CN110304075A - Track of vehicle prediction technique based on Mix-state DBN and Gaussian process - Google Patents

Track of vehicle prediction technique based on Mix-state DBN and Gaussian process Download PDF

Info

Publication number
CN110304075A
CN110304075A CN201910598776.0A CN201910598776A CN110304075A CN 110304075 A CN110304075 A CN 110304075A CN 201910598776 A CN201910598776 A CN 201910598776A CN 110304075 A CN110304075 A CN 110304075A
Authority
CN
China
Prior art keywords
vehicle
driving
variable
dbn
function
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201910598776.0A
Other languages
Chinese (zh)
Other versions
CN110304075B (en
Inventor
罗禹贡
刘金鑫
钟志华
李克强
王庭晗
陈锐
王永胜
徐明畅
于杰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tsinghua University
Original Assignee
Tsinghua University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tsinghua University filed Critical Tsinghua University
Priority to CN201910598776.0A priority Critical patent/CN110304075B/en
Publication of CN110304075A publication Critical patent/CN110304075A/en
Application granted granted Critical
Publication of CN110304075B publication Critical patent/CN110304075B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W50/00—Details of control systems for road vehicle drive control not related to the control of a particular sub-unit, e.g. process diagnostic or vehicle driver interfaces
    • B60W50/0097—Predicting future conditions
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W50/00—Details of control systems for road vehicle drive control not related to the control of a particular sub-unit, e.g. process diagnostic or vehicle driver interfaces
    • B60W2050/0062—Adapting control system settings
    • B60W2050/0075—Automatic parameter input, automatic initialising or calibrating means
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W2556/00—Input parameters relating to data
    • B60W2556/10—Historical data

Abstract

The invention belongs to automatic driving vehicle Context aware and decision-making technic field, in particular to a kind of track of vehicle prediction technique based on Mix-state DBN and Gaussian process.The present invention learns the parameter of MDBN and GP by vehicle nature driving data, multiple vehicle kinematics models are merged using MDBN, the estimated probability of short-term trajectory predictions and driving intention and driving performance is obtained, then, the expression of long-term trajectory predictions and uncertainty in traffic is carried out using GP.This method can either consider the short-term forecast characteristic under vehicle physical motion model, again it is contemplated that vehicle driver's information carries out long-term trajectory predictions and uncertain expression, compared to current track of vehicle prediction technique, present invention incorporates auto model, abstract intention and data-drivens, the scalability of MDBN and GP model is strong, it can be applicable in different Driving Scenes, it can be in conjunction with more effective contextual informations, such as road information, traffic information.

Description

Track of vehicle prediction technique based on Mix-state DBN and Gaussian process
Technical field
It is the invention belongs to automatic driving vehicle Context aware and decision-making technic field, in particular to a kind of based on mixing dynamic The track of vehicle prediction technique of Bayesian network and Gaussian process.
Background technique
Currently, a kind of scheme of the realization Vehicular automatic driving of mainstream is based on " perception-decision-control " framework.It is this Layer-stepping framework uses the thought to personalize, just as people needs the environment with the sense organs such as eye, ear, nose perception periphery；So Afterwards by processing of the brain to perception information, form the understanding and judgement to the environment on periphery, thus make reasonable decision with Planning；Finally by the limbs of people, such as the task that hand, foot execution determine.Obviously, just as brain is the core of human body, The decision system of automatic driving vehicle plays a crucial role.Decision system is the cognition of driving-situation first, needs depth Stratification solution perception information, is constantly estimated, judged and is predicted to the variation of driving environment, secondly, being determined using cognitive information Determine driving behavior and the driving path in planning automatic driving vehicle future in automatic driving vehicle future.Wherein, how to improve certainly The dynamic vehicle that drives correctly estimates, the following rail of judgement and prediction surrounding traffic participant the degree of awareness of driving-situation Mark is the core and challenge of automatic driving vehicle decision system.
For automatic driving vehicle decision system presently, there are problem and challenge, had before part scientific research personnel pass Research is unfolded to the trajectory predictions of all vehicles in automatic driving vehicle driving process, summary has following three categories method: 1, Track of vehicle prediction based on physical model: vehicle is expressed as the transaction constrained by physical rules by it, dynamic according to vehicle The track in mechanics or kinematics model prediction vehicle future；2, based on the track of vehicle prediction of driving intention estimation: it is by vehicle It is expressed as individual behavior entity, according to the intention estimated result to current vehicle, it is corresponding under the driving intention obtains vehicle Future Trajectory；3, the track of vehicle prediction based on depth network: using the natural running data of vehicle, pass through deep learning etc. Method obtains the direct mapping by data to track.
First kind method, the track of vehicle prediction based on physical model, only considers the motion feature of current vehicle, is suitble to short Trajectory predictions in time；Due to the selection of vehicle physical model directly affect trajectory predictions as a result, and vehicle in different travelings Corresponding physical model is different under state, therefore can not accurately carry out trajectory predictions according to single unit vehicle model；Simultaneously as not Consider vehicle driver's information (such as driving intention；Driving performance), the motion profile in the unpredictable long-term range of vehicle；Except this Except, this method has ignored influence of the vehicle-periphery factor to track of vehicle.Second class method is estimated based on driving intention Track of vehicle prediction, it can predict to execute a specific operation by vehicle, and (such as crossing is turned: being slowed down, is turned to, accelerating to turn It is curved) caused by motion profile variation, part implementation method it is also predicted that due to ambient environmental factors variation caused by vehicle Trail change, therefore such methods are able to carry out the prediction of the track of vehicle in Long time scale, still, due to not accounting for vehicle Itself physical motion feature, so the prediction error in the short time is larger.Third class method, the vehicle based on depth network Trajectory predictions, it is the rapid development due to artificial intelligence and big data technology and generates, and relates generally to shot and long term memory net Network (LSTM), depth Bayesian network etc., since the fitting performance of this method is extremely strong, can be realized complex interaction scene (if any Traffic lights or intersection without traffic lights) under track of vehicle prediction.But due to needing a large amount of calibration Data sample does not consider the physical motion feature of vehicle itself, and the generalization ability for the neural network model established is not strong, can solve The disadvantage of the property released difference, it is difficult to complete the track of vehicle prediction under multiple scenes, it is difficult to be found according to mistake output result corresponding Reason, and be difficult to carry out the expression of uncertainty in traffic.
Therefore, need to overcome the trajectory predictions method of all vehicles in a kind of automatic driving vehicle driving process or At least mitigate the drawbacks described above of the prior art.
Summary of the invention
In view of the above technical problems, the object of the present invention is to provide one kind to be based on Mix-state DBN and Gauss mistake The track of vehicle prediction technique of journey, to week in automatic driving vehicle driving process, the track of all vehicles is predicted, to drive automatically The decision-making level for sailing vehicle provides more reliable information.
To achieve the goals above, the present invention provides the following technical scheme that
A kind of track of vehicle prediction technique based on Mix-state DBN and Gaussian process, includes the following steps:
Step 1, building nature driving data library；
Establish automatic driving vehicle sensory perceptual system acquisition surrounding vehicles associated sequence information, road associated sequence information and The test set of traffic associated sequence information, and above- mentioned information are carried out with the training set of driving intention and driving performance calibration；Its In, the test set includes the test set of Mix-state DBN and the test set of Gaussian process；The training set includes The training set of Mix-state DBN and the training set of Gaussian process；
Step 2, using Mix-state DBN, obtain short-term forecast track and the driving intention of surrounding vehicles With the estimated probability of driving performance；
It is selected using driver information, auto model and car status information is as hidden layer variable, with automatic driving vehicle sense Know the surrounding vehicles associated sequence information, road associated sequence information and traffic associated sequence information of system acquisition as observation layer Variable builds Mix-state DBN model；The output of Mix-state DBN is obtained by posterior probability reasoning: The estimated probability of short-term trajectory predictions and driving intention and driving performance based on auto model, and using the output as step 4 input；
Step 3 establishes Gaussian process function under different driving intentions and different driving performance；
It comprises the following steps:
The setting of step 3.1, mean function and covariance function
Using x as input, the expression formula of Gaussian process function is as follows:
F (x)~GP (u (x), Σ (x, x '))
Wherein, mean function u (x) indicates the track of vehicle trend under certain driving intention, therefore mean function can be used To indicate expectation prediction locus；Covariance function Σ (x, x ') not only indicated the variance of different input x itself, but also illustrate (x, x ') Between variance, therefore covariance function can be used to indicate the corresponding uncertainty of expectation prediction locus；
Corresponding mean function u (x) is set according to the track of vehicle under different driving intentions；
Covariance function Σ that square exponential form with noise indicates (x, x ') setting is as follows:
Wherein, σfPoor for signal standards, l is characterized length, σnFor observation noise standard deviation, δ is Kronecker function；
The study of step 3.2, Gaussian process function unknown parameter
The mean function u (x) and covariance function Σ (x, x ') set according to step 3.1, the Gauss for utilizing step 1 to establish The training set of process learns the unknown parameter being related to, according to logarithm marginal likelihood function and its to the inclined of each unknown parameter It leads, obtains parameter learning using the optimization algorithm based on gradient as a result, to establish different driving intentions and driving performance respectively Under Gaussian process function；
Step 4 carries out long-term trajectory predictions and uncertainty table based on Mix-state DBN and Gaussian process Show；
The estimated probability of the driving intention and driving performance that are exported according to Mix-state DBN in step 2 utilizes Maximum probability principle determines corresponding driving intention and driving performance, to determine step according to the driving intention and driving performance Corresponding Gaussian process function in 3；
The short-term forecast track based on auto model that Mix-state DBN in step 2 is exported is as vehicle rail Mark sequence information x1, the Future Trajectory of prediction is x2, then (x1,x2) obey Gaussian Profile it is as follows:
Wherein,Indicate normal function；x1With x2Corresponding mean value letter Number is respectively u1And u2；Covariance matrix Σ is symmetrical matrix, i.e. Σ=ΣT, x1With x2Corresponding covariance function difference For Σ11And Σ22, x1And x2Between corresponding covariance function be Σ12、Σ21, and ∑12=∑21 T；
Then known vehicle track x1Under vehicle future may track x2Conditional probability P (x2|x1) obey Gaussian Profile Expression formula is as follows:
Wherein, u2|1And Σ2|1It is variable x2|1Corresponding mean function and covariance function, and (u2|1,Σ2|1) expression Formula is derived by by following formula:
u2|1=u2+Σ12 TΣ11 -1(x1-u1)
Σ2|1=Σ22-Σ12 TΣ11 -1Σ12
It finally obtains in the long time domain of vehicle under the driving intention and driving performance determined using maximum probability principle not Coming track and uncertainty indicates.
In the step 1, the surrounding vehicles associated sequence information include vehicle location, car speed, vehicle acceleration, Yaw rate, distance and left and right vehicle wheel of the vehicle away from both sides of the road turn to the opening and closing situation of taillight；
The road associated sequence information includes the structure feature and road instruction mark of road；
The traffic associated sequence information includes surrounding traffic Warning Mark and traffic signal light condition.
The step 1 includes the following steps:
1.1, the acquisition of driving data
Acquire the surrounding vehicles associated sequence information, road phase of the automatic driving vehicle sensory perceptual system in normal driving process Close sequence information and traffic associated sequence information；
1.2, driving intention classification and calibration
The scene according to locating for vehicle defines the possible driving intention of surrounding vehicles；For each possible driving intention, The key characterization parameter that selection is classified for driving intention with calibration；It is related then in conjunction with the collected surrounding vehicles of step 1.1 Sequence information determines surrounding vehicles driving intention whithin a period of time and sets corresponding driving intention label；
Wherein, scene locating for the vehicle is obtained by the intelligent guidance system of automatic driving vehicle, scene packet locating for vehicle Include High-speed Circumstance, there is traffic lights intersection scene in hill path scene, city normal straight-ahead operation road scene, city, city without Traffic lights intersection scene；
The driving intention includes along present road straight trip, along present road turning, Zuo Huandao, right lane-change, left-hand bend, the right side It turns, turn around, stopping and starting；
The key characterization parameter classified and demarcated for driving intention includes lateral direction of car position, lateral direction of car speed Degree, yaw rate, vehicle distance and left and right vehicle wheel away from both sides of the road turn to taillight and are opened and closed situation；
1.3, the classification and calibration of driving performance
The possible driving performance of surrounding vehicles is defined first, and selection is classified for driving performance to be joined with the key feature determined Number, according to the collected surrounding vehicles associated sequence information of step 1.1, the surrounding vehicles obtained in conjunction with step 1.2 are at one section In driving intention, determine driving performance of the surrounding vehicles under a certain driving intention and set corresponding driving performance mark Label；
The driving performance includes slowly driving, smooth ride and rapid driving；
The key characterization parameter classified and determined for driving performance includes vehicular longitudinal velocity, lateral direction of car speed Degree, longitudinal acceleration of the vehicle, vehicle lateral acceleration and yaw rate；
1.4, the foundation of training set and test set
Surrounding vehicles associated sequence information, road associated sequence information and the traffic correlated series letter that step 1.1 is acquired Cease the test set as Mix-state DBN MDBN；
Take the leading portion moment of the lateral direction of car position and lengthwise position that correspond under a certain driving intention and driving performance A test set of the sequence information as Gaussian process；
Surrounding vehicles associated sequence information, road associated sequence information and the traffic associated sequence information that step 1.1 acquires Classify through step 1.2 driving intention with after calibration and the classification and calibration of 1.3 driving performances, obtains Mix-state DBN The training set of MDBN；
According to the training set of Mix-state DBN MDBN, will correspond under a certain driving intention and driving performance A training set of the full sequence information of lateral direction of car position and lengthwise position as Gaussian process GP.
In the step 2, connection relationship between the observation layer variable, hidden layer variable and two variables, using driving naturally It sails data and improved adaptive GA-IAGA obtains；The corresponding probability parameter of connection relationship is obtained by maximum likelihood method.
In the step 2, the auto model includes at the uniform velocity model, even acceleration model, at the uniform velocity steering model；
The car status information includes lateral direction of car position, vehicle lateral speed, vehicle lateral acceleration, longitudinal direction of car Position, vehicular longitudinal velocity, longitudinal acceleration of the vehicle and yaw rate.
In the step 2, the hidden layer variable further comprises that other related higher-levels imply variable, other described related height The implicit variable of layer includes the reciprocation between the risk assessment of vehicle and environment, vehicle；
The observation layer variable further comprise with other related higher-levels of hidden layer variable imply variable it is corresponding other Related bottom observation information.
The step 2 specifically comprises the following steps:
2.1, the hidden layer and observation node layer of Mix-state DBN model are defined
Hidden node includes three-layer information: first layer is driver information variables D, and the second layer is auto model selection variable M, third layer are car status information variable S；Hidden layer variables collection H={ D, M, S }, wherein { D, M } is discrete variable, { S } is Continuous variable；
Observing node layer mainly includes one layer of information: the surrounding vehicles correlated series of automatic driving vehicle sensory perceptual system acquisition Information, road associated sequence information and traffic associated sequence information；Observation layer variables collection is { O }, and is continuous variable；
2.2, the optimization of network structure
Given variables set { D, M, S, O } and training set T, learning and finding this instruction of a best match by way of search Practice network N=(B, the θ) of collection T, and uses measurement of the criterion function as matching degree；Wherein, B indicates network structure, θ Indicate the parameter of network；
2.3, the unknown parameter study between network node
The conditional probability of part hidden layer variable is obtained by the method for data statistics, the condition between remaining hidden layer and observed quantity The training set for the Mix-state DBN MDBN that probability is obtained using step 1 carries out parameter learning, is estimated using maximum likelihood Meter method solves, it is known that about likelihood function L (θ)=logP of network parameter θ (D | θ) under training sample D, using being based on gradient Optimization algorithm solve optimal solution
2.4, the posterior probability reasoning under observation layer variable is introduced
Posterior probability refers to the probability distribution P (H | O) of hidden layer variable under known observation layer variable, i.e., obtains in step 1 mixed Under the test set for closing dynamic bayesian network MDBN, corresponding driving intention, driving performance, auto model and vehicle-state are inferred Process, obtain the estimated probability of short-term trajectory predictions and driving intention and driving performance based on auto model.
In the step 2, the posterior probability reasoning is a kind of hypothesis density filtering approximate resoning, and detailed process is as follows:
1) it predicts: obtaining the joint probability distribution at t+1 moment with the probability distribution of t moment；If all discrete hidden layers of t moment Variable { Mt,DtJoint Distribution be
The then discrete hidden layer variable { M at t+1 momentt,Dt,Mt+1,Dt+1Joint Distribution are as follows:
T+1 moment continuous hidden layer variable { St+1Conditional probability are as follows:
2) it updates: introducing the observational variable at t+1 moment to obtain new probability distribution；If the observational variable of auto model is { V }, and meetRemaining observational variable is { E } in { O }, then the discrete hidden layer variable { M at t+1 momentt,Dt, Mt+1,Dt+1Joint Distribution are as follows:
Wherein,
T+1 moment continuous hidden layer variable { St+1Conditional probability are as follows:
3) marginalisation: by the variable marginalisation before the t+1 moment in the t+1 moment probability distribution obtained from above, only retain and work as The probability distribution of variable is inscribed when preceding t+1, for prediction next time and the iteration of update step:
Wherein, D indicates driver information variable, and driver information includes driving intention and driving performance；M indicates vehicle mould Type selects variable, and auto model includes at the uniform velocity model, acceleration model, at the uniform velocity steering model；S indicates car status information variable, Car status information includes that lateral direction of car position, vehicle lateral speed, vehicle lateral acceleration, longitudinal direction of car position, vehicle are vertical To speed, longitudinal acceleration of the vehicle and yaw rate；It utilizesObtain t+1 moment vehicle The estimated probability of model, driving intention and driving performance；It utilizesIt obtains based on the short-term of auto model Prediction locus.
In the step 3.2, maximum likelihood method is selected to carry out the unknown parameter study of Gaussian process.
The detailed process of the step 3.2 are as follows:
The mean function u (x) and covariance function Σ (x, x ') set according to step 3.1, the Gauss for utilizing step 1 to establish Process GP training set DGP(X, Y)={ (xi,yi), i=1:N1The unknown parameter being related to is learnt, wherein N1For training sample This sequence length, if the unknown parameter of mean function and covariance function is Θ={ θi, i=1:N2, wherein N2For parameter Number, logarithm marginal likelihood function logP (Y | X, Θ) are as follows:
Wherein,Indicate normal function；X and Y indicates lateral direction of car position and lengthwise position；uYAnd ∑YFor corresponding variable Y Mean function and covariance function；
Maximum likelihood method is selected, solves to obtain the unknown parameter of Gaussian process by following formula
Parameter learning is obtained using the optimization algorithm based on gradient as a result, wherein to each component θ of unknown parameter ΘiAsk inclined It leads and is shown below:
Wherein γ=∑Y -1(Y-uY),
By the parameter learning of Gaussian process, the Gaussian process letter under different driving intentions and driving performance is established respectively Number.
Compared with prior art, the beneficial effects of the present invention are:
Track of vehicle prediction technique based on Mix-state DBN and Gaussian process of the invention, it is mixed by combining Close dynamic bayesian network (Mixture Dynamic Bayesian Network, abbreviation MDBN) and Gaussian process (Gaussian Process, abbreviation GP) realizes the track of vehicle prediction based on auto model fusion and data study, compared to Estimate currently based on vehicle physical model, based on driving intention, the three categories method based on big data study, the present invention passes through vehicle The parameter of natural driving data study MDBN and GP, merges multiple vehicle kinematics models using MDBN, obtains short-term track The estimated probability of prediction and driving intention and driving performance then carries out long-term trajectory predictions using GP and prediction is uncertain The expression of property.This method can either consider the short-term forecast characteristic under vehicle physical motion model, and vehicle drive can be considered Member's information carries out long-term trajectory predictions and uncertain expression, and compared to current track of vehicle prediction technique, the present invention is combined Auto model, it is abstract be intended to and data-driven, the scalability of MDBN and GP model is strong, can be applicable in different Driving Scenes, It can be in conjunction with more effective contextual informations, such as road information, traffic information.
Detailed description of the invention
Fig. 1 is that the present invention is based on the track of vehicle prediction technique frame diagrams of Mix-state DBN and Gaussian process；
Fig. 2 is Mix-state DBN of the present invention (MDBN) schematic diagram.
Specific embodiment
Invention is further explained with reference to the accompanying drawings and examples.
As shown in Figure 1, the track of vehicle prediction technique based on Mix-state DBN and Gaussian process, including it is as follows Step:
Step 1, building nature driving data library；
Establish automatic driving vehicle sensory perceptual system acquisition surrounding vehicles associated sequence information, road associated sequence information and The test set of traffic associated sequence information, and above- mentioned information are carried out with the training set of driving intention and driving performance calibration；Its In, the test set includes the test set of Mix-state DBN and the test set of Gaussian process；The training set includes The training set of Mix-state DBN and the training set of Gaussian process；
The surrounding vehicles associated sequence information includes vehicle location, car speed, vehicle acceleration, Vehicular yaw angle speed Degree, distance and left and right vehicle wheel of the vehicle away from both sides of the road turn to the opening and closing situation of taillight；
The road associated sequence information includes the structure feature (such as straight way and bend) and road instruction mark of road；
The traffic associated sequence information includes surrounding traffic Warning Mark and traffic signal light condition.
The acquisition of step 1.1, driving data
In order to train the function model in Mix-state DBN model and Gaussian process, need largely to train sample This, it is therefore desirable to acquisition surrounding vehicles associated sequence information of automatic driving vehicle sensory perceptual system, road in normal driving process Road associated sequence information and traffic associated sequence information；Wherein, automatic driving vehicle sensory perceptual system includes that sensor technology is (thunderous Reach, camera, navigator fix) and wireless communication technique (such as V2V, V2I), the data type that sensory perceptual system is handled include object Manage data and image data.For the needs of network model training after cooperating, seen according to Mix-state DBN MDBN The needs for surveying layer acquire and handle to obtain associated sequence information, including are handled to obtain vehicle location, vehicle speed by physical data Degree, vehicle acceleration, yaw rate, distance of the vehicle away from both sides of the road, obtain left and right vehicle wheel by image real time transfer It turns to taillight and is opened and closed situation, the structure feature of road, road instruction mark, surrounding traffic Warning Mark, traffic signal light condition.
Step 1.2, driving intention classification and calibration
The scene according to locating for vehicle defines the possible driving intention of surrounding vehicles；For each possible driving intention, The key characterization parameter that selection is classified for driving intention with calibration；It is related then in conjunction with the collected surrounding vehicles of step 1.1 Sequence information determines surrounding vehicles driving intention whithin a period of time and sets corresponding driving intention label.
Wherein, scene locating for the vehicle can be obtained by the intelligent guidance system of automatic driving vehicle, scene locating for vehicle There are traffic lights intersection scene, city including High-speed Circumstance, hill path scene, city normal straight-ahead operation road scene, city Without traffic lights intersection scene etc..
The driving intention includes along present road straight trip, along present road turning, Zuo Huandao, right lane-change, left-hand bend, the right side It turns, turn around, stopping and starting.
The key characterization parameter classified and demarcated for driving intention includes lateral direction of car position, lateral direction of car speed Degree, yaw rate, vehicle distance and left and right vehicle wheel away from both sides of the road turn to taillight and are opened and closed situation.
For example, surrounding predicts that the possible driving intention of vehicle has along current road if vehicle is under High-speed Circumstance Road straight trip is set to 1,2,3,4 along present road turning, left lane-change and right lane-change, these four types of driving intention labels.For edge The driving intention of present road straight trip, selects the distance of lateral direction of car position and vehicle away from both sides of the road to join as key feature Number；For the driving intention turned along present road, select the distance of yaw rate and vehicle away from both sides of the road as Key characterization parameter；For the driving intention of left lane-change and right lane-change, select lateral direction of car position and vehicle away from both sides of the road Distance is used as key characterization parameter.Assuming that the associated sequence information of certain vehicle whithin a period of time is around step 1.1 acquisition The distance of lateral direction of car position and vehicle away from both sides of the road remains unchanged, then can determine whether the driving of the vehicle during this period of time It is intended to keep straight on along present road, driving intention label is set as 1.
The classification and calibration of step 1.3, driving performance
Since the driving performance of the driver under same intention is different, cause the driving trace of vehicle that notable difference occurs. The possible driving performance of surrounding vehicles is defined first, selection is classified for driving performance and the key characterization parameter of judgement, according to The collected surrounding vehicles associated sequence information of step 1.1, surrounding vehicles the driving whithin a period of time obtained in conjunction with step 1.2 It sails intention, determine driving performance of the surrounding vehicles under a certain driving intention and sets corresponding driving performance label.
The driving performance includes slowly driving, smooth ride and rapid driving；
The key characterization parameter classified and determined for driving performance includes vehicular longitudinal velocity, lateral direction of car speed Degree, longitudinal acceleration of the vehicle, vehicle lateral acceleration and yaw rate.
For example, three classes driving performance label is set to 1,2,3 if being in the vehicle under High-speed Circumstance.For edge The driving intention of present road straight trip, selects vehicular longitudinal velocity, longitudinal acceleration of the vehicle as key characterization parameter；For edge The driving intention of present road turning, selects yaw rate as key characterization parameter；For left lane-change and right lane-change Driving intention, select vehicle lateral speed and vehicle lateral acceleration as key characterization parameter.Assuming that step 1.2 obtains week The driving intention for enclosing certain vehicle is to keep straight on along present road, the correlated series letter of the vehicle that step 1.1 acquires whithin a period of time Breath be vehicular longitudinal velocity be 60km/h or so, longitudinal acceleration of the vehicle fluctuates near zero, then can determine whether the vehicle on edge Driving performance under the driving intention of present road straight trip is slowly to drive, and driving performance label is set as 1.
The foundation of step 1.4, training set and test set
The training set is: the data set in order to carry out network model parameter learning preparation.By surrounding vehicles correlated series Information, road associated sequence information and traffic associated sequence information carry out the calibration of driving intention and driving performance.
The test set is: the data set prepared to verify the validity of the network model parameter after study.It will be all Vehicle associated sequence information, road associated sequence information and traffic associated sequence information are enclosed as the input of network model.
Surrounding vehicles associated sequence information, road associated sequence information and the traffic correlated series letter that step 1.1 is acquired Cease the test set as Mix-state DBN MDBN；
Take the leading portion moment of the lateral direction of car position and lengthwise position that correspond under a certain driving intention and driving performance A test set of the sequence information as Gaussian process；
Surrounding vehicles associated sequence information, road associated sequence information and the traffic associated sequence information that step 1.1 acquires Classify through step 1.2 driving intention with after calibration and the classification and calibration of 1.3 driving performances, obtains Mix-state DBN The training set of MDBN；
According to the training set of Mix-state DBN MDBN, will correspond under a certain driving intention and driving performance A training set of the full sequence information of lateral direction of car position and lengthwise position as Gaussian process GP.
Step 2, using Mix-state DBN, obtain short-term forecast track and the driving intention of surrounding vehicles With the estimated probability of driving performance.
It is selected using driver information, auto model and car status information is as hidden layer variable, with automatic driving vehicle sense Know the surrounding vehicles associated sequence information, road associated sequence information and traffic associated sequence information of system acquisition as observation layer Variable builds Mix-state DBN model；Obtain Mix-state DBN MDBN's by posterior probability reasoning Output: the estimated probability of short-term trajectory predictions and driving intention and driving performance based on auto model, and the output is made For the input of step 4.
Connection relationship between the observation layer variable, hidden layer variable and two variables, utilizes nature driving data and improvement Genetic algorithm obtains；The corresponding probability parameter of connection relationship is obtained by maximum likelihood method.
The auto model includes at the uniform velocity model, even acceleration model, at the uniform velocity steering model.
The car status information includes lateral direction of car position, vehicle lateral speed, vehicle lateral acceleration, longitudinal direction of car Position, vehicular longitudinal velocity, longitudinal acceleration of the vehicle and yaw rate.
Preferably due to the scalability of network node, if it is considered that other high-level semantics can add variable, it is described hidden Layer variable further comprises that other related higher-levels imply variable, and it includes vehicle and environment that other described related higher-levels, which imply variable, Reciprocation between risk assessment, vehicle.The observation layer variable further comprises hidden with other related higher-levels of hidden layer variable Containing other corresponding related bottom observation informations of variable.
The step 2 specifically comprises the following steps:
Step 2.1, the hidden layer for defining Mix-state DBN model and observation node layer
As shown in Fig. 2, Mix-state DBN is mainly by node and Bian Zucheng, node on behalf variable, side, which represents, to be become Dependence between amount, node are divided into hidden node and observation node layer, and the expression of network model is mainly according to locating for vehicle Traffic scene defines hidden layer and observes the quantity and meaning of node layer, is now defined as follows:
1. hidden node includes three-layer information:
First layer is driver information variables D, including driving intention and driving performance, embodies the field locating for a certain vehicle Under scape surrounding vehicles possibility driving intention (such as: the driving intention of High-speed Circumstance have along present road straight trip, along present road Turning, Zuo Huandao, right lane-change) and driving performance (such as slowly driving, smooth ride, rapid driving)；
The second layer is auto model selection variable M, for selecting the auto model under a certain observation sequence.Auto model Including at the uniform velocity model, acceleration model, at the uniform velocity steering model；
Third layer is car status information variable S, selects car status information required for auto model.Vehicle-state letter Breath includes lateral direction of car position, vehicle lateral speed, vehicle lateral acceleration, longitudinal direction of car position, vehicular longitudinal velocity, vehicle Longitudinal acceleration and yaw rate.As shown in Fig. 2, hidden layer variables collection H={ D, M, S }, wherein { D, M } is discrete Variable, { S } are continuous variable.
2. observing node layer mainly includes one layer of information:
Observation information includes the surrounding vehicles associated sequence information of automatic driving vehicle sensory perceptual system acquisition, road correlation sequence Column information and traffic associated sequence information, particular content are shown in step 1.Such as Fig. 2, observation layer variables collection is { O }, and is continuous Variable.
The expression-form of Mix-state DBN as shown in Fig. 2, it is the Bayesian network being unfolded in temporal sequence, Due to increasing car status information variable S, the existing discrete variable of hidden layer variable at this time has continuous variable again, therefore is known as mixing Dynamic bayesian network.
The optimization of step 2.2, network structure
The optimization of network structure is also referred to as the Structure learning of network model, is about the network structure under known network parameters Determine problem, i.e., given variables set { D, M, S, O } and training set T find a best match with by way of search learning The network N of this training set T=(B, θ), and use measurement of the criterion function as matching degree.Wherein, the structure of network Refer to the connection relationship between the network node of definition, B indicates network structure, and θ indicates the parameter of network.
As shown in Fig. 2, if there are certain probability passes between hidden layer variable, between hidden layer variable and observation layer variable System, then connected them with directed edge, otherwise two variables are mutually indepedent, therefore, constitute network by node and directed edge Structure；If this directed edge just imparts the conditional probability between variable, these conditional probabilities there are directed edge between variable It is exactly the parameter of network.The structure and parameter of network can learn to obtain according to data, can also be artificially defined.
It is most of since its calculation amount is relatively large for the structural optimization problems of Mix-state DBN MDBN Dynamic bayesian network structure is artificially defined according to practical problem, especially in the case where network node quantity is more.Here, by In the network node of definition, the strongly connected node in part can pre-define its connection relationship to reduce calculation amount, such as Fig. 2 institute Show, figure interior joint M and node S have High relevancy, can determine a directed edge.When a certain amount of node variable of definition And when giving a certain amount of data, network structure study, such as greedy algorithm mechanism can be carried out by heuritic approach Genetic algorithm, based on Probabilistic Models Evolutionary Algorithms etc..When there are enough data supportings, made using data learning network structure Network structure can preferably coincide with data.
Unknown parameter study between step 2.3, network node
Unknown parameter study between network node is also referred to as the parameter learning of network model, is about under known network structure Network parameter determine problem, wherein the parameter of network is exactly the conditional probability in network model.
For the parameter learning problem of Mix-state DBN MDBN, since the conditional probability of part hidden layer variable can To be obtained by the method for data statistics, to reduce certain calculation amount, the conditional probability between remaining hidden layer and observed quantity can Parameter learning is carried out with the Mix-state DBN MDBN training set obtained using step 1, uses maximum likelihood estimate (Maximum Likelihood Estimation, MLE) is solved, it is known that the likelihood under training sample D about network parameter θ Function L (θ)=logP (D | θ), optimal solution is solved using the optimization algorithm based on gradientIn order to avoid falling into local optimum, initial value can be randomly choosed repeatedly.
The Structure learning and parameter learning collectively form the learning Content of network model, and network model learning Content Theory needs the support of the reasoning content of next step network model, first explains the study of network model for convenience of explanation here Content.
2.4, the posterior probability reasoning under observation layer variable is introduced
The posterior probability reasoning introduced under observation layer variable is also referred to as the reasoning of network model.Mainly solve given network ginseng Posterior probability computational problem under several and structure, wherein posterior probability refers to the probability point of hidden layer variable under known observation layer variable Cloth P (H | O), i.e., in the time serial message of each observational variable known (the i.e. obtained Mix-state DBN MDBN of step 1 Test set) under, infer the process of corresponding driving intention, driving performance, auto model and vehicle-state.Inference method mainly divides For Accurate Reasoning and approximate resoning, Accurate Reasoning algorithm has forward backward algorithm (Forward-Backward Smoothing Algorithm), decompose tree algorithm (Unrolled Junction Tree Algorithm) etc., due to dynamic bayesian network Sequence ductility, NP hard problem is easy to cause using exact algorithm, therefore use suitable approximate algorithm, such as ADM (Moment Matching), importance sampling (Importance Sampling), Markov Monte Carlo (MCMC), particle Filter (Particle Filtering) etc..
For the MDBN frame in Fig. 2, a kind of hypothesis density filtering ADF (Assumed Density is provided here Filtering) approximate resoning process:
1. prediction: obtaining the joint probability distribution at t+1 moment with the probability distribution of t moment.If all discrete hidden layers of t moment Variable { Mt,DtJoint Distribution be
The then discrete hidden layer variable { M at t+1 momentt,Dt,Mt+1,Dt+1Joint Distribution are as follows:
T+1 moment continuous hidden layer variable { St+1Conditional probability are as follows:
2. updating: introducing the observational variable at t+1 moment to obtain new probability distribution.If the observational variable of auto model is { V }, and meetRemaining observational variable is { E } in { O }, then the discrete hidden layer variable { M at t+1 momentt,Dt, Mt+1,Dt+1Joint Distribution are as follows:
Wherein,
T+1 moment continuous hidden layer variable { St+1Conditional probability are as follows:
3. marginalisation: by the variable marginalisation before the t+1 moment in the t+1 moment probability distribution obtained from above, only retaining and work as The probability distribution of variable is inscribed when preceding t+1, for prediction next time and the iteration of update step:
Wherein, D indicates driver information variable, and driver information includes driving intention and driving performance；M indicates vehicle mould Type selects variable, and auto model includes at the uniform velocity model, acceleration model, at the uniform velocity steering model；S indicates car status information variable, Car status information includes that lateral direction of car position, vehicle lateral speed, vehicle lateral acceleration, longitudinal direction of car position, vehicle are vertical To speed, longitudinal acceleration of the vehicle and yaw rate.It utilizesThe available t+1 moment The estimated probability of auto model, driving intention and driving performance；It utilizesIt is available to be based on vehicle mould The short-term forecast track of type.
The meaning of each network node is defined by step 2.1；Step 2.2 has determined the connection relationship of network node；Step Rapid 2.3 have determined the parameter of network, to introduce the posterior probability reasoning under observational variable by step 2.4, obtain based on vehicle The estimated probability of the short-term forecast track of model and driving intention and driving performance, by itself and the Gaussian process phase in step 4 Fusion obtains final prediction locus.
Step 3 establishes Gaussian process function under different driving intentions and different driving performance.
Since the driving performance of the driver under same intention is different, the driving trace generation that will lead to vehicle is obvious poor It is different, therefore according to different driving intention and driving performance, the height under different driving intentions and different driving performances is established respectively This procedure function.
The error and surrounding traffic ring that uncertainty is mainly deposited by vehicle sensory perceptual system for track of vehicle prediction Caused by the dynamic and randomness in border, these uncertain factors may cause track of vehicle prediction result and deviation occur.
Establish the track of vehicle Gaussian process function of different driving performances under these intentions.Gaussian process mainly includes as follows Step:
The setting of step 3.1, mean function and covariance function
Using x as input, the expression formula of Gaussian process function is as follows:
F (x)~GP (u (x), Σ (x, x '))
Wherein mean function u (x) indicates the track of vehicle trend under certain driving intention, therefore can be with mean function For indicating expectation prediction locus；Covariance function Σ (x, x ') not only indicated the variance of different input x itself, but also illustrate (x, X ') between variance, therefore covariance function can be used to indicate the corresponding uncertainty of expectation prediction locus.
Mean function is driven there are many form, such as constant, linear function, various nonlinear functions etc. here according to difference Corresponding mean function u (x) is arranged in the track of vehicle sailed under being intended to, for example under High-speed Circumstance, keeps straight on along present road and drive meaning Mean function under figure can be set as constant, and the mean function under present road turning driving intention can be set as quadratic function, left Mean function can be set as quintic algebra curve form under lane-change or right lane-change driving intention.Covariance function is also referred to as kernel function, often There are many kernel functions, such as linear function, gaussian kernel function, symmetrical kernel function etc., and different kernel functions have different spies Property, consider that the flatness of vehicle driving trace is high, the square exponential form of Selection of kernel function here, in addition to this, due to considering It to the presence of uncertainty in traffic factor, needs to introduce noise variance, is set to white Gaussian noise here, therefore, band noise Covariance function Σ (x, the x ') setting that indicates of square exponential form it is as follows:
Wherein, σfPoor for signal standards, l is characterized length, σnFor observation noise standard deviation, δ is Kronecker Kronecker function.
The study of step 3.2, Gaussian process function unknown parameter
The mean function u (x) and covariance function Σ (x, x ') set according to step 3.1, the Gauss for utilizing step 1 to establish Process GP training set DGP(X, Y)={ (xi,yi), i=1:N1The unknown parameter being related to is learnt, wherein N1For training sample This sequence length, if the unknown parameter of mean function and covariance function is Θ={ θi, i=1:N2, wherein N2For parameter Number, logarithm marginal likelihood function logP (Y | X, Θ) are as follows:
Wherein,Indicate normal function；X and Y indicates lateral direction of car position and lengthwise position；uYAnd ∑YFor corresponding variable Y Mean function and covariance function；
Maximum likelihood method is selected, solves to obtain the unknown parameter of Gaussian process by following formula
Using the available parameter learning of optimization algorithm based on gradient as a result, wherein to each component θ of unknown parameter Θi Local derviation is asked to be shown below:
Wherein γ=∑Y -1(Y-uY), it is the operator for simplified expression.
By the parameter learning of Gaussian process, the Gaussian process under different driving intentions and driving performance can be established respectively Function.Since the algorithm may cause local optimum, it is therefore desirable to several groups of initial parameters be arranged more.
Step 4 carries out long-term trajectory predictions and its not really based on Mix-state DBN MDBN and Gaussian process GP Qualitative representation.
By Mix-state DBN MDBN obtain as a result, being suitable for the prediction in the short time, it is therefore desirable to tie It closes Gaussian process and carries out long-term trajectory predictions in short-term forecast result, while describing uncertainty in traffic.
Here, according to the estimation of the driving intention exported and driving performance of Mix-state DBN MDBN in step 2 Probability determines corresponding driving intention and driving performance using maximum probability principle, to according to the driving intention and drive special Property determine corresponding Gaussian process function in step 3, by step 2 Mix-state DBN MDBN export based on vehicle The short-term forecast track of model is as x1, the Future Trajectory of prediction is x2, then (x1,x2) obey Gaussian Profile it is as follows:
Wherein,Indicate normal function；x1With x2Corresponding mean value letter Number is respectively u1And u2；Covariance matrix Σ is symmetrical matrix, i.e. Σ=ΣT, x1With x2Corresponding covariance function difference For Σ11And Σ22, x1And x2Between corresponding covariance function be Σ12、Σ21, and Σ12=Σ21 T；
Then known vehicle track x1Under vehicle future may track x2Conditional probability P (x2|x1) obey Gaussian Profile Expression formula is as follows:
Wherein, u2|1And Σ2|1It is variable x2|1Corresponding mean function and covariance function, and (u2|1,Σ2|1) expression Formula is derived by by following formula:
u2|1=u2+∑12 T∑11 -1(x1-u1)
∑2|1=∑22-∑12 T∑11 -1∑12
It finally obtains in the long time domain of vehicle under the driving intention and driving performance determined using maximum probability principle not Coming track and uncertainty indicates.
Step 4 combines the transformation of step 2 and step 3 completion from short-term forecast to long-term forecast, while not true to predicting It is qualitative to be indicated.
The track of vehicle prediction technique based on Mix-state DBN and Gaussian process that the present invention designs, mainly has Three big stages:
First stage is definition, reasoning and the relevant data that Mix-state DBN is carried out according to Driving Scene Study, main function is to provide foundation for next stage.One section of time series data of network model observational variable is input to MDBN network model obtains the estimated probability of the driving intention and driving performance in surrounding vehicles this period and based on vehicle The short-term forecast track of model；
Second stage is that corresponding Gaussian process function, main function are established according to different driving intentions and driving performance It is to carry out the long-term trajectory predictions of vehicle and be indicated to uncertainty in traffic.According under a certain driving intention and driving performance Lateral direction of car position and lengthwise position Partial sequence information, calculate known portions track of vehicle sequence information under vehicle not Come possible track.
Phase III be according to the first stage output as a result, in conjunction with second stage content, complete determine driving intention With the expression of trajectory predictions and uncertainty in traffic in the long-term range of vehicle under driving performance.

Claims (10)

1. a kind of track of vehicle prediction technique based on Mix-state DBN and Gaussian process, it is characterised in that: the party Method includes the following steps:
Step 1, building nature driving data library；
Establish surrounding vehicles associated sequence information, road associated sequence information and the traffic of the acquisition of automatic driving vehicle sensory perceptual system The test set of associated sequence information, and above- mentioned information are carried out with the training set of driving intention and driving performance calibration；Wherein, institute Stating test set includes the test set of Mix-state DBN and the test set of Gaussian process；The training set includes that mixing is dynamic The training set of state Bayesian network and the training set of Gaussian process；
Step 2, using Mix-state DBN, obtain surrounding vehicles short-term forecast track and driving intention with drive Sail the estimated probability of characteristic；
It is selected using driver information, auto model and car status information is as hidden layer variable, system is perceived with automatic driving vehicle Surrounding vehicles associated sequence information, road associated sequence information and the traffic associated sequence information acquired of uniting becomes as observation layer Amount, builds Mix-state DBN model；The output of Mix-state DBN: base is obtained by posterior probability reasoning In the short-term trajectory predictions and driving intention of auto model and the estimated probability of driving performance, and using the output as step 4 Input；
Step 3 establishes Gaussian process function under different driving intentions and different driving performance；
It comprises the following steps:
The setting of step 3.1, mean function and covariance function
Using x as input, the expression formula of Gaussian process function is as follows:
F (x)~GP (u (x), Σ (x, x '))
Wherein, mean function u (x) indicates the track of vehicle trend under certain driving intention, therefore mean function can be used to table Show desired prediction locus；Covariance function Σ (x, x ') not only indicated the variance of different input x itself, but also illustrate (x, x ') between Variance, therefore covariance function can be used to indicate the corresponding uncertainty of expectation prediction locus；
Corresponding mean function u (x) is set according to the track of vehicle under different driving intentions；
Covariance function Σ that square exponential form with noise indicates (x, x ') setting is as follows:
Wherein, σfPoor for signal standards, l is characterized length, σnFor observation noise standard deviation, δ is Kronecker function；
The study of step 3.2, Gaussian process function unknown parameter
The mean function u (x) and covariance function Σ (x, x ') set according to step 3.1, the Gaussian process for utilizing step 1 to establish Training set the unknown parameter being related to is learnt, according to logarithm marginal likelihood function and its to the local derviation of each unknown parameter, Parameter learning is obtained using the optimization algorithm based on gradient as a result, to establish under different driving intentions and driving performance respectively Gaussian process function；
Step 4 carries out long-term trajectory predictions and uncertainty expression based on Mix-state DBN and Gaussian process；
The estimated probability of the driving intention and driving performance that are exported according to Mix-state DBN in step 2, utilizes maximum Principle of probability determines corresponding driving intention and driving performance, to be determined in step 3 according to the driving intention and driving performance Corresponding Gaussian process function；
The short-term forecast track based on auto model that Mix-state DBN in step 2 is exported is as track of vehicle sequence Column information x1, the Future Trajectory of prediction is x2, then (x1,x2) obey Gaussian Profile it is as follows:
Wherein,Indicate normal function；x1With x2Corresponding mean function point It Wei not u1And u2；Covariance matrix Σ is symmetrical matrix, i.e. Σ=ΣT, x1With x2Corresponding covariance function is respectively Σ11And Σ22, x1And x2Between corresponding covariance function be Σ12、Σ21, and ∑12=∑21 T；
Then known vehicle track x1Under vehicle future may track x2Conditional probability P (x2|x1) obey Gaussian Profile expression Formula is as follows:
Wherein, u2|1And Σ2|1It is variable x2|1Corresponding mean function and covariance function, and (u2|1,Σ2|1) expression formula by Following formula is derived by:
u2|1=u2+Σ12 TΣ11 -1(x1-u1)
Σ2|1=Σ22-Σ12 TΣ11 -1Σ12
Finally obtain the following rail in the long time domain of vehicle under the driving intention and driving performance determined using maximum probability principle Mark and uncertainty indicate.
2. the track of vehicle prediction technique according to claim 1 based on Mix-state DBN and Gaussian process, It is characterized by: in the step 1,
The surrounding vehicles associated sequence information includes vehicle location, car speed, vehicle acceleration, yaw rate, Distance and left and right vehicle wheel of the vehicle away from both sides of the road turn to the opening and closing situation of taillight；
The road associated sequence information includes the structure feature and road instruction mark of road；
The traffic associated sequence information includes surrounding traffic Warning Mark and traffic signal light condition.
3. the track of vehicle prediction technique according to claim 1 based on Mix-state DBN and Gaussian process, It is characterized by: the step 1 includes the following steps:
1.1, the acquisition of driving data
Acquisition surrounding vehicles associated sequence information of automatic driving vehicle sensory perceptual system, road correlation sequence in normal driving process Column information and traffic associated sequence information；
1.2, driving intention classification and calibration
The scene according to locating for vehicle defines the possible driving intention of surrounding vehicles；For each possible driving intention, selection The key characterization parameter classified for driving intention and demarcated；Then in conjunction with the collected surrounding vehicles correlated series of step 1.1 Information determines surrounding vehicles driving intention whithin a period of time and sets corresponding driving intention label；
Wherein, scene locating for the vehicle is obtained by the intelligent guidance system of automatic driving vehicle, and scene locating for vehicle includes height Fast scene, hill path scene, city normal straight-ahead operation road scene, city have traffic lights intersection scene, city without traffic Signal lamp intersection scene；
The driving intention includes along present road straight trip, along present road turning, Zuo Huandao, right lane-change, left-hand bend, right-hand rotation It is curved, turn around, stop with starting；
The key characterization parameter classified and demarcated for driving intention includes lateral direction of car position, vehicle lateral speed, vehicle Yaw velocity, vehicle distance and left and right vehicle wheel away from both sides of the road turn to taillight and are opened and closed situation；
1.3, the classification and calibration of driving performance
The possible driving performance of surrounding vehicles is defined first, selection is classified for driving performance and the key characterization parameter of judgement, According to the collected surrounding vehicles associated sequence information of step 1.1, the surrounding vehicles obtained in conjunction with step 1.2 are whithin a period of time Driving intention, determine driving performance of the surrounding vehicles under a certain driving intention and set corresponding driving performance label；
The driving performance includes slowly driving, smooth ride and rapid driving；
The key characterization parameter classified and determined for driving performance includes vehicular longitudinal velocity, vehicle lateral speed, vehicle Longitudinal acceleration, vehicle lateral acceleration and yaw rate；
1.4, the foundation of training set and test set
Surrounding vehicles associated sequence information, road associated sequence information and the traffic associated sequence information that step 1.1 is acquired are made For the test set of Mix-state DBN MDBN；
Take the sequence at the leading portion moment of the lateral direction of car position and lengthwise position that correspond under a certain driving intention and driving performance A test set of the information as Gaussian process；
Surrounding vehicles associated sequence information, road associated sequence information and the traffic associated sequence information that step 1.1 acquires are through step Rapid 1.2 driving intention classification obtains Mix-state DBN MDBN with after calibration and the classification and calibration of 1.3 driving performances Training set；
According to the training set of Mix-state DBN MDBN, the vehicle under a certain driving intention and driving performance will be corresponded to A training set of the full sequence information of lateral position and lengthwise position as Gaussian process GP.
4. the track of vehicle prediction technique according to claim 1 based on Mix-state DBN and Gaussian process, It is characterized by: in the step 2,
Connection relationship between the observation layer variable, hidden layer variable and two variables using nature driving data and improves heredity Algorithm obtains；The corresponding probability parameter of connection relationship is obtained by maximum likelihood method.
5. the track of vehicle prediction technique according to claim 1 based on Mix-state DBN and Gaussian process, It is characterized by: in the step 2,
The auto model includes at the uniform velocity model, even acceleration model, at the uniform velocity steering model；
The car status information includes lateral direction of car position, vehicle lateral speed, vehicle lateral acceleration, longitudinal direction of car position It sets, vehicular longitudinal velocity, longitudinal acceleration of the vehicle and yaw rate.
6. the track of vehicle prediction technique according to claim 1 based on Mix-state DBN and Gaussian process, It is characterized by: in the step 2,
The hidden layer variable further comprises that other related higher-levels imply variable, and it includes vehicle that other described related higher-levels, which imply variable, Reciprocation between the risk assessment of environment, vehicle；
The observation layer variable further comprises implying variable to other related higher-levels of hidden layer variable corresponding other being related Bottom observation information.
7. the track of vehicle prediction technique according to claim 1 based on Mix-state DBN and Gaussian process, It is characterized by: the step 2 specifically comprises the following steps:
2.1, the hidden layer and observation node layer of Mix-state DBN model are defined
Hidden node includes three-layer information: first layer is driver information variables D, and the second layer is auto model selection variable M, the Three layers are car status information variable S；Hidden layer variables collection H={ D, M, S }, wherein { D, M } is discrete variable, { S } is continuous Variable；
Observing node layer mainly includes one layer of information: the surrounding vehicles correlated series letter of automatic driving vehicle sensory perceptual system acquisition Breath, road associated sequence information and traffic associated sequence information；Observation layer variables collection is { O }, and is continuous variable；
2.2, the optimization of network structure
Given variables set { D, M, S, O } and training set T, learn with search for by way of find best match this training set The network N of T=(B, θ), and use measurement of the criterion function as matching degree；Wherein, B indicates network structure, and θ is indicated The parameter of network；
2.3, the unknown parameter study between network node
The conditional probability of part hidden layer variable is obtained by the method for data statistics, the conditional probability between remaining hidden layer and observed quantity The training set of the Mix-state DBN MDBN obtained using step 1 carries out parameter learning, uses maximum likelihood estimate To solve, it is known that about likelihood function L (θ)=logP of network parameter θ (D | θ) under training sample D, using based on the excellent of gradient Change algorithm and solves optimal solution
2.4, the posterior probability reasoning under observation layer variable is introduced
Posterior probability refers to the probability distribution P (H | O) of hidden layer variable under known observation layer variable, i.e., dynamic in the mixing that step 1 obtains Under the test set of state Bayesian network MDBN, the mistake of corresponding driving intention, driving performance, auto model and vehicle-state is inferred Journey obtains the estimated probability of short-term trajectory predictions and driving intention and driving performance based on auto model.
8. the track of vehicle prediction technique according to claim 1 based on Mix-state DBN and Gaussian process, It is characterized by: the posterior probability reasoning is a kind of hypothesis density filtering approximate resoning, and detailed process is such as in the step 2 Under:
1) it predicts: obtaining the joint probability distribution at t+1 moment with the probability distribution of t moment；If all discrete hidden layer variables of t moment {Mt,DtJoint Distribution be
The then discrete hidden layer variable { M at t+1 momentt,Dt,Mt+1,Dt+1Joint Distribution are as follows:
T+1 moment continuous hidden layer variable { St+1Conditional probability are as follows:
2) it updates: introducing the observational variable at t+1 moment to obtain new probability distribution；If the observational variable of auto model is { V }, And meetRemaining observational variable is { E } in { O }, then the discrete hidden layer variable { M at t+1 momentt,Dt,Mt+1, Dt+1Joint Distribution are as follows:
Wherein,
T+1 moment continuous hidden layer variable { St+1Conditional probability are as follows:
3) marginalisation: by the variable marginalisation before the t+1 moment in the t+1 moment probability distribution obtained from above, only retain current t+1 When inscribe the probability distribution of variable, for prediction next time and update the iteration of step:
Wherein, D indicates driver information variable, and driver information includes driving intention and driving performance；M indicates auto model choosing Variable is selected, auto model includes at the uniform velocity model, acceleration model, at the uniform velocity steering model；S indicates car status information variable, vehicle Status information includes lateral direction of car position, vehicle lateral speed, vehicle lateral acceleration, longitudinal direction of car position, longitudinal direction of car speed Degree, longitudinal acceleration of the vehicle and yaw rate；It utilizesObtain t+1 moment auto model, The estimated probability of driving intention and driving performance；It utilizesObtain the short-term forecast rail based on auto model Mark.
9. the track of vehicle prediction technique according to claim 1 based on Mix-state DBN and Gaussian process, It is characterized by: selecting maximum likelihood method to carry out the unknown parameter study of Gaussian process in the step 3.2.
10. the track of vehicle prediction technique according to claim 1 based on Mix-state DBN and Gaussian process, It is characterized by: the detailed process of the step 3.2 are as follows:
The mean function u (x) and covariance function Σ (x, x ') set according to step 3.1, the Gaussian process for utilizing step 1 to establish GP training set DGP(X, Y)={ (xi,yi), i=1:N1The unknown parameter being related to is learnt, wherein N1For training sample Sequence length, if the unknown parameter of mean function and covariance function is Θ={ θi, i=1:N2, wherein N2For number of parameters, Its logarithm marginal likelihood function logP (Y | X, Θ) are as follows:
Wherein,Indicate normal function；X and Y indicates lateral direction of car position and lengthwise position；uYAnd ∑YFor the equal of corresponding variable Y Value function and covariance function；
Maximum likelihood method is selected, solves to obtain the unknown parameter of Gaussian process by following formula
Parameter learning is obtained using the optimization algorithm based on gradient as a result, wherein to each component θ of unknown parameter ΘiSeek local derviation such as Shown in following formula:
Wherein γ=∑Y -1(Y-uY),
By the parameter learning of Gaussian process, the Gaussian process function under different driving intentions and driving performance is established respectively.
CN201910598776.0A 2019-07-04 2019-07-04 Vehicle track prediction method based on hybrid dynamic Bayesian network and Gaussian process Active CN110304075B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910598776.0A CN110304075B (en) 2019-07-04 2019-07-04 Vehicle track prediction method based on hybrid dynamic Bayesian network and Gaussian process

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910598776.0A CN110304075B (en) 2019-07-04 2019-07-04 Vehicle track prediction method based on hybrid dynamic Bayesian network and Gaussian process

Publications (2)

Publication Number Publication Date
CN110304075A true CN110304075A (en) 2019-10-08
CN110304075B CN110304075B (en) 2020-06-26

Family

ID=68078121

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910598776.0A Active CN110304075B (en) 2019-07-04 2019-07-04 Vehicle track prediction method based on hybrid dynamic Bayesian network and Gaussian process

Country Status (1)

Country Link
CN (1) CN110304075B (en)

Cited By (35)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110675632A (en) * 2019-11-11 2020-01-10 重庆邮电大学 Vehicle short-time trajectory prediction control method aiming at multi-feature space and data sparseness
CN110834644A (en) * 2019-10-30 2020-02-25 中国第一汽车股份有限公司 Vehicle control method and device, vehicle to be controlled and storage medium
CN110986994A (en) * 2019-11-14 2020-04-10 苏州智加科技有限公司 Automatic lane change intention marking method based on high-noise vehicle track data
CN111046919A (en) * 2019-11-21 2020-04-21 南京航空航天大学 Peripheral dynamic vehicle track prediction system and method integrating behavior intents
CN111353644A (en) * 2020-02-27 2020-06-30 成都美云智享智能科技有限公司 Prediction model generation method of intelligent network cloud platform based on reinforcement learning
CN111452022A (en) * 2020-03-24 2020-07-28 东南大学 Bayesian optimization-based upper limb rehabilitation robot active training reference track complexity adjusting method
CN111681335A (en) * 2020-03-18 2020-09-18 桂林电子科技大学 Automobile track prediction system based on LSTM technology and prediction method thereof
CN111898211A (en) * 2020-08-07 2020-11-06 吉林大学 Intelligent vehicle speed decision method based on deep reinforcement learning and simulation method thereof
CN111930110A (en) * 2020-06-01 2020-11-13 西安理工大学 Intent track prediction method for generating confrontation network by combining society
CN112015842A (en) * 2020-09-02 2020-12-01 中国科学技术大学 Bicycle track prediction automatic driving vehicle risk assessment method and system
CN112036001A (en) * 2020-07-01 2020-12-04 长安大学 Automatic driving test scene construction method, device and equipment and readable storage medium
CN112070808A (en) * 2020-09-01 2020-12-11 三一专用汽车有限责任公司 Trajectory prediction method, apparatus and computer-readable storage medium
CN112258841A (en) * 2020-10-26 2021-01-22 大连大学 Intelligent vehicle risk assessment method based on vehicle track prediction
CN112308171A (en) * 2020-11-23 2021-02-02 浙江天行健智能科技有限公司 Vehicle position prediction modeling method based on simulated driver
CN112365140A (en) * 2020-10-30 2021-02-12 东南大学 Bayesian network-based method for diagnosing influence factors of air-rail link bottleneck of user
CN112364997A (en) * 2020-12-08 2021-02-12 北京三快在线科技有限公司 Method and device for predicting track of obstacle
CN112614335A (en) * 2020-11-17 2021-04-06 南京师范大学 Traffic flow characteristic modal decomposition method based on generation-filtering mechanism
CN112800670A (en) * 2021-01-26 2021-05-14 清华大学 Multi-target structure optimization method and device for driving cognitive model
CN112829744A (en) * 2021-02-09 2021-05-25 清华大学 Vehicle long time domain track prediction method based on longitudinal and transverse coupling
CN112885079A (en) * 2021-01-11 2021-06-01 成都语动未来科技有限公司 Vehicle track prediction method based on global attention and state sharing
CN113525406A (en) * 2020-04-15 2021-10-22 百度（美国）有限责任公司 Bayesian global optimization based parameter tuning for vehicle motion controllers
WO2021213366A1 (en) * 2020-04-23 2021-10-28 华为技术有限公司 Method for optimizing decision-making regulation and control, method for controlling vehicle traveling, and related devices
CN113643542A (en) * 2021-10-13 2021-11-12 北京理工大学 Vehicle track prediction method and system under multi-vehicle interaction environment based on ensemble learning
CN113682302A (en) * 2021-08-03 2021-11-23 中汽创智科技有限公司 Driving state estimation method and device, electronic equipment and storage medium
CN113954830A (en) * 2021-11-10 2022-01-21 江苏大学 Vehicle self-adaptive cruise system and side vehicle merging line identification method thereof
WO2022021661A1 (en) * 2020-07-27 2022-02-03 深圳大学 Gaussian process-based visual positioning method, system, and storage medium
CN114194213A (en) * 2021-12-29 2022-03-18 北京三快在线科技有限公司 Target object trajectory prediction method and device, storage medium and electronic equipment
CN114212105A (en) * 2021-12-16 2022-03-22 中国人民解放军国防科技大学 Interactive vehicle driving intention prediction method and device with high generalization capability
CN114655231A (en) * 2022-04-22 2022-06-24 广东机电职业技术学院 Truck standard driving assistance method and system
CN114715145A (en) * 2022-04-29 2022-07-08 阿波罗智能技术(北京)有限公司 Trajectory prediction method, device and equipment and automatic driving vehicle
CN115755606A (en) * 2022-11-16 2023-03-07 上海友道智途科技有限公司 Carrier controller automatic optimization method, medium and equipment based on Bayesian optimization
CN116257797A (en) * 2022-12-08 2023-06-13 江苏中路交通发展有限公司 Single trip track identification method of motor vehicle based on Gaussian mixture model
CN116975574A (en) * 2023-08-31 2023-10-31 国家海洋环境监测中心 Marine environment heavy metal pollution evaluation method
CN117077042A (en) * 2023-10-17 2023-11-17 北京鑫贝诚科技有限公司 Rural level crossing safety early warning method and system
CN112800670B (en) * 2021-01-26 2024-05-03 清华大学 Multi-target structure optimization method and device for driving cognitive model

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130179382A1 (en) * 2012-01-11 2013-07-11 Honda Research Institute Europe Gmbh Vehicle with computing means for monitoring and predicting traffic participant objects
CN103345577A (en) * 2013-06-27 2013-10-09 江南大学 Probability hypothesis density multi-target tracking method based on variational Bayesian approximation technology
CN106952471A (en) * 2016-01-06 2017-07-14 通用汽车环球科技运作有限责任公司 The prediction of driver intention at intersection
CN107298100A (en) * 2017-05-16 2017-10-27 开易（北京）科技有限公司 A kind of track of vehicle Forecasting Methodology, system based on gauss hybrid models
US20180079422A1 (en) * 2017-11-27 2018-03-22 GM Global Technology Operations LLC Active traffic participant
EP3477333A1 (en) * 2017-10-26 2019-05-01 Continental Automotive GmbH Method and device of determining kinematics of a target

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130179382A1 (en) * 2012-01-11 2013-07-11 Honda Research Institute Europe Gmbh Vehicle with computing means for monitoring and predicting traffic participant objects
CN103345577A (en) * 2013-06-27 2013-10-09 江南大学 Probability hypothesis density multi-target tracking method based on variational Bayesian approximation technology
CN106952471A (en) * 2016-01-06 2017-07-14 通用汽车环球科技运作有限责任公司 The prediction of driver intention at intersection
CN107298100A (en) * 2017-05-16 2017-10-27 开易（北京）科技有限公司 A kind of track of vehicle Forecasting Methodology, system based on gauss hybrid models
EP3477333A1 (en) * 2017-10-26 2019-05-01 Continental Automotive GmbH Method and device of determining kinematics of a target
US20180079422A1 (en) * 2017-11-27 2018-03-22 GM Global Technology Operations LLC Active traffic participant

Cited By (47)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110834644A (en) * 2019-10-30 2020-02-25 中国第一汽车股份有限公司 Vehicle control method and device, vehicle to be controlled and storage medium
CN110675632A (en) * 2019-11-11 2020-01-10 重庆邮电大学 Vehicle short-time trajectory prediction control method aiming at multi-feature space and data sparseness
CN110986994A (en) * 2019-11-14 2020-04-10 苏州智加科技有限公司 Automatic lane change intention marking method based on high-noise vehicle track data
CN110986994B (en) * 2019-11-14 2021-08-03 苏州智加科技有限公司 Automatic lane change intention marking method based on high-noise vehicle track data
CN111046919A (en) * 2019-11-21 2020-04-21 南京航空航天大学 Peripheral dynamic vehicle track prediction system and method integrating behavior intents
CN111353644B (en) * 2020-02-27 2023-04-07 成都美云智享智能科技有限公司 Prediction model generation method of intelligent network cloud platform based on reinforcement learning
CN111353644A (en) * 2020-02-27 2020-06-30 成都美云智享智能科技有限公司 Prediction model generation method of intelligent network cloud platform based on reinforcement learning
CN111681335A (en) * 2020-03-18 2020-09-18 桂林电子科技大学 Automobile track prediction system based on LSTM technology and prediction method thereof
CN111452022A (en) * 2020-03-24 2020-07-28 东南大学 Bayesian optimization-based upper limb rehabilitation robot active training reference track complexity adjusting method
CN111452022B (en) * 2020-03-24 2021-04-13 东南大学 Bayesian optimization-based upper limb rehabilitation robot active training reference track complexity adjusting method
CN113525406A (en) * 2020-04-15 2021-10-22 百度（美国）有限责任公司 Bayesian global optimization based parameter tuning for vehicle motion controllers
WO2021213366A1 (en) * 2020-04-23 2021-10-28 华为技术有限公司 Method for optimizing decision-making regulation and control, method for controlling vehicle traveling, and related devices
CN111930110A (en) * 2020-06-01 2020-11-13 西安理工大学 Intent track prediction method for generating confrontation network by combining society
CN112036001B (en) * 2020-07-01 2024-04-23 长安大学 Automatic driving test scene construction method, device, equipment and readable storage medium
CN112036001A (en) * 2020-07-01 2020-12-04 长安大学 Automatic driving test scene construction method, device and equipment and readable storage medium
US20220168900A1 (en) * 2020-07-27 2022-06-02 Shenzhen University Visual positioning method and system based on gaussian process, and storage medium
WO2022021661A1 (en) * 2020-07-27 2022-02-03 深圳大学 Gaussian process-based visual positioning method, system, and storage medium
CN111898211A (en) * 2020-08-07 2020-11-06 吉林大学 Intelligent vehicle speed decision method based on deep reinforcement learning and simulation method thereof
CN112070808A (en) * 2020-09-01 2020-12-11 三一专用汽车有限责任公司 Trajectory prediction method, apparatus and computer-readable storage medium
CN112015842B (en) * 2020-09-02 2024-02-27 中国科学技术大学 Automatic driving vehicle risk assessment method and system for bicycle track prediction
CN112015842A (en) * 2020-09-02 2020-12-01 中国科学技术大学 Bicycle track prediction automatic driving vehicle risk assessment method and system
CN112258841B (en) * 2020-10-26 2022-08-02 大连大学 Intelligent vehicle risk assessment method based on vehicle track prediction
CN112258841A (en) * 2020-10-26 2021-01-22 大连大学 Intelligent vehicle risk assessment method based on vehicle track prediction
CN112365140A (en) * 2020-10-30 2021-02-12 东南大学 Bayesian network-based method for diagnosing influence factors of air-rail link bottleneck of user
CN112614335A (en) * 2020-11-17 2021-04-06 南京师范大学 Traffic flow characteristic modal decomposition method based on generation-filtering mechanism
CN112614335B (en) * 2020-11-17 2021-12-07 南京师范大学 Traffic flow characteristic modal decomposition method based on generation-filtering mechanism
CN112308171A (en) * 2020-11-23 2021-02-02 浙江天行健智能科技有限公司 Vehicle position prediction modeling method based on simulated driver
CN112364997A (en) * 2020-12-08 2021-02-12 北京三快在线科技有限公司 Method and device for predicting track of obstacle
CN112885079A (en) * 2021-01-11 2021-06-01 成都语动未来科技有限公司 Vehicle track prediction method based on global attention and state sharing
CN112800670A (en) * 2021-01-26 2021-05-14 清华大学 Multi-target structure optimization method and device for driving cognitive model
CN112800670B (en) * 2021-01-26 2024-05-03 清华大学 Multi-target structure optimization method and device for driving cognitive model
CN112829744A (en) * 2021-02-09 2021-05-25 清华大学 Vehicle long time domain track prediction method based on longitudinal and transverse coupling
CN113682302A (en) * 2021-08-03 2021-11-23 中汽创智科技有限公司 Driving state estimation method and device, electronic equipment and storage medium
CN113643542A (en) * 2021-10-13 2021-11-12 北京理工大学 Vehicle track prediction method and system under multi-vehicle interaction environment based on ensemble learning
CN113954830A (en) * 2021-11-10 2022-01-21 江苏大学 Vehicle self-adaptive cruise system and side vehicle merging line identification method thereof
CN114212105A (en) * 2021-12-16 2022-03-22 中国人民解放军国防科技大学 Interactive vehicle driving intention prediction method and device with high generalization capability
CN114212105B (en) * 2021-12-16 2024-03-05 中国人民解放军国防科技大学 Interactive vehicle driving intention prediction method and device with high generalization capability
CN114194213A (en) * 2021-12-29 2022-03-18 北京三快在线科技有限公司 Target object trajectory prediction method and device, storage medium and electronic equipment
CN114655231A (en) * 2022-04-22 2022-06-24 广东机电职业技术学院 Truck standard driving assistance method and system
CN114715145A (en) * 2022-04-29 2022-07-08 阿波罗智能技术(北京)有限公司 Trajectory prediction method, device and equipment and automatic driving vehicle
CN114715145B (en) * 2022-04-29 2023-03-17 阿波罗智能技术(北京)有限公司 Trajectory prediction method, device and equipment and automatic driving vehicle
CN115755606A (en) * 2022-11-16 2023-03-07 上海友道智途科技有限公司 Carrier controller automatic optimization method, medium and equipment based on Bayesian optimization
CN116257797A (en) * 2022-12-08 2023-06-13 江苏中路交通发展有限公司 Single trip track identification method of motor vehicle based on Gaussian mixture model
CN116975574A (en) * 2023-08-31 2023-10-31 国家海洋环境监测中心 Marine environment heavy metal pollution evaluation method
CN116975574B (en) * 2023-08-31 2024-04-16 国家海洋环境监测中心 Marine environment heavy metal pollution evaluation method
CN117077042A (en) * 2023-10-17 2023-11-17 北京鑫贝诚科技有限公司 Rural level crossing safety early warning method and system
CN117077042B (en) * 2023-10-17 2024-01-09 北京鑫贝诚科技有限公司 Rural level crossing safety early warning method and system

Also Published As

Publication number Publication date
CN110304075B (en) 2020-06-26

Similar Documents

Publication Publication Date Title
CN110304075A (en) Track of vehicle prediction technique based on Mix-state DBN and Gaussian process
Yu et al. Deep learning-based traffic safety solution for a mixture of autonomous and manual vehicles in a 5G-enabled intelligent transportation system
CN110796856B (en) Vehicle lane change intention prediction method and training method of lane change intention prediction network
Lee et al. Convolution neural network-based lane change intention prediction of surrounding vehicles for ACC
JP6797254B2 (en) Interaction recognition decision making
Miyajima et al. Driver-behavior modeling using on-road driving data: A new application for behavior signal processing
CN112347567A (en) Vehicle intention and track prediction method
CN109460023A (en) Driver's lane-changing intention recognition methods based on Hidden Markov Model
CN107310550A (en) Road vehicles travel control method and device
CN109726804A (en) A kind of intelligent vehicle driving behavior based on driving prediction field and BP neural network personalizes decision-making technique
CN115056798B (en) Automatic driving vehicle lane change behavior vehicle-road collaborative decision algorithm based on Bayesian game
CN116134292A (en) Tool for performance testing and/or training an autonomous vehicle planner
CN110320916A (en) Consider the autonomous driving vehicle method for planning track and system of occupant's impression
CN113901718A (en) Deep reinforcement learning-based driving collision avoidance optimization method in following state
Wilhelem et al. Energy consumption evaluation based on a personalized driver–vehicle model
CN114516336B (en) Vehicle track prediction method considering road constraint conditions
Chen et al. Platoon separation strategy optimization method based on deep cognition of a driver’s behavior at signalized intersections
CN114954498A (en) Reinforced learning lane change behavior planning method and system based on simulated learning initialization
CN114495486A (en) Microscopic traffic flow prediction system and method based on hierarchical reinforcement learning
Hu et al. A safe driving decision-making methodology based on cascade imitation learning network for automated commercial vehicles
Gadepally Estimation of driver behavior for autonomous vehicle applications
Gutiérrez-Moreno et al. Hybrid decision making for autonomous driving in complex urban scenarios
Islam et al. Enhancing Longitudinal Velocity Control With Attention Mechanism-Based Deep Deterministic Policy Gradient (DDPG) for Safety and Comfort
Yuan et al. Driving Intention Recognition and Speed Prediction at Complex Urban Intersections Considering Traffic Environment
Raffone et al. A Humanized Vehicle Speed Control to Improve the Acceptance of Automated Longitudinal Control

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant