Vehicle track prediction method and system based on interaction potential field and graph convolutional network

Abstract

The invention belongs to the field of automatic driving, and discloses a vehicle trajectory prediction method and system based on an interaction potential field and a graph convolutional network, and the method comprises the steps: obtaining time sequence features based on the historical trajectory data of a vehicle, representing the interaction influence relation between vehicles through the interaction potential field, and extracting the longitudinal and transverse interaction features of the vehicle through combining with the graph convolutional network; and splicing the time sequence features, the vehicle inherent features and the interaction features to form time sequence interaction fusion features, inputting the time sequence interaction fusion features to a decoder model, and finally outputting a high-precision vehicle prediction trajectory. According to the method, the vehicle time sequence features are extracted, the vehicle interaction feature extraction model is constructed by using the interaction potential field and the graph convolutional network, the complex interaction relationship between vehicles is effectively captured, the problem that an existing method neglects external factors and environment vehicle interaction is solved, directivity influence is introduced through the interaction potential field, and the vehicle interaction effect is improved. And the vehicle motion state is integrated, so that the trajectory prediction accuracy is remarkably improved, and the method is suitable for high-precision trajectory prediction requirements in complex traffic scenes.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving, and specifically relates to a vehicle trajectory prediction method and system based on interaction potential field and graph convolutional network. Background Art

[0002] With the intelligent transformation of the automotive industry, intelligent vehicles have become a key development direction in the global transportation field. Through advanced sensors, intelligent algorithms, and communication technologies, they can effectively improve traffic safety and operation efficiency, reduce traffic accidents, and alleviate traffic congestion. Trajectory prediction, as a key front-end link in the decision-making and chassis control of intelligent vehicles, plays a decisive role in the decision-making effect of intelligent vehicles and the feasibility of the planned path.

[0003] Currently, vehicle trajectory prediction methods are mainly divided into model-based methods, reinforcement learning-based methods, and deep learning-based methods. Model-based methods predict trajectories by constructing vehicle kinematic and dynamic models and combining algorithms such as Kalman filtering. These methods only consider the impact of the vehicle's own motion state on the trajectory, ignoring external factors such as surrounding vehicles, traffic regulations, and traffic signs, resulting in large long-term trajectory prediction errors in complex traffic scenarios. Reinforcement learning and deep learning-based methods show certain advantages in dealing with complex scenarios, but most use black-box models to represent vehicle interactions, lacking interpretability. Although graph neural networks can represent the influence relationship between vehicles by constructing nodes and edges, with good robustness and certain interpretability, the adjacency matrix constructed based on the distance between vehicles can only describe the intensity of interaction influence, unable to explain the directionality of interaction influence, and ignoring the impact of vehicle motion on interaction, resulting in the trajectory prediction accuracy being difficult to meet actual requirements. Moreover, in actual traffic scenarios, the driving trajectories of vehicles are affected by a variety of factors. For example, at intersections on urban roads, vehicles not only need to make decisions based on their own driving directions and speeds but also need to consider the driving intentions of surrounding vehicles, the status of traffic lights, and the width of the road. On highways, the following distance and lane-changing behavior of vehicles are also affected by the speeds, accelerations, and traffic flow of surrounding vehicles.

[0004] It can be seen that the existing trajectory prediction methods cannot accurately capture these complex interaction relationships and are difficult to achieve high-precision trajectory prediction. Summary of the Invention

[0005] The present invention provides a vehicle trajectory prediction method and system based on interaction potential field and graph convolutional network to solve the technical problem that the existing trajectory prediction methods cannot accurately capture these complex interaction relationships and are difficult to achieve high-precision trajectory prediction.

[0006] To achieve the above object, the present invention adopts the following technical solutions: A vehicle trajectory prediction method based on an interaction potential field and a graph convolutional network, comprising: Obtaining vehicle temporal features based on vehicle historical trajectory data; After splicing the vehicle temporal features and the vehicle inherent features, inputting them into a pre-constructed vehicle interaction feature extraction model based on an interaction potential field and a graph convolutional network to obtain longitudinal and lateral vehicle interaction features; the interaction potential field is used to characterize the vehicle interaction influence relationship; Inputting the temporal interaction fusion features obtained by splicing the vehicle temporal features, the vehicle inherent features, and the longitudinal and lateral vehicle interaction features into a pre-constructed decoder model, and outputting the vehicle prediction trajectory.

[0007] Further, the obtaining vehicle temporal features based on vehicle historical trajectory data includes: Constructing a vehicle group from the target vehicle and the surrounding environmental vehicles; In the vehicle group, obtaining the vehicle historical trajectory data of the vehicle within the observation horizon; the trajectory features include the vehicle historical position, speed, and acceleration; Inputting the vehicle historical trajectory data into a pre-constructed temporal feature extraction model of a bidirectional long short-term memory network based on an attention mechanism to obtain vehicle trajectory temporal features.

[0008] Further, the inputting the vehicle historical trajectory data into a pre-constructed temporal feature extraction model of a bidirectional long short-term memory network based on an attention mechanism to obtain vehicle trajectory temporal features includes: Inputting the vehicle historical trajectory data into a pre-constructed temporal feature extraction model of a bidirectional long short-term memory network based on an attention mechanism; Combining the propagation processes of the forward and backward LSTMs to extract vehicle trajectory temporal features, and the specific formula is as follows:

[0009] In the formula, are the forward and backward propagations respectively, are the features at time t extracted by the forward and backward LSTMs respectively, is the trajectory feature at time t, is the output of the Bi-LSTM, represents the concatenation operation, and Bi-LSTM represents the temporal feature extraction model of the bidirectional long short-term memory network; Based on the vehicle trajectory temporal features, calculating the weights at each moment through a linear layer and SoftMax, and obtaining the vehicle temporal features through weighted summation and then through an MLP, and the specific formula is as follows:

[0010] In the formula, are the parameters to be trained, is t the hidden feature of the moment trajectory data; is the attention weight of the moment trajectory data, is the vehicle weighted feature tensor, N is the number of vehicles, h is the number of hidden layers of Bi-LSTM.

[0011] Further, before the vehicle temporal features and vehicle inherent features are concatenated and input into a pre-constructed vehicle interaction feature extraction model based on the interaction potential field and graph convolutional network to obtain the longitudinal and lateral vehicle interaction features, it includes: Construct a vehicle interaction feature extraction model based on the interaction potential field and graph convolutional network, specifically as follows: Taking the vehicle as the potential field source, the longitudinal interaction potential field energy of the vehicle is obtained according to the vehicle length and longitudinal vehicle speed. The specific formula is as follows:

[0012] In the formula, is the vehicle j For vehicle i longitudinal potential field intensity, is a parameter related to the longitudinal influence intensity, is the longitudinal correction term; Establish the longitudinal interaction potential field energy of the vehicle according to the vehicle length and longitudinal vehicle speed. The specific formula is as follows:

[0013] In the formula, are respectively the vehicle length and longitudinal speed, is the vehicle and vehicle longitudinal relative distance, is the vehicle longitudinal interaction range; The longitudinal traffic flow difference correction term , used to characterize the directional difference of vehicle interaction. The longitudinal traffic flow difference correction term is expressed as:

[0014] In the formula, is a parameter related to the longitudinal traffic flow difference correction intensity, is a constant, representing the inherent front and rear vehicle influence difference of the longitudinal traffic flow, represents the longitudinal relative speed of the vehicle, respectively represent the vehicle i and vehicle j longitudinal positions; The lateral difference correction term , which is used to characterize the difference in the longitudinal influence caused by the lateral distance of different lanes. The lateral difference correction term is expressed as:

[0015] In the formula, is a parameter related to the lateral difference correction intensity of the longitudinal interaction, are respectively the i and the vehicle j lateral positions of; Based on the longitudinal traffic flow difference correction term and the lateral difference correction term, the longitudinal interaction potential field correction term is obtained, which is expressed as:

[0016] A lateral interaction potential field is established, and the lateral interaction potential field energy of the vehicle is obtained according to the vehicle length and the lateral vehicle speed, specifically as follows:

[0017] In the formula, is the vehicle j For the vehicle i lateral potential field intensity of, is a parameter related to the lateral influence intensity, is the lateral correction term, is a constant, representing the potential lateral interaction energy between vehicles, is the vehicle and the lateral relative distance between vehicle j, is the vehicle lateral interaction range of; Among them, the lateral correction term includes the lateral traffic flow difference correction term and the longitudinal difference correction term , as shown in the following formula:

[0018] In the formula, and are parameters related to the correction intensity, is the lateral relative speed of the vehicle; The interaction influence received by vehicle i is expressed as:

[0019] In the formula, is the longitudinal and lateral influence vector of the vehicle , are respectively the longitudinal and lateral influence potential field values of the vehicle i affected by the vehicle j ; Normalize the longitudinal and lateral influence vectors to obtain the longitudinal and lateral influence weights of the vehicle and construct a longitudinal and lateral influence matrix.

[0020] Further, inputting the vehicle time series features and vehicle inherent features after splicing into a pre-constructed vehicle interaction feature extraction model based on interactive potential field and graph convolutional network to obtain the longitudinal and lateral vehicle interaction features includes: Splice the vehicle time series features and vehicle inherent features to obtain a spliced feature; Use the spliced feature as the input of the vehicle interaction feature extraction model based on interactive potential field and graph convolutional network, and output the longitudinal and lateral vehicle interaction features.

[0021] Further, inputting the time series interaction fusion feature obtained by splicing the vehicle time series features, vehicle inherent features and longitudinal and lateral vehicle interaction features into a pre-constructed decoder model to output the vehicle prediction trajectory includes: Splice the vehicle time series features, vehicle inherent features and longitudinal and lateral vehicle interaction features to obtain a time series interaction fusion feature; Input the time series interaction fusion feature into the pre-constructed decoder model, and use layers to perform dimensionality reduction encoding on the time series interaction fusion feature and extract the implicit features of the driving intention; Through and layers respectively capture the longitudinal and lateral driving intention probabilities, and use the Softmax function to generate the longitudinal and lateral driving intention classifications; Based on the longitudinal and lateral driving intention classifications, respectively use and layers to encode and decode the vehicle features, and finally use layer to predict the future trajectory of the vehicle as the vehicle prediction trajectory output.

[0022] Further, the and layers respectively capture the longitudinal and lateral driving intention probabilities, and use the Softmax function to generate the longitudinal and lateral driving intention classifications. The specific formula is as follows:

[0023] In the formula, is the vehicle feature tensor after dimensionality reduction, is the time series interaction fusion feature, are the longitudinal and lateral driving intentions respectively; Based on the longitudinal and lateral driving intention classifications, respectively use and layers to encode and decode the vehicle features, and finally use The layer predicts the future trajectory of the vehicle as the output of the predicted vehicle trajectory, and the specific formula is as follows:

[0024] In the formula, is the vehicle feature tensor decoded by the layer, is the parameter matrix of the layer, is the predicted trajectory at time, are respectively the predicted vehicle at the longitudinal and lateral positions at time, is the observation horizon, is the future horizon.

[0025] A vehicle trajectory prediction system based on an interaction potential field and a graph convolutional network, comprising: A vehicle time-series feature extraction module, configured to obtain vehicle time-series features based on vehicle historical trajectory data; A longitudinal and lateral interaction feature extraction module, configured to splice the vehicle time-series features and the vehicle inherent features and input them into a pre-constructed vehicle interaction feature extraction model based on an interaction potential field and a graph convolutional network to obtain vehicle longitudinal and lateral interaction features; the interaction potential field is used to characterize the vehicle interaction influence relationship; A decoding module, configured to input the time-series interaction fusion features spliced from the vehicle time-series features, the vehicle inherent features, and the vehicle longitudinal and lateral interaction features into a pre-constructed decoder model, and output the predicted vehicle trajectory.

[0026] A device, comprising: A memory, configured to store a computer program; A processor, configured to implement the steps of the above-mentioned vehicle trajectory prediction method based on an interaction potential field and a graph convolutional network when executing the computer program.

[0027] A computer-readable storage medium stores a computer program, and the computer program is used to implement the steps of the above-mentioned vehicle trajectory prediction method based on an interaction potential field and a graph convolutional network when executed by a processor.

[0028] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a vehicle trajectory prediction method based on an interaction potential field and a graph convolutional network. First, this method obtains temporal features based on vehicle historical trajectory data, and then uses the interaction potential field to characterize the interaction influence relationship between vehicles, and combines the graph convolutional network to extract the longitudinal and lateral interaction features of the vehicles. Subsequently, the temporal features, vehicle inherent features, and interaction features are concatenated to form temporal interaction fusion features, which are input into the decoder model, and finally, a high-precision vehicle prediction trajectory is output. By extracting vehicle temporal features and using the vehicle interaction feature extraction model constructed by the interaction potential field and the graph convolutional network, this method can effectively capture the complex interaction relationships between vehicles, not only solving the problem that existing methods ignore external factors and the interaction with environmental vehicles, but also introducing directional influences through the interaction potential field and comprehensively considering the vehicle motion state, significantly improving the accuracy of trajectory prediction, and being applicable to the high-precision trajectory prediction requirements in complex traffic scenarios.

[0029] Preferably, in the present invention, by constructing a vehicle group and extracting the historical trajectory data of the target vehicle within the observation horizon, this method can more comprehensively capture the driving state of the target vehicle. Using a bidirectional long short-term memory network based on the attention mechanism, the temporal features of the vehicle trajectory are further extracted, providing strong support for subsequent feature fusion and trajectory prediction.

[0030] Preferably, in the present invention, through the combination of forward and backward LSTM and the introduction of the attention mechanism, this method can more effectively extract the key information in the vehicle trajectory and assign different weights to the trajectory features at different moments, thereby improving the accuracy of feature extraction.

[0031] Preferably, in the present invention, a vehicle interaction feature extraction model based on the interaction potential field and the graph convolutional network is constructed, and a traffic flow difference correction term and a lateral difference correction term are introduced to characterize the directionality and difference of the interaction. This model can more accurately describe the interaction relationship between vehicles, providing richer feature information for subsequent trajectory prediction.

[0032] Preferably, in the present invention, by concatenating the vehicle temporal features and the vehicle inherent features and using them as the input of the vehicle interaction feature extraction model based on the interaction potential field and the graph convolutional network, this method can extract more comprehensive vehicle interaction features. These features provide a more accurate basis for subsequent trajectory prediction.

[0033] Preferably, in the present invention, this method obtains temporal interaction fusion features by concatenating the vehicle temporal features, vehicle inherent features, and vehicle longitudinal and lateral interaction features. Using the decoder model to further process and predict these features can output a more accurate vehicle prediction trajectory. At the same time, by introducing driving intention classification and the encoding and decoding process, this method can better capture the driving intention of the vehicle and the future trajectory change trend.

[0034] Preferably, in the present invention, the trajectory prediction process based on longitudinal and lateral driving intention classification includes encoding and decoding vehicle features using MLP and LSTM layers, and finally outputting the trajectory prediction. Through this process, the method can more accurately predict the future driving trajectory of the vehicle, providing more reliable data support for the intelligent transportation system. At the same time, the implementation process of the method also has high flexibility and scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Schematic diagrams of each module in the vehicle trajectory prediction method based on interactive potential field and graph convolutional network provided by the embodiment of the present invention; Figure 2 Driving scenario graph provided by the embodiment of the present invention; Figure 3 Temporal feature extraction model diagram provided by the embodiment of the present invention; Figure 4 Interactive feature extraction model diagram provided by the embodiment of the present invention; Figure 5 Longitudinal and lateral interactive potential field diagram provided by the embodiment of the present invention; Figure 6 Decoder model diagram provided by the embodiment of the present invention; Figure 7 Flowchart of a vehicle trajectory prediction method based on interactive potential field and graph convolutional network provided by the embodiment of the present invention; Figure 8 Flowchart of a vehicle trajectory prediction method based on interactive potential field and graph convolutional network provided by the present invention; Figure 9 Structural schematic diagram of a vehicle trajectory prediction system based on interactive potential field and graph convolutional network provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] To further understand the content of the present invention, the following describes the present invention in detail with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments are only for explaining the present invention and not for limiting it.

[0037] The following explains the technical terms involved in the present invention: Bidirectional Long Short-Term Memory (BiLSTM) is a special recurrent neural network (RNN) designed specifically for processing sequence data. By combining the outputs of two long short-term memory networks (LSTM), one forward and one backward, it can capture bidirectional dependencies in the sequence, thereby considering both past and future context information when processing sequence data.

[0038] The Graph Convolutional Network (GCN) is a neural network model that can process data with a generalized topological graph structure. It combines the topological structure of the graph with node features and extracts the relationships and feature information between nodes through graph convolution operations. Compared with the traditional Convolutional Neural Network (CNN), GCN can directly operate on irregular graph structures without converting them into regular grid data.

[0039] The MLP layer (Multi-Layer Perceptron layer) is an artificial neural network model composed of an input layer, one or more hidden layers, and an output layer, with fully connected neurons between each layer.

[0040] According to the background technology, vehicle trajectory prediction methods can be mainly divided into three types: model-based methods, reinforcement learning-based methods, and deep learning-based methods. Model-based methods construct vehicle kinematic and dynamic models and combine methods such as Kalman filtering for trajectory prediction, but only consider the influence of the vehicle itself on the trajectory, resulting in large long-term trajectory prediction errors. Most reinforcement learning and deep learning-based methods use black-box models to represent vehicle interactions and lack interpretability. Among them, graph neural networks construct nodes and edges to represent the influence relationships between vehicles, have good robustness and interpretability, and have become a current research hotspot. However, the adjacency matrix constructed based on the distance between vehicles can only describe the intensity of interaction influence, cannot explain the directionality of interaction influence, and ignores the influence of vehicle movement on the interaction, resulting in low trajectory prediction accuracy.

[0041] To solve the above problems, this embodiment provides a vehicle trajectory prediction method based on an interaction potential field and a graph convolutional network, which accurately describes the complex and random interaction relationships between vehicles and improves the vehicle trajectory prediction accuracy. In this prediction method, a pre-constructed trajectory time-series feature extraction model (a time-series feature extraction model of a bidirectional long short-term memory network based on an attention mechanism), a vehicle interaction feature extraction model, and a decoder model are adopted; the trajectory time-series feature extraction model comprehensively captures the context features of vehicle historical trajectory information using a Bi-LSTM combined with an attention mechanism. The vehicle interaction feature extraction model analyzes the longitudinal and lateral interaction relationships between vehicles based on an artificial potential field, constructs a longitudinal and lateral influence matrix, and extracts vehicle interaction features through a double-layer graph convolution. Finally, an LSTM-based decoder model is designed to predict the future trajectory of the vehicle by combining time-series features, interaction features, and driving intentions. The present invention has good trajectory prediction effects in sparse and dense traffic scenarios.

[0042] Combined with Figure 1 and Figure 7As shown in the figure, this embodiment provides a vehicle trajectory prediction method based on an interaction potential field and a graph convolutional network. The specific steps are as follows: Use Bi-LSTM to extract the temporal features of the vehicle's historical trajectory from both the forward and reverse directions, and use the attention mechanism to assign different weights to the trajectory data at different times to emphasize the important information in the trajectory, obtaining the weighted temporal feature tensor S.

[0043] Use the artificial potential field to describe the longitudinal and lateral interaction relationships between vehicles, introduce a correction term to achieve directional differences, and use a two-layer graph convolution to extract the longitudinal and lateral interaction features H 1 and H 2 .

[0044] Combine the vehicle trajectory temporal features and interaction features to classify the vehicle driving intention, and then decode the vehicle feature tensor based on LSTM to predict the vehicle's future trajectory .

[0045] Exemplarily, combined with Figures 2 to 6 , the more specific steps of the prediction method provided in this embodiment are as follows: On the first hand, this embodiment establishes a Bi-LSTM temporal feature extraction model based on the attention mechanism, that is.

[0046] Step 1, as shown in the driving scenario Figure 2 , the vehicle group includes the target vehicle (TV) and its surrounding environmental vehicles , where the environmental vehicles include the front vehicle (FV), the rear vehicle (RV), the left front vehicle (LFV), the left neighboring vehicle (LNV), the left rear vehicle (LRV), the right front vehicle (RFV), the right neighboring vehicle (RNV), and the right rear vehicle (RRV). Use the trajectory features of the vehicle within the observation horizon as the input.

[0047]

[0048] The trajectory features of the vehicle at t time are:

[0049] Among them, , , are the longitudinal and lateral positions, longitudinal and lateral speeds, and longitudinal and lateral accelerations of the vehicle i at t time.

[0050] Within the observation horizon, the vehicle historical trajectory data is:

[0051] The vehicle inherent characteristics are defined as:

[0052] Among them, N is the number of vehicles, d is the number of types of inherent characteristics.

[0053] Step 2, the structure of the time series feature extraction model is as Figure 3 shown. Import the vehicle historical trajectory data into Bi-LSTM to extract the forward and backward time series features of the historical trajectory. The working process of LSTM is as follows:

[0054] Among them, is the hidden vector of the trajectory at time is the activation function, respectively represent the input gate, forget gate, and output gate, is the memory cell, is the bias vector, is the weight matrix.

[0055] Combining the propagation processes of forward and backward LSTM, the time series features of the vehicle trajectory extracted by Bi-LSTM are:

[0056] Among them, are forward and backward propagation respectively, are the features at time t extracted by forward and backward LSTM respectively, is the output of Bi-LSTM, represents the concatenation operation.

[0057] Step 3, use the attention mechanism for the output of Bi-LSTM, obtain the hidden features of the trajectory data at each time through a linear layer, and normalize them using the SoftMax layer to obtain the corresponding weights. Finally, perform a weighted sum on the time series features of the vehicle trajectory, and pass through layer to obtain the vehicle time series feature tensor S :

[0058] Among them, are the parameters to be trained, is t the hidden feature of the trajectory data at time is the attention weight of the trajectory data at time is the vehicle weighted feature tensor, h is the tensor length, that is, the number of Bi-LSTM hidden layers.

[0059] In a second aspect, the present invention designs a vehicle interaction feature extraction model based on an interaction potential field and a graph convolutional network, and its structure is as shown in Figure 4 as follows.

[0060] Step 1: Combining the artificial potential field theory, taking the vehicle as the potential field source, the closer the vehicles are to each other, the greater the interaction influence intensity between the vehicles. An interaction potential field is established as follows:

[0061] where, is the potential field intensity of vehicle j For vehicle i ; is the potential field energy; is a parameter related to the influence intensity; is vehicle i and vehicle j ; is the correction term.

[0062] Step 2: For the convenience of analysis, the vehicle interaction is divided into longitudinal interaction and lateral interaction.

[0063] The specific scenario is as shown in Figure 5 . For the longitudinal interaction of vehicles, the longitudinal interaction potential field energy of the vehicle is obtained according to the vehicle length and longitudinal vehicle speed:

[0064] where, are respectively the length and longitudinal speed of vehicle ; is vehicle and vehicle ; is vehicle ;

[0065] Step 3: To characterize the directional difference in the interaction between vehicles, the longitudinal correction term is composed of a longitudinal traffic flow difference correction term and a lateral difference correction term as follows:

[0066] For the longitudinal traffic flow difference correction term, it is used to describe the longitudinal influence difference caused by the longitudinal traffic flow. For example, for vehicles in the same lane, due to the flow characteristics of the traffic flow, the vehicle in front can be regarded as an obstacle to the vehicle behind, and at the same time, the vehicle in front has the right of way of the lane and is less affected by the vehicle behind. Therefore, there is a difference in the interaction influence intensity between the front and rear vehicles. The longitudinal traffic flow difference correction term is as follows:

[0067] Among them, is a parameter related to the correction intensity of the longitudinal traffic flow difference, is a constant, representing the inherent difference in the influence of the front and rear vehicles in the traffic flow, represents the longitudinal relative speed of the vehicle. For the lateral difference correction term, it is used to describe the longitudinal influence difference caused by the lateral distance of different lanes. For example, for vehicles in different lanes, the driver will avoid too close longitudinal distances through acceleration and deceleration operations, indicating that there is also a longitudinal interaction influence among vehicles in different lanes, but it is obviously weaker compared to vehicles in the same lane. The lateral difference correction term is as follows:

[0068] Among them, is a parameter related to the correction intensity of the lateral difference in longitudinal interaction. Step 4, for vehicle lateral interaction, define the vehicle lateral interaction potential field energy as:

[0069] Among them, is a constant, representing the potential lateral interaction energy between vehicles, is the vehicle and the vehicle 's lateral relative distance, is the lateral interaction range of the vehicle .

[0070] As Figure 5 shown, assume the vehicle lateral speed direction as the lateral traffic flow direction and use lateral traffic flow to describe. Different from the overall flow of the longitudinal traffic flow, the lateral traffic flow is only the flow of a single vehicle and is accompanied by the encroachment of the target lane. Therefore, the vehicle in the lateral traffic flow direction is subject to stronger interaction influence. In the lateral interaction potential field, add correction terms for lateral traffic flow difference and longitudinal difference respectively:

[0071] Among them, is the lateral correction term, is the lateral traffic flow difference correction term, is the longitudinal difference correction term, and are parameters related to the correction intensity of the lateral traffic flow difference and longitudinal difference in vehicle lateral interaction, is the vehicle lateral relative speed.

[0072] In summary, the longitudinal and lateral influence potential field values between vehicles are respectively:

[0073] Among them, , are parameters related to the longitudinal and lateral influence potential field strengths respectively.

[0074] Step 4, through the above artificial potential field method, the interactive influence received by the vehicle i can be expressed as:

[0075] Among them, is the longitudinal and lateral influence vector of the vehicle , are respectively the longitudinal and lateral influence potential field values of the vehicle i affected by the vehicle j .

[0076] After normalization, the longitudinal and lateral influence weights of the vehicle are obtained, and the longitudinal and lateral influence matrices and are constructed as follows:

[0077] Among them, is the weight vector of the longitudinal and lateral influence received by vehicle i, including the longitudinal weight and the lateral weight .

[0078] Step 5, the vehicle trajectory time series feature tensor and the vehicle inherent feature are concatenated to form a new feature tensor , which is used as the input of the upper-layer graph convolution module to extract the vehicle longitudinal interaction feature , and the output is used as the input of the lower-layer graph convolution module. The lower-layer graph convolution extracts the vehicle lateral interaction feature , realizing the hierarchical aggregation of the vehicle longitudinal and lateral interaction features, as follows:

[0079] Among them, is the GCN layer, are respectively the vehicle longitudinal and lateral interaction feature tensors obtained by graph convolution, d is the number of vehicle inherent features, is activation function. In this embodiment, the vehicle inherent features include vehicle length, vehicle type, drive form, etc., and can be directly obtained from the vehicle calibration parameters.

[0080] The graph convolution is as follows:

[0081] Among them, the node influence matrix , is the degree matrix of each node, represents the node features after the -th layer of graph convolution,

[0082] Thirdly, the present invention establishes a decoder model considering vehicle driving intention, and its network structure is as Figure 6 shown.

[0083] Step 1: Concatenate the vehicle feature tensor and the longitudinal and lateral interaction feature tensors of the vehicle to form a new feature tensor as the input of the trajectory prediction model to comprehensively reflect the temporal and interaction features of the vehicle trajectory, as shown in the following formula:

[0084] Step 2: Take as the input, and use layers to perform dimensionality reduction encoding on the vehicle feature tensor to extract the implicit features of the driving intention. Then, capture the longitudinal and lateral driving intention probabilities through and layers respectively, and use the Softmax function to generate the longitudinal and lateral driving intention classifications, as shown in the following formula:

[0085] where is the vehicle feature tensor after dimensionality reduction, are the longitudinal and lateral driving intentions respectively. The longitudinal and lateral driving intentions include three longitudinal driving intentions of acceleration, deceleration, and constant speed, and three lateral driving intentions of left lane change, right lane change, and keeping lane.

[0086] Step 3: Combine the longitudinal and lateral driving intentions , and respectively use and layers to encode and decode the vehicle features, and finally predict the future trajectory of the vehicle by layer, as shown in the following formula:

[0087] where is the vehicle feature tensor decoded by layer, is the parameter matrix of layer, is the predicted trajectory, is the predicted trajectory at time respectively represent the predicted vehicle at the longitudinal and lateral trajectories at a moment, which is the future horizon.

[0088] Exemplarily, as Figure 8 shown, this embodiment provides a vehicle trajectory prediction method based on an interaction potential field and a graph convolutional network, including the following steps: Based on the vehicle historical trajectory data, obtain the vehicle time series features; After splicing the vehicle time series features and the vehicle inherent features, input them into a pre-constructed vehicle interaction feature extraction model based on an interaction potential field and a graph convolutional network to obtain the longitudinal and lateral vehicle interaction features; the interaction potential field is used to characterize the vehicle interaction influence relationship; Input the time series interaction fusion features after splicing the vehicle time series features, the vehicle inherent features, and the longitudinal and lateral vehicle interaction features into a pre-constructed decoder model, and output the vehicle prediction trajectory.

[0089] In this embodiment, the obtaining the vehicle time series features based on the vehicle historical trajectory data includes: Construct a vehicle group with the target vehicle and the surrounding environmental vehicles; In the vehicle group, obtain the vehicle historical trajectory data of the vehicle in the observation horizon; the trajectory features include the vehicle historical position, speed, and acceleration; Input the vehicle historical trajectory data into a pre-constructed time series feature extraction model of a bidirectional long short-term memory network based on an attention mechanism to obtain the vehicle trajectory time series features.

[0090] In this embodiment, the inputting the vehicle historical trajectory data into a pre-constructed time series feature extraction model of a bidirectional long short-term memory network based on an attention mechanism to obtain the vehicle trajectory time series features includes: Input the vehicle historical trajectory data into a pre-constructed time series feature extraction model of a bidirectional long short-term memory network based on an attention mechanism; Combined with the propagation processes of the forward and backward LSTMs, extract the vehicle trajectory time series features. The specific formula is as follows:

[0091] In the formula, are the forward and backward propagations respectively, are the features at time t extracted by the forward and backward LSTMs respectively, is the trajectory feature at time t, is the output of the Bi-LSTM, represents the concatenation operation, and Bi-LSTM represents the time series feature extraction model of the bidirectional long short-term memory network; Based on the temporal characteristics of vehicle trajectories, the weights at each moment are calculated through a linear layer and SoftMax, and the vehicle temporal characteristics are obtained through an MLP after weighted summation. The specific formula is as follows:

[0092] In the formula, are the parameters to be trained, is t the hidden feature of the trajectory data at time is the attention weight of the trajectory data at time is the vehicle weighted feature tensor, N is the number of vehicles, h is the number of hidden layers of the Bi-LSTM.

[0093] In this embodiment, before splicing the vehicle temporal characteristics and the vehicle inherent characteristics and inputting them into the pre-constructed vehicle interaction feature extraction model based on the interaction potential field and the graph convolutional network to obtain the longitudinal and lateral vehicle interaction characteristics, it includes: Construct a vehicle interaction feature extraction model based on the interaction potential field and the graph convolutional network, specifically as follows: Regarding the vehicle as a potential field source, the longitudinal interaction potential field energy of the vehicle is obtained according to the vehicle length and the longitudinal vehicle speed. The specific formula is as follows:

[0094] In the formula, is the longitudinal potential field strength of vehicle j For vehicle i ; is a parameter related to the longitudinal influence strength, is the longitudinal correction term; The longitudinal interaction potential field energy of the vehicle is established according to the vehicle length and the longitudinal vehicle speed. The specific formula is as follows:

[0095] In the formula, are respectively the length and the longitudinal speed of vehicle ; is the longitudinal relative distance between vehicle and vehicle ; is the longitudinal interaction range of vehicle ; The longitudinal traffic flow difference correction term , which is used to characterize the directional difference of vehicle interaction. The longitudinal traffic flow difference correction term is expressed as:

[0096] In the formula, is a parameter related to the correction intensity of longitudinal traffic flow differences, is a constant, representing the inherent influence difference between the front and rear vehicles in longitudinal traffic flow, represents the longitudinal relative speed of the vehicle, respectively represent the vehicle i and the vehicle j longitudinal positions; The lateral difference correction term , used to characterize the longitudinal influence difference caused by the lateral distance of different lanes. The lateral difference correction term is expressed as:

[0097] In the formula, is a parameter related to the correction intensity of the lateral difference in longitudinal interaction, respectively are the i and the vehicle j lateral positions; Based on the longitudinal traffic flow difference correction term and the lateral difference correction term, the longitudinal interaction potential field correction term is obtained, and is expressed as:

[0098] Establish a lateral interaction potential field, and obtain the lateral interaction potential field energy of the vehicle according to the vehicle length and the lateral vehicle speed, specifically as follows:

[0099] In the formula, is the lateral potential field intensity of the vehicle j For the vehicle i , is a parameter related to the lateral influence intensity, is the lateral correction term, is a constant, representing the potential lateral interaction energy between vehicles, is the vehicle and the lateral relative distance between vehicle j, is the lateral interaction range of the vehicle ; Among them, the lateral correction term includes the lateral traffic flow difference correction term and the longitudinal difference correction term , as follows:

[0100] In the formula, and are parameters related to the correction intensity, is the lateral relative speed of the vehicle; The interaction influence received by vehicle i is expressed as:

[0101] In the formula, is the longitudinal and lateral influence vector of the vehicle respectively, and are the longitudinal and lateral influence potential field values of the vehicle i when the vehicle j is affected by the vehicle Normalize the longitudinal and lateral influence vectors to obtain the longitudinal and lateral influence weights of the vehicle and construct the longitudinal and lateral influence matrix.

[0102] In this embodiment, the step of splicing the vehicle temporal features and the vehicle inherent features and inputting them into a pre-constructed vehicle interaction feature extraction model based on the interaction potential field and the graph convolutional network to obtain the longitudinal and lateral vehicle interaction features includes: Splice the vehicle temporal features and the vehicle inherent features to obtain the spliced features; Use the spliced features as the input of the vehicle interaction feature extraction model based on the interaction potential field and the graph convolutional network, and output the longitudinal and lateral vehicle interaction features.

[0103] In this embodiment, the step of inputting the temporal interaction fusion features obtained by splicing the vehicle temporal features, the vehicle inherent features and the longitudinal and lateral vehicle interaction features into a pre-constructed decoder model and outputting the vehicle prediction trajectory includes: Splice the vehicle temporal features, the vehicle inherent features and the longitudinal and lateral vehicle interaction features to obtain the temporal interaction fusion features; Input the temporal interaction fusion features into the pre-constructed decoder model, and use layers to perform dimensionality reduction encoding on the temporal interaction fusion features and extract the implicit features of the driving intention; Through and layers, capture the longitudinal and lateral driving intention probabilities respectively, and use the Softmax function to generate the longitudinal and lateral driving intention classifications; Based on the longitudinal and lateral driving intention classifications, use and layers to encode and decode the vehicle features respectively, and finally use layer to predict the future trajectory of the vehicle as the vehicle prediction trajectory output.

[0104] In this embodiment, the step of using and layers to capture the longitudinal and lateral driving intention probabilities respectively and use the Softmax function to generate the longitudinal and lateral driving intention classifications, the specific formula is as follows:

[0105] In the formula, is the vehicle feature tensor after dimensionality reduction, is the time - series interaction fusion feature, are the longitudinal and lateral driving intentions respectively; Based on the classification of longitudinal and lateral driving intentions, and layers are used to encode and decode the vehicle features. Finally, the layer predicts the future trajectory of the vehicle as the output of the vehicle prediction trajectory. The specific formula is as follows:

[0106] In the formula, is the vehicle feature tensor decoded by the layer, is the parameter matrix of the layer, is the predicted trajectory, is the predicted trajectory at the moment, are the predicted longitudinal and lateral positions of the vehicle at the moment respectively, is the observation horizon, is the future horizon.

[0107] As shown in Figure 9 , this embodiment also provides a vehicle trajectory prediction system based on an interaction potential field and a graph convolutional network, including: a vehicle time - series feature extraction module, which is used to obtain vehicle time - series features based on vehicle historical trajectory data; a longitudinal and lateral interaction feature extraction module, which is used to splice the vehicle time - series features and vehicle inherent features and input them into a pre - constructed vehicle interaction feature extraction model based on an interaction potential field and a graph convolutional network to obtain vehicle longitudinal and lateral interaction features; the interaction potential field is used to characterize the vehicle interaction influence relationship; a decoding module, which is used to input the time - series interaction fusion feature spliced from the vehicle time - series features, vehicle inherent features and vehicle longitudinal and lateral interaction features into a pre - constructed decoder model to output the vehicle prediction trajectory.

[0108] The present invention also provides a device, including: a memory for storing a computer program; a processor for implementing the steps of the vehicle trajectory prediction method based on an interaction potential field and a graph convolutional network when executing the computer program.

[0109] When the processor executes the computer program, it implements the steps of the above vehicle trajectory prediction based on the interaction potential field and graph convolutional network. For example: based on the vehicle historical trajectory data, obtain the vehicle temporal features; after splicing the vehicle temporal features and the vehicle inherent features, input them into a pre-constructed vehicle interaction feature extraction model based on the interaction potential field and graph convolutional network to obtain the longitudinal and lateral vehicle interaction features; the interaction potential field is used to characterize the vehicle interaction influence relationship; input the temporal interaction fusion features after splicing the vehicle temporal features, the vehicle inherent features and the longitudinal and lateral vehicle interaction features into a pre-constructed decoder model to output the vehicle predicted trajectory.

[0110] Alternatively, when the processor executes the computer program, it implements the functions of each module in the above system.

[0111] Exemplarily, the computer program can be divided into one or more modules / units. The one or more modules / units are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of completing preset functions, and the instruction segments are used to describe the execution process of the computer program in the vehicle trajectory prediction device based on the interaction potential field and graph convolutional network. For example, the computer program can be divided into a vehicle temporal feature extraction module, a longitudinal and lateral interaction feature extraction module, and a decoding module; the specific functions of each module are as follows: the vehicle temporal feature extraction module is used to obtain the vehicle temporal features based on the vehicle historical trajectory data; the longitudinal and lateral interaction feature extraction module is used to input the spliced vehicle temporal features and the vehicle inherent features into a pre-constructed vehicle interaction feature extraction model based on the interaction potential field and graph convolutional network to obtain the longitudinal and lateral vehicle interaction features; the interaction potential field is used to characterize the vehicle interaction influence relationship; the decoding module is used to input the temporal interaction fusion features after splicing the vehicle temporal features, the vehicle inherent features and the longitudinal and lateral vehicle interaction features into a pre-constructed decoder model to output the vehicle predicted trajectory.

[0112] The vehicle trajectory prediction device based on the interaction potential field and graph convolutional network can be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The vehicle trajectory prediction device based on the interaction potential field and graph convolutional network may include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the above are examples of the vehicle trajectory prediction device based on the interaction potential field and graph convolutional network, and do not constitute a limitation on the vehicle trajectory prediction device based on the interaction potential field and graph convolutional network. It may include more components than the above, or combine some components, or different components. For example, the vehicle trajectory prediction device based on the interaction potential field and graph convolutional network may further include an input / output device, a network access device, a bus, etc.

[0113] The so-called processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The processor is the control center of the vehicle trajectory prediction based on the interaction potential field and graph convolutional network, and connects various parts of the entire vehicle trajectory prediction device based on the interaction potential field and graph convolutional network through various interfaces and lines.

[0114] The memory can be used to store the computer programs and / or modules. By running or executing the computer programs and / or modules stored in the memory, and by calling the data stored in the memory, the processor realizes various functions of the vehicle trajectory prediction device based on the interaction potential field and graph convolutional network.

[0115] The memory may mainly include a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the mobile phone (such as audio data, phone book, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0116] The present invention also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the steps of the vehicle trajectory prediction method based on the interaction potential field and graph convolutional network are realized.

[0117] If the modules / units integrated in the vehicle trajectory prediction system based on the interaction potential field and graph convolutional network are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0118] Based on such understanding, all or part of the processes in the above-mentioned vehicle trajectory prediction method based on the interaction potential field and graph convolutional network of the present invention can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-mentioned vehicle trajectory prediction method based on the interaction potential field and graph convolutional network can be realized. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file or preset intermediate form, etc.

[0119] The computer-readable storage medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0120] It should be noted that the content included in the computer-readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice within the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable storage medium does not include electrical carrier signals and telecommunication signals.

[0121] The present invention provides a vehicle trajectory prediction method based on the interaction potential field and graph convolutional network, which has the following advantages: First, in the historical trajectory time series feature extraction model, due to the adoption of Bi-LSTM and the attention mechanism, the model can extract comprehensive historical trajectory context features. This benefits from the bidirectional information transfer mechanism of Bi-LSTM and the fact that the attention mechanism can pay attention to the importance differences of historical trajectory data at different times.

[0122] Second, in the interaction feature extraction model, by using the artificial potential field to construct the longitudinal and transverse influence matrix of the graph convolutional network and capturing the longitudinal and transverse interaction features of the vehicle through the stacked graph convolutional layers, the interpretability of the model is improved, and the double-layer graph convolutional can extract more comprehensive vehicle interaction feature information.

[0123] Third, in the decoder model, the classification of vehicle driving intentions is incorporated, and the trajectory time series features, longitudinal and transverse interaction features, and intention features are combined as the decoder input, which can improve the accuracy of vehicle trajectory prediction.

[0124] In summary, the present prediction method is based on a temporal feature extraction model of a bidirectional long short-term memory network (Bi-LSTM) with an attention mechanism, comprehensively extracts the temporal features in the vehicle historical trajectory data, and encodes the vehicle with a feature tensor. Secondly, based on an interaction feature extraction model of a graph convolutional network, the vehicle network is constructed as an interaction graph, an artificial potential field is introduced to construct a longitudinal and lateral influence matrix to characterize the vehicle interaction influence relationship, and a double-layer graph convolutional module is used to extract the vehicle interaction features. Finally, the decoder model fuses the vehicle temporal and interaction features, classifies the vehicle driving intention, and uses an LSTM network to decode the vehicle feature tensor to predict the future trajectory of the vehicle. It solves the technical problem that the existing trajectory prediction methods cannot accurately represent the complex interaction relationship between vehicles, and further improves the vehicle trajectory prediction accuracy and the interpretability of the deep learning model.

[0125] The above embodiments are only one of the implementation manners capable of implementing the technical solution of the present invention. The scope of protection required by the present invention is not limited only by this embodiment, but also includes any changes, substitutions and other implementation manners that are easily conceivable by those skilled in the art within the technical scope disclosed by the present invention.

[0126] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solution of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: it is still possible to modify the specific implementation manners of the present invention or make equivalent substitutions, and any modification or equivalent substitution without departing from the spirit and scope of the present invention shall be covered by the protection scope of the claims of the present invention.

Claims

1. A vehicle trajectory prediction method based on interactive potential field and graph convolutional network, characterized in that: include: Based on the historical trajectory data of the vehicle, obtain the time series characteristics of the vehicle; The vehicle time series features and the vehicle inherent features are spliced ​​and input into a pre-built vehicle interaction feature extraction model based on an interaction potential field and a graph convolutional network to obtain the longitudinal and lateral interaction features of the vehicle; the interaction potential field is used to characterize the interaction influence relationship of the vehicle; The time-series interaction fusion features obtained by splicing the vehicle's time-series features, vehicle inherent features, and vehicle longitudinal and lateral interaction features are input into the pre-built decoder model to output the vehicle's predicted trajectory.

2. The vehicle trajectory prediction method based on interactive potential field and graph convolutional network according to claim 1 is characterized in that: The obtaining of vehicle time series characteristics based on vehicle historical trajectory data includes: The target vehicle and the surrounding vehicles around the target vehicle are constructed into a vehicle group; In a vehicle group, obtaining vehicle historical trajectory data of a vehicle in an observation field of view; the trajectory features include the vehicle historical position, speed and acceleration; The vehicle historical trajectory data is input into the pre-built bidirectional long short-term memory network temporal feature extraction model based on the attention mechanism to obtain the vehicle trajectory temporal features.

3. The vehicle trajectory prediction method based on interactive potential field and graph convolutional network according to claim 2 is characterized in that: The vehicle historical trajectory data is input into a pre-built bidirectional long short-term memory network temporal feature extraction model based on an attention mechanism to obtain the vehicle trajectory temporal features, including: Input the vehicle historical trajectory data into the pre-built bidirectional long short-term memory network temporal feature extraction model based on the attention mechanism; Combined with the forward and reverse LSTM propagation process, the vehicle trajectory time series features are extracted. The specific formula is as follows: In the formula, are forward and backward propagation, respectively. They are the features at time t extracted by the forward and reverse LSTM respectively. is the trajectory characteristic at time t, is the output of Bi-LSTM, represents the connection operation, Bi-LSTM represents the bidirectional long short-term memory network temporal feature extraction model; Based on the time series characteristics of vehicle trajectories, the weights of each moment are calculated through the linear layer and SoftMax, and the vehicle time series characteristics are obtained after weighted summation through MLP. The specific formula is as follows: In the formula, are the parameters that need to be trained, for t Hidden features of moment-to-moment trajectory data; for The attention weight of the moment trajectory data, is the vehicle weighted feature tensor, N is the number of vehicles, h is the number of Bi-LSTM hidden layers.

4. The vehicle trajectory prediction method based on interactive potential field and graph convolutional network according to claim 1 is characterized in that: The vehicle time series features and the vehicle inherent features are spliced ​​and input into a pre-built vehicle interaction feature extraction model based on an interactive potential field and a graph convolutional network to obtain the longitudinal and lateral interaction features of the vehicle, including: Construct a vehicle interaction feature extraction model based on interactive potential field and graph convolutional network, as follows: Taking the vehicle as the potential field source, the longitudinal interaction potential field energy of the vehicle is obtained according to the vehicle length and longitudinal speed. The specific formula is as follows: In the formula, For vehicles j For vehicles i The longitudinal potential field strength, is a parameter related to the longitudinal impact intensity, is the longitudinal correction term; The longitudinal interaction potential field energy of the vehicle is established according to the vehicle length and longitudinal speed. The specific formula is as follows: In the formula, For vehicles length and longitudinal velocity, For vehicles and vehicles The longitudinal relative distance For vehicles The vertical interaction range; The longitudinal traffic flow difference correction term , which is used to characterize the directional differences in vehicle interactions, and the longitudinal traffic flow difference correction term is expressed as: In the formula, is a parameter related to the correction intensity of longitudinal traffic flow differences, is a constant, representing the inherent difference in the impact of the front and rear vehicles in the longitudinal traffic flow, represents the longitudinal relative velocity of the vehicle, Respectively represent vehicles i and vehicles j The longitudinal position of The lateral difference correction term , which is used to characterize the longitudinal impact differences caused by different lane lateral distances. The lateral difference correction term is expressed as: In the formula, is a parameter related to the strength of the horizontal difference correction for the vertical interaction, For vehicles i and vehicles j lateral position; The longitudinal interaction potential field correction term is obtained based on the longitudinal traffic flow difference correction term and the lateral difference correction term. , expressed as: A lateral interaction potential field is established, and the energy of the vehicle's lateral interaction potential field is obtained according to the vehicle length and lateral speed, as follows: In the formula, For vehicles j For Vehicles i The transverse potential field strength, is a parameter related to the lateral impact strength, is the lateral correction term, is a constant representing the potential lateral interaction energy between vehicles, For vehicles The lateral relative distance from vehicle j, For vehicles Horizontal interaction range; Among them, the lateral correction term Includes correction for lateral traffic flow differences and the longitudinal difference correction , as follows: In the formula, and is a parameter related to the correction strength, is the lateral relative speed of the vehicle; The interaction effect on vehicle i is expressed as: In the formula, For vehicles The vertical and horizontal influence vectors, For vehicles i Vehicles j The longitudinal and transverse influence potential field values; The longitudinal and lateral influence vectors are normalized to obtain the longitudinal and lateral influence weights of the vehicle and to construct the longitudinal and lateral influence matrix.

5. The vehicle trajectory prediction method based on interactive potential field and graph convolutional network according to claim 1 is characterized in that: The vehicle time series features and the vehicle inherent features are spliced ​​and input into a pre-built vehicle interaction feature extraction model based on an interactive potential field and a graph convolutional network to obtain the longitudinal and lateral interaction features of the vehicle, including: Splicing the vehicle time series features with the vehicle inherent features to obtain splicing features; The splicing features are used as the input of the vehicle interaction feature extraction model based on interaction potential field and graph convolutional network, and the vehicle longitudinal and lateral interaction features are output.

6. The vehicle trajectory prediction method based on interactive potential field and graph convolutional network according to claim 1 is characterized in that: The time-series interaction fusion features obtained by splicing the vehicle time-series features, the vehicle inherent features and the vehicle longitudinal and lateral interaction features are input into the pre-built decoder model to output the vehicle prediction trajectory, including: The vehicle time series features, vehicle inherent features and vehicle longitudinal and lateral interaction features are spliced ​​to obtain the time series interaction fusion features; The temporal interaction fusion features are input into the pre-built decoder model using The layer performs dimensionality reduction encoding on the temporal interaction fusion features to extract the implicit features of driving intention; pass and The layers capture the longitudinal and lateral driving intention probabilities respectively, and use the Softmax function to generate longitudinal and lateral driving intention classifications; Based on the longitudinal and lateral driving intention classification, and The layer encodes and decodes the vehicle features, and finally The layer predicts the future trajectory of the vehicle as the vehicle prediction trajectory output.

7. The vehicle trajectory prediction method based on interactive potential field and graph convolutional network according to claim 6 is characterized in that: Said through and The layers capture the longitudinal and lateral driving intention probabilities respectively, and use the Softmax function to generate the longitudinal and lateral driving intention classification. The specific formula is as follows: In the formula, is the vehicle feature tensor after dimension reduction, is the temporal interactive fusion feature. They are longitudinal and lateral driving intentions respectively; The longitudinal and lateral driving intention classification is respectively adopted and The layer encodes and decodes the vehicle features, and finally The layer predicts the future trajectory of the vehicle as the vehicle prediction trajectory output. The specific formula is as follows: In the formula, for The vehicle feature tensor after layer decoding, for The parameter matrix of the layer, for The predicted trajectory at the moment, The predicted vehicles are exist The vertical and horizontal position at the moment, To observe the horizon, For the future vision.

8. A vehicle trajectory prediction system based on interactive potential field and graph convolutional network, characterized in that: include: A vehicle time series feature extraction module is used to obtain vehicle time series features based on vehicle historical trajectory data; A longitudinal and lateral interaction feature extraction module is used to splice the vehicle time series features and the vehicle inherent features and input them into a pre-built vehicle interaction feature extraction model based on an interaction potential field and a graph convolutional network to obtain the longitudinal and lateral interaction features of the vehicle; the interaction potential field is used to characterize the interaction influence relationship of the vehicle; The decoding module is used to input the time-series interaction fusion features obtained by splicing the vehicle's time-series features, the vehicle's inherent features, and the vehicle's longitudinal and lateral interaction features into a pre-built decoder model, and output the vehicle's predicted trajectory.

9. A device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the vehicle trajectory prediction method based on interactive potential field and graph convolutional network as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it is used to implement the steps of the vehicle trajectory prediction method based on interactive potential field and graph convolutional network as described in any one of claims 1 to 7.

Images