dependent features and the space-time-dependent features can be improved. In addition, through the maximum pooling function, redundant features in the time-dependent features and the space-time-dependent features can be effectively removed to prevent overfitting.

[0041] Further, the method for obtaining the space-time interaction features through the space-time interaction feature extraction module is:

[0042] The original time features of the surrounding vehicle f The original space interaction features The time-dependent feature set The space-time-dependent feature set The surrounding vehicle f is input into the space-time interaction feature extraction module, the surrounding vehicle f is taken as a node f, and an interactive space-time dynamic graph of the node f at time t is constructed, the interactive space-time dynamic graph includes node features and edge features;

[0043] A graph convolutional neural network model is arranged in the space-time interaction feature extraction module;

[0044] Based on the graph convolutional neural network model, the node features and the edge features in the interactive space-time dynamic graph are aggregated respectively to obtain node information aggregation features and edge information aggregation features, and the node information aggregation features and the edge information aggregation features are fused to obtain total information aggregation features of the node f at time t;

[0045] In the spatio-temporal interaction feature extraction module, all total information aggregation features of the node f from the t-T moment to the t moment are updated to obtain a set of spatio-temporal interaction features of the node f from the t-T moment to the t moment

[0046] Further, the construction method of the interactive spatio-temporal dynamic graph is:

[0047] initializing the interactive spatio-temporal dynamic graph;

[0048] concatenating the original time feature and the set of time-dependent features to obtain node features that is At the same time, the original spatial interaction feature is concatenated with the set of spatio-temporal dependent features to obtain edge features that is combining the node features and the edge features to create an adjacency matrix and a mask matrix, and to construct an interactive spatio-temporal dynamic graph.

[0049] Further, the aggregation method of the node information aggregation feature and the edge information aggregation feature is:

[0050] At the t moment, the node feature set of all neighbor nodes of the node f is obtained by using an information transmission mechanism and the edge feature set between the node f and all neighbor nodes Based on the fifth formula, the node information aggregation feature is obtained Based on the sixth formula, the edge information aggregation feature is obtained

[0051] The fifth formula is: wherein, DenseGCN(*) is the graph convolutional neural network model;

[0052] The sixth formula is: wherein, EdgeDNN(*) is another graph convolutional neural network model.

[0053] The beneficial effects of the above step are: first, by aggregating the node features, the influence of the space-time dependent features on the time-dependent features can be further mined. Since the node features contain high-dimensional time-dependent features, directly using ordinary graph convolutional neural networks has the problems of gradient instability and limited receptive field, therefore, the DenseGCN graph convolutional neural network model is used to aggregate the node features. The DenseGCN graph convolutional neural network model is an improved model of the graph convolutional neural network, which solves the problem of gradient instability by adding residual connections between the input and output of each layer, and solves the problem of limited receptive field by using extended convolution instead of ordinary convolution.

[0054] Secondly, by aggregating the edge features, the influence of the time-dependent features on the space-time dependent features can be further mined. By using the EdgeCNN network layer, both local information in the scene and global spatial information can be extracted, and the network model is a differentiable model, which is beneficial to ensure the differentiability of the edge information aggregation features, and further ensure the smoothness and authenticity of the prediction results.

[0055] Further, the method for obtaining the total information aggregation feature of the node f at time t is based on the seventh formula, and the node information aggregation feature is Edge information aggregation feature The total information aggregation feature of the node f at time t is obtained by using the max-pooling function and the nonlinear activation function

[0056] The seventh formula is: Wherein, ReLU(*) is the nonlinear activation function, and MaxPool(*) is the max-pooling function.

[0057] 11. Further, the method for obtaining the space-time interaction feature set by updating the total information aggregation feature is: setting a graph attention mechanism network model in the space-time interaction feature extraction module, updating all total information aggregation features of the node f from time t-T to time t by using the graph attention mechanism network based on the eighth formula, and obtaining the space-time interaction feature

[0058] The eighth formula is: Wherein, The total information aggregation feature set of the node f from time t-T to time t is GAT(*) is the graph attention mechanism network model.

[0059] The beneficial effect of the above step is that when the node information aggregation feature and the edge information aggregation feature are spliced and fused, dimension reduction is realized through the max-pooling function, which can avoid feature redundancy. Through the graph attention mechanism network model, the total information aggregation feature can be updated according to the different contribution degrees of the total information aggregation features at different time points, so as to obtain the final spatio-temporal interaction feature.

[0060] The beneficial effect of the present application is that compared with the traditional trajectory prediction method, in the trajectory prediction method of the present application, in addition to considering the original time features such as the coordinates, speed, acceleration and heading angle of the surrounding vehicles, and the original spatial interaction features such as the relative distance, relative speed and relative acceleration of the vehicles, further consideration is given to the time-dependent features, spatio-temporal dependent features and spatio-temporal interaction features which can represent the driver's driving intention, driving style, driving strategy and the interaction strategy between multiple surrounding vehicles on a higher dimension, so that the interaction between various explicit and implicit factors affecting the trajectories of surrounding vehicles in a complex road traffic environment can be comprehensively considered, and the trajectories of surrounding vehicles can be more accurately predicted. In addition, since a modular feature extraction method is used, it is more beneficial for people to fully understand and master the data analysis and calculation process during the prediction of the trajectories of surrounding vehicles, thereby improving the explainability of the predicted trajectories of the vehicles, and facilitating the debugging and optimization of the vehicle trajectory prediction model, and at the same time, people can fully understand and master the vehicle trajectory prediction process to prevent and avoid the vehicle trajectory prediction model from making wrong predictions of the trajectories of surrounding vehicles under certain conditions, thereby reducing the potential collision risks of vehicles on the road. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 The flowchart of the vehicle trajectory prediction method of the present application; DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be described below in a clear and complete manner. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0063] Embodiment 1:

[0064] The present embodiment provides a vehicle trajectory interaction dynamics-based surrounding vehicle trajectory prediction method, which specifically comprises the following steps:

[0065] The state information of all surrounding vehicles is collected by using a sensor, and the collected information is preprocessed to obtain a respective original time feature set and an original space interaction feature set of each surrounding vehicle corresponding to a respective time, wherein the original time feature set of vehicle f from time t-T to time t is included in all obtained original time feature sets The original space interaction feature set of surrounding vehicles f and g from time t-T to time t is included in all obtained original space interaction feature sets In the original time feature set The original time feature corresponding to vehicle f at time t is composed of the longitudinal coordinate x, the horizontal coordinate y, the speed v, the acceleration a, and the heading angle γ of vehicle f at time t In the original space interaction feature set The original space interaction feature corresponding to vehicle f and vehicle g at time t is composed of the relative distance d, the relative speed Δv, and the relative acceleration Δa between vehicle f and vehicle g at time t

[0066] A time-dependent feature extraction module is established, and an LSTM network layer is set in the time-dependent feature extraction module. The original time feature set is input into the time-dependent feature extraction module. Based on the first formula, the original time feature is processed by LSTM to obtain the hidden feature of the original time feature set The hidden feature Based on the second formula, after processing by the attention mechanism and the max-pooling function, the time-dependent feature set is obtained, wherein the hidden feature is copied to three before being processed by the attention mechanism, and the three hidden features are respectively input into the attention mechanism as query, key, and value for processing. The processed result is processed by the max-pooling function to remove redundant features, and the final extraction of the time-dependent feature set is completed.

[0067] A space-time-dependent feature extraction module is established, and an LSTM network layer is set in the space-time-dependent feature extraction module. The original space interaction feature set is input into the space-time-dependent feature extraction module. Based on the third formula, the original space interaction feature is processed by LSTM to obtain the hidden feature of the original space interaction feature set The hidden feature Based on the fourth formula, after processing by the attention mechanism and the max-pooling function, the space-time-dependent feature set is obtained.​​​ Hidden features Before being processed through the attention mechanism, it self-replicates into three, three implicit features. The query, key, and value are respectively input into the attention mechanism for processing. The processed results are then subjected to max pooling to remove redundant features, thus completing the spatiotemporal dependent feature set. The final extraction;

[0068] A spatiotemporal interaction feature extraction module is established, and a graph convolutional neural network model is set up in the module to initialize the spatiotemporal dynamic graph of the interaction. Then, vehicle f is used as node f in the spatiotemporal dynamic graph of the interaction, and the original time features of vehicle f at time t are extracted. Time-dependent feature set of vehicle v The concatenation is performed, and the concatenated result is used as the node feature of node f at time t. Right now Simultaneously, vehicle g is treated as node g in the interactive spatiotemporal dynamic graph, and the original spatial interaction characteristics of vehicle f and vehicle g at time t are analyzed. and spatiotemporal dependent feature set Perform concatenation, and use the concatenated result as the edge feature between node f and node g. Right now Combining node features With edge features Create an adjacency matrix and a mask matrix to construct an interactive spatiotemporal dynamic graph at time t;

[0069] At time t, using the information transmission mechanism, node f can obtain the node feature set of all neighboring nodes in the interactive spatiotemporal dynamic graph. and the set of edge features between node v and all its neighboring nodes. Based on the fifth formula, the node characteristics of node f at time t are... The set of node features with all neighboring nodes Aggregation is performed to obtain the node information aggregation feature of node f at time t. Based on the sixth formula, the edge feature set of node f at time t The set of edge features between node v and all its neighboring nodes Aggregation is performed to obtain the aggregated edge information features of node f at time t.

[0070] Based on the seventh formula, the node information aggregation characteristics of node f at time t. Features of edge information aggregation By fusing max pooling and nonlinear activation functions, the total information aggregation feature of node f at time t is obtained.

[0071] In the space-time interaction feature extraction module, a graph attention mechanism network model is arranged to aggregate the total information feature set of node f from all interaction space-time dynamic graphs from t-T time to t time The input graph attention mechanism network model obtains the space-time interaction feature of node f based on the eighth formula

[0072] The space-time interaction feature set The time-dependent feature set The space-time-dependent feature set The space-time-dependent feature set

[0073] The final fusion feature is input into the decoder to output the predicted trajectory of vehicle f;

[0074] The above steps are repeated to obtain the predicted trajectories of all surrounding vehicles;

[0075] The first formula is: Wherein, LSTM(*) is a long short-term memory unit, the calculation method is the same as that of the standard LSTM model, c0 is the initial state of the LSTM long short-term memory unit, h0 is the initial state of the LSTM long short-term memory unit hidden state, and c0 and h0 are randomly set;

[0076] The second formula is: Wherein, LeakyReLU(*) is a nonlinear activation function, MaxPool(*) is a maximum pooling function, and MultiheadAtt(*) is an attention mechanism function;

[0077] The third formula is: Wherein, LSTM(*) is a long short-term memory unit, the calculation method is the same as that of the standard LSTM model, c0 is the initial state of the LSTM long short-term memory unit, h0 is the initial state of the LSTM long short-term memory unit hidden state, and c0 and h0 are randomly set;

[0078] The fourth formula is: Wherein, LeakyReLU(*) is a nonlinear activation function, MaxPool(*) is a maximum pooling function, and MultiheadAtt(*) is an attention mechanism function;

[0079] The fifth formula is: Wherein, DenseGCN(*) is the graph convolutional neural network model;

[0080] The sixth formula is: wherein EdgeDNN(*) is another graph convolutional neural network model;

[0081] The seventh formula is: wherein ReLU(*) is the nonlinear activation function, and MaxPool(*) is the max-pooling function.

[0082] The eighth formula is: wherein GAT(*) is a graph attention mechanism network model.

[0083] The above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting a surrounding vehicle trajectory based on vehicle trajectory interaction dynamics, characterized in that, The method comprises the following steps: Step one: collect the state information of all vehicles in the surrounding area using sensors, and preprocess the collected information to obtain the original time feature set of the surrounding vehicles f from t-T time to t time , and the original space interaction feature set of the surrounding vehicles f and g from t-T time to t time ; Step two: build a time-dependent feature extraction module and a space-time-dependent feature extraction module, input the original time feature set into the time-dependent feature extraction module to obtain a time-dependent feature set , input the original space interaction feature set into the space-time-dependent feature extraction module to obtain a space-time-dependent feature set ; Step three: a space-time interaction feature extraction module is established, and the original time feature set , the original space interaction feature set , the time-dependent feature set , and the space-time-dependent feature set are input into the space-time interaction feature extraction module to obtain a space-time interaction feature set ; Step four: concatenating the spatiotemporal interaction feature set , the time-dependent feature set , the spatiotemporal-dependent feature set to obtain the fusion feature Step five: input the fusion features into a trajectory decoder to obtain a predicted trajectory of the surrounding vehicle f; Step six: repeat steps one to five to obtain predicted trajectories of all surrounding vehicles; The method for obtaining the space-time interaction features by the space-time interaction feature extraction module is: the original time feature of the surrounding vehicle f , the original spatial interaction feature , the time-dependent feature set , the space-time dependent feature set In the space-time interaction feature extraction module, the surrounding vehicle f is taken as a node f, an interaction space-time dynamic graph of the node f at time t is constructed, and the interaction space-time dynamic graph includes node features and edge features; A graph convolutional neural network model is arranged in the space-time interaction feature extraction module; Based on the graph convolutional neural network model, the node features and the edge features in the interactive space-time dynamic graph are aggregated respectively to obtain node information aggregation features and edge information aggregation features, and the node information aggregation features and the edge information aggregation features are fused to obtain total information aggregation features of the node f at the time t; In the spatiotemporal interaction feature extraction module, all total information aggregation features of the node f from a t-T moment to a t moment are updated to obtain a spatiotemporal interaction feature set of the node f from the t-T moment to the t moment ; The method for constructing the interactive space-time dynamic graph is: The interactive space-time dynamic graph is initialized; concatenate the original time features with the time-dependent feature set to obtain node features , i.e. [ , ] , and concatenate the original spatial interaction features with the space-time dependent feature set to obtain edge features , i.e. [ , ] , combine the node features and the edge features to create an adjacency matrix and a mask matrix, and construct an interaction space-time dynamic graph; The node information aggregation feature , edge information aggregation feature The aggregation method is: At time t, using an information transmission mechanism, obtain a node feature set of all neighbor nodes of the node f and an edge feature set between the node f and all neighbor nodes , based on a fifth formula, obtain node information aggregation features , based on a sixth formula, obtain edge information aggregation features ; The fifth formula is: = DenseGCN( , ), where DenseGCN( ) is the graph convolutional neural network model; The sixth formula is: = EdgeCNN( , ), wherein EdgeDNN( ) is another of the graph convolutional neural network models. The method for obtaining the total information aggregation feature of the node f at time t is: based on a seventh formula, the node information aggregation feature , edge information aggregation feature The total information aggregation feature of the node f at time t is obtained by using a max-pooling function and a nonlinear activation function ; The seventh formula is: = ReLU(MaxPool([ , ])), wherein ReLU( ) is the nonlinear activation function, and MaxPool( ) is the max-pooling function. Through the update of the total information aggregation feature, a method for obtaining the spatiotemporal interaction feature set is: setting a graph attention mechanism network model in the spatiotemporal interaction feature extraction module, updating all total information aggregation features of node f from t-T time to t time based on the eighth formula by using the graph attention mechanism network, and obtaining the spatiotemporal interaction feature ; The eighth formula is: =GAT( ), wherein, is a total information aggregation feature set of node f from t-T time to t time, GAT( ) is a graph attention mechanism network model.

2. The method of claim 1, wherein, The original time feature set of the surrounding vehicle f at the time t is The original time feature of the surrounding vehicle f at the time t is The original time feature The original time feature of the surrounding vehicle f at the time t is composed of the longitudinal coordinate x, the lateral coordinate y, the speed v, the acceleration a, and the heading angle γ of the surrounding vehicle f at the time t. The original space interaction feature set In the original space interaction feature of the surrounding vehicles f and g at time t is The original space interaction feature Composed of the relative distance d, the relative speed , and the relative acceleration of the surrounding vehicles f and g at time t.

3. The method of claim 1, wherein, The time-dependent feature set is obtained by a time-dependent feature extraction module The method is as follows: An LSTM network layer is arranged in the time-dependent feature extraction module; The original time feature set is input into an LSTM network layer in a time-dependent feature extraction module, and based on a first formula, the hidden features of the original time feature set are obtained ; The original time feature set The implicit feature Based on the second formula, obtain the time-dependent feature set ; The first formula is: = LSTM( , , ), wherein is the initial state of the LSTM network layer, is the initial state of the LSTM network layer hidden state, , are randomly set; The second formula is: = LeakyReLU(MaxPool(MultiheadAtt( ))) where LeakyReLU( ) is a nonlinear activation function, MaxPool( ) is a max-pooling function, and MultiheadAtt( ) is an attention mechanism function.

4. The method of claim 1, wherein, The method for obtaining the space-time dependent features by the space-time dependent feature extraction module is: An LSTM network layer is arranged in the space-time dependent feature extraction module; The original spatial interaction feature set is input into an LSTM network layer in the space-time dependent feature extraction module, and based on a third formula, the original spatial interaction feature set is obtained ; The original spatial interaction feature set The implicit feature Based on the fourth formula, obtain the space-time dependent feature set ; The third formula is: = LSTM( , , ), wherein is the initial state of the LSTM network layer, is the initial state of the LSTM network layer hidden state, , are randomly set; The fourth formula is: = LeakyReLU(MaxPool(MultiheadAtt( ))), where LeakyReLU( ) is a nonlinear activation function, MaxPool( ) is a max-pooling function, and MultiheadAtt( ) is an attention mechanism function.

Images