Summary of the invention
One of them that the invention solves the problems that technical problem is that, the most comprehensively reflects the driving environment of vehicle,
Realization personalizes automatic Pilot.
For achieving the above object, the present invention provides a kind of automatic driving method, including: according to collection
Environment sensing information and operator information set up Vehicular automatic driving data base, and by described vehicle from
Dynamic driving data storehouse is divided into training set and test set;Set up according to the environment sensing information in described training set
The running environment risk field of training, according in the running environment risk field of described training and described training set
Operator information described degree of deep learning model is trained;According to the environment sense in described test set
Know that information sets up the running environment risk field of test, the running environment risk field of described test is inputted institute
State degree of deep learning model, the wagon control variable of output prediction, by the wagon control variable of comparison prediction and
Described degree of deep learning model is tested by the operator information in described test set.
Wherein, training is used or the running environment risk field of test uses following methods to set up: according to quiet
The behavior that potential energy field information, the kinetic energy field information of moving object formation and the driver that only object is formed is formed
Field information sets up running environment risk field;Wherein,
For the running environment risk field of training, the potential energy field information of stationary object formation and moving object
The kinetic energy field information formed determines according to the environment sensing information in described training set, the behavior that driver is formed
Field information determines according to the operator information in described training set;
For the running environment risk field of test, the potential energy field information of stationary object formation and moving object
The kinetic energy field information formed determines according to the environment sensing information in described test set, the behavior that driver is formed
Field information determines according to the operator information in described test set.
In one embodiment, the potential energy field information that stationary object is formed is according to the attribute of stationary object and road
Travel permit part determines;The kinetic energy field information that moving object is formed is according to attribute, kinestate and the road of moving object
Travel permit part determines.
In the case of environment sensing information is obtained by multiple sensor acquisition, the method also includes:
The coordinate system of multiple sensors is changed, to form unified coordinate system;
And/or
Use the same target of the mahalanobis distance association different sensors observation of target, different sensors is seen
The same target surveyed is weighted averagely by probability of happening, as the probability of happening of this same target.
In one embodiment, according in the running environment risk field of described training and described training set
Described degree of deep learning model is trained including by operator information: by the running environment of described training
Operator information input degree of deep learning model in risk field and described training set, the vehicle of output prediction
The loss information of control variable and operator information;Loss information according to described operator information
Revise the parameter of wagon control variable in described degree of deep learning model.
For achieving the above object, the present invention provides a kind of automatic driving device, including: sample is formed
Module, for setting up Vehicular automatic driving data according to the environment sensing information gathered and operator information
Storehouse, and described Vehicular automatic driving data base is divided into training set and test set;Model training module, uses
In setting up the running environment risk field trained according to the environment sensing information in described training set, according to described
The described degree of depth is learnt by the operator information in the running environment risk field of training and described training set
Model is trained;Model measurement module, surveys for setting up according to the environment sensing information in described test set
Running environment risk field on probation, inputs described degree of depth study mould by the running environment risk field of described test
Type, the wagon control variable of output prediction, by the wagon control variable of comparison prediction and described test set
Operator information described degree of deep learning model is tested.
Described model training module includes that unit is set up in the first risk field, for formed according to stationary object
Kinetic energy field information and the behavior field information of driver's formation that potential energy field information, moving object are formed set up instruction
The running environment risk field practiced;Wherein, stationary object formed potential energy field information and moving object formed
Kinetic energy field information determines according to the environment sensing information in described training set, the behavior field information that driver is formed
Determine according to the operator information in described training set;
Described model measurement module includes that unit is set up in the second risk field, for formed according to stationary object
Kinetic energy field information and the behavior field information of driver's formation that potential energy field information, moving object are formed are set up and are surveyed
Running environment risk field on probation;Wherein, stationary object formed potential energy field information and moving object formed
Kinetic energy field information determines according to the environment sensing information in described test set, the behavior field information that driver is formed
Determine according to the operator information in described test set.
Wherein, the potential energy field information that stationary object is formed determines according to attribute and the road conditions of stationary object;
The kinetic energy field information that moving object is formed determines according to attribute, kinestate and the road conditions of moving object.
In the case of environment sensing information is obtained by multiple sensor acquisition, described sample forms module bag
Include: data processing unit and sample form unit;
Described data processing unit, is used for
The coordinate system of multiple sensors is changed, to form unified coordinate system;
And/or
Use the same target of the mahalanobis distance association different sensors observation of target, different sensors is seen
The same target surveyed is weighted averagely by probability of happening, as the probability of happening of this same target;
Described sample forms unit, for building according to the environment sensing information gathered and operator information
Vertical Vehicular automatic driving data base, and described Vehicular automatic driving data base is divided into training set and test
Collection.
Described model training module includes model training unit, for by the running environment wind of described training
Operator information input degree of deep learning model in field, danger and described training set, the vehicle control of output prediction
The loss information of variable processed and operator information;Loss information according to described operator information is repaiied
The parameter of wagon control variable in the most described degree of deep learning model.
Present invention theory based on field sets up the risk model of vehicle running environment, thus utilizes running environment
The driving environment of vehicle is comprehensively reflected in risk field, is advantageously implemented the automatic Pilot under different road environment.
And according to running environment risk field and operator, Vehicular automatic driving model is trained, with study
The experience of outstanding human driver, it is achieved personalize automatic Pilot.
Detailed description of the invention
The present invention proposes a kind of automatic driving method, and the method utilizes the environment sensing information gathered
Set up running environment risk field, train degree of deep learning model according to running environment risk field and operator,
The automatic Pilot of vehicle can be realized, reduce the training difficulty of Vehicular automatic driving model (abbreviation model).
Fig. 1 is the schematic flow sheet of one embodiment of automatic driving method of the present invention.As it is shown in figure 1,
The method comprises the following steps:
Step S102, environment sensing information and operator information according to gathering are set up vehicle and are automatically driven
Sail data base, and Vehicular automatic driving data base is divided into training set and test set, to form sample.Instruction
Practice collection and be used for training pattern, use in the model training stage;Test set is used for verifying the availability of model,
The model measurement stage uses.
Wherein, environment sensing information is by the environmental data of at least one sensor acquisition.Such as, vehicle-mounted
The image of camera acquisition, the some cloud information of laser radar and the target information etc. of millimetre-wave radar, but do not limit
In examples cited.
Wherein, operator information includes the information such as Vehicular turn angle, vehicle plus/minus speed.In order to obtain
Take rich and varied driving data, the driving data of different driver can be selected.Such as, the data of use
Storehouse sample frequency is 10 hertz, selects different driver's driving data of two hours as training set, amounts to
72000 frames, select the different halfhour driving data of driver as test set, 18000 frame altogether.Logical
Cross the driving behavior of the different driver of study, it is possible to achieve the automatic Pilot that personalizes of vehicle.
Step S104, sets up the running environment risk field of training according to the environment sensing information in training set,
According to the operator information in the running environment risk field trained and training set to degree of deep learning model
It is trained.
In one embodiment, it is trained including to degree of deep learning model: by the running environment of training
Operator information input degree of deep learning model in risk field and training set, the wagon control of output prediction
Variable, such as Vehicular turn angle, vehicle plus/minus speed etc., according to wagon control variable and the expectation of prediction
Wagon control variable (determining desired wagon control variable according to operator information) determine driver
The loss information of operation information, according to car in the loss Information revision degree of deep learning model of operator information
The parameter of control variable.Through the iteration of certain number of times, obtain satisfactory degree of deep learning model.Its
In, degree of deep learning model can be such as degree of depth convolutional neural networks model.
Running environment risk field can describe running environment all sidedly, is advantageously implemented under different road environment
Automatic Pilot.
Step S106, sets up the running environment risk field of test according to the environment sensing information in test set,
By the running environment risk field input degree of deep learning model of test, the wagon control variable of output prediction, lead to
Degree of deep learning model is entered by the operator information crossed in the wagon control variable of comparison prediction and test set
Row test.
A kind of exemplary method of testing is described as follows, if prediction wagon control variable and test set in
Operator information between gap less than preset value, that is, the concordance of the two is preferable, the most permissible
Determine that degree of deep learning model can be used.Wherein, wagon control variable such as include Vehicular turn angle, vehicle add/
Deceleration etc..
But, it will be understood by those skilled in the art that above-mentioned method of testing is the most unique.Such as, will
The wagon control variable of prediction is for being controlled vehicle, and can observe vehicle normally travel, if permissible
Normal traveling, it is determined that degree of deep learning model can be used.
Above-described embodiment present invention theory based on field sets up the risk model of vehicle running environment, thus sharp
Comprehensively reflect the driving environment of vehicle with running environment risk field, be advantageously implemented under different road environment
Automatic Pilot.And according to running environment risk field and operator, Vehicular automatic driving model is instructed
Practice, to learn the experience of outstanding human driver, it is achieved personalize automatic Pilot.Additionally, relative to directly
Utilize environment sensing information that Vehicular automatic driving model is trained, reduce Vehicular automatic driving model
Training difficulty.
Present invention also offers a kind of method that degree of deep learning model is trained.Shown in Figure 2
The flow chart that degree of deep learning model is trained, for the data in training set, training process is as follows:
Step S202, is identified by the environment sensing information of at least one sensor acquisition in training set,
Identify the driving-environment informations such as such as stationary object, moving object, road.
As a example by video camera, laser radar, millimetre-wave radar etc., identification process is described below.
According to the image of camera acquisition, identify the target such as lane line and vehicle.Wherein it is possible to use
Lane line in image processing method identification image.The method is marked by image adaptive Threshold segmentation, track
Note line feature point extracts, feature points clustering mates with matching, lane line and the step such as tracking, it is achieved lane line
Accurately identify and tenacious tracking.Wherein it is possible to use the vehicle mesh in the method identification image of machine learning
Mark.The method uses HOG (Histogram of Oriented Gradient, histograms of oriented gradients)
Feature and AdaBoost (a kind of iterative algorithm) cascade classifier training vehicle detection model, and then use
The accurate detection of vehicle detection model realization vehicle target.It will be understood by those skilled in the art that for pedestrian,
The targets such as cyclist, road, road sign are referred to the recognition methods of aforesaid lane line and vehicle target, this
In repeat no more.
Additionally, laser radar can obtain some cloud information (the i.e. space coordinates letter on barrier and road surface on road
Breath).Millimetre-wave radar can the information such as Position And Velocity of acquired disturbance thing (such as the target such as vehicle, fence).
Step S204, optionally, in the case of multiple sensor acquisition environment sensing information, it is also possible to
Carry out Coordinate Conversion and/or data fusion.
Wherein, Coordinate Conversion refers to change the coordinate system of multiple sensors, to form unified seat
Mark system, makes follow-up data fusion be easier to.A kind of method of Coordinate Conversion is such as it may be that sit image
Mark system is converted to camera coordinates, then is transformed into unified by the coordinate system of camera coordinates and other sensors
Vehicle axis system (is such as fixed on the coordinate system from car, zero is in vehicle centroid), it is achieved not simultaneous interpretation
The Coordinate Conversion of sensor perception information.
Wherein, different, such as millimetre-wave radar lateral resolution due to the attribute of different sensors perception information
Low, vision sensor range accuracy differences etc., the present invention uses the mahalanobis distance association different sensors of target to see
The same target surveyed, further for merging different sensors observation, that observes different sensors is same
Target is weighted averagely by probability of happening, as the probability of happening of this same target, thus realizes many sensings
The fusion of device information and effective estimation of observation time of day.Wherein, joint probability number can such as be used
According to association (JPDA, Joint Probability Data Association) method, different sensors is seen
The same target surveyed is weighted averagely by probability of happening.
By above-mentioned Coordinate Conversion or data fusion, road environment information can be identified more accurately.
Step S206, sets up the running environment risk field of training according to the environment sensing information in training set.
The risk field that the invention provides a kind of degree of risk that can comprehensively reflect vehicle running environment is built
Cube method.That is, according to stationary object (as stop vehicle etc.) formed potential energy field information, moving object
Kinetic energy field information and the behavior field information of driver's formation that (such as vehicle and the pedestrian of motion) is formed are set up
Running environment risk field, formula is expressed as follows:
Es=Er+Ev+Ed (1)
Wherein, Es represents running environment risk field, and Er represents the potential energy field information that stationary object is formed, Ev
Representing the kinetic energy field information that moving object is formed, Ed represents the behavior field information that driver is formed.
For the running environment risk field of training, the potential energy field information of stationary object formation and the moving object bodily form
The kinetic energy field information become determines according to the environment sensing information in training set, the behavior field information that driver is formed
Determine according to the operator information in training set.Specifically, potential energy field characterizes stationary object on road
Physical field on traffic safety impact, the size and Orientation of potential energy field field intensity is mainly by stationary object attribute and road
Road conditional decision.Kinetic energy field is to characterize the physical field that on road, traffic safety is affected by moving object, kinetic energy field
The size and Orientation of field intensity is mainly determined by the attribute of moving object, kinestate and road conditions.Behavior field
Being to characterize the physical field that traffic safety is affected by driver behavior pattern, the size of behavior field field intensity is mainly by driving
The behavioral trait of the person of sailing determines.Under the same terms, the driver of radical type usually makes than conservative driver
The driving risk become is big, and its behavior field field intensity is the biggest;The driver that driving efficiency is low is generally high than driving efficiency
Driving behavior field field intensity big.
Fig. 3 illustrates the schematic diagram of the running environment risk field under a typical road environment.For convenience of the degree of depth
The training process practised, can be by risk field discretization the image projecting to two dimension.Wherein, risk field picture
Abscissa represent the horizontal direction of vehicle, vertical coordinate represents the longitudinal direction of vehicle, and image pixel value represents
Degree of risk (such as can quantify to 0 to 255).The most such as can consider left and right vehicle wheel
20 meters, first 100 meters, the scope of latter 50 meters, each pixel represents the length of 0.5 meter, therefore the wind generated
Field, danger gray level image size is 300x80.
Step S208, by the operator information in the running environment risk field trained and training set (i.e.
Supervision message) input degree of deep learning model, the wagon control variable of output prediction.
Wherein, operator information includes the information such as Vehicular turn angle, vehicle plus/minus speed.In order to obtain
Take rich and varied driving data, the driving data of different driver can be selected.
Wherein, degree of deep learning model can be such as degree of depth convolutional neural networks model, and this model includes five layers
Convolutional layer and the full articulamentum of two-layer, the wagon control amount of last layer of output two dimension.
Step S210, wagon control variable and desired wagon control variable according to prediction are (according to driving
Member's operation information determines) determine the loss information of operator information, such as use L2 loss function, root
According to the parameter of wagon control variable in the loss Information revision degree of deep learning model of operator information.
Through the iteration of certain number of times (such as 100,000 times), satisfactory degree of depth study mould can be obtained
Type, thus complete the training process to degree of deep learning model.
Above-described embodiment, theory based on field is set up the risk evaluation model of vehicle running environment, is merged multiple
Sensor information inputs, and sets up comprehensive running environment description system, is advantageously implemented under different road environment
Automatic Pilot.Export in conjunction with the corresponding vehicle operating of vehicle running environment and driver, learn based on the degree of depth
The automatic Pilot model of method study vehicle, can realize the automatic Pilot of vehicle.By the different driver of study
Driving behavior, the automatic Pilot that personalizes of vehicle can be realized.
Present invention also offers a kind of method that degree of deep learning model is tested.Shown in Figure 4
The flow chart that degree of deep learning model is tested, for the data in test set, test process is as follows:
Step S402, is identified by the environment sensing information of at least one sensor acquisition in test set,
Identify the driving-environment informations such as such as stationary object, moving object, road.
Wherein, the recognition methods to the environment sensing information in test set is referred to the ring in training set
The recognition methods (i.e. with reference to step S202) of border perception information, repeats no more here.
Step S404, optionally, in the case of multiple sensor acquisition environment sensing information, it is also possible to
Carry out Coordinate Conversion and/or data fusion.
Wherein, Coordinate Conversion and/or data fusion method to the environment sensing information in test set can be joined
Examine the Coordinate Conversion to the environment sensing information in training set and/or data fusion method (i.e. with reference to step
S204), repeat no more here.
Step S406, sets up the running environment risk field of test according to the environment sensing information in test set.
Wherein, the method for building up of the running environment risk field of test is referred to the running environment wind of training
The method for building up (i.e. with reference to step S206) of field, danger, repeats no more here.
Step S408, the degree of deep learning model that the running environment risk field input training of test is obtained,
The degree of deep learning model i.e. using training processes the running environment risk field of input, the wagon control of output prediction
Variable.
Such as, it is input to train the degree of depth obtained by the risk field gray level image that the size of generation is 300x80
Convolutional neural networks model, obtains the wagon control amount of two dimension by recurrence, and such as Vehicular turn angle, add/
The information such as deceleration.
Step S410, by the operator information in the wagon control variable of comparison prediction and test set
Degree of deep learning model is tested.
A kind of exemplary method of testing is described as follows, if prediction wagon control variable and test set in
Operator information between gap less than preset value, that is, the concordance of the two is preferable, the most permissible
Determine that degree of deep learning model can be used.Wherein, wagon control variable such as include Vehicular turn angle, vehicle add/
Deceleration etc..
If degree of deep learning model can be used, then can according to the degree of deep learning model output wagon control amount (as
Vehicular turn angle, plus/minus speed etc.), use PID (PID) control realization that vehicle is had
Effect controls.
Present invention also offers a kind of automatic driving device, with reference to Fig. 5, this device includes:
Sample forms module 502, for setting up according to the environment sensing information gathered and operator information
Vehicular automatic driving data base, and Vehicular automatic driving data base is divided into training set and test set;
Model training module 504, for setting up the traveling of training according to the environment sensing information in training set
Environmental risk field, according to the operator information in the running environment risk field trained and training set to deeply
Degree learning model is trained;
Model measurement module 506, for setting up the traveling of test according to the environment sensing information in test set
Environmental risk field, by the running environment risk field input degree of deep learning model of test, the vehicle of output prediction
Control variable, by the operator information in the wagon control variable of comparison prediction and test set to the degree of depth
Learning model is tested.
With reference to Fig. 6, in the case of environment sensing information is obtained by multiple sensor acquisition, sample forms mould
Block 502 includes: data processing unit 5022 and sample form unit 5024;
Data processing unit 5022 is for changing the coordinate system of multiple sensors, unified to be formed
Coordinate system;And/or, use the same target of the mahalanobis distance association different sensors observation of target, right
The same target of different sensors observation is weighted averagely by probability of happening, as the generation of this same target
Probability.
Sample forms unit 5024, for building according to the environment sensing information gathered and operator information
Vertical Vehicular automatic driving data base, and Vehicular automatic driving data base is divided into training set and test set.
Wherein, model training module 504 includes that unit 5042 is set up in the first risk field, for according to static
The behavior field that potential energy field information, the kinetic energy field information of moving object formation and the driver that object is formed is formed
Information sets up the running environment risk field of training;Wherein, stationary object formed potential energy field information and motion
The kinetic energy field information that object is formed determines according to the environment sensing information in training set, the behavior that driver is formed
Field information determines according to the operator information in training set.
Wherein, model training module 504 includes model training unit 5044, for the traveling by training
Operator information input degree of deep learning model in environmental risk field and training set, the vehicle of output prediction
The loss information of control variable and operator information;Loss Information revision according to operator information
The parameter of wagon control variable in degree of deep learning model.
Wherein, model measurement module 506 includes that unit 5062 is set up in the second risk field, for according to static
The behavior field that potential energy field information, the kinetic energy field information of moving object formation and the driver that object is formed is formed
Information sets up the running environment risk field of test;Wherein, stationary object formed potential energy field information and motion
The kinetic energy field information that object is formed determines according to the environment sensing information in test set, the behavior that driver is formed
Field information determines according to the operator information in test set.
Wherein, the potential energy field information that stationary object is formed determines according to attribute and the road conditions of stationary object;
The kinetic energy field information that moving object is formed determines according to attribute, kinestate and the road conditions of moving object.
Wherein, model measurement module 506 includes model measurement unit 5064, for the traveling by test
Environmental risk field input degree of deep learning model, the wagon control variable of output prediction, by the car of comparison prediction
Degree of deep learning model is tested by the operator information in control variable and test set.
Present invention theory based on field sets up the risk model of vehicle running environment, thus utilizes running environment
The driving environment of vehicle is comprehensively reflected in risk field, is advantageously implemented the automatic Pilot under different road environment.
And according to running environment risk field and operator, Vehicular automatic driving model is trained, with study
The experience of outstanding human driver, it is achieved personalize automatic Pilot.Additionally, relative to directly utilizing environment sense
Know that Vehicular automatic driving model is trained by information, reduce the training difficulty of Vehicular automatic driving model.
Last it is noted that above example is only in order to illustrate technical scheme, rather than right
It limits.It will be understood by those within the art that: can be to the technology described in foregoing embodiments
Scheme is modified, or wherein portion of techniques feature is carried out equivalent;These amendments or replacement,
The essence not making appropriate technical solution departs from the spirit and scope of various embodiments of the present invention technical scheme.