CN117585017A

CN117585017A - Automatic driving vehicle lane change decision method, device, equipment and storage medium

Info

Publication number: CN117585017A
Application number: CN202311694726.5A
Authority: CN
Inventors: 乔旭强; 褚文博; 周明珂; 沈斌; 王星
Original assignee: Western Science City Intelligent Connected Vehicle Innovation Center Chongqing Co ltd
Current assignee: Western Science City Intelligent Connected Vehicle Innovation Center Chongqing Co ltd
Priority date: 2023-12-11
Filing date: 2023-12-11
Publication date: 2024-02-23

Abstract

The embodiment of the application discloses a lane change decision method, device and equipment for an automatic driving vehicle and a storage medium. The method comprises the following steps: acquiring road information, vehicle running state information and running state information of each obstacle; inputting road information, driving state information of the own vehicle and driving state information of each obstacle into a pre-trained lane change decision model, and correspondingly obtaining predicted track information of the own vehicle and predicted track information of each obstacle by a track predictor of the lane change decision model; the driving style probability sub-model obtains driving style probability corresponding to the own vehicle; the lane change trend probability submodel obtains the lane change trend probability of the own vehicle; the collision risk probability sub-model obtains the collision risk probability of the own vehicle; the dynamic Bayesian decision network sub-model obtains a lane change decision result corresponding to the current moment of the own vehicle based on the driving style probability, the lane change trend probability and the collision risk probability. By applying the scheme provided by the embodiment of the application, the accuracy of channel switching decision can be improved.

Description

Automatic driving vehicle lane change decision method, device, equipment and storage medium

Technical Field

The application relates to the technical field of automatic driving, in particular to a lane change decision method, a lane change decision device, lane change decision equipment and a lane change decision storage medium for an automatic driving vehicle.

Background

Currently, most of the mass-produced L2-L3-level automatic driving vehicles on the market adopt a layered system architecture, namely perception-decision planning-control, and key technologies include: environmental awareness technology, obstacle movement prediction technology, driving task decision and planning, and driving track optimization and control technology. The architecture takes reference to the thinking mode of human beings, and is just like the human beings need to sense the surrounding environment by means of the senses such as eyes, ears, nose and the like, and then the brain processes the sensing information to form understanding and judgment on the surrounding environment, so that reasonable decision and planning are made.

It is clear that the decision making system of an autonomous vehicle plays a crucial role in the overall system. Most of the known decision methods adopt finite state machines based on rules, and the methods only consider the environmental information at the current moment, the relative motion information of the host vehicle and other vehicles and the like when deciding. However, during the actual running of the autopilot, for example, whether a lane change is needed, other factors are also involved, such as the risk of future collision between the host vehicle and the other vehicle, and the style factor of the driver. Therefore, the known lane change decision method has fewer consideration factors, poor output decision safety and comfort, lower accuracy and how to improve the lane change decision accuracy, and becomes a technical problem to be solved urgently.

Disclosure of Invention

The application provides a lane change decision method, device and equipment for an automatic driving vehicle and a storage medium, so as to improve the accuracy of lane change decision. The specific technical scheme is as follows.

In a first aspect, an embodiment of the present application provides a lane change decision method for an automatic driving vehicle, where the method includes:

acquiring road information, vehicle running state information and running state information of each obstacle;

inputting the road information, the self-vehicle driving state information and the obstacle driving state information into a pre-trained lane change decision model, wherein the lane change decision model comprises the following components: the system comprises a track predictor, a driving style probability sub-model, a lane change trend probability sub-model, a collision risk probability sub-model and a dynamic Bayesian decision network sub-model; so that the lane change decision model performs the following operations:

the track predictor correspondingly obtains the predicted track information of the own vehicle and the predicted track information of each obstacle based on the running state information of the own vehicle and the running state information of each obstacle respectively;

the driving style probability sub-model obtains driving style probability corresponding to the own vehicle based on the own vehicle driving state information;

The lane change trend probability sub-model obtains lane change trend probability of the own vehicle based on the running state information of the own vehicle and the road information;

the collision risk probability sub-model obtains the collision risk probability of the own vehicle based on the predicted track information of the own vehicle and the predicted track information of each obstacle;

and the dynamic Bayesian decision network sub-model obtains a lane change decision result corresponding to the current moment of the own vehicle based on the driving style probability, the lane change trend probability and the collision risk probability.

According to the embodiment of the application, when the lane change decision is made on the automatic driving vehicle, the style information of the driver, the collision risk information of the obstacle vehicle and the road information are comprehensively considered, so that the driving environment is continuously estimated, the potential risk is predicted and judged, the driving decision which is safe and meets the individuation of the driver is made, and the accuracy of the lane change decision is improved.

Optionally, the trajectory predictor includes: a track encoding network and a track decoding network; the track predictor correspondingly obtains the predicted track information of the own vehicle and the predicted track information of each obstacle based on the running state information of the own vehicle and the running state information of each obstacle respectively, and the step of correspondingly obtaining the predicted track information of the own vehicle and the predicted track information of each obstacle comprises the following steps:

The track coding network obtains corresponding environment coding vectors based on the self-vehicle running state information and the obstacle running state information respectively;

the track decoder correspondingly obtains the predicted track information of the own vehicle and the predicted track information of each obstacle based on each environment coding vector.

Optionally, the step of obtaining the driving style probability corresponding to the own vehicle based on the driving style probability sub-model of the own vehicle driving state information includes:

the driving style probability sub-model obtains the historical driving style probability of the previous period, and obtains the driving style probability corresponding to the own vehicle based on the historical driving style probability, the speed of the own vehicle and the distance between the front vehicles;

the step of obtaining the lane change trend probability of the own vehicle based on the running state information of the own vehicle and the road information by the lane change trend probability sub-model comprises the following steps:

the lane change trend sub-model obtains the historical lane change trend probability of the previous period, and obtains the lane change trend probability of the own vehicle based on the historical lane change trend probability, the distance between the centroid of the own vehicle and the lane line between the current lane and the target lane and the course angle of the own vehicle;

The step of obtaining the collision risk probability of the own vehicle based on the predicted track information of the own vehicle and the predicted track information of each obstacle by the collision risk probability sub-model comprises the following steps:

the collision risk probability sub-model obtains historical left lane collision risk probability and historical right lane collision risk probability of the previous cycle, determines x-direction coordinates of the own vehicle, x-direction coordinates of a front vehicle and x-direction coordinates of a rear vehicle in the current lane left lane, x-direction coordinates of the front vehicle and x-direction coordinates of the rear vehicle in the current lane right lane based on the predicted track information of the own vehicle and the predicted track information of each obstacle, obtains left lane collision risk probability of the own vehicle according to the historical left lane collision risk probability, and obtains right lane collision risk probability of the own vehicle according to the historical right lane collision risk probability, the x-direction coordinates of the front vehicle and the x-direction coordinates of the rear vehicle in the current lane right lane; the x direction is the direction of travel of the host vehicle.

Optionally, the step of obtaining the lane change decision result corresponding to the current moment of the own vehicle by the dynamic bayesian decision network sub-model based on the driving style probability, the lane change trend probability and the collision risk probability includes:

the dynamic Bayesian decision network sub-model acquires a historical lane change decision result of the previous period;

and the dynamic Bayesian decision network sub-model obtains a lane change decision result corresponding to the current moment of the own vehicle based on the historical lane change decision result, the driving style probability, the lane change trend probability and the collision risk probability.

Optionally, the training process of the lane change decision model includes:

collecting a natural driving data set, and extracting a lane change data set according to the natural driving data set;

training to obtain the trajectory predictor using the natural driving dataset;

acquiring a first labeling result of the lane change data set, and fitting a first mixed Gaussian model based on the lane change data set and the first labeling result to obtain the driving style probability submodel; the first labeling result comprises: cautious, general or aggressive;

obtaining a second labeling result of the lane change data set, and fitting a second Gaussian mixture model based on the lane change data set and the second labeling result to obtain the lane change trend probability submodel; the second labeling result comprises: no lane change trend or lane change trend;

Acquiring a third labeling result of the lane change data set, and fitting a third mixed Gaussian model based on the lane change data set and the third labeling result to obtain the collision risk probability submodel; the third labeling result comprises: the left lane has no collision risk, the left lane has collision risk, the right lane has no collision risk or the right lane has collision risk;

constructing a lane change decision model comprising the trajectory predictor, the driving style probability sub-model, the lane change trend probability sub-model, the collision risk probability sub-model and the dynamic Bayesian decision network sub-model, acquiring a fourth labeling result of the lane change data set, and training the dynamic Bayesian decision network sub-model based on the lane change data set and the fourth labeling result; the fourth labeling result comprises: left lane change, lane keeping, or right lane change.

In a second aspect, an embodiment of the present application provides an automatic driving vehicle lane change decision apparatus, the apparatus including:

the information acquisition module is used for acquiring road information, self-vehicle running state information and running state information of each obstacle;

the lane change decision module is used for inputting the road information, the self-vehicle driving state information and the obstacle driving state information into a lane change decision model trained in advance, and the lane change decision model comprises: the system comprises a track predictor, a driving style probability sub-model, a lane change trend probability sub-model, a collision risk probability sub-model and a dynamic Bayesian decision network sub-model; so that the lane change decision model performs the following operations:

Optionally, the trajectory predictor includes: a track encoding network and a track decoding network; the lane change decision module is specifically configured to:

Optionally, the lane change decision module is specifically configured to:

Optionally, the apparatus further includes: a model training module for:

training to obtain the trajectory predictor using the natural driving dataset;

In a third aspect, an embodiment of the present application provides an electronic device, including: a memory and a processor, the memory and the processor coupled;

the memory is used for storing one or more computer instructions;

the processor is configured to execute the one or more computer instructions to implement the method of automatically driving a vehicle lane change decision as described in the first aspect above.

In a fourth aspect, embodiments of the present application provide a computer-readable storage medium having stored thereon one or more computer instructions executable by a processor to implement the automated driving vehicle lane change decision method as described in the first aspect above.

In a fifth aspect, the present application provides a computer program product comprising a computer program which, when executed by a processor, implements the method of lane change decision for an autonomous vehicle according to the first aspect.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following description will briefly introduce the drawings that are required to be used in the description of the embodiments or the prior art. It is apparent that the drawings in the following description are only some of the embodiments of the present application. Other figures may be derived from these figures without inventive effort for a person of ordinary skill in the art.

FIG. 1 is a schematic diagram of a track predictor according to an embodiment of the present application;

FIG. 2 is a schematic structural diagram of a lane change decision model according to an embodiment of the present disclosure;

fig. 3 is a schematic diagram of driving status information of a host vehicle and surrounding vehicles according to an embodiment of the present application;

Fig. 4 is a schematic flow chart of a lane change decision method for an automatic driving vehicle according to an embodiment of the present application;

fig. 5 is a schematic structural diagram of an automatic driving vehicle lane change decision device according to an embodiment of the present disclosure;

fig. 6 is a schematic structural diagram of an electronic device according to an embodiment of the present application.

Detailed Description

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. It will be apparent that the described embodiments are only some, but not all, of the embodiments of the present application. All other embodiments, which can be made by one of ordinary skill in the art based on the embodiments herein without undue burden, are within the scope of the present application.

It should be noted that the terms "comprising" and "having" and any variations thereof in the embodiments and figures herein are intended to cover a non-exclusive inclusion. A process, method, system, article, or apparatus that comprises a list of steps or elements is not limited to only those listed but may alternatively include other steps or elements not listed or inherent to such process, method, article, or apparatus.

The good decision-making system needs to have excellent driving environment cognition capability, continuously estimate the driving environment, predict and judge potential risks, and simultaneously can take the style of the current driver into consideration to make safe and personalized driving decisions which accord with the driver. How to comprehensively consider driver style information, obstacle vehicle collision risk information and road information in the decision system, accurately predict future tracks of surrounding vehicles, improve safety and comfort of the decision system, and have great significance in promoting development of intelligent automobiles.

The embodiment of the application discloses a lane change decision method, device and equipment for an automatic driving vehicle and a storage medium, which can improve the accuracy of lane change decision. The embodiments of the present application are described in detail below.

Natural driving data under different driving environments are collected through a real vehicle data collection platform, data fragments of left lane change, right lane change and lane keeping conditions are respectively extracted, and a training and testing database is constructed; the lane change process is decomposed, so that main factors influencing the lane change decision of the driver are extracted, a lane change decision model based on reasoning is established by comprehensively considering the style information, the obstacle vehicle motion information, the road information and the like of the driver by utilizing a dynamic Bayesian network, the lane change decision process of the automatic driving vehicle is reasoning like a person, and the interpretability of the driving decision and the humanization of the automatic driving vehicle are improved. And moreover, a motion vehicle track predictor can be constructed based on a data driving method, and the future motion state and risk of the vehicle in the interested area are predicted, so that the result is used as the input of a lane change decision model, and the safety of a decision result is effectively ensured. The method has strong expansibility, can adapt to various driving scenes and has good application prospect in future automatic driving automobiles.

Specifically, a lane-changing decision model may be trained first, where the lane-changing decision model includes: the system comprises a track predictor, a driving style probability sub-model, a lane change trend probability sub-model, a collision risk probability sub-model and a dynamic Bayesian decision network sub-model. The method comprises the steps of firstly training to obtain a track predictor, then fitting a driving style probability sub-model, a lane change trend probability sub-model and a collision risk probability sub-model respectively, and finally training a dynamic Bayesian decision network sub-model based on the trained track predictor, the fitted driving style probability sub-model, lane change trend probability sub-model and the fitted collision risk probability sub-model.

In one implementation, the training process of the lane-change decision model may include: collecting a natural driving data set, and extracting a lane change data set according to the natural driving data set; training by using a natural driving data set to obtain a track predictor; acquiring a first labeling result of the lane change data set, and fitting the first mixed Gaussian model based on the lane change data set and the first labeling result to obtain a driving style probability sub-model; the first labeling result comprises: cautious, general or aggressive; obtaining a second labeling result of the lane change data set, and fitting a second Gaussian mixture model based on the lane change data set and the second labeling result to obtain a lane change trend probability sub-model; the second labeling result comprises: no lane change trend or lane change trend; obtaining a third labeling result of the lane change data set, and fitting a third mixed Gaussian model based on the lane change data set and the third labeling result to obtain a collision risk probability sub-model; the third labeling result comprises: the left lane has no collision risk, the left lane has collision risk, the right lane has no collision risk or the right lane has collision risk; the method comprises the steps of constructing a lane changing decision model comprising a track predictor, a driving style probability sub-model, a lane changing trend probability sub-model, a collision risk probability sub-model and a dynamic Bayesian decision network sub-model, acquiring a fourth labeling result of a lane changing data set, and training the dynamic Bayesian decision network sub-model based on the lane changing data set and the fourth labeling result; the fourth labeling result comprises: left lane change, lane keeping, or right lane change.

The automatic driving vehicle sensing system can be built based on a real vehicle testing platform, natural driving data of a driver in different driving environments are collected to be used as a natural driving data set, driving data under the working conditions of left lane change, right lane change and lane keeping are extracted, the driving data are calibrated, and a training set and a testing set are formulated. The collected information may include a self-vehicle driving state information, an obstacle driving state information, and road information, the self-vehicle driving state information may include a vehicle speed, an acceleration, a steering wheel angle, a yaw rate, position information, etc., the obstacle driving state information may include surrounding vehicle driving state information and static obstacle state information, the surrounding vehicle driving state information may include a speed, an acceleration, and position information, and the environment information may include a number of lanes, a road type, a traffic sign, etc. Or, the real vehicle data acquisition platform can acquire the related information, the environmental information and the host vehicle information of the surrounding vehicles acquired by the driving vehicle sensing system in the normal running process; the driving environment mainly comprises conventional scenes such as expressways, inter-urban express ways, national roads and the like.

When the lane change data set is extracted according to the natural driving data set, the lane change data set can be extracted specifically according to the lane change working condition and the lane keeping working condition.

The lane change working condition refers to a passive lane change working condition, mainly comprises a left lane change and a right lane change, and mainly meets the following requirements:

lane limiting: in the same lane, a host vehicle (own vehicle) has obstacle vehicles in front of the host vehicle to run, namely, the passive lane change behavior is considered, the lane change behavior is considered for one time, and the continuous lane change behavior is ignored;

speed limit: not less than 20km/h and not more than 120km/h;

time limit: the average time required for completing one lane change action on the expressway is 5s, and data 5s before and after a lane change point are extracted to be used as a complete lane change track;

lane changing point: and selecting the intersection point of the vehicle track and the lane line as a lane change point, and extracting data of the vehicle and surrounding vehicles to form a lane change data set.

The lane keeping condition mainly meets the following requirements:

lane limiting: the lane keeps the working condition and the main vehicle and the front vehicle run on the same lane; if other vehicles enter/exit the space between the two vehicles, the lane keeping working condition is finished;

distance limitation: the relative longitudinal distance between the two vehicles is smaller than 120m and larger than 5m;

speed limit: the speed of the main car and the front car is more than 20km/h and less than 120km/h;

time limit: the following duration is more than 30s, and the last 10s of data are extracted to form a lane change data set.

After the channel change data set is extracted, the training set and the test set can be manufactured. Specifically, the three extracted working conditions can be utilized to label the data set into three types, namely a left lane change working condition, a lane keeping working condition and a right lane change working condition, and the three working conditions can be represented by the set {0,1,2 }. The marked data are respectively divided into a training set and a testing set according to the proportion of 8:2 and are used for training and testing a track predictor and a lane change decision model.

Because the data-driven prediction method has high prediction precision and strong fitting capability, a specific physical model is not needed, and the running state of the vehicle in a complex environment can be predicted, in the embodiment of the application, an LSTM (Long Short-Term Memory) can be adopted as a reference model to construct a track predictor, as shown in fig. 1, the track predictor specifically comprises a track coding network (Encoder LSTM) and a track decoding network (DecoderLSTM), wherein the two networks have the same unit number and are 128 units; the input front end of the track coding network constructs an embedded layer (full connection layer) with 64 units for uniformly coding the input, the output of the track coding network is an environment vector with the size of 128x1, and the input of the track decoding network is the output of the track coding network.

Specifically, the input to the track-encoding network is a predicted vehicle's historical track:

wherein x is ^t ＝[x ^t ,y ^t ,v ^t ,a ^t ]，t _h Is the historical time domain length; x is x ^t ,y ^t ,v ^t ,a ^t The longitudinal coordinates, the transverse coordinates, the speed and the acceleration of the predicted vehicle are respectively output into an environment coding vector (context) as a track decodingAn input to the network.

The output of the track decoding network is:

wherein y is ^t+1 ＝[x ^t+1 ,y ^t+1 ]，t _f Is the predicted horizon length, and y is the predicted future travel trajectory of the vehicle.

The track predictor is trained and tested by using the lane change data set, the trained track predictor can predict the tracks of the host vehicle and surrounding vehicles of the interested region, and the prediction result is used for judging the subsequent collision risk of the road.

In the embodiment of the present application, factors affecting lane change decisions may be classified into three categories: driver style factor H ^style Road factor H ^road Dynamic risk factor H ^dyn ＝[H ^dyn1 ,H ^dyn2 ]The above factors correspond to different observation variables O.

Before making lane change decisions according to human driving thinking, a driver evaluates the observed driving conditions according to self driving experience, analyzes the feasibility and possibility of lane change of the vehicle, and executes corresponding driving strategies (hidden states) once specific conditions are met, which are highly compatible with the dynamic Bayesian hidden state reasoning process. According to the channel change decision influencing factors, a channel change decision model shown in fig. 2 can be constructed, wherein rectangular nodes are conditional variables, discrete variables, diamond nodes are decision variables, discrete variables and observation variables, and circular nodes are continuous variables.

As shown in FIG. 2, H ^style The method is characterized in that the method is used for representing the style variable of the driver, is a hidden variable, and has a value set of {0,1,2}, and represents three classes of discreet, general and aggressive drivers respectively; h ^road The channel change trend signal is a hidden variable, and the value is {0,1}, which are respectively expressed as no channel change trend and channel change trend; h ^dyn1 The left lane risk variable is a hidden variable, and the value is {0,1}, which are respectively indicated as no collision risk and collision risk; h ^dyn2 Representing right lane risk variablesThe value of the hidden variable is {0,1}, which is respectively indicated as no collision risk and collision risk.

O is an observation variable corresponding to the hidden variable, and is a continuous variable. G is a decision variable, representing lane change intention, and is a discrete variable, {0,1,2} represents left lane change, lane keeping, right lane change, respectively.

The observed variables corresponding to each hidden variable are as follows:

H ^style ←O ^style ＝[v _e ,L]

wherein v is _e The speed of the main vehicle, L is the distance from the front vehicle, and the specific meaning is shown in figure 3; o (O) ^style The mixture Gaussian distribution is met, and the expression is as follows:

wherein N represents a Gaussian function, pi _k Representing the probability of the kth gaussian,v is respectively _e The average value and variance corresponding to L are used for dividing the lane change data set into three types according to different following distances during lane change, and the three types of lane change data respectively represent cautious type, general type and aggressive type drivers.

And solving the mixed Gaussian model parameters by using the lane change data set and labeling results (cautious, general or aggressive) through an EM (Expectation-maximization) algorithm to obtain three types of driving style probability models.

H ^road ←O ^road ＝[|Δy|,θ]

Wherein, |Δy| is the absolute value of the distance between the centroid position and the boundary line of the target road, θ is the course angle, and the specific meaning is shown in fig. 3; the target road may be an adjacent lane on the same side of the current lane as the heading angle of the host vehicle, for example, when the heading angle is leftward, the target lane may be a left adjacent lane of the current lane, and when the heading angle is rightward, the target lane may be a right adjacent lane of the current lane; o (O) ^road The mixture Gaussian distribution is met, and the expression is as follows:

wherein N represents a Gaussian function, pi _i Representing the probability of the ith gaussian,the mean and variance corresponding to delta y and theta are respectively divided into two categories according to whether lane change is carried out or not, and the two categories respectively represent lane change trend and lane non-change trend.

And solving the parameters of the Gaussian mixture model by using the lane change data set and labeling results (with lane change trend and no lane change trend) through an EM algorithm to obtain two types of lane change trend probability models.

H ^dyn ＝[H ^dyn1 ,H ^dyn2 ]

H ^dyn1 Represents the dynamic risk of the left road, H ^dyn2 Represents the right-hand road dynamic risk, wherein,the specific relationship is shown in fig. 3:

in the method, in the process of the invention,for the relative distance between the host vehicle and the left-hand front vehicle,/->For the relative distance between the host vehicle and the left-lane rear vehicle,/->For the relative distance between the host vehicle and the front vehicle of the right lane,/->O is the relative distance between the main vehicle and the rear vehicle of the right lane ^dyn1 、O ^dyn2 For the corresponding observation variable, +.> The X-direction coordinates of the future main vehicle, the left-lane rear vehicle, the left-lane front vehicle, the right-lane rear vehicle and the right-lane front vehicle are obtained by a track predictor.

O ^dyn1 ，O ^dyn2 All obeys the Gaussian mixture distribution, the minimum safe vehicle distance model is utilized to divide the lane change data set into two categories of collision risk and collision-free risk, the parameters of the Gaussian mixture model are solved through an EM algorithm by utilizing the lane change data set, and the collision risk probability models of the left lane and the right lane are obtained.

The minimum safe distance model is as follows:

minimum safety distance between main vehicle and rear vehicle of target laneExpressed as:

minimum safety distance between main vehicle and front vehicle of target laneRepresented as

In the formula, v _r For the speed of the vehicle behind the target lane, v _e Velocity v of the host vehicle _f For the speed of the front vehicle of the target lane, l _safe Can be set to be 3.5 m, t as a safety threshold _tal For the channel change time, the time can be set to be 4 seconds, t _r Is the sum of the driver response time and the brake response time, 1.1 seconds, t _b The brake rise time can be 0.1 seconds, a _g Is the maximum braking deceleration.

After the lane change decision model is obtained through training, the lane change decision of the automatic driving vehicle can be carried out based on the lane change decision model. Specifically, as shown in fig. 4, the lane change decision method for the automatic driving vehicle provided in the embodiment of the present application may include the following steps:

s410: road information, own vehicle running state information, and obstacle running state information are acquired.

In particular, the desired information may be obtained based on the perception system. The road information may include the number of lanes, road type, traffic sign, etc.; the own vehicle running state information may include a vehicle speed, an acceleration, a steering wheel angle, a yaw rate, position information, etc., the obstacle running state information may include surrounding vehicle running state information and static obstacle state information, and the surrounding vehicle running state information may include a speed, an acceleration, position information, etc.

S420: inputting road information, self-vehicle driving state information and obstacle driving state information into a pre-trained lane change decision model, wherein the lane change decision model comprises the following components: the system comprises a track predictor, a driving style probability sub-model, a lane change trend probability sub-model, a collision risk probability sub-model and a dynamic Bayesian decision network sub-model; so that the lane change decision model performs the following operations: the track predictor correspondingly obtains predicted track information of the own vehicle and predicted track information of each obstacle based on the running state information of the own vehicle and the running state information of each obstacle respectively; the driving style probability sub-model obtains driving style probability corresponding to the own vehicle based on the driving state information of the own vehicle; the lane change trend probability sub-model obtains lane change trend probability of the own vehicle based on the running state information and the road information of the own vehicle; the collision risk probability sub-model obtains the collision risk probability of the own vehicle based on the predicted track information of the own vehicle and the predicted track information of each obstacle; the dynamic Bayesian decision network sub-model obtains a lane change decision result corresponding to the current moment of the own vehicle based on the driving style probability, the lane change trend probability and the collision risk probability.

Wherein, the trajectory predictor may include: a track encoding network and a track decoding network; the track predictor may correspondingly obtain predicted track information of the own vehicle and predicted track information of each obstacle based on the running state information of the own vehicle and the running state information of each obstacle, respectively, and the step of predicting the track information of each obstacle may specifically include: the track coding network obtains corresponding environment coding vectors based on the running state information of the vehicle and the running state information of each obstacle respectively; the track decoder correspondingly obtains the predicted track information of the own vehicle and the predicted track information of each obstacle based on each environment coding vector.

When the driving style probability sub-model obtains the driving style probability corresponding to the own vehicle based on the driving state information of the own vehicle, the historical driving style probability of the previous period can be obtained, and the driving style probability corresponding to the own vehicle is obtained based on the historical driving style probability, the speed of the own vehicle and the distance between the front vehicles.

When the lane change trend sub-model obtains the lane change trend probability of the own vehicle based on the running state information and the road information of the own vehicle, the historical lane change trend probability of the previous period can be obtained, and the lane change trend probability of the own vehicle is obtained based on the historical lane change trend probability, the distance of the center of mass of the own vehicle relative to the lane line between the current lane and the target lane and the course angle of the own vehicle.

When the collision risk probability sub-model obtains the collision risk probability of the own vehicle based on the predicted track information of the own vehicle and the predicted track information of each obstacle, the historical left lane collision risk probability and the historical right lane collision risk probability of the own vehicle can be obtained, the x-direction coordinate of the own vehicle after the preset time period, the x-direction coordinate of the own vehicle in the left lane of the current lane and the x-direction coordinate of the rear vehicle, the x-direction coordinate of the own vehicle in the right lane of the current lane and the x-direction coordinate of the rear vehicle are determined based on the predicted track information of the own vehicle and the predicted track information of each obstacle, and the right lane collision risk probability of the own vehicle is obtained according to the historical left lane collision risk probability, the x-direction coordinate of the own vehicle in the left lane of the current lane and the x-direction coordinate of the rear vehicle; the x direction is the direction of travel of the vehicle.

The dynamic Bayesian decision network sub-model can acquire a historical lane change decision result of the previous period when acquiring a lane change decision result corresponding to the current moment of the own vehicle based on the driving style probability, the lane change trend probability and the collision risk probability; and obtaining a lane change decision result corresponding to the current moment of the vehicle based on the historical lane change decision result, the driving style probability, the lane change trend probability and the collision risk probability.

The lane change decision process described above is described in detail with reference to fig. 2, and in fig. 2, the probability relationship between the decision variable and the condition variable may be represented as follows:is equivalent to

In the method, in the process of the invention,is the decision variable at the current moment, +.>Is a decision variable at the previous time, is a discrete variable, and {0,1,2} represents a left lane change, a lane keeping, and a right lane change, respectively. That is, the driving decision at the present time is subjected to the previous timeThe decision impact and the impact of driver characteristic factors, road factors and driving environment dynamic risk factors at the current moment. The network updates the probability distribution values of the remaining variables with the observed data received at each instant.

Specifically, the lane-change decision model includes three layers of nodes: a hiding layer, an observing layer and a decision layer.

Hidden layer variables include driver style variables, road factor variables, and dynamic risk factor variables. The driver style variables refer to the differences in the following speed and following distance of different drivers before making lane change decisions. Road factors refer to the relative distance difference of the vehicle relative to the side line of the lane, and represent the lane change trend. Dynamic risk factors refer to the risk of collision of a host vehicle with a target lane obstacle vehicle, including the risk of collision with a front and rear vehicle of the target lane, characterizing interaction with an environmental obstacle vehicle.

The observation layer variable corresponds to the hidden layer variable and comprises self-vehicle information, surrounding obstacle vehicle information and road related sequence information which are acquired by an automatic driving vehicle sensing system.

The decision layer is influenced by the decision of the previous moment and the driver style factors, road factors and driving environment dynamic risk factors of the current moment, and represents the decision made by the driver at the current moment after comprehensively considering the factors.

The inference network (lane change decision model) of FIG. 2 includes three types of probability relationships, namely the state transition probabilities P (H) _t |H _t-1 ) Conditional probability P (O) between a conditional variable and an observed variable _t |H _t ) Probability relation P (G _t |G _t-1 ,H _t )。

Probability of state transition of conditional variable P (H _t |H _t-1 ) The method specifically comprises the following steps: the matrix dimension is 3×3 dimension, 2×2 dimensionDimension, 2 x 2 dimension. Let the condition variable node be lambda _style ＝λ _dyn1 ＝λ _dyn2 ＝λ _road Probability of =0.9 maintains the current state according to 1- λ _style 、1-λ _dyn1 、1-λ _dyn2 、1-λ _road To other states.

The transition probability relation of all nodes in the conditional variable set is that

The observed variable is considered to be independently distributed given the condition variable, then the conditional probability P (O) _t |H _t ) Has the following relation

The observed variable is obtained by an EM (effective magnetic resonance) algorithm after obeying the Gaussian mixture distribution.

Probability relation P (G) _t |G _t-1 ,H _t ) Which is equivalent toIs a 3 x 2 matrix, from the conditional variables P (H _t ) And the decision P (G) _t-1 ) And (5) deducing and acquiring.

For decision variables P (G _t ) And updating the probability distribution of each variable by receiving the observed quantity at each moment, thereby completing the reasoning process.

By introducing a posterior joint probability at a previous instantAnd a priori distribution of the current moment->Obtained by the current momentObservation variable and conditional probability calculation of posterior distribution at this moment +.>

the joint probability distribution of the hidden variable and the decision variable at the moment t is as follows

The prior distribution at the time t can be further calculated

With a priori distribution at time tUpdating posterior distribution at the moment according to the observation variable and conditional probability at the moment

Given the condition variables, the observation variables are assumed to be independently distributed, so the condition probability relation can be calculated by the following formula

Opposite typeAnd performing accumulation summation to obtain the current moment strategy of the decision node G, thereby completing the reasoning process.

Fig. 5 is a schematic structural diagram of an automatic driving vehicle lane change decision device according to an embodiment of the present application, where the device includes:

an information acquisition module 510 for acquiring road information, vehicle running state information, and running state information of each obstacle;

the lane change decision module 520 is configured to input the road information, the driving status information of the vehicle, and the driving status information of each obstacle into a pre-trained lane change decision model, where the lane change decision model includes: the system comprises a track predictor, a driving style probability sub-model, a lane change trend probability sub-model, a collision risk probability sub-model and a dynamic Bayesian decision network sub-model; so that the lane change decision model performs the following operations:

Optionally, the trajectory predictor includes: a track encoding network and a track decoding network; the lane change decision module 520 is specifically configured to:

Optionally, the lane change decision module 520 is specifically configured to:

Optionally, the apparatus further includes: a model training module for:

training to obtain the trajectory predictor using the natural driving dataset;

The device embodiment corresponds to the method embodiment, and has the same technical effects as the method embodiment, and the specific description refers to the method embodiment. The apparatus embodiments are based on the method embodiments, and specific descriptions may be referred to in the method embodiment section, which is not repeated herein.

Next, referring to fig. 6, fig. 6 is a schematic structural diagram of a server according to an embodiment of the present application, where the server includes:

one or more processors 40;

the processor 40 is coupled to a storage means 41, which storage means 41 is adapted to store one or more programs,

when executed by the one or more processors 40, causes the electronic device to implement a solution for an automated driving vehicle lane change decision method as described in fig. 4.

The embodiment of the application also provides a computer readable storage medium, on which a computer program is stored, the program being executed by a processor to implement the automatic driving vehicle lane change decision method as described in fig. 4.

The present application provides a computer program product comprising a computer program which when executed by a processor implements an autonomous vehicle lane change decision method as described in fig. 4.

Those of ordinary skill in the art will appreciate that: the figures are schematic representations of one embodiment only and the modules or flows in the figures are not necessarily required to practice the present application.

Those of ordinary skill in the art will appreciate that: the modules in the apparatus of the embodiments may be distributed in the apparatus of the embodiments according to the description of the embodiments, or may be located in one or more apparatuses different from the present embodiments with corresponding changes. The modules of the above embodiments may be combined into one module, or may be further split into a plurality of sub-modules.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present application, and are not limiting thereof; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some of the technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the corresponding technical solutions.

Claims

1. A lane change decision method for an autonomous vehicle, the method comprising:

2. The method of claim 1, wherein the trajectory predictor comprises: a track encoding network and a track decoding network; the track predictor correspondingly obtains the predicted track information of the own vehicle and the predicted track information of each obstacle based on the running state information of the own vehicle and the running state information of each obstacle respectively, and the step of correspondingly obtaining the predicted track information of the own vehicle and the predicted track information of each obstacle comprises the following steps:

3. The method of claim 1, wherein the driving style probability sub-model is based on the vehicle driving state information, and the step of obtaining the driving style probability corresponding to the vehicle comprises:

4. A method according to any one of claims 1-3, wherein the step of obtaining the lane change decision result corresponding to the current time of the host vehicle based on the driving style probability, the lane change trend probability, and the collision risk probability by the dynamic bayesian decision network sub-model comprises:

5. A method according to any one of claims 1-3, wherein the training process of the lane-change decision model comprises:

training to obtain the trajectory predictor using the natural driving dataset;

6. An automatic driving vehicle lane change decision apparatus, the apparatus comprising:

7. The apparatus of claim 6, wherein the trajectory predictor comprises: a track encoding network and a track decoding network; the lane change decision module is specifically configured to:

8. The apparatus of claim 6, wherein the lane change decision module is specifically configured to:

9. An electronic device, comprising: a memory and a processor, the memory and the processor coupled;

the memory is used for storing one or more computer instructions;

the processor is configured to execute the one or more computer instructions to implement the autonomous vehicle lane change decision method of any of claims 1 to 5.

10. A computer readable storage medium having stored thereon one or more computer instructions, wherein the instructions are executed by a processor to implement the automated driving vehicle lane change decision method of any of claims 1 to 5.