CN110304075B - Vehicle track prediction method based on hybrid dynamic Bayesian network and Gaussian process - Google Patents

Abstract

The invention belongs to the technical field of environment cognition and decision making of an automatic driving vehicle, and particularly relates to a vehicle track prediction method based on a hybrid dynamic Bayesian network and a Gaussian process. The method learns the parameters of the MDBN and the GP through the natural driving data of the vehicle, utilizes the MDBN to fuse a plurality of vehicle kinematic models to obtain the short-term track prediction and the estimation probability of the driving intention and the driving characteristic, and then utilizes the GP to predict the long-term track and express the prediction uncertainty. Compared with the conventional vehicle trajectory prediction method, the method combines the vehicle model, the abstract intention and the data drive, and the MDBN and GP models have strong expansibility, can be suitable for different driving scenes and can combine more effective situation information, such as road information and traffic information.

Description

Vehicle track prediction method based on hybrid dynamic Bayesian network and Gaussian process

Technical Field

The invention belongs to the technical field of environment cognition and decision making of an automatic driving vehicle, and particularly relates to a vehicle track prediction method based on a hybrid dynamic Bayesian network and a Gaussian process.

Background

Currently, one of the mainstream solutions for implementing automatic driving of a vehicle is based on a "perception-decision-control" architecture. The layered architecture adopts the anthropomorphic thought, just like the need of human to sense the surrounding environment by using the sense organs such as eyes, ears, nose and the like; then, the brain processes the perception information to form understanding and judgment of the surrounding environment, so that reasonable decision and planning are made; finally, the determined task is performed by the body of the person, such as the hand, the foot, etc. Clearly, the decision making system for autonomous vehicles plays a crucial role as the brain is the core of the human body. The decision-making system firstly realizes the cognition of the driving situation, needs to deeply understand the perception information, continuously estimates, judges and predicts the change of the driving environment, and secondly utilizes the cognition information to determine the future driving behavior of the automatic driving vehicle and plan the future driving path of the automatic driving vehicle. How to improve the cognitive degree of the automatic driving vehicle on the driving situation and correctly estimate, judge and predict the future track of the surrounding traffic participants is the core and challenge of the automatic driving vehicle decision-making system.

Aiming at the existing difficult problems and challenges of an automatic driving vehicle decision-making system, partial scientific researchers have already developed research on the track prediction of vehicles around the week in the driving process of the automatic driving vehicle, and the following three major methods are summarized: 1. vehicle trajectory prediction based on physical models: the method comprises the steps of representing a vehicle as a dynamic entity constrained by physical rules, and predicting the future track of the vehicle according to a vehicle dynamic or kinematic model; 2. vehicle trajectory prediction based on driving intent estimation: the method comprises the steps of representing a vehicle as an independent behavior entity, and obtaining a corresponding future track of the vehicle under a driving intention according to an intention estimation result of the current vehicle; 3. vehicle trajectory prediction based on deep network: the direct mapping from the data to the track is obtained by using the natural driving data of the vehicle and a deep learning method and the like.

The first method is based on vehicle track prediction of a physical model, only considers the motion characteristics of the current vehicle, and is suitable for track prediction in a short time; because the selection of the vehicle physical model directly influences the track prediction result, and the corresponding physical models of the vehicle in different driving states are different, the track prediction cannot be accurately carried out according to a single vehicle model; meanwhile, because the information of the vehicle driver (such as driving intention; driving characteristics) is not considered, the motion trail of the vehicle in a long-term range cannot be predicted; in addition, the method ignores the influence of environmental factors around the vehicle on the trajectory of the vehicle. The second type of method, which is based on the vehicle trajectory prediction estimated by the driving intention, can predict the change of the vehicle trajectory caused by the vehicle performing a specific operation (such as intersection turning: deceleration, steering, acceleration turning), and some implementation methods can also predict the change of the vehicle trajectory caused by the change of the surrounding environmental factors, so that this type of method can predict the vehicle trajectory in a long time range, but the prediction error in a short time is large because the physical motion characteristics of the vehicle itself are not considered. The third kind of method, vehicle track prediction based on deep network, it is produced because of artificial intelligence and big data technology's rapid development, mainly involve long short term memory network (LSTM), deep Bayesian network, etc., because the fitting performance of this method is extremely strong, can realize the vehicle track prediction under the complicated interactive scene (such as the intersection of the traffic signal lamp or having no traffic signal lamp). However, because a large number of calibration data samples are needed, the physical motion characteristics of the vehicle are not considered, and the established neural network model has the defects of weak generalization capability and poor interpretability, the vehicle trajectory prediction under a plurality of scenes is difficult to complete, the corresponding reason is difficult to find according to the wrong output result, and the prediction uncertainty is difficult to represent.

Accordingly, there is a need for a method of predicting the trajectory of a vehicle around a week during travel of an autonomous vehicle that overcomes or at least mitigates the above-mentioned deficiencies of the prior art.

Disclosure of Invention

In view of the above technical problems, an object of the present invention is to provide a vehicle trajectory prediction method based on a hybrid dynamic bayesian network and a gaussian process, which predicts the trajectory of a vehicle around the week during the driving process of an autonomous vehicle and provides reliable information for a decision layer of the autonomous vehicle.

In order to achieve the purpose, the invention provides the following technical scheme:

a vehicle track prediction method based on a hybrid dynamic Bayesian network and a Gaussian process comprises the following steps:

step 1, constructing a natural driving database;

establishing a test set of peripheral vehicle related sequence information, road related sequence information and traffic related sequence information acquired by an automatic driving vehicle sensing system, and a training set for calibrating driving intention and driving characteristics of the information; wherein the test set comprises a test set of a hybrid dynamic Bayesian network and a test set of a Gaussian process; the training set comprises a training set of a hybrid dynamic Bayesian network and a training set of a Gaussian process;

step 2, obtaining short-term predicted trajectories of surrounding vehicles and estimated probabilities of driving intentions and driving characteristics by adopting a hybrid dynamic Bayesian network;

taking driver information, vehicle model selection and vehicle state information as hidden layer variables, taking surrounding vehicle related sequence information, road related sequence information and traffic related sequence information acquired by an automatic driving vehicle sensing system as observation layer variables, and building a hybrid dynamic Bayesian network model; and obtaining the output of the hybrid dynamic Bayesian network through posterior probability inference: predicting short-term trajectory based on the vehicle model and the estimated probability of driving intention and driving characteristics, and taking the output as the input of step 4;

step 3, establishing Gaussian process functions under different driving intentions and different driving characteristics;

comprises the following steps:

step 3.1 setting of mean function and covariance function

With x as an input, the expression of the gaussian process function is as follows:

f(x)～GP(u(x),Σ(x,x’))

the mean function u (x) represents the vehicle track trend under a certain driving intention, so that the mean function can be used for representing the expected predicted track; the covariance function Σ (x, x ') represents both the variance of the different inputs x themselves and the variance between (x, x'), and therefore the covariance function can be used to represent the uncertainty corresponding to the desired predicted trajectory;

setting corresponding mean functions u (x) according to vehicle tracks under different driving intentions;

the covariance function Σ (x, x') represented in the form of a noisy mean-square index is set as follows:

wherein σ_fIs the signal standard deviation, l is the characteristic length, σ_nTo observe the noise standard deviation, δ is the kronecker function;

step 3.2, learning of unknown parameters of Gaussian process function

Learning related unknown parameters by using the training set of the Gaussian process established in the step 1 according to the mean function u (x) and the covariance function Σ (x, x') set in the step 3.1, and obtaining parameter learning results by using a gradient-based optimization algorithm according to the logarithmic marginal likelihood function and the partial derivatives of the unknown parameters thereof, thereby respectively establishing Gaussian process functions under different driving intents and driving characteristics;

step 4, long-term trajectory prediction and uncertainty representation are carried out based on a hybrid dynamic Bayesian network and a Gaussian process;

determining the corresponding driving intention and driving characteristics by using a maximum probability principle according to the driving intention and the estimated probability of the driving characteristics output by the hybrid dynamic Bayesian network in the step 2, thereby determining the corresponding Gaussian process function in the step 3 according to the driving intention and the driving characteristics;

using the short-term predicted track based on the vehicle model output by the hybrid dynamic Bayesian network in the step 2 as vehicle track sequence information x₁Predicted future trajectory is x₂Then (x)₁,x₂) The gaussian distribution obeyed is as follows:

wherein,

represents a normal function;

x₁and x₂The respective corresponding mean functions are respectively u₁And u₂(ii) a The covariance matrix Σ is a symmetric matrix, i.e., Σ ═ Σ^T，x₁And x₂The respective covariance functions are ∑₁₁Sum-sigma₂₂，x₁And x₂The corresponding covariance function is ∑₁₂、Σ₂₁And ∑₁₂＝∑₂₁ ^T；

Then the vehicle trajectory x is known₁Future possible trajectory x of vehicle₂Conditional probability P (x) of₂|x₁) The expression of gaussian distribution obeyed is as follows:

wherein u is_2|1Sum-sigma_2|1Is a variable x_2|1Corresponding mean function and covariance function, and (u)_2|1,Σ_2|1) The expression (c) is derived from the following formula:

u_2|1＝u₂+Σ₁₂ ^TΣ₁₁ ^-1(x₁-u₁)

Σ_2|1＝Σ₂₂-Σ₁₂ ^TΣ₁₁ ^-1Σ₁₂

and finally obtaining the future track and uncertainty representation thereof in the long-term domain of the vehicle under the driving intention and the driving characteristics determined by using the maximum probability principle.

In the step 1, the peripheral vehicle related sequence information comprises vehicle position, vehicle speed, vehicle acceleration, vehicle yaw rate, distance between the vehicle and two sides of the road and opening and closing conditions of left and right steering tail lamps of the vehicle;

the road related sequence information comprises structural characteristics of roads and road indication signs;

the traffic-related sequence information includes surrounding traffic indication signs and traffic light states.

The step 1 comprises the following steps:

1.1 acquisition of Driving data

Collecting vehicle related sequence information, road related sequence information and traffic related sequence information around an automatic driving vehicle sensing system in a normal driving process;

1.2 classifying and calibrating driving intentions

Defining possible driving intentions of surrounding vehicles according to the scene of the vehicles; for each possible driving intention, selecting key characteristic parameters for classifying and calibrating the driving intention; then, the driving intention of the surrounding vehicles within a period of time is judged and corresponding driving intention labels are set by combining the related sequence information of the surrounding vehicles acquired in the step 1.1;

the scene of the vehicle is obtained by an intelligent navigation system of the automatic driving vehicle, and comprises a high-speed scene, a mountain road scene, an urban normal straight road scene, an urban traffic signal lamp intersection scene and an urban traffic signal lamp-free intersection scene;

the driving intention comprises straight running along the current road, turning along the current road, left lane changing, right lane changing, left turning, right turning, turning around, parking and starting;

the key characteristic parameters for classifying and calibrating the driving intention comprise the transverse position of the vehicle, the transverse speed of the vehicle, the yaw speed of the vehicle, the distance between the vehicle and the two sides of the road and the opening and closing conditions of left and right steering tail lamps of the vehicle;

1.3 Classification and calibration of Driving characteristics

Firstly, defining possible driving characteristics of surrounding vehicles, selecting key characteristic parameters for classification and judgment of the driving characteristics, judging the driving characteristics of the surrounding vehicles under a certain driving intention and setting corresponding driving characteristic labels according to the related sequence information of the surrounding vehicles acquired in the step 1.1 and the driving intention of the surrounding vehicles in a period of time acquired in the step 1.2;

the driving characteristics include slow driving, smooth driving, and jerky driving;

the key characteristic parameters for the classification and determination of the driving characteristics comprise a vehicle longitudinal speed, a vehicle lateral speed, a vehicle longitudinal acceleration, a vehicle lateral acceleration and a vehicle yaw rate;

1.4 creation of training set and test set

Taking the peripheral vehicle related sequence information, the road related sequence information and the traffic related sequence information collected in the step 1.1 as a test set of the hybrid dynamic Bayesian network MDBN;

taking sequence information of front section time corresponding to a transverse position and a longitudinal position of a vehicle under a certain driving intention and driving characteristics as a test set of a Gaussian process;

1.1, obtaining a training set of a hybrid dynamic Bayesian network (MDBN) after the surrounding vehicle related sequence information, the road related sequence information and the traffic related sequence information which are collected in the step 1.1 are classified and calibrated according to the driving intention in the step 1.2 and the driving characteristics in the step 1.3;

according to the training set of the hybrid dynamic Bayesian network MDBN, all sequence information corresponding to the transverse position and the longitudinal position of the vehicle under a certain driving intention and driving characteristics is used as one training set of the Gaussian process GP.

In the step 2, the observation layer variable, the hidden layer variable and the connection relation between the two variables are obtained by utilizing natural driving data and an improved genetic algorithm; and obtaining the probability parameter corresponding to the connection relation by a maximum likelihood method.

In the step 2, the vehicle model comprises a uniform speed model, a uniform acceleration model and a uniform speed steering model;

the vehicle state information includes a vehicle lateral position, a vehicle lateral velocity, a vehicle lateral acceleration, a vehicle longitudinal position, a vehicle longitudinal velocity, a vehicle longitudinal acceleration, and a vehicle yaw rate.

In the step 2, the hidden variables further include other related high-level hidden variables, where the other related high-level hidden variables include risk assessment of the vehicle and the environment and interaction between the vehicles;

the observation layer variables further comprise other relevant bottom layer observation information corresponding to other relevant high layer hidden variables of the hidden layer variables.

The step 2 specifically comprises the following steps:

2.1 defining hidden and observation layer nodes of hybrid dynamic Bayesian network model

Hidden nodes contain three layers of information: the first layer is a driver information variable D, the second layer is a vehicle model selection variable M, and the third layer is a vehicle state information variable S; the hidden layer variable set H is { D, M, S }, wherein { D, M } is a discrete variable, and { S } is a continuous variable;

the observation layer node mainly comprises one layer of information: the related sequence information of surrounding vehicles, the related sequence information of roads and the related sequence information of traffic, which are collected by an automatic driving vehicle sensing system; the observation layer variable set is { O }, and is a continuous variable;

2.2 optimization of network architecture

Given a variable set { D, M, S, O } and a training set T, finding a network N ═ B, θ which best matches the training set T by means of learning and searching, and using a criterion function as a measure of the degree of matching; wherein, B represents a network structure, and theta represents a parameter of the network;

2.3 learning of unknown parameters between network nodes

The conditional probability of part hidden layer variables is obtained through a data statistics method, the conditional probability between the rest hidden layers and the observed quantity is subjected to parameter learning by using the training set of the mixed dynamic Bayesian network MDBN obtained in the step 1, the solution is achieved by using a maximum likelihood estimation method, the likelihood function L (theta) related to the network parameter theta under the known training sample D is logP (D | theta), and the optimal solution is solved by using a gradient-based optimization algorithm

2.4 posterior probabilistic reasoning with introduction of Observation layer variables

The posterior probability refers to the probability distribution P (H | O) of hidden layer variables under the known observation layer variables, that is, the process of deducing the corresponding driving intention, driving characteristics, vehicle model and vehicle state under the test set of the hybrid dynamic bayesian network MDBN obtained in step 1, and the short-term trajectory prediction based on the vehicle model and the estimation probability of the driving intention and the driving characteristics are obtained.

In the step 2, the posterior probability inference is assumed density filtering approximation inference, and the specific process is as follows:

1) and (3) prediction: obtaining the joint probability distribution of the t +1 moment by using the probability distribution of the t moment; all discrete hidden layer variables { M) at time t_t,D_tThe joint distribution of

Discrete hidden layer variable M at time t +1_t,D_t,M_t+1,D_t+1The joint distribution is:

continuous hidden layer variable (S) at time t +1_t+1The conditional probabilities are:

2) updating: introducing an observation variable at the t +1 moment to obtain new probability distribution; let the observed variable of the vehicle model be { V }, and satisfy

The rest of the observed variables in the O are E, and the discrete hidden layer variable M at the t +1 moment_t,D_t,M_t+1,D_t+1The joint distribution of the { is:

wherein,

continuous hidden layer variable (S) at time t +1_t+1The conditional probabilities are:

3) marginalizing: marginalizing the variables before the t +1 moment in the obtained t +1 moment probability distribution, and only keeping the probability distribution of the variables at the current t +1 moment so as to be used for the iteration of the next prediction and updating step:

wherein D represents a driver information variable, the driver information including a driving intention and a driving characteristic; m represents a vehicle model selection variable, and the vehicle model comprises a constant speed model, an acceleration model and a constant speed steering model; s represents a vehicle state information variable, and the vehicle state information comprises a vehicle transverse position, a vehicle transverse speed, a vehicle transverse acceleration, a vehicle longitudinal position, a vehicle longitudinal speed, a vehicle longitudinal acceleration and a vehicle yaw rate; by using

Obtaining the estimated probability of the vehicle model, the driving intention and the driving characteristic at the moment of t + 1; by using

A short-term predicted trajectory based on the vehicle model is obtained.

In the step 3.2, the maximum likelihood method is selected to learn the unknown parameters in the gaussian process.

The specific process of the step 3.2 is as follows:

according to the mean function u (x) and the covariance function Σ (x, x') set in step 3.1, the gaussian process GP training set D established in step 1 is used_GP(X,Y)＝{(x_i,y_i),i＝1:N₁Learning the unknown parameters involved, where N₁For training sequences of samplesLength, let the unknown parameters of the mean function and covariance function be Θ ═ θ_i,i＝1:N₂In which N is₂For the number of parameters, the log marginal likelihood function logP (Y | X, Θ) is:

wherein,

represents a normal function; x and Y represent the lateral and longitudinal vehicle position; u. of_YAnd ∑_YA mean function and a covariance function corresponding to the variable Y;

selecting a maximum likelihood method, and solving to obtain unknown parameters of the Gaussian process by the following formula

Obtaining a parameter learning result by utilizing a gradient-based optimization algorithm, wherein each component theta of an unknown parameter theta is subjected to_iThe derivation is shown as follows:

wherein gamma is ∑_Y ^-1(Y-u_Y)，

And respectively establishing Gaussian process functions under different driving intentions and driving characteristics through parameter learning of the Gaussian process.

Compared with the prior art, the invention has the beneficial effects that:

the vehicle track prediction method based on the hybrid Dynamic Bayesian Network and the Gaussian Process realizes vehicle track prediction based on vehicle model fusion and data learning by combining the hybrid Dynamic Bayesian Network (MDBN) and the Gaussian Process (GP). Compared with the conventional vehicle trajectory prediction method, the method combines the vehicle model, the abstract intention and the data drive, and the MDBN and GP models have strong expansibility, can be suitable for different driving scenes and can combine more effective situation information, such as road information and traffic information.

Drawings

FIG. 1 is a block diagram of a vehicle trajectory prediction method based on a hybrid dynamic Bayesian network and Gaussian process in accordance with the present invention;

FIG. 2 is a schematic diagram of the hybrid dynamic Bayesian network (MDBN) of the present invention.

Detailed Description

The invention is further illustrated with reference to the following figures and examples.

As shown in fig. 1, the vehicle trajectory prediction method based on the hybrid dynamic bayesian network and the gaussian process includes the following steps:

step 1, constructing a natural driving database;

establishing a test set of peripheral vehicle related sequence information, road related sequence information and traffic related sequence information acquired by an automatic driving vehicle sensing system, and a training set for calibrating driving intention and driving characteristics of the information; wherein the test set comprises a test set of a hybrid dynamic Bayesian network and a test set of a Gaussian process; the training set comprises a training set of a hybrid dynamic Bayesian network and a training set of a Gaussian process;

the peripheral vehicle related sequence information comprises vehicle position, vehicle speed, vehicle acceleration, vehicle yaw velocity, the distance between the vehicle and the two sides of the road and the opening and closing conditions of left and right steering tail lamps of the vehicle;

the road-related sequence information comprises structural characteristics (such as straight roads and curved roads) of the road and road indication signs;

the traffic-related sequence information includes surrounding traffic indication signs and traffic light states.

Step 1.1, acquisition of Driving data

In order to train a hybrid dynamic Bayesian network model and a function model in a Gaussian process, a large number of training samples are needed, so that surrounding vehicle related sequence information, road related sequence information and traffic related sequence information of an automatic driving vehicle sensing system in a normal driving process need to be collected; the automatic driving vehicle perception system comprises sensor technologies (such as radar, a camera and navigation positioning) and wireless communication technologies (such as V2V and V2I), and the data types processed by the perception system comprise physical data and image data. In order to match with the requirement of network model training, according to the requirement of a hybrid dynamic Bayesian network MDBN observation layer, relevant sequence information is acquired and processed, wherein the relevant sequence information comprises the vehicle position, the vehicle speed, the vehicle acceleration, the vehicle yaw velocity and the distance between the vehicle and the two sides of a road, and the opening and closing conditions of the left and right steering tail lamps of the vehicle, the structural characteristics of the road, the road indication sign, the surrounding traffic indication sign and the traffic signal lamp state are acquired through image data processing.

Step 1.2, classifying and calibrating driving intentions

Defining possible driving intentions of surrounding vehicles according to the scene of the vehicles; for each possible driving intention, selecting key characteristic parameters for classifying and calibrating the driving intention; and then, combining the related sequence information of the surrounding vehicles collected in the step 1.1, judging the driving intention of the surrounding vehicles in a period of time and setting corresponding driving intention labels.

The scene of the vehicle can be obtained by an intelligent navigation system of the automatic driving vehicle, and the scene of the vehicle comprises a high-speed scene, a mountain road scene, a normal urban straight road scene, an urban traffic signal intersection scene, an urban non-traffic signal intersection scene and the like.

The driving intention comprises straight running along the current road, turning along the current road, left lane changing, right lane changing, left turning, right turning, turning around, parking and starting.

The key characteristic parameters for the classification and calibration of the driving intention comprise the transverse position of the vehicle, the transverse speed of the vehicle, the yaw rate of the vehicle, the distance between the vehicle and the two sides of the road and the opening and closing conditions of left and right steering tail lamps of the vehicle.

For example, if the vehicle is in a high speed scene, the driving intentions of the surrounding predicted vehicle are straight along the current road, turn along the current road, change lane left and change lane right, and the four types of driving intention labels are set to 1, 2, 3 and 4 respectively. Aiming at the driving intention of straight running along the current road, selecting the transverse position of the vehicle and the distance between the vehicle and the two sides of the road as key characteristic parameters; aiming at the driving intention of turning along the current road, selecting the yaw velocity of the vehicle and the distance between the vehicle and the two sides of the road as key characteristic parameters; and aiming at the driving intentions of the left lane changing and the right lane changing, selecting the transverse position of the vehicle and the distance between the vehicle and the two sides of the road as key characteristic parameters. Assuming that the relevant sequence information of a certain surrounding vehicle in a period of time collected in step 1.1 is that both the lateral position of the vehicle and the distance from the vehicle to both sides of the road remain unchanged, it can be determined that the driving intention of the vehicle in the period of time is to go straight along the current road, and the driving intention tag is set to 1.

Step 1.3, classification and calibration of driving characteristics

The driving characteristics of the drivers are different under the same intention, so that the driving tracks of the vehicles are obviously different. Firstly, defining possible driving characteristics of the surrounding vehicle, selecting key characteristic parameters for classification and judgment of the driving characteristics, judging the driving characteristics of the surrounding vehicle under a certain driving intention and setting a corresponding driving characteristic label according to the related sequence information of the surrounding vehicle acquired in the step 1.1 and the driving intention of the surrounding vehicle in a period of time acquired in the step 1.2.

The driving characteristics include slow driving, smooth driving, and jerky driving;

the key characteristic parameters for the classification and determination of the driving characteristics include a vehicle longitudinal speed, a vehicle lateral speed, a vehicle longitudinal acceleration, a vehicle lateral acceleration, and a vehicle yaw rate.

For example, if the vehicle is in a high-speed scene, the three types of driving characteristic tags are set to 1, 2, and 3, respectively. Aiming at the driving intention of straight running along the current road, selecting the longitudinal speed and the longitudinal acceleration of the vehicle as key characteristic parameters; aiming at the driving intention of turning along the current road, selecting the yaw angular speed of the vehicle as a key characteristic parameter; and selecting the lateral speed and the lateral acceleration of the vehicle as key characteristic parameters aiming at the driving intentions of the left lane changing and the right lane changing. Assuming that the driving intention of a certain surrounding vehicle obtained in step 1.2 is straight-ahead along the current road, and the relevant sequence information of the vehicle collected in step 1.1 in a period of time is that the longitudinal speed of the vehicle is about 60km/h, and the longitudinal acceleration of the vehicle fluctuates around a zero value, it can be determined that the driving characteristic of the vehicle under the driving intention straight-ahead along the current road is slow driving, and the driving characteristic label is set to 1.

Step 1.4, establishment of training set and testing set

The training set is: a data set prepared for network model parameter learning. And calibrating the driving intention and the driving characteristics of the surrounding vehicle related sequence information, the road related sequence information and the traffic related sequence information.

The test set is: a data set prepared for verifying the validity of the learned network model parameters. The surrounding vehicle-related sequence information, the road-related sequence information, and the traffic-related sequence information are taken as inputs of the network model.

Taking the peripheral vehicle related sequence information, the road related sequence information and the traffic related sequence information collected in the step 1.1 as a test set of the hybrid dynamic Bayesian network MDBN;

taking sequence information of front section time corresponding to a transverse position and a longitudinal position of a vehicle under a certain driving intention and driving characteristics as a test set of a Gaussian process;

1.1, obtaining a training set of a hybrid dynamic Bayesian network (MDBN) after the surrounding vehicle related sequence information, the road related sequence information and the traffic related sequence information which are collected in the step 1.1 are classified and calibrated according to the driving intention in the step 1.2 and the driving characteristics in the step 1.3;

according to the training set of the hybrid dynamic Bayesian network MDBN, all sequence information corresponding to the transverse position and the longitudinal position of the vehicle under a certain driving intention and driving characteristics is used as one training set of the Gaussian process GP.

And 2, obtaining short-term predicted tracks of surrounding vehicles and estimated probabilities of driving intentions and driving characteristics by adopting a hybrid dynamic Bayesian network.

Taking driver information, vehicle model selection and vehicle state information as hidden layer variables, taking surrounding vehicle related sequence information, road related sequence information and traffic related sequence information acquired by an automatic driving vehicle sensing system as observation layer variables, and building a hybrid dynamic Bayesian network model; and obtaining the output of the mixed dynamic Bayesian network MDBN through posterior probability inference: the short-term trajectory prediction based on the vehicle model and the estimated probabilities of driving intent and driving characteristics are used as input to step 4.

The observation layer variable, the hidden layer variable and the connection relation between the two variables are obtained by utilizing natural driving data and an improved genetic algorithm; and obtaining the probability parameter corresponding to the connection relation by a maximum likelihood method.

The vehicle model comprises a uniform speed model, a uniform acceleration model and a uniform speed steering model.

The vehicle state information includes a vehicle lateral position, a vehicle lateral velocity, a vehicle lateral acceleration, a vehicle longitudinal position, a vehicle longitudinal velocity, a vehicle longitudinal acceleration, and a vehicle yaw rate.

Preferably, due to the scalability of the network node, variables can be added if other high-level semantics are considered, the hidden variables further including other relevant high-level hidden variables including risk assessment of the vehicle and the environment, interaction between vehicles. The observation layer variables further comprise other relevant bottom layer observation information corresponding to other relevant high layer hidden variables of the hidden layer variables.

The step 2 specifically comprises the following steps:

step 2.1, defining hidden layer and observation layer nodes of hybrid dynamic Bayesian network model

As shown in fig. 2, the hybrid dynamic bayesian network mainly comprises nodes and edges, where the nodes represent variables, the edges represent dependency relationships among the variables, the nodes are divided into hidden layer nodes and observation layer nodes, and the representation of the network model mainly defines the number and meaning of the hidden layer nodes and the observation layer nodes according to the traffic scene where the vehicle is located, and is defined as follows:

① hidden nodes contain three layers of information:

the first layer is a driver information variable D which comprises driving intentions and driving characteristics and reflects the possible driving intentions (such as the driving intentions of a high-speed scene comprise straight running along the current road, turning along the current road, left lane changing and right lane changing) and the driving characteristics (such as slow driving, smooth driving and jerky driving) of surrounding vehicles under the scene that a certain vehicle is positioned;

the second layer is a vehicle model selection variable M for selecting a vehicle model under a certain observation sequence. The vehicle model comprises a constant speed model, an acceleration model and a constant speed steering model;

the third level is a vehicle state information variable S, which selects the vehicle state information required by the vehicle model. The vehicle state information includes a vehicle lateral position, a vehicle lateral velocity, a vehicle lateral acceleration, a vehicle longitudinal position, a vehicle longitudinal velocity, a vehicle longitudinal acceleration, and a vehicle yaw rate. As shown in fig. 2, the hidden layer variable set H ═ D, M, S, where { D, M } is a discrete variable and { S } is a continuous variable.

② the observation layer node mainly contains one layer of information:

the observation information comprises surrounding vehicle related sequence information, road related sequence information and traffic related sequence information which are collected by an automatic driving vehicle perception system, and the specific content of the observation information is shown in step 1. As in FIG. 2, the set of observation layer variables is { O }, and is a continuous variable.

The expression form of the hybrid dynamic bayesian network is shown in fig. 2, which is a bayesian network developed according to a time sequence, and since a vehicle state information variable S is added, a hidden layer variable at this time has both a discrete variable and a continuous variable, the hybrid dynamic bayesian network is called.

Step 2.2 optimization of network architecture

Optimization of network structure, also called structure learning of network model, is a network structure determination problem under known network parameters, that is, given a variable set { D, M, S, O } and a training set T, a network N ═ B, θ, which best matches the training set T, is found by means of learning and searching, and a criterion function is used as a measure of the degree of matching. The structure of the network refers to the connection relationship between defined network nodes, B represents the network structure, and θ represents the parameters of the network.

As shown in fig. 2, if there are certain probability relations between hidden layer variables, between hidden layer variables and observation layer variables, they are linked by directed edges, otherwise, the two variables are independent from each other, so that a network structure is formed by nodes and directed edges; if there is a directed edge between the variables, this directed edge gives conditional probabilities between the variables, which are parameters of the network. The structure and parameters of the network can be obtained according to data learning, and can also be artificially defined.

For the structure optimization problem of the hybrid dynamic bayesian network MDBN, most dynamic bayesian network structures are artificially defined according to practical problems due to the relatively large calculation amount, especially under the condition of a large number of network nodes. Here, since some nodes with strong association in the defined network nodes may define their connection relationship in advance to reduce the amount of computation, as shown in fig. 2, the node M and the node S in the graph have strong association, and a directed edge may be determined. When a certain amount of node variables are defined and a certain amount of data is given, network structure learning can be performed through a heuristic algorithm, such as a genetic algorithm of a greedy algorithm mechanism, a probabilistic model-based evolution algorithm, and the like. When there is sufficient data support, the network structure is learned using the data so that the network structure can better fit the data.

Step 2.3, learning of unknown parameters between network nodes

The unknown parameter learning among the network nodes is also called the parameter learning of the network model, and is a problem about determining the network parameters under the known network structure, wherein the parameters of the network are the conditional probability in the network model.

For the parameter learning problem of the hybrid dynamic bayesian network MDBN, the conditional probability of the partial hidden layer variables can be obtained by a data statistics method, so that a certain amount of calculation is reduced, the conditional probabilities between the rest hidden layers and the observed quantity can be learned by using the training set of the hybrid dynamic bayesian network MDBN obtained in step 1, a Maximum Likelihood Estimation (MLE) method is used to solve the problem, the Likelihood function L (θ) related to the network parameter θ under the known training sample D is logP (D | θ), and an optimal solution is solved by using a gradient-based optimization algorithm

To avoid falling into the local optimum, the initial value may be randomly selected a number of times.

The structure learning and the parameter learning jointly form the learning content of the network model, and the theory of the learning content of the network model needs to be supported by the reasoning content of the network model in the next step, wherein the learning content of the network model is explained first for convenience of explanation.

2.4 posterior probabilistic reasoning with introduction of Observation layer variables

The posterior probabilistic reasoning introduced under the observation layer variables is also called reasoning of the network model. The method mainly solves the problem of calculating the posterior probability under given network parameters and structures, wherein the posterior probability refers to the probability distribution P (H | O) of hidden variables under known observation layer variables, namely the process of deducing corresponding driving intentions, driving characteristics, vehicle models and vehicle states under the condition of known time sequence information (namely the mixed dynamic Bayesian network MDBN test set obtained in the step 1) of each observation variable. The inference method mainly includes precise inference and approximate inference, and the precise inference Algorithm is a Forward-Backward learning Algorithm (Forward-Backward learning Algorithm), a decomposition Tree Algorithm (Unrolled Junction Tree Algorithm), etc. because of the sequence ductility of the dynamic bayesian network, the use of the precise Algorithm easily causes NP hard problem, so the suitable approximation Algorithm is used, such as adm (moment matching), Importance Sampling (immunity Sampling), markov monte carlo (MCMC), Particle Filtering (Particle Filtering), etc.

For the MDBN framework in fig. 2, a hypothetical density filtering adf (assumed density filtering) approximation reasoning process is presented here:

① predicting, obtaining the joint probability distribution of t +1 time by the probability distribution of t time, setting all discrete hidden layer variables { M ] of t time_t,D_tThe joint distribution of

Discrete hidden layer variable M at time t +1_t,D_t,M_t+1,D_t+1The joint distribution is:

continuous hidden layer variable (S) at time t +1_t+1The conditional probabilities are:

② updating, introducing the observed variable at t +1 moment to obtain new probability distribution, setting the observed variable of the vehicle model as V, and satisfying

The rest of the observed variables in the O are E, and the discrete hidden layer variable M at the t +1 moment_t,D_t,M_t+1,D_t+1The joint distribution of the { is:

wherein,

continuous hidden layer variable (S) at time t +1_t+1The conditional probabilities are:

③ marginalizing the variables before the t +1 moment in the t +1 moment probability distribution obtained from the above, and only keeping the probability distribution of the variables at the current t +1 moment, so as to be used for the iteration of the next prediction and update steps:

wherein D represents a driver information variable, the driver information including a driving intention and a driving characteristic; m represents a vehicle model selection variable, and the vehicle model comprises a constant speed model, an acceleration model and a constant speed steering model; s denotes a vehicle state information variable, and the vehicle state information includes a vehicle lateral position, a vehicle lateral speed, a vehicle lateral acceleration, a vehicle longitudinal position, a vehicle longitudinal speed, a vehicle longitudinal acceleration, and a vehicle yaw rate. By using

The estimated probability of the vehicle model, the driving intention and the driving characteristic at the moment of t +1 can be obtained; by using

A short term predicted trajectory based on the vehicle model may be obtained.

The meaning of each network node is defined through step 2.1; step 2.2, determining the connection relation of the network nodes; and 2.3, determining parameters of the network, so that the posterior probability reasoning under the observation variables is introduced in the step 2.4 to obtain a short-term predicted trajectory based on the vehicle model and the estimated probabilities of the driving intention and the driving characteristics, and the short-term predicted trajectory and the estimated probabilities are fused with the Gaussian process in the step 4 to obtain a final predicted trajectory.

And 3, establishing Gaussian process functions under different driving intentions and different driving characteristics.

Since the driving characteristics of the drivers are different under the same intention, the driving tracks of the vehicles are obviously different, and therefore Gaussian process functions under different driving intentions and different driving characteristics are respectively established according to the different driving intentions and driving characteristics.

Uncertainty in vehicle trajectory prediction is mainly caused by errors existing in a vehicle sensing system and the dynamic and random properties of surrounding traffic environments, and these uncertain factors may cause deviation of vehicle trajectory prediction results.

A gaussian process function of the vehicle trajectory for different driving characteristics under these intentions is established. The Gaussian process mainly comprises the following steps:

step 3.1 setting of mean function and covariance function

With x as an input, the expression of the gaussian process function is as follows:

f(x)～GP(u(x),Σ(x,x’))

wherein the mean function u (x) represents a vehicle trajectory trend under a certain driving intention, and thus the mean function can be used to represent a desired predicted trajectory; the covariance function Σ (x, x ') represents both the variance of the different inputs x themselves and the variance between (x, x'), and thus the covariance function can be used to represent the uncertainty corresponding to the desired predicted trajectory.

The mean function has various forms, such as a constant, a linear function, various non-linear functions, and the like, and the corresponding mean function u (x) is set according to the vehicle tracks under different driving intentions, such as in a high-speed scene, the mean function under the straight driving intention along the current road can be set to be a constant, the mean function under the turning driving intention along the current road can be set to be a quadratic function, and the mean function under the left lane changing or right lane changing driving intention can be set to be a quintic polynomial form. The covariance function is also called a kernel function, and there are various commonly used kernel functions, such as a linear function, a gaussian kernel function, a symmetric kernel function, etc., different kernel functions have different characteristics, and the kernel function here selects a mean square index form in consideration of high smoothness of a vehicle running track, and besides, since a noise variable needs to be introduced in consideration of existence of a prediction uncertainty factor, and it is set as white gaussian noise here, the covariance function Σ (x, x') expressed in the form of a noisy mean square index is set as follows:

wherein σ_fIs the signal standard deviation, l is the characteristic length, σ_nTo observe the noise standard deviation, δ is the Kronecker function of Kronecker.

Step 3.2, learning of unknown parameters of Gaussian process function

According to the mean function u (x) and the covariance function Σ (x, x') set in step 3.1, the gaussian process GP training set D established in step 1 is used_GP(X,Y)＝{(x_i,y_i),i＝1:N₁Learning the unknown parameters involved, where N₁For the sequence length of the training samples, let the unknown parameters of the mean function and covariance function be Θ ═ θ_i,i＝1:N₂In which N is₂For the number of parameters, the log marginal likelihood function logP (Y | X, Θ) is:

wherein,

represents a normal function; x and Y represent the lateral and longitudinal vehicle position; u. of_YAnd ∑_YA mean function and a covariance function corresponding to the variable Y;

selecting a maximum likelihood method, and solving to obtain unknown parameters of the Gaussian process by the following formula

The parameter learning result can be obtained by utilizing a gradient-based optimization algorithm, wherein each component theta of the unknown parameter theta is subjected to_iThe derivation is shown as follows:

wherein gamma is ∑_Y ^-1(Y-u_Y) It is an operator used to simplify the expression.

Through parameter learning of the Gaussian process, Gaussian process functions under different driving intentions and driving characteristics can be established respectively. Since this algorithm may result in local optima, several sets of initial parameters need to be set more.

And 4, performing long-term trajectory prediction and uncertainty representation thereof based on the mixed dynamic Bayesian network MDBN and the Gaussian process GP.

The results obtained by the hybrid dynamic bayesian network MDBN are suitable for prediction in a short time, and therefore long-term trajectory prediction needs to be performed on short-term prediction results in combination with the gaussian process, while describing prediction uncertainty.

Here, the driving intention and the driving characteristics are determined using the maximum probability principle based on the estimated probabilities of the driving intention and the driving characteristics output from the hybrid dynamic bayesian network MDBN in step 2, so that the short-term predicted trajectory based on the vehicle model output from the hybrid dynamic bayesian network MDBN in step 2 is taken as x based on the corresponding gaussian process function in step 3 determined from the driving intention and the driving characteristics₁Predicted future trajectory is x₂Then (x)₁,x₂) The gaussian distribution obeyed is as follows:

wherein,

represents a normal function;

x₁and x₂The respective corresponding mean functions are respectively u₁And u₂(ii) a The covariance matrix Σ is a symmetric matrix, i.e., Σ ═ Σ^T，x₁And x₂The respective covariance functions are ∑₁₁Sum-sigma₂₂，x₁And x₂The corresponding covariance function is ∑₁₂、Σ₂₁And sigma₁₂＝Σ₂₁ ^T；

Then the vehicle trajectory x is known₁Future possible trajectory x of vehicle₂Conditional probability P (x) of₂|x₁) The expression of gaussian distribution obeyed is as follows:

wherein u is_2|1Sum-sigma_2|1Is a variable x_2|1Corresponding mean function and covariance function, and (u)_2|1,Σ_2|1) The expression (c) is derived from the following formula:

u_2|1＝u₂+∑₁₂ ^T∑₁₁ ^-1(x₁-u₁)

∑_2|1＝∑₂₂-∑₁₂ ^T∑₁₁ ^-1∑₁₂

and finally obtaining the future track and uncertainty representation thereof in the long-term domain of the vehicle under the driving intention and the driving characteristics determined by using the maximum probability principle.

Step 4 combines step 2 with step 3 to complete the transition from short-term prediction to long-term prediction while representing the prediction uncertainty.

The vehicle track prediction method based on the hybrid dynamic Bayesian network and the Gaussian process mainly comprises three stages:

the first stage is to define, reason and relevant data learning of the hybrid dynamic Bayesian network according to the driving scene, and the main function of the first stage is to provide basis for the next stage. Inputting a section of time sequence data of the network model observation variable into the MDBN network model to obtain the estimated probability of the driving intention and the driving characteristic of surrounding vehicles in the period and the short-term predicted track based on the vehicle model;

the second stage is to establish corresponding Gaussian process functions according to different driving intentions and driving characteristics, and the main functions of the Gaussian process functions are to predict the long-term track of the vehicle and express prediction uncertainty. And calculating the possible future track of the vehicle under the known partial vehicle track sequence information according to the partial sequence information of the transverse position and the longitudinal position of the vehicle under a certain driving intention and driving characteristics.

And the third stage is to complete the determination of the driving intention and the prediction of the trajectory and the representation of the prediction uncertainty in the long-term range of the vehicle under the driving characteristics according to the output result of the first stage and the content of the second stage.

Claims (9)

1. A vehicle track prediction method based on a hybrid dynamic Bayesian network and a Gaussian process is characterized in that: the method comprises the following steps:

step 1, constructing a natural driving database;

establishing a test set of peripheral vehicle related sequence information, road related sequence information and traffic related sequence information acquired by an automatic driving vehicle sensing system, and a training set for calibrating driving intention and driving characteristics of the information; the peripheral vehicle related sequence information comprises a vehicle position, a vehicle speed, a vehicle acceleration, a vehicle yaw rate, a distance between the vehicle and two sides of a road and the opening and closing conditions of left and right steering tail lamps of the vehicle; the road related sequence information comprises structural characteristics of roads and road indication signs; the traffic related sequence information comprises surrounding traffic indication signs and traffic light states; the test set comprises a test set of a hybrid dynamic Bayesian network and a test set of a Gaussian process; the training set comprises a training set of a hybrid dynamic Bayesian network and a training set of a Gaussian process;

step 2, obtaining short-term predicted trajectories of surrounding vehicles and estimated probabilities of driving intentions and driving characteristics by adopting a hybrid dynamic Bayesian network;

taking driver information, vehicle model selection and vehicle state information as hidden layer variables, taking surrounding vehicle related sequence information, road related sequence information and traffic related sequence information acquired by an automatic driving vehicle sensing system as observation layer variables, and building a hybrid dynamic Bayesian network model; and obtaining the output of the hybrid dynamic Bayesian network through posterior probability inference: predicting short-term trajectory based on the vehicle model and the estimated probability of driving intention and driving characteristics, and taking the output as the input of step 4;

step 3, establishing Gaussian process functions under different driving intentions and different driving characteristics;

comprises the following steps:

step 3.1 setting of mean function and covariance function

With x as an input, the expression of the gaussian process function is as follows:

f(x)～GP(u(x),Σ(x,x’))

the mean function u (x) represents the vehicle track trend under a certain driving intention, so that the mean function can be used for representing the expected predicted track; the covariance function Σ (x, x ') represents both the variance of the different inputs x themselves and the variance between (x, x'), and therefore the covariance function can be used to represent the uncertainty corresponding to the desired predicted trajectory;

setting corresponding mean functions u (x) according to vehicle tracks under different driving intentions;

the covariance function Σ (x, x') represented in the form of a noisy mean-square index is set as follows:

wherein σ_fIs the signal standard deviation, l is the characteristic length, σ_nFor observing noiseStandard deviation, δ is the kronecker function;

step 3.2, learning of unknown parameters of Gaussian process function

Learning related unknown parameters by using the training set of the Gaussian process established in the step 1 according to the mean function u (x) and the covariance function Σ (x, x') set in the step 3.1, and obtaining parameter learning results by using a gradient-based optimization algorithm according to the logarithmic marginal likelihood function and the partial derivatives of the unknown parameters thereof, thereby respectively establishing Gaussian process functions under different driving intents and driving characteristics;

step 4, long-term trajectory prediction and uncertainty representation are carried out based on a hybrid dynamic Bayesian network and a Gaussian process;

determining the corresponding driving intention and driving characteristics by using a maximum probability principle according to the driving intention and the estimated probability of the driving characteristics output by the hybrid dynamic Bayesian network in the step 2, thereby determining the corresponding Gaussian process function in the step 3 according to the driving intention and the driving characteristics;

using the short-term predicted track based on the vehicle model output by the hybrid dynamic Bayesian network in the step 2 as vehicle track sequence information x₁Predicted future trajectory is x₂Then (x)₁,x₂) The gaussian distribution obeyed is as follows:

wherein,

represents a normal function;

x₁and x₂The respective corresponding mean functions are respectively u₁And u₂(ii) a The covariance matrix Σ is a symmetric matrix, i.e., Σ ═ Σ^T，x₁And x₂The respective covariance functions are ∑₁₁Sum-sigma₂₂，x₁And x₂Corresponding to each otherCovariance function of ∑₁₂、Σ₂₁And ∑₁₂＝∑₂₁ ^T；

Then the vehicle trajectory x is known₁Future possible trajectory x of vehicle₂Conditional probability P (x) of₂|x₁) The expression of gaussian distribution obeyed is as follows:

wherein u is_2|1Sum-sigma_2|1Is a variable x_2|1Corresponding mean function and covariance function, and (u)_2|1,Σ_2|1) The expression (c) is derived from the following formula:

u_2|1＝u₂+Σ₁₂ ^TΣ₁₁ ^-1(x₁-u₁)

Σ_2|1＝Σ₂₂-Σ₁₂ ^TΣ₁₁ ^-1Σ₁₂

and finally obtaining the future track and uncertainty representation thereof in the long-term domain of the vehicle under the driving intention and the driving characteristics determined by using the maximum probability principle.

2. The hybrid dynamic bayesian network and gaussian process based vehicle trajectory prediction method of claim 1, wherein: the step 1 comprises the following steps:

1.1 acquisition of Driving data

Collecting vehicle related sequence information, road related sequence information and traffic related sequence information around an automatic driving vehicle sensing system in a normal driving process;

1.2 classifying and calibrating driving intentions

Defining possible driving intentions of surrounding vehicles according to the scene of the vehicles; for each possible driving intention, selecting key characteristic parameters for classifying and calibrating the driving intention; then, the driving intention of the surrounding vehicles within a period of time is judged and corresponding driving intention labels are set by combining the related sequence information of the surrounding vehicles acquired in the step 1.1;

the scene of the vehicle is obtained by an intelligent navigation system of the automatic driving vehicle, and comprises a high-speed scene, a mountain road scene, an urban normal straight road scene, an urban traffic signal lamp intersection scene and an urban traffic signal lamp-free intersection scene;

the driving intention comprises straight running along the current road, turning along the current road, left lane changing, right lane changing, left turning, right turning, turning around, parking and starting;

the key characteristic parameters for classifying and calibrating the driving intention comprise the transverse position of the vehicle, the transverse speed of the vehicle, the yaw speed of the vehicle, the distance between the vehicle and the two sides of the road and the opening and closing conditions of left and right steering tail lamps of the vehicle;

1.3 Classification and calibration of Driving characteristics

Firstly, defining possible driving characteristics of surrounding vehicles, selecting key characteristic parameters for classification and judgment of the driving characteristics, judging the driving characteristics of the surrounding vehicles under a certain driving intention and setting corresponding driving characteristic labels according to the related sequence information of the surrounding vehicles acquired in the step 1.1 and the driving intention of the surrounding vehicles in a period of time acquired in the step 1.2;

the driving characteristics include slow driving, smooth driving, and jerky driving;

the key characteristic parameters for the classification and determination of the driving characteristics comprise a vehicle longitudinal speed, a vehicle lateral speed, a vehicle longitudinal acceleration, a vehicle lateral acceleration and a vehicle yaw rate;

1.4 creation of training set and test set

Taking the peripheral vehicle related sequence information, the road related sequence information and the traffic related sequence information collected in the step 1.1 as a test set of the hybrid dynamic Bayesian network MDBN;

taking sequence information of front section time corresponding to a transverse position and a longitudinal position of a vehicle under a certain driving intention and driving characteristics as a test set of a Gaussian process;

1.1, obtaining a training set of a hybrid dynamic Bayesian network (MDBN) after the surrounding vehicle related sequence information, the road related sequence information and the traffic related sequence information which are collected in the step 1.1 are classified and calibrated according to the driving intention in the step 1.2 and the driving characteristics in the step 1.3;

according to the training set of the hybrid dynamic Bayesian network MDBN, all sequence information corresponding to the transverse position and the longitudinal position of the vehicle under a certain driving intention and driving characteristics is used as one training set of the Gaussian process GP.

3. The hybrid dynamic bayesian network and gaussian process based vehicle trajectory prediction method of claim 1, wherein: in the step 2, in the step of processing,

the observation layer variable, the hidden layer variable and the connection relation between the two variables are obtained by utilizing natural driving data and an improved genetic algorithm; and obtaining the probability parameter corresponding to the connection relation by a maximum likelihood method.

4. The hybrid dynamic bayesian network and gaussian process based vehicle trajectory prediction method of claim 1, wherein: in the step 2, in the step of processing,

the vehicle model comprises a uniform speed model, a uniform acceleration model and a uniform speed steering model;

the vehicle state information includes a vehicle lateral position, a vehicle lateral velocity, a vehicle lateral acceleration, a vehicle longitudinal position, a vehicle longitudinal velocity, a vehicle longitudinal acceleration, and a vehicle yaw rate.

5. The hybrid dynamic bayesian network and gaussian process based vehicle trajectory prediction method of claim 1, wherein: in the step 2, in the step of processing,

the hidden variables further comprise other related high-level hidden variables, and the other related high-level hidden variables comprise risk assessment of the vehicle and the environment and interaction between the vehicles;

the observation layer variables further comprise other relevant bottom layer observation information corresponding to other relevant high layer hidden variables of the hidden layer variables.

6. The hybrid dynamic bayesian network and gaussian process based vehicle trajectory prediction method of claim 1, wherein: the step 2 specifically comprises the following steps:

2.1 defining hidden and observation layer nodes of hybrid dynamic Bayesian network model

Hidden nodes contain three layers of information: the first layer is a driver information variable D, the second layer is a vehicle model selection variable M, and the third layer is a vehicle state information variable S; the hidden layer variable set H is { D, M, S }, wherein { D, M } is a discrete variable, and { S } is a continuous variable;

the observation layer node mainly comprises one layer of information: the related sequence information of surrounding vehicles, the related sequence information of roads and the related sequence information of traffic, which are collected by an automatic driving vehicle sensing system; the observation layer variable set is { O }, and is a continuous variable;

2.2 optimization of network architecture

Given a variable set { D, M, S, O } and a training set T, finding a network N ═ B, θ which best matches the training set T by means of learning and searching, and using a criterion function as a measure of the degree of matching; wherein, B represents a network structure, and theta represents a parameter of the network;

2.3 learning of unknown parameters between network nodes

The conditional probability of part hidden layer variables is obtained through a data statistics method, the conditional probability between the rest hidden layers and the observed quantity is subjected to parameter learning by using the training set of the mixed dynamic Bayesian network MDBN obtained in the step 1, the solution is achieved by using a maximum likelihood estimation method, the likelihood function L (theta) related to the network parameter theta under the known training sample D is logP (D | theta), and the optimal solution is solved by using a gradient-based optimization algorithm

2.4 posterior probabilistic reasoning with introduction of Observation layer variables

The posterior probability refers to the probability distribution P (H | O) of hidden layer variables under the known observation layer variables, that is, the process of deducing the corresponding driving intention, driving characteristics, vehicle model and vehicle state under the test set of the hybrid dynamic bayesian network MDBN obtained in step 1, and the short-term trajectory prediction based on the vehicle model and the estimation probability of the driving intention and the driving characteristics are obtained.

7. The hybrid dynamic bayesian network and gaussian process based vehicle trajectory prediction method of claim 1, wherein: in the step 2, the posterior probability inference is assumed density filtering approximation inference, and the specific process is as follows:

1) and (3) prediction: obtaining the joint probability distribution of the t +1 moment by using the probability distribution of the t moment; all discrete hidden layer variables { M) at time t_t,D_tThe joint distribution of

Discrete hidden layer variable M at time t +1_t,D_t,M_t+1,D_t+1The joint distribution is:

continuous hidden layer variable (S) at time t +1_t+1The conditional probabilities are:

2) updating: introducing an observation variable at the t +1 moment to obtain new probability distribution; let the observed variable of the vehicle model be { V }, and satisfy

The rest of the observed variables in the O are E, and the discrete hidden layer variable M at the t +1 moment_t,D_t,M_t+1,D_t+1The joint distribution of the { is:

wherein,

continuous hidden layer variable (S) at time t +1_t+1The conditional probabilities are:

3) marginalizing: marginalizing the variables before the t +1 moment in the obtained t +1 moment probability distribution, and only keeping the probability distribution of the variables at the current t +1 moment so as to be used for the iteration of the next prediction and updating step:

wherein D represents a driver information variable, the driver information including a driving intention and a driving characteristic; m represents a vehicle model selection variable, and the vehicle model comprises a constant speed model, an acceleration model and a constant speed steering model; s represents a vehicle state information variable, and the vehicle state information comprises a vehicle transverse position, a vehicle transverse speed, a vehicle transverse acceleration, a vehicle longitudinal position, a vehicle longitudinal speed, a vehicle longitudinal acceleration and a vehicle yaw rate; by using

Obtaining the estimated probability of the vehicle model, the driving intention and the driving characteristic at the moment of t + 1; by using

A short-term predicted trajectory based on the vehicle model is obtained.

8. The hybrid dynamic bayesian network and gaussian process based vehicle trajectory prediction method of claim 1, wherein: in the step 3.2, the maximum likelihood method is selected to learn the unknown parameters in the gaussian process.

9. The hybrid dynamic bayesian network and gaussian process based vehicle trajectory prediction method of claim 1, wherein: the specific process of the step 3.2 is as follows:

according to the mean function u (x) and the covariance function Σ (x, x') set in step 3.1, the gaussian process GP training set D established in step 1 is used_GP(X,Y)＝{(x_i,y_i),i＝1:N₁Learning the unknown parameters involved, where N₁For the sequence length of the training samples, let the unknown parameters of the mean function and covariance function be Θ ═ θ_i,i＝1:N₂In which N is₂For the number of parameters, the log marginal likelihood function logP (Y | X, Θ) is:

wherein,

represents a normal function; x and Y represent the lateral and longitudinal vehicle position; u. of_YAnd ∑_YA mean function and a covariance function corresponding to the variable Y;

selecting a maximum likelihood method, and solving to obtain unknown parameters of the Gaussian process by the following formula

Obtaining a parameter learning result by utilizing a gradient-based optimization algorithm, wherein each component theta of an unknown parameter theta is subjected to_iThe derivation is shown as follows:

wherein gamma is ∑_Y ^-1(Y-u_Y)，

And respectively establishing Gaussian process functions under different driving intentions and driving characteristics through parameter learning of the Gaussian process.

Images