Driving behavior prediction method based on informer neural network under safety potential field

通过在驾驶行为预测中结合安全势场模型和informer神经网络,解决了现有技术中预测精度和安全性不足的问题,实现了更准确和可解释的驾驶行为预测,提升了自动驾驶车辆的实时 Sex and safety.

CN115547040BActive Publication Date: 2024-11-22HENAN UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211137976.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-09-19
Publication Date
2024-11-22
Estimated Expiration
2042-09-19

AI Technical Summary

Technical Problem

The existing driving behavior prediction methods have shortcomings in prediction accuracy and safety, especially in long-term trajectory prediction and real-time performance, and lack interpretability, which cannot meet the safety driving needs of autonomous vehicles.

Method used

Using the driving behavior prediction method based on the informer neural network under the security potential field, the security potential field model is combined with the informer model, and the potential space-time representation is input into the informer neural network for prediction, and the security potential is passed Field-range constraint optimization model to improve prediction accuracy and safety.

Benefits of technology

It improves the accuracy and safety of driving behavior prediction, enhances the interpretability of trajectory prediction, realizes accurate prediction of surrounding vehicle intentions and multi-channel trajectories, and improves the real-time and safety of autonomous vehicles.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115547040B_ABST
    Figure CN115547040B_ABST
Patent Text Reader

Abstract

The present invention proposes a driving behavior prediction method based on an informer neural network under a safety potential field, including preprocessing data in a motion data set to obtain trajectory sequence features; encoding the trajectory sequence features into a potential spatiotemporal representation; inputting the potential spatiotemporal representation into an informer neural network to predict the future driving intention and driving trajectory of the target vehicle; establishing safety potential field models in the following and lane changing situations respectively, and calculating the safety potential field range; constructing a joint loss function of the safety potential field range constraint, predicted driving intention and driving trajectory, and optimizing the prediction model using an Adam optimizer; and using a test set to predict the prediction model. The present invention organically combines the safety potential field model and the informer model, can realize the intention prediction of surrounding vehicles and multi-channel trajectory prediction, and has excellent performance in terms of real-time, accuracy and safety of driving behavior prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of driving behavior prediction of an autonomous driving vehicle, and in particular to a driving behavior prediction method based on an informer neural network under a safety potential field. Background Art

[0002] The intelligence of automobiles is developing rapidly. Although L2-level autonomous driving technology can automatically control steering and acceleration or deceleration in complex traffic environments with mixed human and unmanned driving, L2-level autonomous driving technology does not have the ability to understand the environment and is too conservative in the process of lane changing, resulting in its speed being fast and slow, making traffic flow inefficient and unable to drive or change lanes as smoothly and safely as human drivers. Natural drivers can correctly estimate whether the vehicle meets the lane change conditions and the driving risks faced during the lane change process based on driving experience, ensuring that the vehicle executes fast and safe lane change decisions. Autonomous driving technology above L3 requires that autonomous vehicles can accurately perceive and predict the behavior of surrounding road participants in order to plan safe and comfortable proactive actions, avoid conflicts with others due to emergency reactive decisions, and minimize possible risks.

[0003] In recent years, the methods used for driving behavior prediction can be divided into two categories: traditional dynamic model methods and deep learning methods. Traditional dynamic models are completely based on observed historical trajectories for prediction. Some works use Logistic regression, support vector machines or hidden Markov models to consider the driver's maneuvering intentions, while others use Kalman filters to explain the uncertainty in trajectory prediction. Such methods often have certain limitations in prediction accuracy. Deep learning methods such as recurrent neural networks (RNNs) and their variants, long short-term memory networks (LSTMs) and graph neural networks (GNNs) are used for driving behavior prediction, but these models, as a step-by-step reasoning paradigm, are often affected by dynamic decoding cumulative errors. The prediction accuracy decreases with the increase of the predicted trajectory length, and their memory constraints and gradient vanishing problems also affect the performance of driving behavior prediction. At the same time, these models lack interpretability considerations and only rely on the characteristics of the data to grasp the laws of the vehicle's driving behavior and then make predictions, which will bring certain safety hazards. It is crucial to solve the defects and deficiencies in traditional dynamic model methods and deep learning methods in the current field and improve the prediction accuracy of driving behavior while ensuring rationality and safety, which is crucial to ensure the safe driving of autonomous vehicles.

[0004] The Chinese invention patent with application number 202210279062.5 discloses a driving behavior prediction method and system based on a bidirectional long short-term memory network, which solves the technical problem of inaccurate prediction of the driver's driving behavior. The key point of its technical solution is to use mature on-board sensors to realize the function of predicting the driver's future longitudinal and lateral driving behavior in a short time window, with high prediction accuracy, and realize the prediction of the driver's future longitudinal and lateral driving behavior under a lightweight on-board processor, with low economic cost. It can widely serve the vehicle assisted driving system, improve the safety of the assisted driving system, enhance the auxiliary system's understanding of the driver, help to achieve safe driving, and has practicality and commercial value. At the same time, the deep learning method based on big data has high prediction accuracy, robustness and generalization ability, which is conducive to the in-depth development of advanced driving assistance systems and is widely applicable to driver behavior prediction under different vehicles. However, this method uses an inherent long short-term memory network and is affected by the cumulative error of dynamic decoding. The prediction accuracy decreases with the increase of the predicted trajectory length, and the prediction time increases with the increase of the predicted trajectory length. It cannot meet the accuracy and real-time requirements of long-term driving behavior prediction. Moreover, this method only relies on fuzzy reasoning of big data, without clear safety constraints, and cannot meet the safety and explainability of driving behavior prediction. Summary of the invention

[0005] In view of the technical problem that the existing driving behavior prediction methods have low prediction accuracy and cannot meet the needs, the present invention proposes a driving behavior prediction method based on an informer neural network under a safety potential field, which organically combines the safety potential field model and the informer model to improve the accuracy and safety of driving behavior prediction. The method is suitable for predicting the future intentions and multi-channel trajectories of surrounding vehicles during the driving process of an autonomous driving vehicle.

[0006] In order to achieve the above object, the technical solution of the present invention is implemented as follows: a driving behavior prediction method based on informer neural network under a safety potential field, characterized in that the steps are as follows:

[0007] S1: Preprocess the data in the motion dataset to obtain trajectory sequence features;

[0008] S2: Encode the trajectory sequence features obtained in step S1 into a latent spatiotemporal representation;

[0009] S3: Input the latent spatiotemporal representation obtained in step S2 into the informer neural network to predict the future driving intention and driving trajectory of the target vehicle;

[0010] S4: Considering the safety and feasibility of vehicle movement, safety potential field models are established for the following and lane changing situations, and the safety potential field range is calculated;

[0011] S5: construct a joint loss function of the safety potential field range constraint in step S4, the driving intention predicted in step S3, and the vehicle trajectory, and solve the loss function by gradient descent method, and use Adam optimizer to optimize the informer model containing the safety potential field to obtain a prediction model;

[0012] S6: Save the trained prediction model, use the test set to predict the prediction model, and obtain the future trajectory coordinates of the vehicle.

[0013] Preferably, the preprocessing method of step S1 is:

[0014] S1.1: Eliminate abnormal data in the motion dataset;

[0015] S1.2: Select the lateral and longitudinal coordinates, velocity, acceleration and lane number of the target vehicle in the motion data set as the features of the target vehicle, and select the lateral and longitudinal relative displacements of the vehicles around the target vehicle as the features of the surrounding vehicles;

[0016] S1.3: Extract the trajectory feature set of the target vehicle within the time k to k+n as S;

[0017] S1.4: Calculate the mean and variance of each channel of the multi-channel sequence S;

[0018] S1.5: Perform Z-Score normalization on each channel of the trajectory feature set S to obtain the final trajectory sequence feature

[0019] S1.6: Randomly divide the motion dataset into training and test sets.

[0020] Preferably, the Z-Score standardization method is:

[0021]

[0022] Among them, the mean variance Trajectory feature set S = (S k ,S k+1 ,S k+2 ,...,S k+n-1 ), where S k is the trajectory feature sequence of the target vehicle and its surrounding vehicles at time k, and the sequence S k ={x,y,v x ,v y ,a x ,a y ,id,Δx,Δy}, where x and y represent the horizontal and vertical coordinates of the target vehicle at time k, respectively, and v x ,vy Represent the lateral and longitudinal velocities of the target vehicle at time k, respectively, and a x ,a y They represent the lateral and longitudinal accelerations of the target vehicle at time k, id represents the lane where the target vehicle is located, and Δx and Δy represent the lateral and longitudinal relative displacements of the surrounding vehicles relative to the target vehicle.

[0023] Preferably, the abnormal data in step S1.1 include missing values ​​and values ​​with incorrect information recording in the data file; by setting a python program to scan all the data in the motion data set, the missing values ​​are screened out, and then the missing values ​​are replaced by the average value, maximum value, minimum value or probability estimate of the original data source; the values ​​with incorrect information recording are checked by logical rules, and the constraints between different attributes are used to eliminate the values ​​with incorrect information recording; the sequence S in step S1.4 k The 9 features in are calculated in parallel; in step S1.6, the training set and the test set are divided into a ratio of 7:3.

[0024] Preferably, the encoding in step S2 is implemented as follows:

[0025] S2.1: Use a one-dimensional convolutional layer of a neural network to encode the time frame of the trajectory feature set S into a potential vector representing the time attribute, and obtain a time encoding representation; use a fully connected layer of a neural network to encode the trajectory sequence features Represented as a spatiotemporal vector of linked context data, a feature coding representation is obtained; the relative position information of the trajectory sequence features input at each moment is positionally encoded to obtain a position coding representation;

[0026] S2.2: Perform matrix addition operation on the time coding representation, feature coding representation and position coding representation to obtain the potential spatiotemporal representation.

[0027] Preferably, the method for predicting the future driving intention and driving trajectory of the target vehicle in step S3 is:

[0028] S3.1: Input the latent spatiotemporal representation generated in step S2 into the encoder of the informer neural network, and obtain further trajectory feature representation E through the encoder's multi-head probabilistic sparse attention layer and fully connected layer;

[0029] S3.2: Input the latent spatiotemporal representation generated in step S2 into the decoder of the informer neural network, and obtain the trajectory feature representation M through the masked multi-head attention layer of the decoder;

[0030] S3.3: The trajectory feature representation E of step S3.1 and the trajectory feature representation M of step S3.2 are passed through the interactive multi-head probabilistic sparse attention layer and the fully connected layer to obtain the final output D of the informer neural network;

[0031] S3.4: Output D outputs the predicted driving intention through the intention prediction head, and output D outputs the predicted driving trajectory through the trajectory prediction head.

[0032] Preferably, the calculation process of the neural network fully connected layer is:

[0033]

[0034]

[0035] Among them, i refers to the number of sample indicators, j refers to the number of neurons in the middle layer; w ij represents the trainable weight, b j represents the trainable offset value, θ j represents the output of the fully connected layer, and f(x) represents the activation function;

[0036] The position encoding is expressed as:

[0037] PE(pos,2i)=sin(pos / 100002i / d model )

[0038] PE(pos,2i+1)=cos(pos / 100002i / d model )

[0039] Among them, pos represents the current moment, i represents the dimension of the encoding, and d model Designated as 512;

[0040] The calculation method of the S3.1 encoder is:

[0041]

[0042]

[0043]

[0044] in, Represent the trajectory sequence features The generated vector representation, W i and W o Represents different weight matrices; d k represents the feature dimension, softmax represents the probability mapping function, Attention represents the attention matrix calculation function, Concat represents the merging operation, head i represents the i-th attention head, and h represents the number of attention heads.

[0045] Preferably, the safety potential field model in the following car situation is:

[0046]

[0047] Where vx represents the real-time speed of the target vehicle, ξ and λ are adjustable hyperparameters, and a vpt represents the threshold for measuring whether the vehicle is in an accelerating state, (x, y) is the coordinate of the surrounding vehicles, (x 0 ,y 0 ) is the coordinate of the autonomous driving vehicle, σ x,1 and σ y Represent the variance of the two-dimensional Gaussian distribution; σ x,2 is the Rayleigh distribution of the attenuation process of the acceleration forward tilt effect in the x-axis direction The variance of; vx represents the real-time speed of the target vehicle, amax represents the constant of the vehicle's braking acceleration; a x Indicates the acceleration of the target vehicle on the x-axis;

[0048] The safety potential field model in the lane-changing situation is:

[0049]

[0050]

[0051] in, Indicates the steering angle of the vehicle; (x * ,y * ) is the deflection value corresponding to the original coordinate (x, y) after the turn; a y is the lateral acceleration of the target vehicle; s y The lateral displacement required for the vehicle to change lanes;

[0052] The trajectory sequence features obtained in step S1 are input into the safety potential field model of the target vehicle to calculate the safety potential field range of the target vehicle.

[0053] Preferably, the implementation method of step S5 is:

[0054] S5.1: Use the cross entropy function as the loss function for the predicted intent:

[0055]

[0056] in, is the probability distribution of the predicted intent, I∈{I s ,I l ,I r} represents the vehicle's true intention, i.e., going straight, turning left or right;

[0057] S5.2: Use the mean square error function as the loss function for the predicted trajectory:

[0058]

[0059] in, is the predicted future trajectory point, F is the actual future trajectory point, T pred Represents the number of predicted frames;

[0060] S5.3: Use the mean square error function as the loss function for the safety potential field range constraint:

[0061]

[0062] in, represents the range of the safety potential field under the predicted future state, and E v ∈{E vs ,E vc} represents the real safety potential field range;

[0063] S5.4: The final joint loss function is:

[0064]

[0065] Among them, ω 1 ,ω 2 and ω 3 is a hyperparameter;

[0066] S5.5: Adjust all hyperparameters ω in step S5.4 1 ,ω 2 and ω 3 To a suitable range, use Adam optimization and training set optimization to optimize the back propagation process of the informer model containing the safety potential field until the joint loss function Terminate the training when you stop descending.

[0067] Preferably, a grid search method combined with a heuristic algorithm is used to adjust the hyperparameter ω 1 ,ω 2 and ω 3 That is, within a reasonable data range, the hyperparameters are adjusted in sequence according to historical adjustment experience so that the prediction accuracy reaches the required range.

[0068] Beneficial effects of the present invention: The present invention introduces a transformer-based deep learning informer network as the backbone model for driving behavior prediction. Compared with the long short-term memory model and its variants, the model can accurately capture the long-term dependencies of the trajectory sequence through the probabilistic sparse attention mechanism, improve the accuracy of long-term trajectory prediction, and reduce the complexity of the model, thereby improving the real-time performance of driving behavior prediction. At the same time, the present invention establishes a safety potential field model for the vehicle according to the real-time status of the vehicle. While predicting the trajectory, the rationality and safety of the trajectory are ensured by solving the vehicle's dynamic potential field, which increases the explainability of lane change prediction and trajectory prediction. Specifically:

[0069] 1. In order to address the problem of low long-term trajectory prediction accuracy, the present invention introduces an informer model based on transformer improvement in the field of autonomous driving behavior prediction. This model can effectively capture the long-term dependency coupling relationship of trajectory sequences and improve the accuracy of long-term trajectory prediction through a probabilistic sparse attention mechanism. At the same time, the informer network replaces the traditional attention mechanism with a sparse probabilistic attention mechanism, which greatly reduces the complexity and improves the real-time performance of model prediction, which is crucial for the safe driving of autonomous vehicles.

[0070] 2. The present invention improves the safety and explainability of trajectory prediction by introducing a safety potential field model. According to the real-time state of the vehicle, the present invention limits the state of the vehicle (displacement, speed, acceleration, etc.) to a safe range by modeling the safety potential field in the following and lane changing situations, thereby avoiding the occurrence of dangerous behaviors such as collision and speeding due to unreasonable predictions. The present invention places the safety potential field model in the loss function, and uses gradient descent to enable the vehicle to learn the safe range of the vehicle state, thereby improving the safety and accuracy of driving behavior prediction.

[0071] 3. The present invention organically combines the safety potential field model and the informer model, which can realize the intention prediction and multi-channel trajectory prediction of surrounding vehicles, and has excellent performance in the real-time, accuracy and safety of driving behavior prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0073] Figure 1 It is a flow chart of the present invention.

[0074] Figure 2 This is a model application diagram of the informer network of the present invention.

[0075] Figure 3 Schematic diagram of equipotential surface projection of the following model and lane changing model of the present invention, wherein (a) is the following model during uniform linear motion, (b) is the following model during accelerated linear motion, (c) is the lane changing model during uniform motion, and (d) is the lane changing model during accelerated motion. DETAILED DESCRIPTION

[0076] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] like Figure 1 As shown in the figure, a driving behavior prediction method based on an informer neural network under a safety potential field is firstly established according to the real-time state of the vehicle, and the constraint range of the safety potential field model is learned by gradient descent, thereby limiting the state of the vehicle (displacement, speed, acceleration, etc.) to a safe range. Then, the informer model is used to accurately predict the vehicle intention and multi-channel trajectory, and improve the accuracy, safety and interpretability of driving behavior prediction. Specifically, the following steps are included:

[0078] Step S1: Preprocess the data in the motion datasets of the HighD dataset and the NGSIM dataset to obtain trajectory sequence features.

[0079] The step S1 is specifically as follows:

[0080] S1.1: Eliminate abnormal data in the HighD dataset and NGSIM dataset.

[0081] There are two types of abnormal data, namely missing values ​​in the data file and values ​​with incorrect information records (such as lane ID recognition errors caused by low camera resolution). Missing values ​​can lead to calculation errors and cause model gradient updates to fail. The present invention sets a python program to scan all data, screen out missing values, and then replace the missing values ​​with the average, maximum, minimum or probability estimate of the original data source. Incorrect values ​​of information records can lead to unreasonable prediction results (such as generating incorrect lane change prediction results when lanes cannot be changed). The present invention checks data values ​​through logical rules (common sense rules, data-specific rules, etc.) and uses constraints between different attributes to eliminate these abnormal values. After eliminating abnormal data, the model training will be more stable and the prediction logic will be more accurate, thereby avoiding unreasonable prediction results.

[0082] S1.2: The lateral and longitudinal coordinates, speed, acceleration, and lane number of the target vehicle in the motion data set are selected as the features of the target vehicle. These features contain rich information about the vehicle state, which can accurately describe the real situation of the target vehicle during driving and facilitate accurate prediction of future trajectories. The lateral and longitudinal relative displacements of the vehicles around the target vehicle are selected as the features of the surrounding vehicles. Relative displacement is a basic consideration for safe driving of vehicles, which is conducive to the target vehicle learning the appropriate lane change timing through relative displacement.

[0083] S1.3: Extract the trajectory feature set of the target vehicle within the time k to k+n as S = (S k ,S k+1 ,S k+2 ,...,S k+n-1 ), where S k is the trajectory feature sequence of the target vehicle and its surrounding vehicles at time k. The present invention converts the sequence S k Expressed as: S k ={x,y,v x ,v y ,a x ,a y ,id,Δx,Δy}, where x and y represent the horizontal and vertical coordinates of the target vehicle at time k, respectively, and v x ,v y Represent the lateral and longitudinal velocities of the target vehicle at time k, respectively, and a x ,a y They represent the lateral and longitudinal accelerations of the target vehicle at time k, id represents the lane where the target vehicle is located, and Δx and Δy represent the lateral and longitudinal relative displacements of the surrounding vehicles relative to the target vehicle.

[0084] S1.4: Calculate the mean and variance of each channel of the multi-channel sequence S (sequence S k The 9 features in are calculated in parallel).

[0085]

[0086]

[0087] S1.5: Considering that the displacement in the driving direction increases linearly, and the data size of different lanes in the vertical direction is also different. In order to accurately describe the vehicle status (lane change or straight driving) of different sections and avoid the influence of individual extreme trajectories on the model prediction performance, the Z-Score normalization is performed on each channel of the trajectory feature set S to obtain the final trajectory sequence feature. for:

[0088]

[0089] S1.6: The data set is randomly divided into a training set and a test set in a ratio of 7:3. The training set is used to train the model parameters, and the test set is used to test the accuracy and error of the model to verify the effectiveness of the model.

[0090] S2: Encode the trajectory sequence features obtained in step S1 into latent spatiotemporal representations. The training set and the test set are processed in the same way.

[0091] The step S2 is specifically as follows:

[0092] S2.1: Use a neural network one-dimensional convolutional layer to encode the time frame of the trajectory feature set S into a potential vector representing the time attribute, and use a neural network fully connected layer to encode the trajectory sequence features It is represented as a spatiotemporal vector of connected context data. The following formula is the calculation process of the fully connected layer:

[0093]

[0094]

[0095] Among them, i refers to the number of sample indicators, and j refers to the number of neurons in the middle layer. ij represents the trainable weight, b j represents the trainable offset value, θ j represents the output of the fully connected layer, and f(x) represents the activation function.

[0096] The relative position information of the input feature at each moment is represented by position encoding using the following formula:.

[0097] PE(pos,2i)=sin(pos / 100002i / d model )

[0098] PE(pos,2i+1)=cos(pos / 100002i / dmodel )

[0099] Among them, pos represents the current moment, i represents the dimension of the encoding, and d model It is specified as 512. Therefore each dimension of the positional encoding corresponds to a sine wave.

[0100] S2.2: Perform matrix addition operation on the time coding representation, feature coding representation and position coding representation to obtain the potential spatiotemporal representation.

[0101] Matrix addition is the addition of each corresponding point in two matrices. This is a parallel operation and can calculate the result quickly.

[0102] Figure 2 The flowchart of the present invention using the informer network in driving behavior prediction application is summarized. Taking real-time performance and computational complexity into consideration, the present invention only designs 6 encoders and 1 decoder, where the encoder and decoder share the same input trajectory features.

[0103] S3: Input the latent spatiotemporal representation obtained in step S2 into the informer neural network to predict the future driving intention and trajectory of the target vehicle.

[0104] The step S3 is specifically as follows:

[0105] S3.1: The latent spatiotemporal representation generated in step S2 is input into the encoder of the informer neural network, and the trajectory feature representation E is further obtained through the encoder multi-head probabilistic sparse attention layer and fully connected layer as shown in the following formula.

[0106]

[0107]

[0108]

[0109] in, Represent the trajectory sequence features The generated vector representation, W i and W o Represents different weight matrices. k represents the feature dimension, softmax represents the probability mapping function, Attention represents the attention matrix calculation function, Concat represents the merging operation, head i represents the i-th attention head, and h represents the number of attention heads. The encoder multi-head probabilistic sparse attention layer can extract local and global features of the trajectory, which can better predict driving behavior.

[0110] S3.2: Input the latent spatiotemporal representation generated in step S2 into the decoder of the informer neural network, and obtain the trajectory feature representation M through the masked multi-head attention layer of the decoder. Because the input and output of the decoder of the informer neural network are aligned, the role of the masked multi-head attention layer is to pre-fill the position of the data to be predicted with 0 so as to generate a certain length of predicted data at a time.

[0111] S3.3: The trajectory feature representation E of step S3.1 and the trajectory feature representation M of step S3.2 are passed through the interactive multi-head probabilistic sparse attention layer and the fully connected layer to obtain the final informer model output D.

[0112] The interactive multi-head probabilistic sparse attention layer replaces the vectors Q and K of the multi-head probabilistic sparse attention layer above with the Q and K output by the encoder and the V in the decoder for multi-head attention calculation. This can fuse the information of the encoder and decoder for more accurate trajectory prediction.

[0113] Figure 2 The one-dimensional convolutional pooling module and the linear layer in the process the output of the previous network. Their function is to further represent the trajectory information and prepare for the final driving behavior.

[0114] S3.4: Output D outputs the predicted driving intention through the intention prediction head, and output D outputs the predicted trajectory through the trajectory prediction head.

[0115] The intention includes the intention to go straight, the intention to change lanes left and the intention to change lanes right. The trajectory represents the driving coordinates of the target vehicle in the future (usually within 5 seconds). In the present invention, the driving intention and driving trajectory prediction are combined into the driving behavior prediction. The driving behavior prediction provides a guarantee for the safe driving of the autonomous driving vehicle. It predicts the intention and trajectory of the target vehicle and reasonably plans its own driving trajectory.

[0116] S4: Considering the safety and feasibility of vehicle movement, safety potential field models are established for the following and lane changing situations, and the safety potential field range is calculated.

[0117] The step S4 is specifically as follows:

[0118] S4.1: Safety distance for collision avoidance in following situations Defined as Among them, vx represents the real-time speed of the target vehicle, amax represents the constant of the vehicle's braking acceleration, which is generally taken as 6m / s 2 .

[0119] S4.2: When the vehicle is moving straight ahead at a uniform speed, a two-dimensional Gaussian distribution is used to describe the safety potential field model. Because when an object participates in multiple sub-movements at the same time, each sub-movement is independent of each other, so the two sub-movements in the horizontal direction and the vertical direction are independent of each other. According to the mathematical characteristics of the two-dimensional Gaussian distribution and referring to the concept of pseudo-distance, the potential field range is corrected, and finally the following formula of the following model under uniform straight-ahead movement is obtained:

[0120]

[0121] Among them, ξ and λ are the hyperparameters adjusted by the safety potential field model, (x, y) is the coordinate of the surrounding vehicles, (x 0 ,y 0 ) is the coordinate of the autonomous vehicle. E v represents the potential field of the car-following model under uniform straight-line driving, σ x,1 and σ y Represents the variance of a two-dimensional Gaussian distribution.

[0122] Figure 3 (a) is an example of the equipotential projection of the following model under uniform speed straight driving, where the coordinate (0, 0) is the center of mass of the target vehicle. It can be seen that in the uniform speed straight driving scenario, the safety potential fields in front of and behind the target vehicle and to the left and right are symmetrical, that is, the possibility of the car being threatened is also symmetrical.

[0123] S4.3: When the target vehicle is accelerating, the impact on the vehicle in front of the target vehicle is much greater than that on the vehicle behind it. In order to reflect this difference, Rayleigh distribution is used in the x-axis direction to describe the attenuation process of the acceleration forward tilt effect: At the same time, due to the independence of the lateral speed and the longitudinal speed, the formula for the following model under accelerated motion is established as follows:

[0124]

[0125] Among them, σ x,2 is the variance of Rayleigh distribution, ξ and λ are unknown parameters.

[0126] Figure 3 (b) is an example of the equipotential projection of the car-following model under accelerated linear motion, where the coordinate (0, 0) is the center of mass of the target vehicle. It can be seen that the potential field of the car-following model under accelerated linear motion has a forward-leaning trend, that is, the impact on the vehicles behind the target vehicle is much smaller than that on the front.

[0127] S4.4: Set the collision avoidance safety distance in step S4.1 Substituted into the safety potential field model formulas in step S4.2 and step S4.3 respectively, the final safety potential field following model is constructed as follows:

[0128]

[0129] Among them, ξ, λ, a vpt is a parameter to be determined, a vpt Indicates the threshold for measuring whether the vehicle is in an accelerating state. The values ​​of these parameters are obtained by comparing and analyzing the quality of the actual model training and the actual vehicle conditions. In specific applications, ξ and λ take values ​​within (0, 1), and a vpt The value is in the range of (0, 0.8). It can be seen that the degree of influence of surrounding vehicles on the autonomous vehicle is related to the state of the autonomous vehicle itself and the distance between the surrounding vehicles and the autonomous vehicle: the potential field under the following model decreases with the increase of the braking acceleration amax, increases with the increase of the real-time speed vx, and expands with the increase of ax in the presence of ax, and decreases with the increase of the displacement difference in the x-direction and y-direction with the surrounding vehicles.

[0130] S4.5: The lane change safety distance model is expressed as follows:

[0131] in, is the longitudinal safety distance for vehicle collision avoidance, a y is the lateral acceleration of the target vehicle; s y The lateral displacement required for the vehicle to change lanes.

[0132] S4.6: Considering the vehicle deflection during the lane change process, the present invention adds a rotation component, (x * ,y * ) is the corresponding deflection value after the original coordinate (x, y) turns, and the final safety potential field formula under the lane change model is established as follows:

[0133]

[0134]

[0135] in, Indicates the steering angle of the vehicle.

[0136] Figure 3 (c) is an example of equipotential projection of the lane-changing model under uniform vehicle motion. Figure 3 (d) is an example of the equipotential projection of the lane-changing model under vehicle acceleration, where the coordinate (0, 0) is the center of mass of the target vehicle.

[0137] S4.7: Input the trajectory sequence features obtained in step S1 into the safety potential field model of the target vehicle to calculate the safety potential field range of the target vehicle.

[0138] S5: Construct a joint loss function of the safety potential field range constraint in step S4, the driving intention predicted in step S3, and the vehicle trajectory, and solve the loss function by gradient descent method. Use Adam optimizer to optimize the informer model containing the safety potential field to obtain a prediction model.

[0139] The step S5 is specifically as follows:

[0140] S5.1: Use the cross entropy function as the loss function for the predicted intent:

[0141]

[0142] in, is the probability distribution of the predicted intention, which is a probability value between 0 and 1, such as 30% probability of changing lanes to the left, 40% probability of changing lanes to the right, and 30% probability of going straight. s ,I l ,I r} represents the vehicle's true intention, i.e., going straight, turning left or turning right.

[0143] S5.2: Use the mean square error function as the loss function for the predicted trajectory:

[0144]

[0145] in, is the predicted future trajectory point, F is the actual future trajectory point, T pred Represents the number of frames to be predicted.

[0146] S5.3: Use the mean square error function as the loss function for the safety potential field range constraint:

[0147]

[0148] in, represents the range of the safety potential field under the predicted future state, and E v Represents the real safety potential field range, E v ∈{E vs ,E vc}, select E according to the maneuvering conditions v .

[0149] S5.4: The predicted driving intention and trajectory are the motivation for model learning, and the safety potential field is a tool to assist trajectory prediction. Therefore, the present invention adds ω 1 ,ω 2 and ω 3 Hyperparameters are used to ensure that the driving lane change intention and trajectory prediction tasks dominate the gradient back propagation. The final joint loss function is defined as:

[0150]

[0151] S5.5: Adjust all hyperparameters in step S5.4 to appropriate ranges, and use the Adam optimizer to optimize the model backpropagation process until the joint loss function no longer decreases and terminates the training.

[0152] The parameters are adjusted by combining the grid search method with the heuristic algorithm. That is, within a reasonable data range, the parameters are adjusted sequentially according to historical adjustment experience so that the model prediction accuracy reaches the required range.

[0153] S6: Save the trained prediction model, use the test set to test and predict the prediction model, and obtain the future trajectory coordinates of the vehicle.

[0154] The present invention improves the safety and interpretability of trajectory prediction by introducing a safety potential field model. According to the real-time state of the vehicle, the present invention limits the state of the vehicle (displacement, speed, acceleration, etc.) to a safe range by modeling the safety potential field in the following and lane changing situations, so as to avoid the occurrence of dangerous behaviors such as collision and speeding caused by unreasonable predictions. The present invention puts the safety potential field model in the loss function, and enables the vehicle to learn the safety range of the vehicle state by gradient descent, thereby improving the safety and accuracy of driving behavior prediction. In order to cope with the problem of low long-term trajectory prediction accuracy, the present invention introduces an informer model based on transformer improvement in the field of autonomous driving behavior prediction, which can effectively capture the long-term dependent coupling relationship of the trajectory sequence and improve the accuracy of long-term trajectory prediction through a probabilistic sparse attention mechanism. At the same time, the informer replaces the traditional attention mechanism with a sparse probabilistic attention mechanism, which greatly reduces the complexity and improves the real-time performance of the model prediction, which is crucial for the safe driving of autonomous driving vehicles. The present invention organically combines the safety potential field model and the informer model, and can realize the intention prediction of surrounding vehicles and multi-channel trajectory prediction. It has excellent performance in real-time, accuracy and safety of driving behavior prediction.

[0155] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A driving behavior prediction method based on informer neural network under safety potential field, characterized in that: The steps are as follows: S1: Preprocess the data in the motion dataset to obtain trajectory sequence features; S2: Encode the trajectory sequence features obtained in step S1 into a latent spatiotemporal representation; S3: Input the latent spatiotemporal representation obtained in step S2 into the informer neural network to predict the future driving intention and driving trajectory of the target vehicle; S4: Considering the safety and feasibility of vehicle movement, safety potential field models are established for the following and lane changing situations, and the safety potential field range is calculated; The safety potential field model in the following situation is: Where vx represents the real-time speed of the target vehicle, ξ and λ are adjustable hyperparameters, and a vpt represents the threshold for measuring whether the vehicle is in an accelerating state, (x, y) is the coordinate of the surrounding vehicles, (x0, y0) is the coordinate of the autonomous driving vehicle, σ x,1 and σ y Represent the variance of the two-dimensional Gaussian distribution; σ x,2 is the Rayleigh distribution of the attenuation process of the acceleration forward tilt effect in the x-axis direction The variance of x>0; vx represents the real-time speed of the target vehicle, amax represents the constant of the vehicle's braking acceleration; a x Indicates the acceleration of the target vehicle on the x-axis; The safety potential field model in the lane-changing situation is: in, Indicates the steering angle of the vehicle; (x * ,y * ) is the deflection value corresponding to the original coordinate (x, y) after the turn; a y is the lateral acceleration of the target vehicle; s y is the lateral displacement required for the vehicle to change lanes; a vpt Indicates the threshold for measuring whether the vehicle is in an accelerating state; S5: construct a joint loss function of the safety potential field range constraint in step S4, the driving intention predicted in step S3, and the vehicle trajectory, and solve the loss function by gradient descent method, and use Adam optimizer to optimize the informer model containing the safety potential field to obtain a prediction model; S6: Save the trained prediction model, use the test set to predict the prediction model, and obtain the future trajectory coordinates of the vehicle.

2. The driving behavior prediction method based on informer neural network under safety potential field according to claim 1 is characterized in that: The pre-processing method of step S1 is: S1.1: Eliminate abnormal data in the motion dataset; S1.2: Select the lateral and longitudinal coordinates, velocity, acceleration and lane number of the target vehicle in the motion data set as the features of the target vehicle, and select the lateral and longitudinal relative displacements of the vehicles around the target vehicle as the features of the surrounding vehicles; S1.3: Extract the trajectory feature set of the target vehicle within the time k to k+n as S; S1.4: Calculate the mean and variance of each channel of the multi-channel sequence S; S1.5: Perform Z-Score normalization on each channel of the trajectory feature set S to obtain the final trajectory sequence feature S1.6: Randomly divide the motion dataset into training and test sets.

3. The driving behavior prediction method based on informer neural network under safety potential field according to claim 2 is characterized in that: The Z-Score standardization method is: Among them, the mean variance Trajectory feature set S = (Sk, Sk +1 ,Sk +2 ,...,Sk + n-1), where S k is the trajectory feature sequence of the target vehicle and its surrounding vehicles at time k, and the sequence S k ={x,y,v x ,v y ,a x ,a y ,id,Δx,Δy}, where x and y represent the horizontal and vertical coordinates of the target vehicle at time k, respectively, and v x ,v y Respectively represent the lateral and longitudinal speeds of the target vehicle at time k, a x ,a y They represent the lateral and longitudinal accelerations of the target vehicle at time k, id represents the lane where the target vehicle is located, and Δx and Δy represent the lateral and longitudinal relative displacements of the surrounding vehicles relative to the target vehicle.

4. The driving behavior prediction method based on informer neural network under safety potential field according to claim 2 or 3 is characterized in that: The abnormal data in step S1.1 include missing values ​​and values ​​with incorrect information records in the data file; by setting a python program to scan all the data in the motion data set, the missing values ​​are screened out, and then the missing values ​​are replaced by the average value, maximum value, minimum value or probability estimate of the original data source; the values ​​with incorrect information records are checked by logical rules, and the constraints between different attributes are used to eliminate the values ​​with incorrect information records; in step S1.4, the sequence S k The 9 features in are calculated in parallel; in step S1.6, the training set and the test set are divided into a ratio of 7:

3.

5. The driving behavior prediction method based on informer neural network under safety potential field according to claim 4 is characterized in that: The implementation method of the encoding in step S2 is: S2.1: Use a one-dimensional convolutional layer of a neural network to encode the time frame of the trajectory feature set S into a potential vector representing the time attribute, and obtain a time encoding representation; use a fully connected layer of a neural network to encode the trajectory sequence features Represented as a spatiotemporal vector of linked context data, a feature coding representation is obtained; the relative position information of the trajectory sequence features input at each moment is positionally encoded to obtain a position coding representation; S2.2: Perform matrix addition operation on the time coding representation, feature coding representation and position coding representation to obtain the potential spatiotemporal representation.

6. The driving behavior prediction method based on informer neural network under safety potential field according to claim 5 is characterized in that: The method for predicting the future driving intention and driving trajectory of the target vehicle in step S3 is: S3.1: Input the latent spatiotemporal representation generated in step S2 into the encoder of the informer neural network, and obtain further trajectory feature representation E through the encoder's multi-head probabilistic sparse attention layer and fully connected layer; S3.2: Input the latent spatiotemporal representation generated in step S2 into the decoder of the informer neural network, and obtain the trajectory feature representation M through the masked multi-head attention layer of the decoder; S3.3: The trajectory feature representation E of step S3.1 and the trajectory feature representation M of step S3.2 are passed through the interactive multi-head probabilistic sparse attention layer and the fully connected layer to obtain the final output D of the informer neural network; S3.4: Output D outputs the predicted driving intention through the intention prediction head, and output D outputs the predicted driving trajectory through the trajectory prediction head.

7. The driving behavior prediction method based on informer neural network under safety potential field according to claim 6 is characterized in that: The calculation process of the fully connected layer of the neural network is: Among them, i refers to the number of sample indicators, j refers to the number of neurons in the middle layer; wij represents the trainable weight, bj represents the trainable offset value, θj represents the output of the fully connected layer, and f(x) represents the activation function; The position encoding is expressed as: Among them, pos represents the current moment, i represents the dimension of the encoding, and d model Designated as 512; The calculation method of the S3.1 encoder is: in, K and V represent the trajectory sequence features. The generated vector representation, Wi and Wo represent different weight matrices; dk represents the feature dimension, softmax represents the probability mapping function, Attention represents the attention matrix calculation function, Concat represents the merging operation, headi represents the i-th attention head, and h represents the number of attention heads.

8. The driving behavior prediction method based on informer neural network under the safety potential field according to any one of claims 4 to 7, characterized in that: The trajectory sequence features obtained in step S1 are input into the safety potential field model of the target vehicle to calculate the safety potential field range of the target vehicle.

9. The driving behavior prediction method based on informer neural network under safety potential field according to claim 8 is characterized in that: The implementation method of step S5 is: S5.1: Use the cross entropy function as the loss function for the predicted intent: in, is the probability distribution of the predicted intention, I∈{Is,Il,Ir} represents the real intention of the vehicle, i.e., going straight, turning left or right; S5.2: Use the mean square error function as the loss function for the predicted trajectory: in, is the predicted future trajectory point, F is the actual future trajectory point, T pred Represents the number of predicted frames; S5.3: Use the mean square error function as the loss function for the safety potential field range constraint: in, represents the range of the safety potential field under the predicted future state, while Ev∈{Evs,Evc} represents the real range of the safety potential field; S5.4: The final joint loss function is: Among them, ω1, ω2 and ω3 are hyperparameters; S5.5: Adjust all hyperparameters ω1, ω2, and ω3 in step S5.4 to appropriate ranges, and use Adam optimization and training set optimization to optimize the back propagation process of the informer model containing the safety potential field until the joint loss function Terminate the training when you stop descending.

10. The driving behavior prediction method based on informer neural network under safety potential field according to claim 9 is characterized in that: The grid search method combined with the heuristic algorithm is used to adjust the hyperparameters ω1, ω2 and ω3. That is, within a reasonable data range, the hyperparameters are adjusted in sequence according to historical adjustment experience so that the prediction accuracy reaches the required range.