CN109739245A - One kind being based on unpiloted end to end model appraisal procedure and device - Google Patents
One kind being based on unpiloted end to end model appraisal procedure and device Download PDFInfo
- Publication number
- CN109739245A CN109739245A CN201910122748.1A CN201910122748A CN109739245A CN 109739245 A CN109739245 A CN 109739245A CN 201910122748 A CN201910122748 A CN 201910122748A CN 109739245 A CN109739245 A CN 109739245A
- Authority
- CN
- China
- Prior art keywords
- data
- model
- driving vehicle
- target
- travel speed
- 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.)
- Pending
Links
Landscapes
- Control Of Driving Devices And Active Controlling Of Vehicle (AREA)
Abstract
This application discloses one kind to be based on unpiloted end to end model appraisal procedure and device, this method comprises: getting first object data to be assessed, second target data, after third target data, second target data and third target data are input to the rate conversion model constructed in advance, calculate the travel speed of target automatic driving vehicle, then the travel speed and first object data are input to vehicle kinematics model, obtain the driving trace of target automatic driving vehicle, the driving trace and both fixed tracks are compared again, and according to comparing result, the output result of end-to-end model is assessed, to obtain assessment result.It can be seen that, the application is first with the rate conversion model constructed in advance, and vehicle kinematics model, E2E model output result is converted to the driving trace of automatic driving vehicle, the driving trace and both fixed tracks are compared again, so as to carry out accurate evaluation according to output result of the comparing result to E2E model.
Description
Technical field
This application involves field of artificial intelligence, more particularly to one kind to be based on unpiloted end to end model assessment side
Method and device.
Background technique
Automatic driving vehicle is a kind of novel intelligent automobile, mainly passes through electronic control unit (Electronic
Control Unit, abbreviation ECU) etc. vehicle-mounted terminal equipments in vehicle various pieces carry out accurately control with calculate point
Analysis, to realize the fully automatic operation of vehicle, reaches the unpiloted target of vehicle.
Currently, realization is preferable in order to enable the operating status of automatic driving vehicle can drive vehicle closer to artificial
Unmanned effect, determine nobody generally by preparatory trained end-to-end (End To End, abbreviation E2E) model
Required steering wheel angle, accelerator open degree and brake aperture when driving vehicle driving, then, then control automatic driving vehicle with
The determining steering wheel angle of the E2E model, accelerator open degree, brake aperture travel on road.
It but since the output of E2E model is steering wheel angle, accelerator open degree and brake aperture, and is not unmanned
The driving trace of vehicle, so can not intuitively, explicitly embody automatic driving vehicle and artificial driving vehicle according to these three values
The difference of driving status.Therefore, also just can not accurate evaluation go out the training effect based on unmanned E2E model, so,
In order to enable automatic driving vehicle is in the steering wheel angle, accelerator open degree, brake aperture determined using E2E model in road uplink
When sailing, it can be realized better unmanned effect (i.e. closer to artificial driving condition), need the training to the E2E model
Effect carries out accurate evaluation, and there is no a kind of methods that can carry out accurate evaluation to E2E model at present, therefore, how real
Now to the accurate evaluation of E2E model, and then realize preferably unmanned effect, it has also become urgent problem to be solved.
Summary of the invention
To solve the above problems, this application provides one kind to be based on unpiloted end to end model appraisal procedure and dress
It sets, specific technical solution is as follows:
In a first aspect, this application provides one kind to be based on unpiloted end to end model appraisal procedure, comprising:
Obtain first object data to be assessed, the second target data and third target data, the first object number
According to the steering wheel angle of the target automatic driving vehicle exported for end-to-end model, second target data is described end-to-end
The accelerator open degree of the target automatic driving vehicle of model output, the third target data are end-to-end model output
The target automatic driving vehicle brake aperture；
Second target data and the third target data are input to the rate conversion model constructed in advance, in terms of
Calculate the travel speed of the target automatic driving vehicle；
The travel speed of the first object data and the target automatic driving vehicle is input to vehicle kinematics mould
Type, to obtain the driving trace of the target automatic driving vehicle；
The driving trace of the target automatic driving vehicle is compared with both fixed tracks, and according to comparing result, it is right
The output result of the end-to-end model is assessed, to obtain assessment result.
Optionally, the rate conversion model is constructed, comprising:
Obtain artificial the first training data, the second training data and the corresponding actual travel speed for driving vehicle, institute
The accelerator open degree that the first training data is the artificial driving vehicle is stated, second training data is the artificial driving vehicle
Brake aperture；
Extract the data characteristics of first training data and second training data；
According to the data characteristics and corresponding actual travel speed of first training data and second training data
Degree is trained initial velocity transformation model, generates the rate conversion model.
Optionally, the initial velocity transformation model connects entirely comprising the first full articulamentum, shot and long term memory network layer, second
Connect layer.
Optionally, the rate conversion model is constructed, comprising:
Using given objective function, the rate conversion model is constructed；
Wherein, the objective function is difference of two squares loss function.
Optionally, the method also includes:
Obtain artificial the first test data, the second test data and the corresponding actual travel speed for driving vehicle；
Extract the data characteristics of first test data and second test data；
The data characteristics of first test data and second test data is inputted into the rate conversion model, meter
Calculate first test data and the corresponding test travel speed of second test data of the artificial driving vehicle；
When the test travel speed and the actual travel speed are inconsistent, by first test data and described
Second test data is re-used as first training data and second training data respectively, to the rate conversion model
It is updated.
Second aspect, the application provide a kind of based on unpiloted end to end model device, comprising:
Target data acquiring unit, for obtaining first object data to be assessed, the second target data and third mesh
Mark data, the first object data are the steering wheel angle of the target automatic driving vehicle of end-to-end model output, described the
Two target datas are the accelerator open degree of the target automatic driving vehicle of end-to-end model output, the third number of targets
According to the brake aperture of the target automatic driving vehicle exported for the end-to-end model；
Travel speed acquiring unit, for second target data and the third target data to be input to preparatory structure
The rate conversion model built, to calculate the travel speed of the target automatic driving vehicle；
Driving trace acquiring unit, for the first object data and the traveling of the target automatic driving vehicle are fast
Degree is input to vehicle kinematics model, to obtain the driving trace of the target automatic driving vehicle；
Assessment result obtaining unit, for carrying out pair the driving trace of the target automatic driving vehicle and both fixed tracks
Than, and according to comparing result, the output result of the end-to-end model is assessed, to obtain assessment result.
Optionally, described device further include:
Training data acquiring unit, for obtain artificial the first training data for driving vehicle, the second training data and
Corresponding actual travel speed, first training data are the accelerator open degree of the artificial driving vehicle, second training
Data are the brake aperture of the artificial driving vehicle；
Fisrt feature extraction unit, the data for extracting first training data and second training data are special
Sign；
First model generation unit, for the data characteristics according to first training data and second training data
And corresponding actual travel speed is trained initial velocity transformation model, generates the rate conversion model.
Optionally, the initial velocity transformation model connects entirely comprising the first full articulamentum, shot and long term memory network layer, second
Connect layer.
Optionally, described device further include:
Second model generation unit, for constructing the rate conversion model using given objective function；
Wherein, the objective function is difference of two squares loss function.
Optionally, device further include:
Test data acquiring unit, for obtain artificial the first test data for driving vehicle, the second test data and
Corresponding actual travel speed；
Second feature extraction unit, the data for extracting first test data and second test data are special
Sign；
Test speed acquiring unit, for the data characteristics of first test data and second test data is defeated
Enter the rate conversion model, calculates first test data and second test data of the artificial driving vehicle
Corresponding test travel speed；
Transformation model updating unit is used for when the test travel speed and the actual travel speed are inconsistent, will
First test data and second test data are re-used as first training data and second training respectively
Data are updated the rate conversion model.
It is provided by the embodiments of the present application a kind of based on unpiloted end to end model appraisal procedure and device, it is getting
It, can be by the second target data and third after first object data to be assessed, the second target data and third target data
Target data is input to the rate conversion model constructed in advance, to calculate the travel speed of target automatic driving vehicle, wherein
First object data refer to the steering wheel angle of the target automatic driving vehicle of end-to-end model output, and the second target data refers to
Be end-to-end model output target automatic driving vehicle accelerator open degree, third target data refers to that end-to-end model is defeated
The brake aperture of target automatic driving vehicle out, then, then by the traveling of first object data and target automatic driving vehicle
Speed is input to vehicle kinematics model, to obtain the driving trace of target automatic driving vehicle, in turn, can by target nobody
The driving trace for driving vehicle is compared with both fixed tracks, and according to comparing result, to the output result of end-to-end model into
Row assessment, to obtain assessment result.As it can be seen that the embodiment of the present application is first with the rate conversion model constructed in advance, Yi Jiche
Steering wheel angle, accelerator open degree, brake aperture that E2E model exports are converted to automatic driving vehicle by kinematics model
Then driving trace compares the driving trace and both fixed tracks, so as to according to comparing result to E2E model
It exports result and carries out accurate evaluation, and then realize preferably unmanned effect.
Detailed description of the invention
In order to illustrate the technical solutions in the embodiments of the present application or in the prior art more clearly, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is the application
Some embodiments for those of ordinary skill in the art without creative efforts, can also basis
These attached drawings obtain other attached drawings.
Fig. 1 is a kind of process signal based on unpiloted end to end model appraisal procedure provided by the embodiments of the present application
Figure；
Fig. 2 is the flow diagram of building rate conversion model provided by the embodiments of the present application；
Fig. 3 is the structural schematic diagram of initial velocity transformation model provided by the embodiments of the present application；
Fig. 4 is a kind of flow diagram of rate conversion model verification method provided by the embodiments of the present application；
Fig. 5 is a kind of composition signal that device is assessed based on unpiloted end to end model provided by the embodiments of the present application
Figure.
Specific embodiment
In general, E2E model is applied to unmanned field, is to realize better unmanned effect, that is,
The operating status of automatic driving vehicle is enabled to drive vehicle closer to artificial.When being trained to E2E model, first
It is that it is end-to-end to be input to this by shooting the video image when artificially driving vehicle operation on both fixed tracks as input data
Model, by model output prediction vehicle run when steering wheel angle, accelerator open degree and brake opening value, and with it is preparatory
Actual steering wheel corner, accelerator open degree and brake aperture compare in the artificial driving procedure of record, according to comparing result
Model parameter is updated, gets the model of E2E through excessive training in rotation, but since the output of E2E model is that steering wheel angle, throttle are opened
Degree and brake aperture, and it is not the driving trace of automatic driving vehicle, so can not be intuitive, specific according to these three values
Embody automatic driving vehicle and the artificial difference for driving vehicle running state.Therefore, also just can not accurate evaluation go out this and be based on
The training effect of unmanned E2E model.
But in order to enable automatic driving vehicle is in the steering wheel angle, accelerator open degree, brake aperture determined using E2E model
On road when driving, it can be realized better unmanned effect, then it is accurate to need to carry out the training effect of the E2E model
Assessment, and there is no a kind of methods that can carry out accurate evaluation to E2E model therefore how to realize to E2E model at present
Accurate evaluation, and then realize preferably unmanned effect, it has also become urgent problem to be solved
To solve the above problems, the embodiment of the present application provides one kind based on unpiloted end to end model assessment side
Method, after getting first object data to be assessed, the second target data and third target data, by the second target data
It is input to the rate conversion model constructed in advance with third target data, calculates the travel speed of target automatic driving vehicle,
Wherein, first object data refer to the steering wheel angle of the target automatic driving vehicle of end-to-end model output, the second target
Data refer to the accelerator open degree of the target automatic driving vehicle of end-to-end model output, and third target data refers to end-to-end
The brake aperture of the target automatic driving vehicle of model output, then, then by first object data and calculated target nobody
The travel speed for driving vehicle is input to vehicle kinematics model, obtains the driving trace of target automatic driving vehicle, in turn, can
To compare the driving trace and both fixed tracks, and according to comparing result, the output result of E2E model is accurately commented
Estimate.As it can be seen that the embodiment of the present application is first with the rate conversion model and vehicle kinematics model constructed in advance, by E2E mould
Then the driving trace that type output result is converted to automatic driving vehicle compares the driving trace and both fixed tracks, from
And accurate evaluation can be carried out according to output result of the comparing result to E2E model, and then realize preferably unmanned effect.
To keep the purposes, technical schemes and advantages of the embodiment of the present application clearer, below in conjunction with the embodiment of the present application
In attached drawing, the technical scheme in the embodiment of the application is clearly and completely described, it is clear that described embodiment is
Some embodiments of the present application, instead of all the embodiments.Based on the embodiment in the application, those of ordinary skill in the art
Every other embodiment obtained without making creative work, shall fall in the protection scope of this application.
First embodiment
It is that a kind of process based on unpiloted end to end model appraisal procedure provided in this embodiment is shown referring to Fig. 1
It is intended to, method includes the following steps:
S101: first object data to be assessed, the second target data and third target data are obtained.
In the present embodiment, in order to based on unpiloted E2E model progress accurate evaluation, it is necessary first to get
The output data of model is based on the target data, target automatic driving vehicle can be in road as target data to be assessed
Upper carry out normally travel, it should be noted that the target data to be assessed include first object data, the second target data with
And third target data, wherein first object data refer to that the steering wheel of the target automatic driving vehicle of E2E model output turns
Angle, the second target data refer to the accelerator open degree of the target automatic driving vehicle of E2E model output, and third target data refers to
It is the brake aperture of the target automatic driving vehicle of E2E model output.
Second target data and third target data: being input to the rate conversion model constructed in advance by S102, to calculate
The travel speed of target automatic driving vehicle out.
In the present embodiment, by step S101 get first object data to be assessed, the second target data and
Third target data can after getting accelerator open degree and the brake aperture of the target automatic driving vehicle of E2E model output
To be handled using the data characteristics extracting method of existing or future appearance it, for example, can use linear discriminant analysis
(Linear Discriminate Analysis, abbreviation LDA) feature extracting method, from the second target data and third number of targets
The feature vector that can characterize each data semantic information is extracted in, then using the two feature vectors as input data,
It is input to the rate conversion model constructed in advance, the second target data and third target data pair can be predicted by the model
The travel speed for the target automatic driving vehicle answered.
It should be noted that needing to construct a rate conversion model in advance, specifically constructing to realize this step S102
Journey can be found in the related introduction of subsequent second embodiment.
S103: being input to vehicle kinematics model for the travel speed of first object data and target automatic driving vehicle,
To obtain the driving trace of target automatic driving vehicle.
In the present embodiment, by step S102 predict the second target data and the corresponding target of third target data without
After people drives the travel speed of vehicle, can by the travel speed and first object data (i.e. the target of E2E model output nobody
Drive the steering wheel angle of vehicle) it is input in vehicle kinematics model, go out target automatic driving vehicle using the models fitting
Driving trace, which more can intuitively embody the training effect of E2E model.
It should be noted that this implementation use vehicle kinematics model can be any of existing or future appearance can
Fit the vehicle kinematics model of target vehicle driving trace using travel speed and steering wheel angle, specific implementation process with
Existing method is consistent, and details are not described herein by the application.
S104: the driving trace of target automatic driving vehicle and both fixed tracks are compared, and according to comparing result, right
The output result of end-to-end model is assessed, to obtain assessment result.
In the present embodiment, after the driving trace that target automatic driving vehicle is obtained by step S103, that is, pass through E2E mould
It, can be by the driving trace and set traveling rail after the output prediction of result of type goes out the driving trace of target automatic driving vehicle
Mark compares, and judges whether error between the two meets preset condition, if satisfied, then showing that target is unmanned
Vehicle and the artificial vehicle running state that drives are closely similar, and then the assessment result of available model are as follows: E2E model output knot
The accuracy rate of fruit is higher, and model training effect is preferable；Conversely, if the driving trace of the target automatic driving vehicle predicted with both
Error between fixed driving trace is unsatisfactory for preset condition, then shows target automatic driving vehicle and artificial driving vehicle row
It is larger to sail state difference, that is, is not carried out preferable unmanned effect, and then the assessment result of available model are as follows: E2E
The accuracy rate that model exports result is lower, and model training effect is poor.
To sum up, provided in this embodiment a kind of based on unpiloted end to end model appraisal procedure, it is to be evaluated getting
It, can be by the second target data and third target after first object data, the second target data and the third target data estimated
Data are input to the rate conversion model constructed in advance, to calculate the travel speed of target automatic driving vehicle, wherein first
Target data refers to the steering wheel angle of the target automatic driving vehicle of end-to-end model output, and the second target data refers to
The accelerator open degree of the target automatic driving vehicle of end-to-end model output, third target data refer to end-to-end model output
The brake aperture of target automatic driving vehicle, then, then by the travel speed of first object data and target automatic driving vehicle
It is input to vehicle kinematics model, it, in turn, can be unmanned by target to obtain the driving trace of target automatic driving vehicle
The driving trace of vehicle is compared with both fixed tracks, and according to comparing result, is commented the output result of end-to-end model
Estimate, to obtain assessment result.As it can be seen that the embodiment of the present application is transported first with the rate conversion model and vehicle constructed in advance
It is dynamic to learn model, steering wheel angle, accelerator open degree, brake aperture that E2E model exports are converted to the traveling of automatic driving vehicle
Then track compares the driving trace and both fixed tracks, so as to the output according to comparing result to E2E model
As a result accurate evaluation is carried out, and then realizes preferably unmanned effect.
Second embodiment
The specific building process of the rate conversion model referred in first embodiment will be introduced in the present embodiment.It utilizes
The rate conversion model constructed in advance, can accurately calculate the travel speed of target automatic driving vehicle.
Referring to fig. 2, it illustrates the flow diagram of building rate conversion model provided in this embodiment, which includes
Following steps:
S201: artificial the first training data for driving vehicle, the second training data and corresponding actual travel speed are obtained
Degree.
In the present embodiment, it in order to construct rate conversion model, needs to carry out a large amount of preparation in advance, firstly, needing
Collect the artificial accelerator open degree for driving vehicle and brake aperture, and respectively as the first training data and the second training data,
It is also desirable to collect the actual travel speed of each accelerator open degree artificial driving vehicle corresponding with aperture of braking, and will collect
The each accelerator open degree and brake opening value of the artificial driving vehicle arrived are respectively as sample data, to training speed modulus of conversion
Type.
S202: the data characteristics of the first training data and the second training data is extracted.
In the present embodiment, artificial the first training data for driving vehicle, the second training number are got by step S201
Accordingly and after corresponding actual travel speed, it can not be directly used in trained formation speed transformation model, but need to extract
The data characteristics of one training data, the second training data and corresponding actual travel speed, wherein the first training data and
The extraction of the data characteristics of two training datas refers to the first training data and the second training data being converted into one group of tool respectively
Have a feature vector of obvious physics, realize dimensionality reduction effect, in characteristic extraction procedure, can use LDA feature extracting method into
Row extracts, and then can use the data characteristics of the first training data and the second training data that extract, and training obtains speed
Transformation model.
S203: according to the data characteristics and corresponding actual travel speed pair of the first training data and the second training data
Initial velocity transformation model is trained, formation speed transformation model.
In the present embodiment, the data characteristics of the first training data and the second training data is extracted by step S202
Afterwards, further, it can be obtained according to the data characteristics of first training data and the second training data and by step S201
The actual travel speed of the corresponding artificial driving vehicle of the first training data and the second training data got turns initial velocity
Mold changing type is trained, and then formation speed transformation model.
Wherein, the first full articulamentum (Fully is contained in the model framework of initial velocity transformation model
Connection), shot and long term memory network (long Short Term Memory, abbreviation LSTM) layer, the second full articulamentum, such as
Shown in Fig. 3, the model framework of initial velocity transformation model further includes input layer (Input Tensor) and output layer (Output
Tensor), the connection direction between each layer is as shown in the arrow direction in Fig. 3, i.e. the model framework of initial velocity transformation model
For " the full articulamentum of the input layer -> the first -> full articulamentum -> output layer of shot and long term memory network layer -> the second ".In the present embodiment,
The vector dimension of input layer input is 2, is respectively used to the data characteristics of input the first training data and the second training data；LSTM
The step-length of network layer is 20, and the data characteristics pair of the first training data and the second training data can be generated by LSTM network layer
The forward direction hidden layer characterization answeredAnd backward hidden layer characterizationAnd then the two can be spliced into a feature vector,
This feature vector can characterize the contextual information of the first training data and the second training data, so can according to this feature to
Amount predicts the artificial travel speed for driving vehicle；The vector dimension of output layer output is 1, for exporting predict first
The data characteristics of the travel speed of training data and the corresponding artificial driving vehicle of the second training data.
In turn, one group of sample data can be successively extracted from model training data, and initial velocity shown in Fig. 3 is converted
Model carries out multiwheel models training, until meeting training termination condition, at this point, i.e. formation speed transformation model.
It specifically, can be by the second target data and third number of targets in first embodiment when carrying out epicycle training
According to the first training data and the second training data in the sample data for replacing with epicycle extraction, turned by current initial velocity
It is corresponding artificial can to predict the sample data according to the implementation procedure of step S102 in first embodiment for mold changing type
Drive the travel speed of vehicle.It is then possible to which the travel speed is carried out with the corresponding artificial actual travel speed for driving vehicle
Compare, and model parameter is updated according to the difference of the two, specifically, back-propagation algorithm can be used, model is joined
Number is updated, until meeting preset condition, then stops the update of model parameter, completes the training of rate conversion model, raw
At a trained rate conversion model.
In the training process, a kind of to be optionally achieved in that, it can use given objective function, turn to construct speed
Mold changing type, wherein objective function can be difference of two squares loss function.It further, can be in conjunction with the mode of chain rule derivation
Carry out undated parameter.Specifically, it when being trained using objective function to rate conversion model, can be damaged according to objective function
The model parameter of rate conversion model is constantly updated in the variation of mistake value, until objective function penalty values are met the requirements, than
Such as amplitude of variation very little, then stop the update of model parameter, completes the training of rate conversion model.It should be noted that this implementation
The calculation method of difference of two squares loss function and chain rule derivation that example uses is consistent with existing method, and specific implementation process can join
See existing scheme, details are not described herein.
Through the foregoing embodiment, artificial the first training data for driving vehicle, the second training data and right be can use
The actual travel speed training formation speed transformation model answered then further can use artificial the first survey for driving vehicle
Examination data, the second test data and corresponding actual travel speed verify the rate conversion model of generation.
Rate conversion model verification method provided by the embodiments of the present application is introduced with reference to the accompanying drawing.
Referring to fig. 4, it illustrates a kind of flow charts of rate conversion model verification method provided by the embodiments of the present application, such as
Shown in Fig. 4, this method comprises:
S401: artificial the first test data for driving vehicle, the second test data and corresponding actual travel speed are obtained
Degree.
In practical applications, rate conversion model is verified in order to realize, it is necessary first to obtain artificial driving vehicle
The first test data, the second test data and corresponding actual travel speed, wherein artificial the first test for driving vehicle
Data, the second test data refer to can be used to carry out rate conversion model verifying artificial driving vehicle accelerator open degree and
Brake aperture is getting artificial the first test data for driving vehicle, the second test data and corresponding actual travel speed
After degree, step 402 can be continued to execute.
S402: the data characteristics of the first test data and second test data is extracted.
In practical applications, by step S401, artificial the first test data for driving vehicle, the second test number are got
Accordingly and after corresponding actual travel speed, it can not be directly used in verifying speed transformation model, but need to extract the first survey
Try the data characteristics of data, the second test data and corresponding actual travel speed, wherein the first test data and second is surveyed
The extraction for trying the data characteristics of data refers to for the first test data and the second test data being converted into one group respectively with bright
The feature vector of aobvious physics realizes dimensionality reduction effect, for example, in characteristic extraction procedure, can use LDA feature extracting method into
Row extracts, and then can use the data characteristics of the first test data and the second test data that extract, the speed verified
Spend transformation model.
S403: by the data characteristics input speed transformation model of the first test data and the second test data, people is calculated
For the first test data and the corresponding test travel speed of the second test data for driving vehicle.
During specific implementation, the data for extracting the first test data and the second test data by step 402 are special
After sign, further, the data characteristics of the first test data and the second test data can be input to rate conversion model, with
The travel speed of the first test data and the corresponding artificial driving vehicle of the second test data is predicted, and then step can be continued to execute
Rapid S404.
S404: when testing travel speed and actual travel speed is inconsistent, by the first test data and the second test number
According to the first training data and the second training data is re-used as respectively, rate conversion model is updated.
In practical applications, it by step S403, predicts the first test data and the second test data is corresponding artificial
After the travel speed for driving vehicle, when the travel speed artificial driving vehicle corresponding with the first test data and the second test data
Actual travel speed it is inconsistent when, the first test data and the second test data can be re-used as respectively first training
Data and the second training data, are updated rate conversion model.
Through the foregoing embodiment, can use using artificial the first test data for driving vehicle, the second test data with
And corresponding actual travel speed effectively verifies rate conversion model, when the travel speed predicted according to verify data
When inconsistent with actual travel speed, renewal speed transformation model can be adjusted in time, and then help to improve to model evaluation
Precision and accuracy.
To sum up, using rate conversion model made of the present embodiment training, the second target data of characterization and the be can use
The data characteristics of the semantic information of three target datas calculates the travel speed of target automatic driving vehicle, accurately in order to rear
It is continuous to carry out accurate evaluation using output result of the travel speed to E2E model, and then realize preferably unmanned effect.
3rd embodiment
The present embodiment one kind will be introduced based on unpiloted end to end model assessment device, and related content please join
See above method embodiment.
It is that a kind of composition for assessing device based on unpiloted end to end model provided in this embodiment shows referring to Fig. 5
It is intended to, which includes:
Target data acquiring unit 501, for obtaining first object data to be assessed, the second target data and third
Target data, the first object data are the steering wheel angle of the target automatic driving vehicle of end-to-end model output, described
Second target data is the accelerator open degree of the target automatic driving vehicle of end-to-end model output, the third target
Data are the brake aperture of the target automatic driving vehicle of end-to-end model output；
Travel speed acquiring unit 502, it is pre- for second target data and the third target data to be input to
The rate conversion model first constructed, to calculate the travel speed of the target automatic driving vehicle；
Driving trace acquiring unit 503, for by the row of the first object data and the target automatic driving vehicle
It sails speed and is input to vehicle kinematics model, to obtain the driving trace of the target automatic driving vehicle；
Assessment result obtaining unit 504, for by the driving trace of the target automatic driving vehicle and both fixed tracks into
Row comparison, and according to comparing result, the output result of the end-to-end model is assessed, to obtain assessment result.
In a kind of implementation of the present embodiment, described device further include:
Training data acquiring unit, for obtain artificial the first training data for driving vehicle, the second training data and
Corresponding actual travel speed, first training data are the accelerator open degree of the artificial driving vehicle, second training
Data are the brake aperture of the artificial driving vehicle；
Fisrt feature extraction unit, the data for extracting first training data and second training data are special
Sign；
First model generation unit, for the data characteristics according to first training data and second training data
And corresponding actual travel speed is trained initial velocity transformation model, generates the rate conversion model.
In a kind of implementation of the present embodiment, the initial velocity transformation model includes the first full articulamentum, length
Phase memory network layer, the second full articulamentum.
In a kind of implementation of the present embodiment, described device further include:
Second model generation unit, for constructing the rate conversion model using given objective function；
Wherein, the objective function is difference of two squares loss function.
In a kind of implementation of the present embodiment, described device further include:
Test data acquiring unit, for obtain artificial the first test data for driving vehicle, the second test data and
Corresponding actual travel speed；
Second feature extraction unit, the data for extracting first test data and second test data are special
Sign；
Test speed acquiring unit, for the data characteristics of first test data and second test data is defeated
Enter the rate conversion model, calculates first test data and second test data of the artificial driving vehicle
Corresponding test travel speed；
Transformation model updating unit is used for when the test travel speed and the actual travel speed are inconsistent, will
First test data and second test data are re-used as first training data and second training respectively
Data are updated the rate conversion model.
To sum up, provided in this embodiment a kind of based on unpiloted end to end model assessment device, it is to be evaluated getting
It, can be by the second target data and third target after first object data, the second target data and the third target data estimated
Data are input to the rate conversion model constructed in advance, to calculate the travel speed of target automatic driving vehicle, wherein first
Target data refers to the steering wheel angle of the target automatic driving vehicle of end-to-end model output, and the second target data refers to
The accelerator open degree of the target automatic driving vehicle of end-to-end model output, third target data refer to end-to-end model output
The brake aperture of target automatic driving vehicle, then, then by the travel speed of first object data and target automatic driving vehicle
It is input to vehicle kinematics model, it, in turn, can be unmanned by target to obtain the driving trace of target automatic driving vehicle
The driving trace of vehicle is compared with both fixed tracks, and according to comparing result, is commented the output result of end-to-end model
Estimate, to obtain assessment result.As it can be seen that the embodiment of the present application is transported first with the rate conversion model and vehicle constructed in advance
It is dynamic to learn model, steering wheel angle, accelerator open degree, brake aperture that E2E model exports are converted to the traveling of automatic driving vehicle
Then track compares the driving trace and both fixed tracks, so as to the output according to comparing result to E2E model
As a result accurate evaluation is carried out, and then realizes preferably unmanned effect.
As seen through the above description of the embodiments, those skilled in the art can be understood that above-mentioned implementation
All or part of the steps in example method can be realized by means of software and necessary general hardware platform.Based on such
Understand, substantially the part that contributes to existing technology can be in the form of software products in other words for the technical solution of the application
It embodies, which can store in storage medium, such as ROM/RAM, magnetic disk, CD, including several
Instruction is used so that a computer equipment (can be the network communications such as personal computer, server, or Media Gateway
Equipment, etc.) execute method described in certain parts of each embodiment of the application or embodiment.
It should be noted that each embodiment in this specification is described in a progressive manner, each embodiment emphasis is said
Bright is the difference from other embodiments, and the same or similar parts in each embodiment may refer to each other.For reality
For applying device disclosed in example, since it is corresponded to the methods disclosed in the examples, so being described relatively simple, related place
Referring to method part illustration.
It should also be noted that, herein, relational terms such as first and second and the like are used merely to one
Entity or operation are distinguished with another entity or operation, without necessarily requiring or implying between these entities or operation
There are any actual relationship or orders.Moreover, the terms "include", "comprise" or its any other variant are intended to contain
Lid non-exclusive inclusion, so that the process, method, article or equipment including a series of elements is not only wanted including those
Element, but also including other elements that are not explicitly listed, or further include for this process, method, article or equipment
Intrinsic element.In the absence of more restrictions, the element limited by sentence "including a ...", it is not excluded that
There is also other identical elements in process, method, article or equipment including the element.
The foregoing description of the disclosed embodiments makes professional and technical personnel in the field can be realized or use the application.
Various modifications to these embodiments will be readily apparent to those skilled in the art, as defined herein
General Principle can be realized in other embodiments without departing from the spirit or scope of the application.Therefore, the application
It is not intended to be limited to the embodiments shown herein, and is to fit to and the principles and novel features disclosed herein phase one
The widest scope of cause.
Claims (10)
1. one kind is based on unpiloted end to end model appraisal procedure characterized by comprising
First object data to be assessed, the second target data and third target data are obtained, the first object data are
The steering wheel angle of the target automatic driving vehicle of end-to-end model output, second target data are the end-to-end model
The accelerator open degree of the target automatic driving vehicle of output, the institute that the third target data exports for the end-to-end model
State the brake aperture of target automatic driving vehicle；
Second target data and the third target data are input to the rate conversion model constructed in advance, to calculate
The travel speed of the target automatic driving vehicle；
The travel speed of the first object data and the target automatic driving vehicle is input to vehicle kinematics model, with
Obtain the driving trace of the target automatic driving vehicle；
The driving trace of the target automatic driving vehicle is compared with both fixed tracks, and according to comparing result, to described
The output result of end-to-end model is assessed, to obtain assessment result.
2. the method according to claim 1, wherein constructing the rate conversion model, comprising:
Obtain artificial the first training data, the second training data and the corresponding actual travel speed for driving vehicle, described the
One training data is the accelerator open degree of the artificial driving vehicle, and second training data is the brake of the artificial driving vehicle
Vehicle aperture；
Extract the data characteristics of first training data and second training data；
According to the data characteristics and corresponding actual travel speed pair of first training data and second training data
Initial velocity transformation model is trained, and generates the rate conversion model.
3. according to the method described in claim 2, it is characterized in that, the initial velocity transformation model includes the first full connection
Layer, shot and long term memory network layer, the second full articulamentum.
4. according to the method in claim 2 or 3, which is characterized in that construct the rate conversion model, comprising:
Using given objective function, the rate conversion model is constructed；
Wherein, the objective function is difference of two squares loss function.
5. according to the method in claim 2 or 3, which is characterized in that the method also includes:
Obtain artificial the first test data, the second test data and the corresponding actual travel speed for driving vehicle；
Extract the data characteristics of first test data and second test data；
The data characteristics of first test data and second test data is inputted into the rate conversion model, is calculated
First test data and the corresponding test travel speed of second test data of the artificial driving vehicle；
When the test travel speed and the actual travel speed are inconsistent, by first test data and described second
Test data is re-used as first training data and second training data respectively, carries out to the rate conversion model
It updates.
6. one kind is based on unpiloted end to end model device characterized by comprising
Target data acquiring unit, for obtaining first object data to be assessed, the second target data and third number of targets
According to the first object data are the steering wheel angle of the target automatic driving vehicle of end-to-end model output, second mesh
The accelerator open degree for the target automatic driving vehicle that data are end-to-end model output is marked, the third target data is
The brake aperture of the target automatic driving vehicle of the end-to-end model output；
Travel speed acquiring unit constructs in advance for second target data and the third target data to be input to
Rate conversion model, to calculate the travel speed of the target automatic driving vehicle；
Driving trace acquiring unit, for the travel speed of the first object data and the target automatic driving vehicle is defeated
Enter to vehicle kinematics model, to obtain the driving trace of the target automatic driving vehicle；
Assessment result obtaining unit, for the driving trace of the target automatic driving vehicle to be compared with both fixed tracks,
And according to comparing result, the output result of the end-to-end model is assessed, to obtain assessment result.
7. device according to claim 6, which is characterized in that described device further include:
Training data acquiring unit, for obtaining artificial the first training data, the second training data and the correspondence for driving vehicle
Actual travel speed, first training data be the artificial driving vehicle accelerator open degree, second training data
For the brake aperture of the artificial driving vehicle；
Fisrt feature extraction unit, for extracting the data characteristics of first training data and second training data；
First model generation unit, for according to the data characteristics of first training data and second training data and
Corresponding actual travel speed is trained initial velocity transformation model, generates the rate conversion model.
8. device according to claim 7, which is characterized in that the initial velocity transformation model includes the first full connection
Layer, shot and long term memory network layer, the second full articulamentum.
9. device according to claim 7 or 8, which is characterized in that described device further include:
Second model generation unit, for constructing the rate conversion model using given objective function；
Wherein, the objective function is difference of two squares loss function.
10. device according to claim 7 or 8, which is characterized in that described device further include:
Test data acquiring unit, for obtaining artificial the first test data, the second test data and the correspondence for driving vehicle
Actual travel speed；
Second feature extraction unit, for extracting the data characteristics of first test data and second test data；
Test speed acquiring unit, for the data characteristics of first test data and second test data to be inputted institute
Rate conversion model is stated, first test data and second test data for calculating the artificial driving vehicle are corresponding
Test travel speed；
Transformation model updating unit is used for when the test travel speed and the actual travel speed are inconsistent, will be described
First test data and second test data are re-used as first training data and second training data respectively,
The rate conversion model is updated.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910122748.1A CN109739245A (en) | 2019-02-19 | 2019-02-19 | One kind being based on unpiloted end to end model appraisal procedure and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910122748.1A CN109739245A (en) | 2019-02-19 | 2019-02-19 | One kind being based on unpiloted end to end model appraisal procedure and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109739245A true CN109739245A (en) | 2019-05-10 |
Family
ID=66367706
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910122748.1A Pending CN109739245A (en) | 2019-02-19 | 2019-02-19 | One kind being based on unpiloted end to end model appraisal procedure and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109739245A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110406530A (en) * | 2019-07-02 | 2019-11-05 | 宁波吉利汽车研究开发有限公司 | A kind of automatic Pilot method, apparatus, equipment and vehicle |
| CN112346982A (en) * | 2020-11-11 | 2021-02-09 | 东软睿驰汽车技术（沈阳）有限公司 | Test case generation method and device and electronic equipment |
| CN114415672A (en) * | 2021-01-19 | 2022-04-29 | 百度（美国）有限责任公司 | Dynamic model evaluation for autonomously driven vehicles |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1775601A (en) * | 2005-11-18 | 2006-05-24 | 吉林大学 | Vehicle driving trace predicating and lane deviation evaluating method |
| CN107450511A (en) * | 2017-06-28 | 2017-12-08 | 百度在线网络技术（北京）有限公司 | Assess method, apparatus, equipment and the computer-readable storage medium of wagon control model |
| CN107563332A (en) * | 2017-09-05 | 2018-01-09 | 百度在线网络技术（北京）有限公司 | For the method and apparatus for the driving behavior for determining unmanned vehicle |
| CN107697070A (en) * | 2017-09-05 | 2018-02-16 | 百度在线网络技术（北京）有限公司 | Driving behavior Forecasting Methodology and device, unmanned vehicle |
| CN107783943A (en) * | 2017-09-05 | 2018-03-09 | 百度在线网络技术（北京）有限公司 | A kind of appraisal procedure and device of the longitudinally controlled model of end-to-end automated driving system |
| CN108573357A (en) * | 2018-05-08 | 2018-09-25 | 清华大学 | Driving risk real time evaluating method based on equivalent force and its device |
| CN109189985A (en) * | 2018-08-17 | 2019-01-11 | 北京达佳互联信息技术有限公司 | Text style processing method, device, electronic equipment and storage medium |
| CN109285348A (en) * | 2018-10-26 | 2019-01-29 | 深圳大学 | A kind of vehicle behavior recognition methods and system based on two-way length memory network in short-term |
-
2019
- 2019-02-19 CN CN201910122748.1A patent/CN109739245A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1775601A (en) * | 2005-11-18 | 2006-05-24 | 吉林大学 | Vehicle driving trace predicating and lane deviation evaluating method |
| CN107450511A (en) * | 2017-06-28 | 2017-12-08 | 百度在线网络技术（北京）有限公司 | Assess method, apparatus, equipment and the computer-readable storage medium of wagon control model |
| CN107563332A (en) * | 2017-09-05 | 2018-01-09 | 百度在线网络技术（北京）有限公司 | For the method and apparatus for the driving behavior for determining unmanned vehicle |
| CN107697070A (en) * | 2017-09-05 | 2018-02-16 | 百度在线网络技术（北京）有限公司 | Driving behavior Forecasting Methodology and device, unmanned vehicle |
| CN107783943A (en) * | 2017-09-05 | 2018-03-09 | 百度在线网络技术（北京）有限公司 | A kind of appraisal procedure and device of the longitudinally controlled model of end-to-end automated driving system |
| CN108573357A (en) * | 2018-05-08 | 2018-09-25 | 清华大学 | Driving risk real time evaluating method based on equivalent force and its device |
| CN109189985A (en) * | 2018-08-17 | 2019-01-11 | 北京达佳互联信息技术有限公司 | Text style processing method, device, electronic equipment and storage medium |
| CN109285348A (en) * | 2018-10-26 | 2019-01-29 | 深圳大学 | A kind of vehicle behavior recognition methods and system based on two-way length memory network in short-term |
Non-Patent Citations (2)
| Title |
|---|
| 李国豪: "《基于3D CNN-DDPG端到端无人驾驶控制》", 《电子设计工程》 * |
| 金凡: "《基于时空递归神经网络的智能车端到端决策研究》", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110406530A (en) * | 2019-07-02 | 2019-11-05 | 宁波吉利汽车研究开发有限公司 | A kind of automatic Pilot method, apparatus, equipment and vehicle |
| CN112346982A (en) * | 2020-11-11 | 2021-02-09 | 东软睿驰汽车技术（沈阳）有限公司 | Test case generation method and device and electronic equipment |
| CN114415672A (en) * | 2021-01-19 | 2022-04-29 | 百度（美国）有限责任公司 | Dynamic model evaluation for autonomously driven vehicles |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| JP6869562B2 (en) | A method of tracking an object using a CNN including a tracking network, and a device using it {METHOD FOR TRACKING OBJECT BY USING CONVOLUTIONAL NEURAL NETWORK INCLUDING TRACKING NETWORK AND COMPUTING | |
| CN109739245A (en) | One kind being based on unpiloted end to end model appraisal procedure and device | |
| CN103049792B (en) | Deep-neural-network distinguish pre-training | |
| CN109829577A (en) | Track train operating status prediction technique based on deep neural network structural model | |
| US10671077B2 (en) | System and method for full-stack verification of autonomous agents | |
| Bi et al. | A data-driven model for lane-changing in traffic simulation. | |
| CN107697070A (en) | Driving behavior Forecasting Methodology and device, unmanned vehicle | |
| CN107310550A (en) | Road vehicles travel control method and device | |
| CN113704956A (en) | Urban road online microscopic simulation method and system based on digital twin technology | |
| CN110555476B (en) | Intelligent vehicle lane change track prediction method suitable for man-machine hybrid driving environment | |
| Zhou et al. | Learning the Car‐following Behavior of Drivers Using Maximum Entropy Deep Inverse Reinforcement Learning | |
| CN114194211B (en) | Automatic driving method and device, electronic equipment and storage medium | |
| Bach et al. | Reactive-replay approach for verification and validation of closed-loop control systems in early development | |
| JP2022089806A (en) | Modeling operation profiles of vehicle | |
| Yang et al. | Longitudinal tracking control of vehicle platooning using DDPG-based PID | |
| Zhu et al. | Interaction-aware cut-in trajectory prediction and risk assessment in mixed traffic | |
| Amirian et al. | Data-driven crowd simulation with generative adversarial networks | |
| CN115742855A (en) | Electric automobile remaining mileage prediction method and device, electric automobile and medium | |
| Ramakrishna et al. | Anti-carla: An adversarial testing framework for autonomous vehicles in carla | |
| CN115496201A (en) | Train accurate parking control method based on deep reinforcement learning | |
| CN116861370B (en) | Motion data processing method for underground explosion-proof rubber-tyred vehicle | |
| CN111507499B (en) | Method, device and system for constructing model for prediction and testing method | |
| CN111881121A (en) | Automatic driving data filling method and device | |
| CN111341102A (en) | Motion primitive library construction method and device and motion primitive connection method and device | |
| CN115499467B (en) | Intelligent network vehicle connection test platform based on digital twinning and building method and system thereof |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190510 |