Vehicle trajectory prediction method considering step-by-step bidirectional interaction

Abstract

According to the vehicle trajectory prediction method considering gradual bidirectional interaction provided by the invention, the network model capable of fully extracting bidirectional interaction information between vehicles is constructed, the trajectory and spatial position information of the predicted vehicle and other surrounding vehicles are extracted step by step as input, motion coding is performed on the vehicle trajectory by using a coding architecture, and the vehicle trajectory prediction accuracy is improved. According to the method, the future trajectory of the vehicle is accurately predicted in combination with the gradual bidirectional interaction relationship between the vehicles, the complex mutual influence between the vehicles is fully considered, different observation tensors are constructed for the predicted vehicle and the surrounding vehicles according to the respective conditions, and the space information between the vehicles is extracted by using a convolutional pooling network. According to the method, interaction information between different vehicles is reasonably processed, finally, the extracted interaction information is input into a decoder, the future trajectory of the predicted vehicle is obtained in multiple segments, the influence of the future trajectory of the predicted vehicle and the future trajectory of surrounding vehicles on each other is fully considered, and the trajectory prediction precision is improved.

Description

Technical Field

[0001] The present invention belongs to the field of automobile trajectory prediction and provides a vehicle trajectory prediction method considering step-by-step two-way interaction; specifically, a model design that predicts segmented predictions and integrates vehicle future trajectory interactions is used to predict the future trajectories of other vehicles around a main vehicle in a driving environment. Background Art

[0002] When it comes to the intelligence and automation of modern transportation systems, vehicle trajectory prediction is one of the key research areas. However, due to the influence of technology, society and ethics, the coexistence of autonomous and manually driven cars on the road will become the norm for a long time in the future. In traffic scenarios, vehicle spatial interaction is very important. Vehicles are not just isolated entities. They often need to interact with other vehicles around them on the road. These interactions not only affect the driving path of each vehicle, but also directly determine the efficiency and safety of the overall transportation system. The core of two-way interaction lies in understanding and predicting the mutual influence between traffic participants and their impact on behavior. For example, when a vehicle intends to change lanes, it needs to take into account the speed and position of other vehicles in the current lane, as well as whether there is enough space in the adjacent lane. This multi-faceted spatial interaction requires an accurate prediction model to analyze so that the model can accurately predict the trajectory of the target vehicle. Therefore, how to ensure the accuracy of the prediction of the future trajectory of the predicted vehicle under the condition that the predicted vehicle continues to interact with the surrounding vehicles is a research difficulty that needs to be solved urgently. Summary of the invention

[0003] In order to achieve accurate prediction of vehicle trajectory when a predicted vehicle continuously interacts with surrounding vehicles, the present invention provides a vehicle trajectory prediction method taking into account step-by-step two-way interaction.

[0004] The present invention is achieved by adopting the following technical solutions:

[0005] The present invention provides a vehicle trajectory prediction method considering step-by-step two-way interaction, which fully extracts the two-way interaction information between vehicles by gradually extracting the trajectory and spatial position information of the predicted vehicle and other surrounding vehicles and inputting them into the established network model, and accurately predicts the future trajectory of the vehicle in multiple segments. The method is characterized in that the specific steps of the method are as follows:

[0006] Step 1: Process the vehicle trajectory dataset:

[0007] First, the position information of the vehicles in the data set is extracted in chronological order, that is, the absolute coordinates (x, y) of the horizontal and vertical positions of the vehicles. Then, all vehicle coordinates are further processed, including vehicle trajectory extraction, observation tensor construction, and relative coordinate conversion between vehicles.

[0008] In vehicle trajectory extraction, the absolute coordinates of the predicted vehicle’s historical trajectory TargHist and the absolute coordinates of the future real trajectory TargFut are defined as follows:

[0009]

[0010] in, is the absolute horizontal and vertical coordinates of the predicted vehicle at time t, is the absolute lateral coordinate of the predicted vehicle at time t, is the absolute longitudinal coordinate of the predicted vehicle at time t, t∈{1,2,...,T obs}, T obs is the observation time of the historical trajectory of the predicted vehicle, and is also the current moment, is the real horizontal and vertical absolute coordinates of the predicted vehicle at the future k time, which is used to calculate the loss value of the model. is the absolute lateral coordinate of the predicted vehicle at the next k moments, is the absolute longitudinal coordinate of the predicted vehicle at the kth moment in the future, k∈{1,2,...,T pred}, T pred Predict the duration of the future trajectory of the predicted vehicle;

[0011] The absolute coordinates of the historical trajectories of the vehicles around the predicted vehicle, NbrsHist, and the absolute coordinates of the future real trajectories, NbrsFut, are defined as follows:

[0012]

[0013] Wherein, n∈N is the number of surrounding vehicles of the predicted vehicle, N is a positive integer greater than 0, and the surrounding vehicles of the predicted vehicle are hereinafter referred to as surrounding vehicles; is the set of absolute horizontal and vertical coordinates of n surrounding vehicles at time t. Specifically, is the set of the absolute lateral coordinates of the n surrounding vehicles at time t, is the set of absolute longitudinal coordinates of n surrounding vehicles at time t, t∈{1,2,...,T obs '}, T obs ' is the observation time of the historical trajectory of the surrounding vehicles. Because the model requires that the input predicted vehicle and the historical trajectory observation time of the surrounding vehicles are the same, is the set of the real horizontal and vertical absolute coordinates of the n surrounding vehicles at the next k time. Specifically, is the set of real lateral absolute coordinates of n surrounding vehicles at the next k moments, is the set of real longitudinal absolute coordinates of n surrounding vehicles at the next k time, k∈{1,2,...,T pred '}, Tpred ' is the predicted duration of the future trajectory of the surrounding vehicles. Because the model requires that the output predicted vehicle and the predicted duration of the future trajectory of the surrounding vehicles are the same, T pred '=T pred ;

[0014] In the construction of the observation tensor, the environment within 30 meters in front and behind and 5 meters on the left and right of the predicted vehicle is taken, and a 5×25 two-dimensional space grid is established with the center of mass of the predicted vehicle as the center. The 5 in 5×25 represents the lateral observation range of the predicted vehicle, and the 25 in 5×25 represents the longitudinal observation range of the predicted vehicle. The center of mass of the predicted vehicle is located at the center of the two-dimensional space grid, that is, the 3rd grid in the horizontal direction and the 13th grid in the longitudinal direction. The length of each small square in the two-dimensional space grid is defined as 2.4 meters and the width is 2 meters. The orientation of the surrounding vehicles is recorded in the two-dimensional space grid; in this 5×25 two-dimensional space grid, the small squares with vehicles are set to TRUE, and the small squares without vehicles are set to FALSE. The number of surrounding vehicles is n, and then the observation tensor CentGrid∈R of the surrounding vehicle orientation represented by TRUE and FALSE is obtained. 5×25 , through the observation tensor CentGrid∈R 5×25 Ability to extract spatial interaction features between the predicted vehicle and surrounding vehicles;

[0015] In the conversion of relative coordinates between vehicles, the absolute coordinates of the predicted historical trajectory of the vehicle TargHist and the absolute coordinates of the future real trajectory TargFut are respectively subtracted from the current time T obs The absolute horizontal and vertical coordinates of the predicted vehicle under Get the relative historical trajectory coordinates TargCentHist of the predicted vehicle after conversion and the relative future trajectory coordinates TargCentFut of the predicted vehicle after conversion:

[0016]

[0017]

[0018] Among them, TargCentHist is used for model input, and TargCentFut is used for model calculation loss value. is the relative historical trajectory coordinate of the predicted vehicle at time t after transformation, The absolute horizontal and vertical coordinates of the predicted vehicle at time t, is the current time T obs The absolute horizontal and vertical coordinates of the predicted vehicle under obs}, is the relative future trajectory coordinate of the predicted vehicle at the next k time after transformation, is the converted relative future trajectory abscissa of the predicted vehicle at the next k time, is the relative future trajectory ordinate of the predicted vehicle at the future k time after transformation, k∈{1,2,...,T pred};

[0019] Subtract the absolute coordinates of the historical trajectories NbrsHist and the absolute coordinates of the future trajectories NbrsFut of the surrounding vehicles from the current time T obs The predicted absolute lateral and longitudinal coordinates of the vehicle under Get the relative historical trajectory coordinates NbrsCentHist of the surrounding vehicles after conversion and the relative future trajectory coordinates NbrsCentFut of the surrounding vehicles after conversion:

[0020]

[0021] Among them, NbrsCentHist is used for model input, and NbrsCentFut is used for model calculation loss value. is the relative historical trajectory coordinates of the surrounding vehicles at time t after transformation, t'∈{1,2,...,T obs '}; is the relative future trajectory coordinates of the surrounding vehicles after transformation at the future k' time, is the relative future trajectory abscissa of the surrounding vehicles after transformation at the future k' time, The relative future trajectory ordinate of the transformed surrounding vehicles at the future k' time, k'∈{1,2,...,T pred '};

[0022] Step 2: Design a vehicle trajectory prediction model

[0023] The trajectory prediction model includes the predicted vehicle trajectory prediction model TargModel and the surrounding vehicle trajectory prediction model NbrsModel. The predicted vehicle trajectory prediction model TargMode and the surrounding vehicle trajectory prediction model NbrsModel have the same model structure, but different parameters in the model. The predicted vehicle trajectory prediction model TargMode and the surrounding vehicle trajectory prediction model NbrsModel both include a vehicle motion trajectory information extraction module and a vehicle trajectory prediction module:

[0024] For the vehicle motion trajectory information extraction module, first convert the predicted vehicle relative historical trajectory coordinates in step 1 Relative historical trajectory coordinates of surrounding vehicles The dimensions are transformed into and Will Input into the one-dimensional convolutional neural network Conv1D3 to extract the motion information of the predicted vehicle. Input to another one-dimensional convolutional neural network Conv1D3' to extract the motion information of surrounding vehicles and obtain the motion embedding TargCentHist of the predicted vehicle emb The motion of the surrounding vehicles is embedded in NbrsCentHist emb :

[0025]

[0026] in, Encode information for the predicted vehicle's motion, To encode the motion information of surrounding vehicles, Conv1D3 and Conv1D3' are both one-dimensional convolutional layers with a convolution kernel size of 3, 2 input channels, 32 output features, and a step size of 1. is the weight parameter of Conv1D3, is the weight parameter of Conv1D3', is the bias parameter of Conv1D3, is the bias parameter of Conv1D3';

[0027] The motion encoding information of the vehicle to be predicted The dimension is converted to Encode the motion information of surrounding vehicles The dimension is converted to Then input it into the encoder to get the encoding information of the predicted vehicle and coded information of surrounding vehicles

[0028]

[0029] in, is the hidden state of the predicted vehicle, is the hidden state of the surrounding vehicles, Encoder is the motion encoder of the predicted vehicle, Encoder' is the motion encoder of the surrounding vehicles, and the encoder is composed of a long short-term memory network, namely, an LSTM network. The number of input features of the LSTM network in the encoder is 32, the number of hidden features is 64, and the number of layers is 1. is the weight parameter of Encoder, is the weight parameter of Encoder', is the bias parameter of Encoder, is the bias parameter of Encoder';

[0030] The motion encoding information of the vehicle to be predicted Aggregate the encoded information through the fully connected layer:

[0031]

[0032] in, is the motion information encoding of the predicted vehicle with a feature number of 32, φ is the embedding function of the fully connected layer, is the weight parameter of the fully connected layer that aggregates the predicted vehicle encoding information, The bias parameters of the fully connected layer that aggregates the predicted vehicle encoding information;

[0033] Encode the motion information of surrounding vehicles Fill in the observation tensor CentGrid∈R in step 1 according to its position in the grid 5×25 , that is, Fill to CentGrid∈R 5×25 The positions that are TRUE in the grid are filled with 0, while the positions that are False are filled with 0. If there is only one position in the grid that is TRUE, the position relationship of the vehicle in the grid is not considered and the Fill in this position and Fill to CentGrid∈R 5×25 After that, we get the observed feature tensor CentGridH, and then use CentGridH as input and use the two-dimensional graph convolution Conv2D 3×3 To extract the features of the spatial interaction information between the predicted vehicle and its surrounding vehicles:

[0034]

[0035] Among them, NbrsGridConv∈R 64×5×25 Conv2D is the spatial interaction feature between cars. 3×3 It is a two-dimensional convolution layer with a convolution kernel size of 3×3, the number of input features and the number of output features are both 64, and the step size is 1. For Conv2D 3×3 The weight parameter, For Conv2D 3×3 The bias parameters of

[0036] The spatial interaction feature NbrsGridConv between vehicles is pooled to reduce the dimension, retain key information and improve the generalization of the model:

[0037]

[0038] Where NbrsGridPool∈R 64×9×1, MaxPool 3×3 is a maximum pooling layer with a pooling kernel of 3×3 and a stride of 2. is the weight parameter, is the bias parameter;

[0039] In order to further enhance the prediction network's ability to understand and model spatial feature data, NbrsGridPool is further reconvolved and repooled. That is, NbrsGridPool is input into a two-dimensional convolution layer with a convolution kernel of 3×1, and then input into a maximum pooling layer with a pooling kernel of 2×1. Finally, the spatial interaction information between vehicles is obtained. And transform the dimension into

[0040] The vehicle trajectory prediction module predicts the future time length T by obtaining vehicle motion information and inter-vehicle space information pred The specific operation is as follows: Encode the motion information of the predicted vehicle Spatial interaction information between vehicles Splice and copy T pred The vehicle information aggregation feature with a feature number of 112 is obtained. The aggregate feature Input into the trajectory prediction module to obtain the future predicted trajectory Pred of the predicted vehicle:

[0041]

[0042] In the formula,

[0043] is the predicted vehicle at T obs+m The relative predicted trajectory coordinates at time, is the predicted vehicle at T obs+m The horizontal coordinate of the relative predicted trajectory at time, is the predicted vehicle at T obs+m The relative predicted trajectory ordinate at time, m∈{1,2,...,T pred}, Decoder is a decoder, which is composed of a long short-term memory network LSTM network. The number of input features of the LSTM network in the decoder is 112, the number of hidden features is 128, and the number of layers is 1. is the weight parameter of the decoder, is the bias parameter of the decoder, φ is the embedding function of the fully connected layer, is the weight parameter of the fully connected layer, is the bias parameter of the fully connected layer, the input feature number is 128, and the output dimension is 2;

[0044] Step 3: Construct loss function

[0045] The expression of the loss function Loss is:

[0046]

[0047] in, is the relative real trajectory abscissa of the predicted vehicle at time p, is the relative true trajectory ordinate of the predicted vehicle at time p, is the horizontal coordinate of the relative predicted trajectory of the predicted vehicle at time p output by the model, The ordinate of the relative predicted trajectory of the predicted vehicle at time p output by the model;

[0048] Step 4: Step-by-step two-way interactive prediction trajectory

[0049] The predicted vehicle is denoted as Car targ , set the prediction Car targ The duration of the future trajectory is T, and the duration of the predicted future trajectory of the surrounding vehicles is T', because the model requires the input of the predicted future trajectory of the surrounding vehicles to be Car targ half of Car targ The historical observation time is T obs , the historical observation time of surrounding vehicles is T obs ', because the model requires the input Car targ The same as the historical observation time of the surrounding vehicles, so T obs '=T obs ;

[0050] The model predicts Car in two stages targ Future trajectory:

[0051] In the first stage, the model outputs Car targ future The relative predicted trajectory in the time domain and the relative predicted trajectory of surrounding vehicles in the future 0~T' domain;

[0052] In the second stage, the model outputs Car targ future Relative predicted trajectory in the time domain;

[0053] Finally, the prediction time domain is and The relative prediction trajectories of the two predicted vehicles are spliced ​​in chronological order to obtain Car targ The complete future relative prediction trajectory of duration T;

[0054] First, obtain the first predicted trajectory of the predicted vehicle:

[0055] Car targ The observation tensor CentGrid∈R 5×25 、Car targ The relative historical trajectory coordinates TargCentHist of the vehicle and the relative historical trajectory coordinates NbrsCentHist of the surrounding vehicles are input into the predicted vehicle trajectory prediction model TargModel to obtain Car targ future Prediction trajectory in time domain TargPred:

[0056]

[0057] in, For Car targ In the T obs +a predicted trajectory coordinates at time, For Car targ In the T obs +a is the predicted horizontal coordinate at time point, For Car targ In the T obs +a The predicted ordinate at time,

[0058] Next, for Car targ Relative coordinate transformation with surrounding vehicles:

[0059] Select all surrounding vehicles in turn, and record the selected surrounding vehicles as Car g , where g∈N represents the gth surrounding vehicle, let Car g The original relative historical trajectory coordinates SelHist are:

[0060]

[0061] in, For Car g The original relative historical trajectory coordinates at time b, b∈{1,2,...,T obs'}, Car g The current time T obs 'Location below Set it as the origin of coordinates and establish a new reference coordinate system. g The historical trajectory coordinates of Car are converted into the coordinates in the new reference coordinate system, and the g The new relative history track coordinates TurnHist:

[0062]

[0063] Cartarg The original relative historical trajectory coordinates are converted into the new relative historical trajectory coordinates TargNewHist in the new reference coordinate system:

[0064]

[0065] Among them, TargCentHist is Car targ The original relative historical trajectory coordinates;

[0066] Then, get Car g The observation tensor SelCentGrid:

[0067] In Car targ The observation tensor Centgrid∈R 5×25 Based on the grid Car g The small square at the location is set to TRUE, and the contents of the other small squares are all set to FALSE, and Car g The observation tensor SelCentGrid∈R 5×25 ;

[0068] Next, select all surrounding vehicles in turn and obtain their next predicted trajectories:

[0069] Car targ The number of surrounding vehicles is n, and a surrounding car Car is selected in turn each time g , and perform the surrounding vehicle trajectory prediction process:

[0070] The Car obtained in the previous step g The observation tensor SelCentGrid∈R 5×25 、Car g The new relative historical trajectory coordinates TurnHist, Car targ The new relative historical trajectory coordinates TargNewHist are input into the surrounding vehicle trajectory prediction model NbrsModel to obtain Car g The predicted trajectory in the future 0~T' time domain with its own center of mass as the origin of the reference system;

[0071] After the above surrounding vehicle trajectory prediction process is performed n' times, the predicted trajectory coordinates SelNbrsPred of all surrounding vehicles in the new reference system with their respective center of mass as the reference system origin can be obtained. Because the surrounding vehicle trajectory prediction process needs to be performed as many times as there are vehicles, so n'=n:

[0072]

[0073] in, is the predicted trajectory coordinates of the surrounding vehicles at the future time c, is the predicted trajectory abscissa of the surrounding vehicles at time c in the future, is the predicted trajectory ordinate of the surrounding vehicles at time c in the future, c∈{1,2,...,T'};

[0074] Next, convert SelNbrsPred to Car targ is the predicted trajectory coordinate of the reference system:

[0075] SelNbrsPred is based on Car g The trajectory coordinates in the reference system of the origin center are converted into Car targ The coordinates in the reference system of Add them together and you get Car targ The next predicted trajectory coordinates of the surrounding vehicles in the reference frame NbrsPred:

[0076]

[0077] in, For Car g In T obs 'The original relative historical trajectory coordinates at the moment;

[0078] Finally, the next predicted trajectory coordinates of the predicted vehicle and surrounding vehicles are replaced into their respective old historical trajectories to obtain the new historical trajectory coordinates:

[0079] Discard the old history track TargCentHist track points, add TargPred to TargCentHist as the latest The trajectory points at the moment are updated to obtain the new historical trajectory TargRepHist:

[0080]

[0081] in, For the latest The predicted trajectory points of the predicted vehicle at the moment, is the new historical trajectory of the predicted vehicle at time d,

[0082] And the latest trajectory point of the predicted vehicle As the new reference system origin and establish a new reference coordinate system, the new relative historical trajectory TargCentHist2 of the predicted vehicle is obtained:

[0083]

[0084] in, is the new relative historical trajectory coordinate of the predicted vehicle at time e in the new reference coordinate system, and

[0085] Similarly, the predicted trajectory NbrsPred of the surrounding vehicles is used to replace the first 0 to T' trajectory points in the old historical trajectory NbrsCentHist to obtain the updated new historical trajectory NbrsRepHist:

[0086]

[0087] in, is the predicted trajectory point of the surrounding vehicles at the latest T' moment, is the new historical trajectory of the surrounding vehicles at time f, f∈{T'+1,T'+2...,T obs '-1,T obs ',T obs '+1,...,T obs '+T'};

[0088] And the latest trajectory point of the predicted vehicle As the new origin coordinates, we get the new relative historical trajectory NbrsCentHist2 of the surrounding vehicles in the new reference coordinate system:

[0089]

[0090] in, For surrounding vehicles is the new relative historical trajectory coordinate of the origin at time l, and l∈[T'+1,T'+2,...,T obs '+T'-1,T obs '+T'];

[0091] The predicted vehicle's new relative historical trajectory TargCentHist2, the surrounding vehicles' new relative historical trajectory NbrsCentHist2 and the observation tensor CentGrid are input into the predicted vehicle trajectory prediction model TargModel to obtain the predicted vehicle's future trajectory. Prediction trajectory in time domain TargPred2:

[0092]

[0093] in, The predicted vehicle is obs +p time predicted trajectory coordinates, The predicted vehicle is obs +p time prediction horizontal coordinate, The predicted vehicle isobs +The predicted ordinate at time p,

[0094] Finally, the first-stage predicted trajectory TargPred and the second-stage predicted trajectory TargPred2 of the predicted vehicle are concatenated to obtain the complete predicted trajectory Pred of the predicted vehicle:

[0095]

[0096] in, The predicted vehicle is obs +q time predicted trajectory coordinates, The predicted vehicle is obs +q time prediction horizontal coordinate, The predicted vehicle is obs +q time prediction ordinate,

[0097] Compared with the prior art, the present invention has the following beneficial effects:

[0098] The present invention provides a vehicle trajectory prediction method considering step-by-step two-way interaction, constructs a network model that can fully extract two-way interaction information between vehicles, gradually extracts the trajectory and spatial position information of the predicted vehicle and other surrounding vehicles as input, uses a coding architecture to perform motion coding on the vehicle trajectory, and combines the step-by-step two-way interaction relationship between vehicles to accurately predict the future trajectory of the vehicle. In addition, the complex mutual influence between vehicles is fully considered, different observation tensors are constructed for the predicted vehicle and the surrounding vehicles according to their respective situations, and a convolutional pooling network is used to extract spatial information between vehicles. The interaction information between different vehicles is reasonably processed. Finally, the extracted interaction information is input into a decoder, and the future trajectory of the predicted vehicle is obtained in multiple segments, and the influence of the future trajectories of the predicted vehicle and the surrounding vehicles on each other is fully considered, thereby improving the accuracy of trajectory prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] The present invention will be further described below in conjunction with the accompanying drawings:

[0100] Figure 1 An overall flow chart of a vehicle trajectory prediction method considering step-by-step two-way interaction provided by the present invention.

[0101] Figure 2 A schematic diagram of the framework of a vehicle trajectory prediction model in step 2 of a vehicle trajectory prediction method considering step-by-step two-way interaction provided by the present invention. DETAILED DESCRIPTION

[0102] The present invention is described in detail below in conjunction with the accompanying drawings:

[0103] The present invention provides a vehicle trajectory prediction method considering step-by-step two-way interaction, which fully extracts the two-way interaction information between vehicles by gradually extracting the trajectory and spatial position information of the predicted vehicle and other surrounding vehicles and inputting them into the established network model, and accurately predicts the future trajectory of the vehicle in multiple segments. The method is characterized in that the specific steps of the method are as follows:

[0104] Step 1: Process the vehicle trajectory dataset:

[0105] First, the position information of the vehicles in the data set is extracted in chronological order, that is, the absolute coordinates (x, y) of the horizontal and vertical positions of the vehicles. Then, all vehicle coordinates are further processed, including vehicle trajectory extraction, observation tensor construction, and relative coordinate conversion between vehicles.

[0106] In vehicle trajectory extraction, the absolute coordinates of the predicted vehicle’s historical trajectory TargHist and the absolute coordinates of the future real trajectory TargFut are defined as follows:

[0107]

[0108] in, is the absolute horizontal and vertical coordinates of the predicted vehicle at time t, is the absolute lateral coordinate of the predicted vehicle at time t, is the absolute longitudinal coordinate of the predicted vehicle at time t, t∈{1,2,...,T obs}, T obs is the observation time of the historical trajectory of the predicted vehicle, and is also the current moment, is the real horizontal and vertical absolute coordinates of the predicted vehicle at the future k time, which is used to calculate the loss value of the model. is the absolute lateral coordinate of the predicted vehicle at the next k moments, is the absolute longitudinal coordinate of the predicted vehicle at the kth moment in the future, k∈{1,2,...,T pred}, T pred Predict the duration of the future trajectory of the predicted vehicle;

[0109] The absolute coordinates of the historical trajectories of the vehicles around the predicted vehicle, NbrsHist, and the absolute coordinates of the future real trajectories, NbrsFut, are defined as follows:

[0110]

[0111]

[0112] Wherein, n∈N is the number of surrounding vehicles of the predicted vehicle, N is a positive integer greater than 0, and the surrounding vehicles of the predicted vehicle are hereinafter referred to as surrounding vehicles; is the set of absolute horizontal and vertical coordinates of n surrounding vehicles at time t. Specifically, is the set of the absolute lateral coordinates of the n surrounding vehicles at time t, is the set of absolute longitudinal coordinates of n surrounding vehicles at time t, t∈{1,2,...,T obs '}, T obs ' is the observation time of the historical trajectory of the surrounding vehicles. Because the model requires that the observation time of the historical trajectory of the input predicted vehicle and the surrounding vehicles is the same, T obs '=T obs , is the set of the real horizontal and vertical absolute coordinates of the n surrounding vehicles at the next k time. Specifically, is the set of real lateral absolute coordinates of n surrounding vehicles at the next k moments, is the set of real longitudinal absolute coordinates of n surrounding vehicles at the next k time, k∈{1,2,...,T pred '}, T pred ' is the predicted duration of the future trajectory of the surrounding vehicles. Because the model requires that the output predicted vehicle and the predicted duration of the future trajectory of the surrounding vehicles are the same, T pred '=T pred ;

[0113] In the construction of the observation tensor, the environment within 30 meters in front and behind and 5 meters on the left and right of the predicted vehicle is taken, and a 5×25 two-dimensional space grid is established with the center of mass of the predicted vehicle as the center. The 5 in 5×25 represents the lateral observation range of the predicted vehicle, and the 25 in 5×25 represents the longitudinal observation range of the predicted vehicle. The center of mass of the predicted vehicle is located at the center of the two-dimensional space grid, that is, the 3rd grid in the horizontal direction and the 13th grid in the longitudinal direction. The length of each small square in the two-dimensional space grid is defined as 2.4 meters and the width is 2 meters. The orientation of the surrounding vehicles is recorded in the two-dimensional space grid; in this 5×25 two-dimensional space grid, the small squares with vehicles are set to TRUE, and the small squares without vehicles are set to FALSE. The number of surrounding vehicles is n, and then the observation tensor CentGrid∈R of the surrounding vehicle orientation represented by TRUE and FALSE is obtained. 5×25 , through the observation tensor CentGrid∈R 5×25 Ability to extract spatial interaction features between the predicted vehicle and surrounding vehicles;

[0114] In the conversion of relative coordinates between vehicles, the absolute coordinates of the predicted historical trajectory of the vehicle TargHist and the absolute coordinates of the future real trajectory TargFut are respectively subtracted from the current time T obs The absolute horizontal and vertical coordinates of the predicted vehicle under Get the relative historical trajectory coordinates TargCentHist of the predicted vehicle after conversion and the relative future trajectory coordinates TargCentFut of the predicted vehicle after conversion:

[0115]

[0116] Among them, TargCentHist is used for model input, and TargCentFut is used for model calculation loss value. is the relative historical trajectory coordinate of the predicted vehicle at time t after transformation, The absolute horizontal and vertical coordinates of the predicted vehicle at time t, is the current time T obs The absolute horizontal and vertical coordinates of the predicted vehicle under obs}, is the relative future trajectory coordinate of the predicted vehicle at the next k time after transformation, is the converted relative future trajectory abscissa of the predicted vehicle at the next k time, is the relative future trajectory ordinate of the predicted vehicle at the future k time after transformation, k∈{1,2,...,T pred};

[0117] Subtract the absolute coordinates of the historical trajectories NbrsHist and the absolute coordinates of the future trajectories NbrsFut of the surrounding vehicles from the current time T obs The predicted absolute lateral and longitudinal coordinates of the vehicle under Get the relative historical trajectory coordinates NbrsCentHist of the surrounding vehicles after conversion and the relative future trajectory coordinates NbrsCentFut of the surrounding vehicles after conversion:

[0118]

[0119] Among them, NbrsCentHist is used for model input, and NbrsCentFut is used for model calculation loss value. is the relative historical trajectory coordinates of the surrounding vehicles at time t after transformation, t'∈{1,2,...,T obs '}; is the relative future trajectory coordinates of the surrounding vehicles after transformation at the future k' time, is the relative future trajectory abscissa of the surrounding vehicles after transformation at the future k' time, The relative future trajectory ordinate of the transformed surrounding vehicles at the future k' time, k'∈{1,2,...,T pred '};

[0120] Step 2: Design a vehicle trajectory prediction model

[0121] exist Figure 2 The framework diagram of the trajectory prediction model is shown in Figure 1. The trajectory prediction model includes the predicted vehicle trajectory prediction model TargModel and the surrounding vehicle trajectory prediction model NbrsModel. The predicted vehicle trajectory prediction model TargMode and the surrounding vehicle trajectory prediction model NbrsModel have the same model structure, but different parameters in the model. Both the predicted vehicle trajectory prediction model TargMode and the surrounding vehicle trajectory prediction model NbrsModel include a vehicle motion trajectory information extraction module and a vehicle trajectory prediction module:

[0122] For the vehicle motion trajectory information extraction module, first convert the predicted vehicle relative historical trajectory coordinates in step 1 Relative historical trajectory coordinates of surrounding vehicles The dimensions are transformed into and Will Input into the one-dimensional convolutional neural network Conv1D3 to extract the motion information of the predicted vehicle. Input to another one-dimensional convolutional neural network Conv1D3' to extract the motion information of surrounding vehicles and obtain the motion embedding TargCentHist of the predicted vehicle emb The motion of the surrounding vehicles is embedded in NbrsCentHist emb :

[0123]

[0124] in, Encode information for the predicted vehicle's motion, To encode the motion information of surrounding vehicles, Conv1D3 and Conv1D3' are both one-dimensional convolutional layers with a convolution kernel size of 3, 2 input channels, 32 output features, and a step size of 1. is the weight parameter of Conv1D3, is the weight parameter of Conv1D3', is the bias parameter of Conv1D3, is the bias parameter of Conv1D3';

[0125] The motion encoding information of the vehicle to be predicted The dimension is converted to Encode the motion information of surrounding vehicles The dimension is converted to Then input it into the encoder to get the encoding information of the predicted vehicle and coded information of surrounding vehicles

[0126]

[0127] in, is the hidden state of the predicted vehicle, is the hidden state of the surrounding vehicles, Encoder is the motion encoder of the predicted vehicle, Encoder' is the motion encoder of the surrounding vehicles, and the encoder is composed of a long short-term memory network, namely, an LSTM network. The number of input features of the LSTM network in the encoder is 32, the number of hidden features is 64, and the number of layers is 1. is the weight parameter of Encoder, is the weight parameter of Encoder', is the bias parameter of Encoder, is the bias parameter of Encoder';

[0128] The motion encoding information of the vehicle to be predicted Aggregate the encoded information through the fully connected layer:

[0129]

[0130] in, is the motion information encoding of the predicted vehicle with a feature number of 32, φ is the embedding function of the fully connected layer, is the weight parameter of the fully connected layer that aggregates the predicted vehicle encoding information, The bias parameters of the fully connected layer that aggregates the predicted vehicle encoding information;

[0131] Encode the motion information of surrounding vehicles Fill in the observation tensor CentGrid∈R in step 1 according to its position in the grid 5×25 , that is, Fill to CentGrid∈R 5×25 The positions that are TRUE in the grid are filled with 0, while the positions that are False are filled with 0. If there is only one position in the grid that is TRUE, the position relationship of the vehicle in the grid is not considered and the Fill in this position and Fill to CentGrid∈R 5×25 After that, we get the observed feature tensor CentGridH, and then use CentGridH as input and use the two-dimensional graph convolution Conv2D 3×3 To extract the features of the spatial interaction information between the predicted vehicle and its surrounding vehicles:

[0132]

[0133] Among them, NbrsGridConv∈R 64×5×25 Conv2D is the spatial interaction feature between cars. 3×3 It is a two-dimensional convolution layer with a convolution kernel size of 3×3, the number of input features and the number of output features are both 64, and the step size is 1. For Conv2D 3×3 The weight parameter, For Conv2D 3×3 The bias parameters of

[0134] The spatial interaction feature NbrsGridConv between vehicles is pooled to reduce the dimension, retain key information and improve the generalization of the model:

[0135]

[0136] Where NbrsGridPool∈R 64×9×1 , MaxPool 3×3 is a maximum pooling layer with a pooling kernel of 3×3 and a stride of 2. is the weight parameter, is the bias parameter;

[0137] In order to further enhance the prediction network's ability to understand and model spatial feature data, NbrsGridPool is further reconvolved and repooled. That is, NbrsGridPool is input into a two-dimensional convolution layer with a convolution kernel of 3×1, and then input into a maximum pooling layer with a pooling kernel of 2×1. Finally, the spatial interaction information between vehicles is obtained. And transform the dimension into

[0138] The vehicle trajectory prediction module predicts the future time length T by obtaining vehicle motion information and inter-vehicle space information pred The specific operation is as follows: Encode the motion information of the predicted vehicle Spatial interaction information between vehicles Splice and copy T pred The vehicle information aggregation feature with a feature number of 112 is obtained. The aggregate feature Input into the trajectory prediction module to obtain the future predicted trajectory Pred of the predicted vehicle:

[0139]

[0140] In the formula,

[0141] is the predicted vehicle at T obs+mThe relative predicted trajectory coordinates at time, is the predicted vehicle at T obs+m The horizontal coordinate of the relative predicted trajectory at time, is the predicted vehicle at T obs+m The relative predicted trajectory ordinate at time, m∈{1,2,...,T pred}, Decoder is a decoder, which is composed of a long short-term memory network LSTM network. The number of input features of the LSTM network in the decoder is 112, the number of hidden features is 128, and the number of layers is 1. is the weight parameter of the decoder, is the bias parameter of the decoder, φ is the embedding function of the fully connected layer, is the weight parameter of the fully connected layer, is the bias parameter of the fully connected layer, the input feature number is 128, and the output dimension is 2;

[0142] Step 3: Construct loss function

[0143] The expression of the loss function Loss is:

[0144]

[0145] in, is the relative real trajectory abscissa of the predicted vehicle at time p, is the relative true trajectory ordinate of the predicted vehicle at time p, is the horizontal coordinate of the relative predicted trajectory of the predicted vehicle at time p output by the model, The ordinate of the relative predicted trajectory of the predicted vehicle at time p output by the model;

[0146] Step 4: Step-by-step two-way interactive prediction trajectory

[0147] The process of the step-by-step two-way interactive trajectory prediction method is as follows: Figure 1 As shown, the predicted vehicle is recorded as Car targ , set the prediction Car targ The duration of the future trajectory is T, and the duration of the predicted future trajectory of the surrounding vehicles is T', because the model requires the input of the predicted future trajectory of the surrounding vehicles to be Car targ half of Car targ The historical observation time is T obs , the historical observation time of surrounding vehicles is T obs ', because the model requires the input Car targ The same as the historical observation time of the surrounding vehicles, so T obs '=T obs ;

[0148] The model predicts Car in two stages targ Future trajectory:

[0149] In the first stage, the model outputs Car targ future The relative predicted trajectory in the time domain and the relative predicted trajectory of surrounding vehicles in the future 0~T' domain;

[0150] In the second stage, the model outputs Car targ future Relative predicted trajectory in the time domain;

[0151] Finally, the prediction time domain is and The relative prediction trajectories of the two predicted vehicles are spliced ​​in chronological order to obtain Car targ The complete future relative prediction trajectory of duration T;

[0152] First, obtain the first predicted trajectory of the predicted vehicle:

[0153] Car targ The observation tensor CentGrid∈R 5×25 、Car targ The relative historical trajectory coordinates TargCentHist of the vehicle and the relative historical trajectory coordinates NbrsCentHist of the surrounding vehicles are input into the predicted vehicle trajectory prediction model TargModel to obtain Car targ future Prediction trajectory in time domain TargPred:

[0154]

[0155] in, For Car targ In the T obs +a predicted trajectory coordinates at time, For Car targ In the T obs +a is the predicted horizontal coordinate at time point, For Car targ In the T obs +a The predicted ordinate at time,

[0156] Next, for Car targ Relative coordinate transformation with surrounding vehicles:

[0157] Select all surrounding vehicles in turn, and record the selected surrounding vehicles as Car g , where g∈N represents the gth surrounding vehicle, let Car gThe original relative historical trajectory coordinates SelHist are:

[0158]

[0159] in, For Car g The original relative historical trajectory coordinates at time b, b∈{1,2,...,T obs'}, Car g The current time T obs 'Location below Set it as the origin of coordinates and establish a new reference coordinate system. g The historical trajectory coordinates of Car are converted into the coordinates in the new reference coordinate system, and the g The new relative history track coordinates TurnHist:

[0160]

[0161] Car targ The original relative historical trajectory coordinates are converted into the new relative historical trajectory coordinates TargNewHist in the new reference coordinate system:

[0162]

[0163] Among them, TargCentHist is Car targ The original relative historical trajectory coordinates;

[0164] Then, get Car g The observation tensor SelCentGrid:

[0165] In Car targ The observation tensor Centgrid∈R 5×25 Based on the grid Car g The small square at the location is set to TRUE, and the contents of the other small squares are all set to FALSE, and Car g The observation tensor SelCentGrid∈R 5×25 ;

[0166] Next, select all surrounding vehicles in turn and obtain their next predicted trajectories:

[0167] Car targ The number of surrounding vehicles is n, and a surrounding car Car is selected in turn each time g , and perform the surrounding vehicle trajectory prediction process:

[0168] The Car obtained in the previous step g The observation tensor SelCentGrid∈R5×25 、Car g The new relative historical trajectory coordinates TurnHist, Car targ The new relative historical trajectory coordinates TargNewHist are input into the surrounding vehicle trajectory prediction model NbrsModel to obtain Car g The predicted trajectory in the future 0~T' time domain with its own center of mass as the origin of the reference system;

[0169] After the above surrounding vehicle trajectory prediction process is performed n' times, the predicted trajectory coordinates SelNbrsPred of all surrounding vehicles in the new reference system with their respective center of mass as the reference system origin can be obtained. Because the surrounding vehicle trajectory prediction process needs to be performed as many times as there are vehicles, so n'=n:

[0170]

[0171] in, is the predicted trajectory coordinates of the surrounding vehicles at the future time c, is the predicted trajectory abscissa of the surrounding vehicles at time c in the future, is the predicted trajectory ordinate of the surrounding vehicles at time c in the future, c∈{1,2,...,T'};

[0172] Next, convert SelNbrsPred to Car targ is the predicted trajectory coordinate of the reference system:

[0173] SelNbrsPred is based on Car g The trajectory coordinates in the reference system of the origin center are converted into Car targ The coordinates in the reference system of Add them together and you get Car targ The next predicted trajectory coordinates of the surrounding vehicles in the reference frame NbrsPred:

[0174]

[0175] in, For Car g In T obs 'The original relative historical trajectory coordinates at the moment;

[0176] Finally, the next predicted trajectory coordinates of the predicted vehicle and surrounding vehicles are replaced into their respective old historical trajectories to obtain the new historical trajectory coordinates:

[0177] Discard the old history track TargCentHist track points, add TargPred to TargCentHist as the latest The trajectory points at the moment are updated to obtain the new historical trajectory TargRepHist:

[0178]

[0179] in, For the latest The predicted trajectory points of the predicted vehicle at the moment, is the new historical trajectory of the predicted vehicle at time d,

[0180] And the latest trajectory point of the predicted vehicle As the new reference system origin and establish a new reference coordinate system, the new relative historical trajectory TargCentHist2 of the predicted vehicle is obtained:

[0181]

[0182] in, is the new relative historical trajectory coordinate of the predicted vehicle at time e in the new reference coordinate system, and

[0183] Similarly, the predicted trajectory NbrsPred of the surrounding vehicles is used to replace the first 0 to T' trajectory points in the old historical trajectory NbrsCentHist to obtain the updated new historical trajectory NbrsRepHist:

[0184]

[0185] in, is the predicted trajectory point of the surrounding vehicles at the latest T' moment, is the new historical trajectory of the surrounding vehicles at time f, f∈{T'+1,T'+2...,T obs '-1,T obs ',T obs '+1,...,T obs '+T'};

[0186] And the latest trajectory point of the predicted vehicle As the new origin coordinates, we get the new relative historical trajectory NbrsCentHist2 of the surrounding vehicles in the new reference coordinate system:

[0187]

[0188] in, For surrounding vehicles is the new relative historical trajectory coordinate of the origin at time l, and l∈[T'+1,T'+2,...,T obs '+T'-1,T obs '+T'];

[0189] The predicted vehicle's new relative historical trajectory TargCentHist2, the surrounding vehicles' new relative historical trajectory NbrsCentHist2 and the observation tensor CentGrid are input into the predicted vehicle trajectory prediction model TargModel to obtain the predicted vehicle's future trajectory. Prediction trajectory in time domain TargPred2:

[0190]

[0191] in, The predicted vehicle is obs +p time predicted trajectory coordinates, The predicted vehicle is obs +p time prediction horizontal coordinate, The predicted vehicle is obs +The predicted ordinate at time p,

[0192] Finally, the first-stage predicted trajectory TargPred and the second-stage predicted trajectory TargPred2 of the predicted vehicle are concatenated to obtain the complete predicted trajectory Pred of the predicted vehicle:

[0193]

[0194] in, The predicted vehicle is obs +q time predicted trajectory coordinates, The predicted vehicle is obs +q time prediction horizontal coordinate, The predicted vehicle is obs +q time prediction ordinate,

Claims

1. A vehicle trajectory prediction method considering step-by-step two-way interaction, which fully extracts the two-way interaction information between vehicles by gradually extracting the trajectory and spatial position information of the predicted vehicle and other surrounding vehicles and inputting them into the established network model, and accurately predicts the future trajectory of the vehicle in multiple segments, characterized in that: The specific steps of this method are as follows: Step 1: Process the vehicle trajectory dataset: First, the position information of the vehicles in the data set is extracted in chronological order, that is, the absolute coordinates (x, y) of the horizontal and vertical positions of the vehicles. Then, all vehicle coordinates are further processed, including vehicle trajectory extraction, observation tensor construction, and relative coordinate conversion between vehicles. In vehicle trajectory extraction, the absolute coordinates of the predicted vehicle’s historical trajectory TargHist and the absolute coordinates of the future real trajectory TargFut are defined as follows: in, is the absolute horizontal and vertical coordinates of the predicted vehicle at time t, is the absolute lateral coordinate of the predicted vehicle at time t, is the absolute longitudinal coordinate of the predicted vehicle at time t, t∈{1,2,...,T obs }, T obs is the observation time of the historical trajectory of the predicted vehicle, and is also the current moment, is the real horizontal and vertical absolute coordinates of the predicted vehicle at the future k moment, is the absolute lateral coordinate of the predicted vehicle at the next k moments, is the absolute longitudinal coordinate of the predicted vehicle at the kth moment in the future, k∈{1,2,...,T pred }, T pred Predict the duration of the future trajectory of the predicted vehicle; The absolute coordinates of the historical trajectories of the vehicles around the predicted vehicle, NbrsHist, and the absolute coordinates of the future real trajectories, NbrsFut, are defined as follows: Wherein, n∈N is the number of surrounding vehicles of the predicted vehicle, N is a positive integer greater than 0, and the surrounding vehicles of the predicted vehicle are hereinafter referred to as surrounding vehicles; is the set of absolute horizontal and vertical coordinates of n surrounding vehicles at time t. Specifically, is the set of the absolute lateral coordinates of the n surrounding vehicles at time t, is the set of absolute longitudinal coordinates of n surrounding vehicles at time t, t∈{1,2,...,T obs '}, T obs ' is the observation time of the historical trajectory of the surrounding vehicles. Because the model requires that the observation time of the historical trajectory of the input predicted vehicle and the surrounding vehicles is the same, T obs '=T obs , is the set of the real horizontal and vertical absolute coordinates of the n surrounding vehicles at the next k time. Specifically, is the set of real lateral absolute coordinates of n surrounding vehicles at the next k moments, is the set of real longitudinal absolute coordinates of n surrounding vehicles at the next k time, k∈{1,2,...,T pred '}, T pred ' is the predicted duration of the future trajectory of the surrounding vehicles. Because the model requires that the output predicted vehicle and the predicted duration of the future trajectory of the surrounding vehicles are the same, T pred '=T pred ; In the construction of the observation tensor, the environment within 30 meters in front and behind and 5 meters on the left and right of the predicted vehicle is taken, and a 5×25 two-dimensional space grid is established with the center of mass of the predicted vehicle as the center. The 5 in 5×25 represents the lateral observation range of the predicted vehicle, and the 25 in 5×25 represents the longitudinal observation range of the predicted vehicle. The center of mass of the predicted vehicle is located at the center of the two-dimensional space grid, that is, the 3rd grid in the horizontal direction and the 13th grid in the longitudinal direction. The length of each small square in the two-dimensional space grid is defined as 2.4 meters and the width is 2 meters. The orientation of the surrounding vehicles is recorded in the two-dimensional space grid; in this 5×25 two-dimensional space grid, the small squares with vehicles are set to TRUE, and the small squares without vehicles are set to FALSE. The number of surrounding vehicles is n, and then the observation tensor CentGrid∈R of the surrounding vehicle orientation represented by TRUE and FALSE is obtained. 5×25 , through the observation tensor CentGrid∈R 5×25 Ability to extract spatial interaction features between the predicted vehicle and surrounding vehicles; In the conversion of relative coordinates between vehicles, the absolute coordinates of the predicted historical trajectory of the vehicle TargHist and the absolute coordinates of the future real trajectory TargFut are respectively subtracted from the current time T obs The absolute horizontal and vertical coordinates of the predicted vehicle under Get the relative historical trajectory coordinates TargCentHist of the predicted vehicle after conversion and the relative future trajectory coordinates TargCentFut of the predicted vehicle after conversion: Among them, TargCentHist is used for model input, and TargCentFut is used for model calculation loss value. is the relative historical trajectory coordinate of the predicted vehicle at time t after transformation, The absolute horizontal and vertical coordinates of the predicted vehicle at time t, is the current time T obs The absolute horizontal and vertical coordinates of the predicted vehicle under obs }, is the relative future trajectory coordinate of the predicted vehicle at the next k time after transformation, is the converted relative future trajectory abscissa of the predicted vehicle at the next k time, is the relative future trajectory ordinate of the predicted vehicle at the future k time after transformation, k∈{1,2,...,T pred }; Subtract the absolute coordinates of the historical trajectories NbrsHist and the absolute coordinates of the future trajectories NbrsFut of the surrounding vehicles from the current time T obs The predicted absolute horizontal and vertical coordinates of the vehicle under Get the relative historical trajectory coordinates NbrsCentHist of the surrounding vehicles after conversion and the relative future trajectory coordinates NbrsCentFut of the surrounding vehicles after conversion: Among them, NbrsCentHist is used for model input, and NbrsCentFut is used for model calculation loss value. is the relative historical trajectory coordinates of the surrounding vehicles at time t after transformation, t'∈{1,2,...,T obs '}; is the relative future trajectory coordinates of the surrounding vehicles after transformation at the future k' time, is the relative future trajectory abscissa of the surrounding vehicles after transformation at the future k' time, The relative future trajectory ordinate of the transformed surrounding vehicles at the future k' time, k'∈{1,2,...,T pred '}; Step 2: Design a vehicle trajectory prediction model The trajectory prediction model includes the predicted vehicle trajectory prediction model TargModel and the surrounding vehicle trajectory prediction model NbrsModel. The predicted vehicle trajectory prediction model TargMode and the surrounding vehicle trajectory prediction model NbrsModel have the same model structure, but different parameters in the model. The predicted vehicle trajectory prediction model TargMode and the surrounding vehicle trajectory prediction model NbrsModel both include a vehicle motion trajectory information extraction module and a vehicle trajectory prediction module: For the vehicle motion trajectory information extraction module, first convert the predicted vehicle relative historical trajectory coordinates in step 1 Relative historical trajectory coordinates of surrounding vehicles The dimensions are transformed into and Will Input into the one-dimensional convolutional neural network Conv1D3 to extract the motion information of the predicted vehicle. Input to another one-dimensional convolutional neural network Conv1D3' to extract the motion information of surrounding vehicles and obtain the motion embedding TargCentHist of the predicted vehicle emb The motion of the surrounding vehicles is embedded in NbrsCentHist emb : in, Encode information for the predicted vehicle's motion, To encode the motion information of surrounding vehicles, Conv1D3 and Conv1D3' are both one-dimensional convolutional layers with a convolution kernel size of 3, 2 input channels, 32 output features, and a step size of 1. is the weight parameter of Conv1D3, is the weight parameter of Conv1D3', is the bias parameter of Conv1D3, is the bias parameter of Conv1D3'; The motion encoding information of the vehicle to be predicted The dimension is converted to Encode the motion information of surrounding vehicles The dimension is converted to Then input it into the encoder to get the encoding information of the predicted vehicle and coded information of surrounding vehicles in, is the hidden state of the predicted vehicle, is the hidden state of the surrounding vehicles, Encoder is the motion encoder of the predicted vehicle, Encoder' is the motion encoder of the surrounding vehicles, and the encoder is composed of a long short-term memory network, namely, an LSTM network. The number of input features of the LSTM network in the encoder is 32, the number of hidden features is 64, and the number of layers is 1. is the weight parameter of Encoder, is the weight parameter of Encoder', is the bias parameter of Encoder, is the bias parameter of Encoder'; The motion encoding information of the vehicle to be predicted Aggregate the encoded information through the fully connected layer: in, is the motion information encoding of the predicted vehicle with a feature number of 32, φ is the embedding function of the fully connected layer, is the weight parameter of the fully connected layer that aggregates the predicted vehicle encoding information, The bias parameters of the fully connected layer that aggregates the predicted vehicle encoding information; Encode the motion information of surrounding vehicles Fill in the observation tensor CentGrid∈R in step 1 according to its position in the grid 5×25 , that is, Fill to CentGrid∈R 5×25 The positions that are TRUE in the grid are filled with 0, while the positions that are False are filled with 0. If there is only one position in the grid that is TRUE, the position relationship of the vehicle in the grid is not considered and the Fill in this position and Fill to CentGrid∈R 5×25 After that, we get the observed feature tensor CentGridH, and then use CentGridH as input and use the two-dimensional graph convolution Conv2D 3×3 To extract the features of the spatial interaction information between the predicted vehicle and its surrounding vehicles: Among them, NbrsGridConv∈R 64×5×25 Conv2D is the spatial interaction feature between cars. 3×3 It is a two-dimensional convolution layer with a convolution kernel size of 3×3, the number of input features and the number of output features are both 64, and the step size is 1. For Conv2D 3×3 The weight parameter, For Conv2D 3×3 The bias parameters of The spatial interaction feature NbrsGridConv between vehicles is pooled to reduce the dimension, retain key information and improve the generalization of the model: Where NbrsGridPool∈R 64×9×1 , MaxPool 3×3 is a maximum pooling layer with a pooling kernel of 3×3 and a stride of 2. is the weight parameter, is the bias parameter; In order to further enhance the prediction network's ability to understand and model spatial feature data, NbrsGridPool is further reconvolved and repooled. That is, NbrsGridPool is input into a two-dimensional convolution layer with a convolution kernel of 3×1, and then input into a maximum pooling layer with a pooling kernel of 2×1. Finally, the spatial interaction information between vehicles is obtained. And transform the dimension into The vehicle trajectory prediction module predicts the future time length T by obtaining vehicle motion information and inter-vehicle space information pred The specific operation is as follows: Encode the motion information of the predicted vehicle Spatial interaction information between vehicles Splice and copy T pred The vehicle information aggregation feature with a feature number of 112 is obtained. The aggregate feature Input into the trajectory prediction module to obtain the future predicted trajectory Pred of the predicted vehicle: In the formula, is the predicted vehicle at T obs+m The relative predicted trajectory coordinates at time, is the predicted vehicle at T obs+m The horizontal coordinate of the relative predicted trajectory at time, is the predicted vehicle at T obs+m The relative predicted trajectory ordinate at time, m∈{1,2,...,T pred }, Decoder is a decoder, which is composed of a long short-term memory network LSTM network. The number of input features of the LSTM network in the decoder is 112, the number of hidden features is 128, and the number of layers is 1. is the weight parameter of the decoder, is the bias parameter of the decoder, φ is the embedding function of the fully connected layer, is the weight parameter of the fully connected layer, is the bias parameter of the fully connected layer, the input feature number is 128, and the output dimension is 2; Step 3: Construct loss function The expression of the loss function Loss is: in, is the relative real trajectory abscissa of the predicted vehicle at time p, is the relative true trajectory ordinate of the predicted vehicle at time p, is the horizontal coordinate of the relative predicted trajectory of the predicted vehicle at time p output by the model, The ordinate of the relative predicted trajectory of the predicted vehicle at time p output by the model; Step 4: Step-by-step two-way interactive prediction trajectory The predicted vehicle is denoted as Car targ , set the prediction Car targ The duration of the future trajectory is T, and the duration of the predicted future trajectory of the surrounding vehicles is T', because the model requires the input of the predicted future trajectory of the surrounding vehicles to be Car targ half of Car targ The historical observation time is T obs , the historical observation time of surrounding vehicles is T obs ', because the model requires the input Car targ The same as the historical observation time of the surrounding vehicles, so T obs '=T obs ; The model predicts Car in two stages targ Future trajectory: In the first stage, the model outputs Car targ future The relative predicted trajectory in the time domain and the relative predicted trajectory of surrounding vehicles in the future 0~T' domain; In the second stage, the model outputs Car targ future Relative predicted trajectory in the time domain; Finally, the prediction time domain is and The relative prediction trajectories of the two predicted vehicles are spliced ​​in chronological order to obtain Car targ The complete future relative prediction trajectory of duration T; First, obtain the first predicted trajectory of the predicted vehicle: Car targ The observation tensor CentGrid∈R 5×25 、Car targ The relative historical trajectory coordinates TargCentHist of the vehicle and the relative historical trajectory coordinates NbrsCentHist of the surrounding vehicles are input into the predicted vehicle trajectory prediction model TargModel to obtain Car targ future Prediction trajectory in time domain TargPred: in, For Car targ In the T obs +a predicted trajectory coordinates at time, For Car targ In the T obs +a is the predicted horizontal coordinate at time point, For Car targ In the T obs +a The predicted ordinate at time, Next, for Car targ Relative coordinate transformation with surrounding vehicles: Select all surrounding vehicles in turn, and record the selected surrounding vehicles as Car g , where g∈N represents the gth surrounding vehicle, let Car g The original relative historical trajectory coordinates SelHist are: in, For Car g The original relative historical trajectory coordinates at time b, b∈{1,2,...,T obs' }, Car g The current time T obs 'Location below Set it as the origin of coordinates and establish a new reference coordinate system. g The historical trajectory coordinates of Car are converted into the coordinates in the new reference coordinate system, and the g The new relative history track coordinates TurnHist: Car targ The original relative historical trajectory coordinates are converted into the new relative historical trajectory coordinates TargNewHist in the new reference coordinate system: Among them, TargCentHist is Car targ The original relative historical trajectory coordinates; Then, get Car g The observation tensor SelCentGrid: In Car targ The observation tensor Centgrid∈R 5×25 Based on the grid Car g The small square at the location is set to TRUE, and the contents of the other small squares are all set to FALSE, and Car g The observation tensor SelCentGrid∈R 5×25 ; Next, select all surrounding vehicles in turn and obtain their next predicted trajectories: Car targ The number of surrounding vehicles is n, and a surrounding car Car is selected in turn each time g , and perform the surrounding vehicle trajectory prediction process: The Car obtained in the previous step g The observation tensor SelCentGrid∈R 5×25 、Car g The new relative historical trajectory coordinates TurnHist, Car targ The new relative historical trajectory coordinates TargNewHist are input into the surrounding vehicle trajectory prediction model NbrsModel to obtain Car g The predicted trajectory in the future 0~T' time domain with its own center of mass as the origin of the reference system; After the above surrounding vehicle trajectory prediction process is performed n' times, the predicted trajectory coordinates SelNbrsPred of all surrounding vehicles in the new reference system with their respective center of mass as the reference system origin can be obtained. Because the surrounding vehicle trajectory prediction process needs to be performed as many times as there are vehicles, so n'=n: in, is the predicted trajectory coordinates of the surrounding vehicles at the future time c, is the predicted trajectory abscissa of the surrounding vehicles at time c in the future, is the predicted trajectory ordinate of the surrounding vehicles at time c in the future, c∈{1,2,...,T'}; Next, convert SelNbrsPred to Car targ is the predicted trajectory coordinate of the reference system: SelNbrsPred is based on Car g The trajectory coordinates in the reference system of the origin center are converted into Car targ The coordinates in the reference system of Add them together and you get Car targ The next predicted trajectory coordinates of the surrounding vehicles in the reference frame NbrsPred: in, For Car g In T obs 'The original relative historical trajectory coordinates at the moment; Finally, the next predicted trajectory coordinates of the predicted vehicle and surrounding vehicles are replaced into their respective old historical trajectories to obtain the new historical trajectory coordinates: Discard the old history track TargCentHist track points, add TargPred to TargCentHist as the latest The trajectory points at the moment are updated to obtain the new historical trajectory TargRepHist: in, For the latest The predicted trajectory points of the predicted vehicle at the moment, is the new historical trajectory of the predicted vehicle at time d, And the latest trajectory point of the predicted vehicle As the new reference system origin and establish a new reference coordinate system, the new relative historical trajectory TargCentHist2 of the predicted vehicle is obtained: in, is the new relative historical trajectory coordinate of the predicted vehicle at time e in the new reference coordinate system, and Similarly, the predicted trajectory NbrsPred of the surrounding vehicles is used to replace the first 0 to T' trajectory points in the old historical trajectory NbrsCentHist to obtain the updated new historical trajectory NbrsRepHist: in, is the predicted trajectory point of the surrounding vehicles at the latest T' moment, is the new historical trajectory of the surrounding vehicles at time f, f∈{T'+1,T'+2...,T obs '-1,T obs ',T obs '+1,...,T obs '+T'}; And the latest trajectory point of the predicted vehicle As the new origin coordinates, we get the new relative historical trajectory NbrsCentHist2 of the surrounding vehicles in the new reference coordinate system: in, For surrounding vehicles is the new relative historical trajectory coordinate of the origin at time l, and l∈[T'+1,T'+2,...,T obs '+T'-1,T obs '+T']; The predicted vehicle's new relative historical trajectory TargCentHist2, the surrounding vehicles' new relative historical trajectory NbrsCentHist2 and the observation tensor CentGrid are input into the predicted vehicle trajectory prediction model TargModel to obtain the predicted vehicle's future trajectory. Prediction trajectory in time domain TargPred2: in, The predicted vehicle is obs +p time predicted trajectory coordinates, The predicted vehicle is obs +p time prediction horizontal coordinate, The predicted vehicle is obs +The predicted ordinate at time p, Finally, the first-stage predicted trajectory TargPred and the second-stage predicted trajectory TargPred2 of the predicted vehicle are concatenated to obtain the complete predicted trajectory Pred of the predicted vehicle: in, The predicted vehicle is obs +q time predicted trajectory coordinates, The predicted vehicle is obs +q time prediction horizontal coordinate, The predicted vehicle is obs +q time prediction ordinate,

Images