CN113561974A

CN113561974A - Collision risk prediction method based on vehicle behavior interaction and road structure coupling

Info

Publication number: CN113561974A
Application number: CN202110983185.2A
Authority: CN
Inventors: 王建强; 崔明阳; 杨路; 黄荷叶; 林学武; 许庆
Original assignee: Tsinghua University
Current assignee: Tsinghua University
Priority date: 2021-08-25
Filing date: 2021-08-25
Publication date: 2021-10-29
Anticipated expiration: 2041-08-25
Also published as: CN113561974B

Abstract

The application discloses a collision risk prediction method and device based on vehicle behavior interaction and road structure coupling, which are based on road structure classification and potential double-vehicle collision identification and project to two types of basic interactive collision scene models; aiming at vehicle-vehicle conflicts, an intention identification model based on a dynamic Bayesian network is established, and the conditional probability relation between the vehicle passing intention and the environmental situation and the driving behavior is described; based on observable information, carrying out probability inference on environment situation and behavior semantics; and training and optimizing model parameters of the dynamic Bayesian network based on natural driving data by adopting an EM algorithm. And based on the pass intention identification result, predicting the vehicle running track and the time-space distribution thereof by using a Gaussian process regression algorithm, and evaluating the collision risk of the two vehicles. Therefore, under complex traffic scenes such as ramp junction, intersection passage and the like, collision risk prediction of multi-vehicle behavior interactive coupling is considered, the method can be widely applied to species conflict scenes, and the driving safety of intelligent vehicles is improved.

Description

Collision risk prediction method based on vehicle behavior interaction and road structure coupling

Technical Field

The application relates to the technical field of environment cognition of intelligent driving vehicles, in particular to a collision risk prediction method and device based on vehicle behavior interaction and road structure coupling.

Background

The intelligent driving technology is a basic technology for realizing a safer and more efficient intelligent traffic system in the future, and belongs to the field of hot research focused by all countries. Risk assessment is a key technology of intelligent driving, and the function of the risk assessment is to analyze driving risks based on environment perception information fed back by a sensor so as to provide a decision basis for subsequent driving decisions.

The risk of collision is an important component of the driving risk. In an actual traffic system, a large number of potential conflicts exist among vehicles, for example, in scenes such as intersections, remissions and the like, behaviors of conflicting vehicles are interactively influenced with each other, and the behavior intentions of the conflicting vehicles have the characteristics of time variation, uncertainty and difficulty in direct observation. The existing collision risk prediction method usually uses a single vehicle as an analysis object, or can not give interpretable analysis to the vehicle interaction process, so that the collision risk under the vehicle-vehicle collision scene is difficult to effectively predict.

Disclosure of Invention

The present application is directed to solving, at least to some extent, one of the technical problems in the related art.

Therefore, an object of the present application is to provide a collision risk prediction method coupled with a road structure based on vehicle behavior interaction, which can identify a vehicle potential collision relationship and predict a collision risk thereof by using a vehicle historical track and road structure information as inputs, and provide a basis for an intelligent vehicle to further make a behavior decision.

It is another object of the application to propose a collision risk prediction device coupled to a road structure based on vehicle behaviour interaction.

In order to achieve the above object, an embodiment of an aspect of the present application provides a collision risk prediction method coupled with a road structure based on vehicle behavior interaction, where the method includes the following steps:

identifying potential double-vehicle conflicts based on a road structure, and projecting the double-vehicle conflicts to a basic interactive conflict scene model;

establishing an intention identification model based on a dynamic Bayesian network according to the double-vehicle conflict so as to describe the conditional probability relation between the vehicle passing intention and the environmental situation and driving behavior;

respectively establishing probability graph models according to the environment situation and the semantic behavior so as to respectively utilize observable scene physical information and two-vehicle motion information to carry out probability inference on the environment situation and the semantic behavior;

performing parameter pre-calibration on the intention identification model parameters based on experience, and performing parameter learning based on an EM algorithm and natural driving data;

and outputting a two-vehicle passing intention identification result based on the intention identification model, predicting the motion tracks of the two vehicles by using a Gaussian process regression algorithm, and outputting a collision risk evaluation result based on the Gaussian distribution overlapping degree of the positions of the two vehicles at all times.

In order to achieve the above object, another embodiment of the present application provides a collision risk prediction apparatus coupled with a road structure based on vehicle behavior interaction, including:

the projection module is used for identifying potential double-vehicle conflicts based on a road structure and projecting the double-vehicle conflicts to a basic interactive conflict scene model;

the modeling module is used for establishing an intention identification model based on a dynamic Bayesian network according to the double-vehicle conflict so as to describe the conditional probability relation between the vehicle passing intention and the environmental situation and driving behavior;

the inference module is used for respectively establishing probability graph models according to the environment situation and the semantic behavior so as to respectively utilize observable scene physical information and two-vehicle motion information to carry out probability inference on the environment situation and the semantic behavior, taking the environment situation and the semantic behavior as observation input and carrying out probability inference on the passing intention of the conflicting vehicles based on a dynamic Bayesian network;

the training module is used for performing parameter pre-calibration on the intention identification model parameters based on experience, performing parameter learning based on an EM algorithm and natural driving data, classifying natural driving data sets based on indexes such as intention and the like, and respectively training corresponding Gaussian process regression models for subsequent trajectory prediction;

and the prediction module is used for outputting a two-vehicle passing intention recognition result based on the intention recognition model, predicting the motion tracks of the two vehicles by using a Gaussian process regression algorithm, and outputting a collision risk evaluation result based on the Gaussian distribution overlapping degree of each moment position.

The collision risk prediction method and device based on vehicle behavior interaction and road structure coupling provide a vehicle intention identification and track prediction framework considering vehicle behavior interaction and road structure coupling, and the frame is used for quantitatively predicting collision risks in a vehicle collision scene. The framework considers the influence of the environment situation and the vehicle behavior on the intention of a driver in a fusion manner, and then quantitatively evaluates the collision risk based on the space-time distribution and the contact ratio of the predicted two vehicle tracks, and provides a basis for the subsequent decision making process of the intelligent vehicle. Based on the simulation of human interaction process and the training of natural driving data, the vehicle motion prediction under complex conflict scenes can be realized, and the description capability of the model and the generalization capability of different scene migration applications are further enhanced through the coupling with the road structure.

Additional aspects and advantages of the present application will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the present application.

Drawings

The foregoing and/or additional aspects and advantages of the present application will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:

FIG. 1 is a flow diagram of a method for collision risk prediction coupled with a road structure based on vehicle behavior interaction according to one embodiment of the present application;

FIG. 2 is a logic block diagram of a collision risk prediction method coupled with a road structure based on vehicle behavior interaction according to one embodiment of the present application;

fig. 3 is a schematic diagram of a projection process and output information of a real conflict scene into two types of basic conflict scenes according to an embodiment of the present application;

FIG. 4 is a schematic diagram of a dynamic Bayesian network for vehicle intent inference according to one embodiment of the present application;

FIG. 5 is a probabilistic pictorial illustration for environmental situation inference and behavioral semantic inference according to an embodiment of the present application;

FIG. 6 is a diagram illustrating environmental aspects involved in two types of basic conflict scenarios, according to an embodiment of the present application;

FIG. 7 is a table relating to behavior semantics in two types of basic conflict scenarios, according to one embodiment of the present application;

fig. 8 is a schematic structural diagram of a collision risk prediction device coupled with a road structure based on vehicle behavior interaction according to an embodiment of the present application.

Detailed Description

Reference will now be made in detail to embodiments of the present application, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the drawings are exemplary and intended to be used for explaining the present application and should not be construed as limiting the present application.

In the field of intelligent driving, much research has been conducted on vehicle-to-vehicle collision risks. Modeling for vehicle motion mainly includes two categories based on a fixed motion model and based on a behavior intention. The fixed motion model assumes a motion model with a fixed mode such as a constant speed, a constant acceleration, a constant steering angle, a longitudinal acceleration and the like, and generates collision risk prediction indexes such as TTC (time To collision), THW (time headway) and the like based on the motion model. Compared with the prior art, the behavior intention-based method further combines the road structure in the traffic environment to judge the behavior intention of the vehicle, further predicts the future movement of the vehicle and analyzes the collision risk. The motion model prediction based on the behavior intention is higher in accuracy and more consistent with the cognition of people on collision risks. However, most of the current methods only consider the behavior of the individual vehicles, and there is no analysis of the interaction effect between the vehicles. From the cause analysis of traffic accidents, when there is a potential conflict between vehicles (e.g. ramp junction, intersection without signal light, etc.), a common cause of a collision accident is failure of the interaction process (e.g. no other party is observed, or an overly aggressive strategy is generated at the same time). Conversely, a cooperation is formed based on effective vehicle-to-vehicle interaction, and collision can be avoided by determining the sequence of passing through the collision zones. Therefore, the motion model based on behavioral intent should further consider vehicle-vehicle interaction processes to more effectively predict collision risk. In order to better define the traffic scene studied by the application, referring to related works at home and abroad, the application defines an interactive conflict scene as a traffic scene in which when two vehicles face a potential space-time driving conflict, the two vehicles need to generate a first-last sequence through conflict points (or regions) in a coordinated manner to avoid the occurrence of the collision. And the front vehicle and the rear vehicle in scenes such as rear-end collision have clear active-passive relation and do not form an interactive scene. The uncertainty of the vehicle behavior in a non-interactive scene is smaller, and the existing collision risk method can better realize the function. Therefore, the method and the device are mainly oriented to two interactive scenes, namely, the convergent conflict and the cross conflict.

The evaluation index of the collision risk includes safety indexes such as a safe distance, a potential energy field, and a collision probability in addition to the time indexes represented by TTC and THW. The collision probability index is more suitable for the quantitative description of the collision risk according to the uncertainty characteristics of the vehicle intention and the uncertainty of the future position prediction algorithm.

The following describes a collision risk prediction method and device based on vehicle behavior interaction and road structure coupling according to an embodiment of the present application with reference to the drawings.

A proposed collision risk prediction method coupled with a road structure based on vehicle behavior interaction according to an embodiment of the present application will first be described with reference to the accompanying drawings.

Fig. 1 is a flow chart of a collision risk prediction method coupled with a road structure based on vehicle behavior interaction according to one embodiment of the present application.

Fig. 2 is a logic block diagram of a collision risk prediction method coupled with a road structure based on vehicle behavior interaction according to an embodiment of the present application.

As shown in fig. 1 and 2, the collision risk prediction method coupled with a road structure based on vehicle behavior interaction includes the following steps:

and step S1, identifying potential double-vehicle conflicts based on the road structure, and projecting the double-vehicle conflicts to the basic interactive conflict scene model.

Optionally, in an embodiment of the present application, identifying a potential two-vehicle conflict based on a road structure, and projecting the two-vehicle conflict to a basic interactive conflict scene model includes: based on the road structure, establishing projection on a Frenet coordinate system by taking a lane central line as a reference line to form an interactive conflict scene model of one of an influx type conflict scene model and a cross type conflict scene model; identifying vehicle pairings that constitute an interactive conflict scenario; and initializing an intention identification model according to the confidence degree of the prior passing intention of the two conflicting vehicles.

Optionally, in an embodiment of the present application, identifying vehicle pairs that constitute an interactive conflict scenario includes: and constructing a probability map, taking the states and environmental information of the target vehicle and the surrounding vehicles as input, outputting the collision strength of the target vehicle and the surrounding vehicles, and forming a double-vehicle collision scene with the vehicle with the highest collision strength.

In an embodiment of the present application, potential two-vehicle conflicts are identified based on road structure and projected into the basic conflict models of two types, i.e., the merge conflict and the cross conflict, as shown in fig. 3.

As shown in fig. 3, the present application mainly relates to two basic scenarios, namely, an ingress conflict and a cross conflict. Firstly, a Frenet coordinate system is established based on a vehicle reference track defined by a road structure, and an actual conflict scene is projected to two types of basic conflict models. Geometric information of the real road can be lost in the Frenet coordinate transformation process, and a projection function of the reference track and the curvature of the road on the reference track is established for storing the curvature characteristics of the road. Then, pairwise pairing is carried out on the vehicle conflict relations existing in the scene, and the conflict intensity is used as a criterion in the pairing process.

Further, in step S1, the two-vehicle conflict relationship is modeled in the following manner:

step S11, based on the road structure, a projection on the Frenet coordinate system is established with the lane center line as a reference line, and a basic collision model of either convergence or intersection is constructed. And storing the curvature of the real road along the reference track for probability inference of subsequent environment situation and behavior semantics.

In step S12, vehicle pairings that constitute an interactive conflict scenario are identified. And constructing a probability map, taking the states and environmental information of the target vehicle and the surrounding vehicles as input, outputting the collision strength of the target vehicle and the surrounding vehicles, and forming a double-vehicle collision scene with the vehicle with the highest collision strength. Wherein the probability map is based on the speed v of the target vehicle and the surrounding vehicle₁,v₂Distance l from the reference bump point₁,l₂For input and output, the two vehicles have a priority passing intention Pr₀S confidence of₁,s₂And s is_m＝P(Pr₀|v₁,v₂,l₁,l₂) And m is 1 and 2. The conflict strength can be expressed as the product C-s of the intentions of the two vehicles passing through preferentially₁×s₂。

Step S13, based on the identified main conflict object of the target car, using the probability chart to output the front-back passing intention S of the two cars₁,(1-s₁),s₂,(1-s₂) As an initial value for the subsequent two-car intent inference.

Step S13, with the confidence S of the prior passing intention of two conflicting vehicles₁,s₂And initializing a subsequent intent inference network oriented to the interaction process.

And step S2, establishing an intention identification model based on the dynamic Bayesian network according to the double-vehicle conflict so as to describe the conditional probability relation between the vehicle passing intention and the environmental situation and driving behavior.

Optionally, in an embodiment of the present application, building a dynamic bayesian network-based intention recognition model according to a two-vehicle conflict to describe a conditional probability relationship between a vehicle passing intention and an environmental situation and a driving behavior includes: simulating a human behavior interaction process, and constructing a dynamic Bayesian network, wherein the dynamic Bayesian network comprises an environment situation, a driver intention and a vehicle behavior, and an inference target of the dynamic Bayesian network is the driver intention; establishing a directed connection relation of each variable in the dynamic Bayesian network to form a directed acyclic graph; pre-calibrating the dynamic Bayesian network parameters based on empirical general knowledge; and (4) based on all measurable observation information on the time sequence, carrying out hidden variable probability inference in the dynamic Bayesian network, and outputting the confidence coefficient of the prior passing intention of the two conflicting vehicles.

Specifically, an intention recognition model based on a dynamic Bayesian network is established for researched vehicle-vehicle conflicts. The model infers the confidence that two vehicles pass through the conflict region preferentially (or lags) to themselves (i.e., the two-vehicle traffic intention) based on the environmental situation and behavior semantics during vehicle driving. The dynamic Bayesian network mainly comprises three elements: environmental situation, driver intent and vehicle behavior, where driver intent is an inferred goal of the network.

The influence factors of the passing intention include two types: environmental situation and historical behavior of both parties. The environment situation is coupled with the road structure and the motion state of the two vehicles and is used for describing whether a condition that the vehicle executes a certain behavior is met in a scene, for example, whether the time distance between the vehicle entering the vehicle and the vehicle behind meets the entering requirement or not; the behavior semantics of the vehicle correspond to a semantized behavior having a specific meaning, such as "give way to slow down" or the like.

Further, in step S2, the confidence of the two-vehicle front-rear passing intention during the interaction process is estimated in the following manner.

Step S21, simulating human behavior interaction process, and establishing a dynamic bayesian network as shown in fig. 4. Implicit variables in a dynamic bayesian network can be divided into three layers: the first layer is an environmental situation and is used as an environmental basis for behavior decision of two vehicles; the second layer is a first-second passing intention confidence coefficient and describes the strength of a first-second passing intention of the vehicle in the scene; the third layer is the interactive behavior taken by the vehicle, which includes two types of behavior (MA) and Request (MR), and can be further divided into two types of horizontal behavior and vertical behavior. The explicit variables are observable scene physical information, such as road structure, relative motion state of vehicles, and the like. The explicit variables are used for deducing two types of information of environment situation and behavior semantics and are respectively provided with corresponding deduction models.

Step S22, establishing a Directed connection relationship between the variables in the dynamic bayesian network, and forming a Directed Acyclic Graph (DAG). When variable a points to variable B, the representative node a is the node B parent and has a conditional probability parameter P (B | a).

As shown in FIG. 4, the directed acyclic graph DAG created in the present application includes observation information O^tEnvironment situation P^t _l(1, 2,3, … represent environment situations) and behavior semantics a^t _l(1, 2,3, … represent semantic behaviors), and the intention of two vehicles to pass through. In this application, the logical relationship between hidden variables can be expressed as: the behavior at this moment is determined by the environmental situation, the traffic intention and the behavior at the previous moment, and the traffic intention at this moment is determined by the environmental situation at this moment, the behaviors of both parties at the previous moment and the intention at the previous moment. The environmental situation at this moment is determined by the environmental situation at the previous moment and the behaviors of both parties.

And step S23, performing initial calibration on the network parameters based on the common experience. The calibration logic is based on causal relationship, and the initialization is based on the state { f of a father node_nThe conditional probability P (C ═ C) of inferring the state of the child nodes_m|{f_n}). Taking the cut-in scenario as an example, when the distance between two vehicles is large, the cut-in vehicle has a high possibility of performing cut-in (for example, the parameter P (behavior ═ cut | { f) may be pre-calibrated_n})＝0.7)。

And step S24, deducing environment situation and behavior semantics based on the input information of 30 moments, and further accurately deducing the intention of the two vehicles. The intent inference process may be expressed as:

s_m ^k＝P(Pr₀|E^t-29～E^t,A^t-29～A^t),m＝1,2；k＝(t-29)～t.

the inference process adopts a Forward-Backward Algorithm (Forward-Backward Algorithm) which can fuse Forward and Backward probabilistic inference of reaching a target time and give an accurate inference result of the target time.

And step S3, respectively establishing probability map models according to the environmental situation and the semantic behavior so as to respectively utilize observable scene physical information and two-vehicle motion information to carry out probability inference on the environmental situation and the semantic behavior.

Optionally, in an embodiment of the present application, respectively establishing a probability map model according to the environmental situation and the semantic behavior to perform probability inference on the environmental situation and the semantic behavior by using observable scene physical information and two-vehicle motion information, respectively, includes: establishing a first probability map model for identifying the environment situation, wherein the first probability map model takes road structure information, positions and speeds of two vehicles as input and takes whether the vehicle has the environment situation meeting the behavior condition as output; and establishing a second probability map model for identifying the behavior semantics of the conflict vehicles, wherein the second probability map model takes the motion states of the two vehicles, the road curvatures corresponding to the positions of the two vehicles, the environmental situation result at the moment as input and the behavior semantics identification result as output.

Optionally, in an embodiment of the present application, the probabilistic graphical model employs a junction tree algorithm for probabilistic inference.

Specifically, an environmental situation and behavior semantic recognition model is constructed. The identification model is constructed based on a probability map model, and probability inference is carried out on two kinds of information, namely environment situation and semantic behavior, by respectively utilizing observable scene information and two-vehicle motion information. Because the two types of information have corresponding observable information, the parameter calibration of the probability map model can be performed based on the statistical analysis of the observable information.

Further, in step S3, the environmental situation and semantic behavior are recognized in the following manner:

step S31, establishing a probabilistic graphical model as shown in fig. 5 (a) for identifying the environmental situation P_n. The model takes road structure information, two vehicle positions and speeds as input respectively, and whether the vehicle has the function of the vehicle or notAnd outputting the environment situation meeting the behavior condition. The entry conflict and the cross conflict respectively have different environment situations, and the definitions of the environment situations comprise safe vehicle distance, dynamic gap and the like.

Step S32, establishing a probability map model as shown in (b) of FIG. 5 for identifying the semantic behavior A of the vehicle_n. The model respectively takes the motion states of two vehicles, the curvature of a road corresponding to the position of the two vehicles, the environmental situation result at the moment as input and the behavior semantic identification result as output.

During the interaction process, the vehicle behavior semantics can be classified based on the following dimensions: vehicle dimensions for executing actions (e.g., cut-in conflict, cut-in behavior-back behavior); from the direction dimension of the action, it can be divided into horizontal behavior-vertical behavior (especially for incoming conflicts, two vehicles do not have fixed conflict points, so that both horizontal and vertical behaviors may have specific semantics); the follower dimension includes two categories, behavior achievement (MA) and behavior Request (MR). Wherein the behavior MA means that the vehicle will take some driving behavior (e.g. perform cut-in) without the other party changing the pass intention; the request MR means that the vehicle takes some action to request the other side to change the passing intention (for example, the cut-in vehicle requests the vehicle to pass through the card position and then gives the vehicle to pass).

In steps S31 and S32, a Junction Tree Algorithm (Junction Tree Algorithm) is adopted for accurate inference. The environmental situation and behavior semantic inference process may be expressed as:

P(E^t _l＝a)＝P(E^t _l＝a|O^t)

P(A^t _l＝b)＝P(A^t _l＝b|O^t,E^t _l)

the result output by the inference is the confidence of various environment situations and behavior semantics. And taking the confidence coefficient as input information of the dynamic Bayesian network for a subsequent intention inference process.

And step S4, performing parameter pre-calibration on the intention identification model parameters based on experience, and performing parameter learning based on an EM algorithm and natural driving data.

For the dynamic Bayesian network for intention inference, the process comprises two steps, firstly, parameter pre-calibration is carried out based on experience, and parameter learning is carried out based on EM algorithm and natural driving data.

In particular, the dynamic bayesian network parameters are trained using the EM algorithm. The EM algorithm has the advantages that under the condition of data loss (partial hidden variables lack truth value labels), the optimal probability network parameters aiming at the data set are obtained based on continuous iteration of variable inference and parameter optimization, the effective identification of the two-vehicle intention based on semantic observation information is realized through optimization,

and step S5, outputting a two-vehicle passing intention identification result based on the intention identification model, predicting the moving tracks of the two vehicles by using a Gaussian process regression algorithm, and outputting a collision risk evaluation result based on the Gaussian distribution overlapping degree of the positions at all times.

Optionally, in an embodiment of the present application, outputting a two-vehicle passing intention recognition result based on an intention recognition model, performing two-vehicle motion trajectory prediction by using a gaussian process regression algorithm, and outputting a collision risk assessment result based on a gaussian distribution overlapping degree of positions at each time, where the method includes: classifying vehicle tracks in the natural driving data based on the intention recognition result to construct a training set; training a Gaussian process regression algorithm model based on training set classification; predicting the future track of the vehicle according to the intention recognition result and a Gaussian process regression algorithm model; and according to the predicted future track of the vehicle, identifying the future collision risk of the two vehicles through the overlapping degree of Gaussian distribution of the vehicle positions.

Specifically, based on the identified behavior intention, a Gaussian Regression (GPR) algorithm is used for predicting the vehicle running track and the space-time distribution thereof, and finally the collision risk of the two vehicles is output.

Specifically, the two-vehicle trajectory and the collision risk are predicted in the following manner.

In step S51, to realize the trajectory prediction based on the intention, the training set needs to be divided first. The training set division basis comprises: final passage order, and interaction strength classification based on two-car passage order (weak-medium-strong). The passing order can be directly marked based on the final passing order of the two vehicles in the track data, the interaction strength is classified based on the situation index and the intention index, and the optimal classification is realized by adopting a K-Means algorithm. In the classified data, the weaker the interaction strength is, the higher the safety situation and the smaller the change of the intentions of the two vehicles are in the interaction process; on the contrary, the interaction intensity is higher, the safety situation is lower in the interaction process, and more intention changes may exist in the two vehicles.

And step S52, training the GPR model based on the training set classification. The GPR model can be expressed as:

f～GP(u,K)

u＝{m(t_i),t_i＝1:T}

K＝{k(t_i,t_j),i,j＝1:N}

where u is a mean vector (representing the predicted position at each time instant) and K is a covariance matrix describing the distribution of the predicted positions. In this application, u and K use a polynomial model and a square exponential kernel model, respectively. Based on the training set data of each category, a Maximum Likelihood estimation algorithm (MLE) is used to train the corresponding GPR model.

In step S53, a future trajectory in the prediction time domain is input based on the historical trajectory. Firstly, based on input historical data, the probability graph model is applied to identify the behavior intentions of two vehicles, and interaction strength and weakness attributes are judged based on historical tracks. Based on the discrimination result, a corresponding GPR model is selected to predict the trajectory.

Considering that the initial stage of prediction is short in interaction time between two vehicles, less in input information and low in prediction accuracy based on GPR, a method of fusing a GPR prediction result and a Constant Acceleration model (CA) is adopted in the first 1s of prediction, namely, the average value of the two models is taken as the prediction position.

And step S54, predicting collision risks based on the contact ratio of the future tracks of the two vehicles.

First, the potential collision locations of two vehicles are identified. Potential collision points include: the total of 8 points are right front, right back, right left, right, left front, left back, right front and right back. Based on the predicted positions of two vehicles, toTaking the closest point as a potential collision point, calculating the Gaussian distribution Overlap ratio (OLR) of the position of the potential collision point, and recording the Overlap ratio at the predicted time t as OLR_t。

Thereafter, the collision risk prediction result is represented by the prediction time t (expressed as the predicted number of steps) and the corresponding OLR_tThe coupling is given. The smaller the predicted time t is, the lower the position overlap ratio OLR_tThe higher the corresponding collision risk. Thereby, the collision risk R_colliExpressed as:

where T is the total predicted step size, c_tAn attenuation coefficient of less than or equal to 1. c. C_tThe smaller R_colliThe more concern is near-term collision risk and vice versa the increased concern is about long-term risk.

By the method, the collision risk prediction of multi-vehicle behavior interactive coupling can be considered under the complex traffic scenes of ramp entry, intersection passage and the like, and the method can be widely applied to various collision scenes and is beneficial to improving the driving safety of intelligent vehicles.

In order to further analyze the probability of collision between two vehicles in a potential collision and output a collision risk, a collision risk prediction method based on vehicle behavior interaction and road structure coupling according to the present application is described below with reference to a specific embodiment.

1): potential two-vehicle conflicts are identified based on the road structure and projected into two basic conflict models, namely an incoming conflict model and a cross conflict model, as shown in fig. 3.

Specifically, 1), the two-vehicle conflict relationship is modeled in the following manner:

1-1), based on the road structure, using the lane central line as a reference line to establish the projection of the traffic scene in the Frenet coordinate system, and forming one of the two types of basic conflicts shown in FIG. 3. The projection of the coordinates of the points under the Cartesian coordinate system to the Frenet coordinate system comprises two steps: (1) calculating the shortest distance projection point of the point to be solved on the reference line; (2) on a reference lineStarting from the origin, the length of the reference line (longitudinal distance D) from the origin to the projection point is calculated₁) And the distance from the projected point to the desired point (lateral distance D)₂). In this case, the coordinate of the determined point in the Frenet coordinate system is (D)₁，D₂)。

After the projection is finished, the curvature information of the original road needs to be stored and recorded as a function C (D) taking the longitudinal distance of the Frenet coordinate system as an independent variable₁)。

1-2), identifying vehicle pairings that constitute an interactive conflict scenario. In the pairing process, a research object can be selected first, and then the vehicle with the highest collision strength C with the object can be searched as a collision object.

In the present application, the collision strength C is defined as the product of the intentions of the two vehicles to pass through the collision area preferentially (i.e. when both the vehicles tend to consider that the vehicle will pass through the collision area preferentially by the comparing party, the collision strength is greater). Building speed v of target vehicle-environment vehicle₁,v₂Distance l from the reference bump point₁,l₂Is an input probability map and respectively outputs two vehicles with prior passing intention Pr₀S confidence of₁,s₂(i.e. s)_m＝P(Pr₀|v₁,v₂,l₁,l₂) And m is 1, 2). At this time, the collision strength is expressed as C ═ s₁×s₂。

For calibrating the probability map parameters, a supervised learning method can be adopted, namely, the vehicles which finally actually form the conflict are used as true values, and the maximum likelihood estimation algorithm (MLE) is used for parameter calibration.

1-3) with confidence of prior intention of two conflicting vehicles₁,s₂And initializing a subsequent intent inference network oriented to the interaction process.

2): aiming at the researched double-vehicle conflict, a dynamic Bayesian network is established for identifying the confidence degrees of the drivers of the two vehicles for the drivers to pass through the conflict region preferentially. The dynamic Bayesian network mainly comprises three elements: environmental situation, driver intent and vehicle behavior, where driver intent is an inferred goal of the network.

In 2), the confidence of the front-back pass intention of the two vehicles in the interaction process is estimated in the following way:

2-1), simulating human behavior interaction process, and constructing a dynamic Bayesian network shown in FIG. 4. In terms of node information definition, the network mainly comprises three layers of nodes: (1) and the environment situation node represents whether the conflict vehicle has the condition for executing the priority traffic. For example, in a cut-in scene, a cut-in vehicle can merge lanes only under conditions such as a safe inter-vehicle distance. As shown in fig. 6, in the influx conflict, the situation information includes four types: p1, cutting into the longitudinal running space of the vehicle; p2, cutting into the longitudinal running space between the vehicle and the front environment vehicle; p3, longitudinal running space between colliding cars; p4, direct collision vehicle and front environmental vehicle dynamic clearance. In cross-type conflicts, the situational information contains three categories: p1, time difference between two conflicting vehicles to conflict point; p2, longitudinal running space of left collision vehicle; p3: the lower part conflicts the longitudinal running space of the vehicle. The situation information is discrete hidden variables and comprises two values of 'conditional' and 'unconditional'. (2) The intention nodes represent the conflict vehicles respectively have confidence levels of passing through the conflict areas preferentially. The intention information is a discrete hidden variable and comprises two values of 'priority pass' and 'give way'. (3) And the behavior node describes the meaning of the running track of the conflict vehicle. Based on classification dimensions of conflict vehicles, horizontal-vertical, request-execution, line preemption-line giving and the like, behavior semantics are shown in a table. Each behavior semantic forms a discrete hidden variable and comprises two values of 'having the behavior' and 'not having the behavior'.

In the three-layer node, the environmental situation and the behavior semantics can be inferred based on observable physical information such as the vehicle position and the vehicle speed, and the observation and inference process is described in step S3. The intention does not directly correspond to observable information, and the intention is inferred based on two types of hidden variables of environment and behavior, and is an inferred result finally output by the dynamic bayesian network in the step S2.

2-2), forming a directed acyclic graph based on the three layers of nodes. As shown in FIG. 4, the environmental situation at time t is denoted as E^t _l(1, 2,3, … represent each environment situation), the behavior semantics are represented as a^t _l(1, 2,3, … stand for semantic behaviors) and all measurable observed information is represented as O^t. When variable a points to variable B, the representative node a is the node B parent and has a conditional probability parameter P (B | a).

2-3) pre-calibrating the dynamic Bayesian network parameters. The calibration process is carried out based on experience, and the meaning of the calibration parameter is based on the state { f of the father node_nDeducing child node states C_mConditional probability of P (C ═ C)_m|{f_n}). Taking the cut-in scenario as an example, when the distance between two vehicles is large, it can be determined based on experience that the situation security of the cut-in vehicle for performing cut-in is high (P (situation | { f |) (safety |)_n0.8) }); when the confidence of the cut-in vehicle passing first is high, the confidence of the cut-in vehicle passing later is high, and the situation is highly safe, the probability that the host vehicle takes the cut-in behavior is high (P (behavior is cut | { f)_n})＝0.7))。

2-4), based on all measurable observation information in the time sequence, realizing hidden variable probability inference in the dynamic Bayesian network, and finally outputting the confidence degree of the prior passing intention of the two conflicting vehicles.

Based on the dynamic Bayesian network, two types of hidden variables of environment situation and behavior semantics can be expressed as follows:

P(E^t _l＝a)＝P(E^t _l＝a|O^t)

P(A^t _l＝b)＝P(A^t _l＝b|O^t,s₁ ^t-1,s₂ ^t-1)

the application assumes that 30 periods of observation information is input, and finally, the confidence degrees of the prior passing intentions of two vehicles in 30 periods are output:

s_m ^k＝P(Pr₀|E^t-29～E^t,A^t-29～A^t),m＝1,2；k＝(t-29)～t.

in the implementation and test process of the application, an inference algorithm based on a forward-backward algorithm is adopted. Other common dynamic bayesian network inference algorithms can also be used to solve the inference problem of the present application.

3): aiming at the environmental situation and the vehicle behavior in the step 2), respectively establishing an observer based on a probability map model. The observer outputs discrete situation evaluation and behavior semantics based on measurable scene physical information (such as relative positions of two vehicles, vehicle speed and the like).

In 3), the semantics of the environment situation and the behavior are recognized in the following way:

3-1), establishing a probability map model as shown in (a) of fig. 5 for identifying various environment situations. The input information includes static physical parameters (such as the end position of a ramp merging area, the curvature of a road and the like) and dynamic physical information (including the positions, speeds, accelerations and the like of all collision vehicles and environmental vehicles at all times) of a scene projected to a Frenet coordinate system, and the output estimation result is whether the researched environmental situation has conditions for the vehicles to execute preferential traffic (such as whether a large enough gap exists between a straight collision vehicle and a front environmental vehicle in a cut scene).

In order to calibrate the model parameters of the probability map, the model parameters of the probability map can be optimally designed based on the research on the vehicle-vehicle collision process in the real natural driving data, such as the statistics of the distribution of information such as relative distance, relative vehicle speed and the like in the gathering process.

3-2), establishing a probability graph model shown in (b) of FIG. 5, and identifying the conflict vehicle behavior semantic A_n. The model respectively takes the relative motion information of the two vehicles and the inferred environment situation as input, and identifies corresponding behavior semantics. The behavior semantic classification is shown in fig. 7.

The classification mode of the behavior semantics comprises conflict vehicle, horizontal-vertical, line preemption-line giving-line requesting, and behavior request-behavior realization. The difference between the action Request (MR) and the action Achievement (MA) of the present application is shown in: in the MR behavior, the vehicle does not have a condition for executing a pass intention, and the behavior is intended to request the other party to change the intention, so as to create a condition for the own-party behavior. In the MA behavior, the condition is already satisfied, and the vehicle execution behavior does not involve a change in the intention of the counterpart. For example, in the process of importing, assuming that an importing vehicle transversely cuts in without safe importing, the behavior semantics should be interpreted as a heuristic requesting a subsequent direct vehicle to yield. Conversely, a lateral cut-in behavior under a safe merge condition should be interpreted as it executing its "priority traffic" intent.

Based on the behavior semantic classification method, a probability graph model is established and model parameters are predefined. Further, model parameters may be optimized based on research and statistics of vehicle-to-vehicle collision processes in real nature driving data.

4): the dynamic Bayesian network parameter calibration method comprises the following two steps of firstly performing parameter pre-calibration based on experience and performing parameter learning based on an EM algorithm and natural driving data.

Model parameters in a dynamic Bayesian network are trained based on an EM algorithm, different from the two types of probability map observation models in 3), and the dynamic Bayesian model for intention inference is difficult to obtain direct observable data statistics, so that direct calibration is difficult. The advantage of the EM algorithm is that in the case of data loss (hidden variables have no truth labels), the optimal probabilistic network parameters for the data set are obtained based on the continuous iteration of the variable inference and parameter optimization processes.

5): based on the intention identification result, a Gaussian Process Regression (GPR) is used for predicting the future tracks of the two vehicles and evaluating the collision risks of the two vehicles.

5-1), in order to train the required trajectory prediction model, the vehicle trajectory in the natural driving data needs to be classified first. Based on the output of the intent recognition model, the intent combinations of two conflicting vehicles include four categories, look-ahead, look-yielded, yield-look-ahead, yield-yielded. Wherein, the probability correspondence of each class is the product of confidence degrees of corresponding intentions of two vehicles, namely P₁＝(s₁×s₂)，P₂＝((1-s₁)×s₂)，P₃＝(s₁×(1-s₂))，P₄＝((1-s₁)×(1-s₂) Take the combination with the highest probability as the classification result at that time, i.e., P_m＝max(P₁,P₂,P₃,P₄)。

5-2), training the GPR model based on the training set classification. The GPR model can be expressed as:

f～GP(u,K)

u＝{m(t_i),t_i＝1:T}

K＝{κ(t_i,t_j),i,j＝1:N}

where u is a mean vector (representing the predicted position at each time instant) and K is a covariance matrix describing the distribution of the predicted positions. In the test and validation of the present application, u and K use a quintic polynomial model and a square exponential kernel model, respectively.

In the training process, the input track is based on P_mSegmentation is performed and used to train the GPR model corresponding to the combination of intents, respectively.

5-3), predicting the future track of the vehicle based on the intention recognition result and the GPR model. In the prediction process, the historical track of the vehicle and the intention identification result are used as input, and track prediction is carried out on the basis of each intention combination and the corresponding GPR model. The output of the prediction model is the position distribution of the two vehicles at each time in the future under each intention combination and the probability of each intention combination.

In addition, considering that the initial stage of prediction is short in interaction time between two vehicles, less in input information and low in prediction accuracy based on GPR, a method of fusing a GPR prediction result and a Constant Acceleration model (CA) is adopted in the first 1s of prediction, namely, the average value of the two models is taken as the prediction position.

5-4), and identifying the future collision risk of the two vehicles based on the track prediction. Potential collision points include: the total of 8 points are right front, right back, right left, right, left front, left back, right front and right back. Based on the predicted positions of the two vehicles, the closest point of the two vehicles is taken as a potential collision point, the Gaussian distribution Overlap ratio (OLR) of the potential collision point is calculated, and the Overlap ratio at the predicted time t is recorded as OLR_t。

Since the predicted output contains a trajectory prediction based on a combination of four intents, the overlap ratio calculation of the final output is expressed as:

According to the collision risk prediction method based on vehicle behavior interaction and road structure coupling, a vehicle intention identification and track prediction framework considering the vehicle behavior interaction and the road structure in a coupling mode is provided and used for quantitatively predicting collision risks in a vehicle collision scene. The framework considers the influence of the environment situation and the vehicle behavior on the intention of a driver in a fusion manner, and then the collision risk is quantitatively evaluated based on the space-time distribution and the contact ratio of the tracks of the two vehicles, so that a basis is provided for the subsequent decision making process of the intelligent vehicle. The method is based on simulation of human interaction process and training of natural driving data, vehicle motion prediction under complex conflict scenes can be achieved, and description capacity of a model and generalization capacity of migration application of different scenes are further enhanced through coupling with a road structure.

Next, a collision risk prediction apparatus coupled with a road structure based on vehicle behavior interaction according to an embodiment of the present application will be described with reference to the drawings.

As shown in fig. 8, the collision risk prediction apparatus coupled with a road structure based on vehicle behavior interaction includes: projection module 100, modeling module 200, inference module 300, training module 400, and prediction module 500.

And the projection module 100 is used for identifying potential double-vehicle conflicts based on the road structure and projecting the double-vehicle conflicts to the basic interactive conflict scene model.

And the modeling module 200 is used for establishing an intention identification model based on the dynamic Bayesian network according to the double-vehicle conflict so as to describe the conditional probability relation between the vehicle passing intention and the environmental situation and the driving behavior.

The inference module 300 is configured to respectively establish a probabilistic graph model according to the environmental situation and the semantic behavior, so as to respectively utilize observable scene physical information and two-vehicle motion information to perform probabilistic inference on the environmental situation and the semantic behavior, take the environmental situation and the semantic behavior as observation inputs, and perform probabilistic inference on the traffic intention of a conflicting vehicle based on a dynamic bayesian network.

The training module 400 is configured to perform parameter pre-calibration on the intention identification model parameters based on experience, perform parameter learning based on the EM algorithm and the natural driving data, classify the natural driving data set based on the intention and other indexes, and respectively train corresponding gaussian process regression models for subsequent trajectory prediction.

The prediction module 500 is configured to output a two-vehicle passing intention recognition result based on the intention recognition model, perform two-vehicle motion trajectory prediction by using a gaussian process regression algorithm, and output a collision risk evaluation result based on the gaussian distribution overlapping degree of each time position.

Optionally, in an embodiment of the present application, the projection module is further configured to, based on a road structure, establish a projection on a Frenet coordinate system with a lane center line as a reference line, and form an interactive collision scene model of one of the import collision scene model and the cross collision scene model; identifying vehicle pairings that constitute an interactive conflict scenario; and initializing an intention identification model according to the confidence degree of the prior passing intention of the two conflicting vehicles.

It should be noted that the foregoing explanation of the method embodiment is also applicable to the apparatus of this embodiment, and is not repeated herein.

According to the collision risk prediction device based on vehicle behavior interaction and road structure coupling, a vehicle intention identification and track prediction framework considering the vehicle behavior interaction and the road structure in a coupling mode is provided and used for quantitatively predicting collision risks in a vehicle collision scene. The framework considers the influence of the environment situation and the vehicle behavior on the intention of a driver in a fusion manner, and then the collision risk is quantitatively evaluated based on the space-time distribution and the contact ratio of the tracks of the two vehicles, so that a basis is provided for the subsequent decision making process of the intelligent vehicle. Based on the simulation of human interaction process and the training of natural driving data, the vehicle motion prediction under complex conflict scenes can be realized, and the description capability of the model and the generalization capability of different scene migration applications are further enhanced through the coupling with the road structure.

Furthermore, the terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present application, "plurality" means at least two, e.g., two, three, etc., unless specifically limited otherwise.

In the description herein, reference to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the application. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.

Although embodiments of the present application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present application, and that variations, modifications, substitutions and alterations may be made to the above embodiments by those of ordinary skill in the art within the scope of the present application.

Claims

1. A method for collision risk prediction based on vehicle behavior interaction coupled with a road structure, comprising the steps of:

2. The method of claim 1, wherein identifying potential two-vehicle conflicts based on road structure, projecting the two-vehicle conflicts to a basic interactive conflict scenario model, comprises:

based on the road structure, establishing projection on a Frenet coordinate system by taking a lane central line as a reference line to form an interactive conflict scene model of one of an influx type conflict scene model and a cross type conflict scene model;

identifying vehicle pairings that constitute an interactive conflict scenario;

and initializing the intention identification model with the confidence of the prior passing intention of the two conflicting vehicles.

3. The method of claim 2, wherein the identifying vehicle pairs that constitute an interactive conflict scenario comprises:

and constructing a probability map, taking the states and environmental information of the target vehicle and the surrounding vehicles as input, outputting the collision strength of the target vehicle and the surrounding vehicles, and forming a double-vehicle collision scene with the vehicle with the highest collision strength.

4. The method of claim 1, wherein building a dynamic Bayesian network-based intention recognition model according to the two-vehicle conflict to describe a conditional probability relationship between vehicle traffic intention and environmental situation and driving behavior comprises:

simulating a human behavior interaction process, and constructing a dynamic Bayesian network, wherein the dynamic Bayesian network comprises an environment situation, a driver intention and a vehicle behavior, and an inference target of the dynamic Bayesian network is the driver intention;

establishing a directed connection relation of each variable in the dynamic Bayesian network to form a directed acyclic graph;

pre-calibrating the dynamic Bayesian network parameters based on experience common knowledge, and training and optimizing the network parameters based on real driving data;

and based on all measurable observation information on the time sequence, carrying out hidden variable probability inference in the dynamic Bayesian network, and outputting the confidence coefficient of the prior passing intention of the two conflicting vehicles.

5. The method of claim 1, wherein the establishing probability map models according to the environmental situation and the semantic behavior respectively to perform probability inference on the environmental situation and the semantic behavior by using observable scene physical information and two-vehicle motion information respectively comprises:

establishing a first probability map model for identifying the environment situation, wherein the first probability map model takes road structure information, positions and speeds of two vehicles as input and takes whether the vehicle has the environment situation meeting the behavior condition as output;

and establishing a second probability map model for identifying the behavior semantics of the conflict vehicles, wherein the second probability map model takes the motion states of the two vehicles, the road curvatures corresponding to the positions of the two vehicles, the environmental situation result at the moment as input and the behavior semantics identification result as output.

6. The method of claim 1, wherein the outputting of the two-vehicle passing intention recognition result based on the intention recognition model, the performing of the two-vehicle motion trajectory prediction by using a gaussian process regression algorithm, and the outputting of the collision risk assessment result based on the gaussian distribution overlapping degree of each time position comprises:

classifying vehicle tracks in the natural driving data based on the intention recognition result to construct a training set;

training a Gaussian process regression algorithm model based on the training set classification;

predicting the future track of the vehicle according to the intention recognition result and the Gaussian process regression algorithm model;

and according to the predicted future track of the vehicle, identifying the future collision risk of the two vehicles through the overlapping degree of Gaussian distribution of the vehicles.

7. The method of claim 5, wherein the probabilistic graphical model employs a junction tree algorithm for probabilistic inference; the dynamic bayesian network uses a forward-backward algorithm for inference.

8. A collision risk prediction apparatus coupled to a road structure based on vehicle behavior interaction, comprising:

the training module is used for performing parameter pre-calibration on the intention identification model parameters based on experience, performing parameter learning based on an EM (effective electromagnetic) algorithm and natural driving data, classifying the natural driving data set based on intention indexes, and respectively training corresponding Gaussian process regression models for subsequent trajectory prediction;

9. The device of claim 8, wherein the projection module is further configured to, based on the road structure, establish a projection on a Frenet coordinate system with a lane center line as a reference line to form an interactive collision scenario model of one of the imported collision scenario model and the cross collision scenario model; identifying vehicle pairings that constitute an interactive conflict scenario; and initializing the intention identification model with the confidence of the prior passing intention of the two conflicting vehicles.

10. The apparatus of claim 8, wherein the identifying vehicle pairings that comprise an interactive conflict scenario comprises: