CN106875511A - A kind of method for learning driving style based on own coding regularization network - Google Patents
A kind of method for learning driving style based on own coding regularization network Download PDFInfo
- Publication number
- CN106875511A CN106875511A CN201710124624.8A CN201710124624A CN106875511A CN 106875511 A CN106875511 A CN 106875511A CN 201710124624 A CN201710124624 A CN 201710124624A CN 106875511 A CN106875511 A CN 106875511A
- Authority
- CN
- China
- Prior art keywords
- coding
- layer
- stroke
- driver
- regularization
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/2431—Multiple classes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G07—CHECKING-DEVICES
- G07C—TIME OR ATTENDANCE REGISTERS; REGISTERING OR INDICATING THE WORKING OF MACHINES; GENERATING RANDOM NUMBERS; VOTING OR LOTTERY APPARATUS; ARRANGEMENTS, SYSTEMS OR APPARATUS FOR CHECKING NOT PROVIDED FOR ELSEWHERE
- G07C5/00—Registering or indicating the working of vehicles
- G07C5/08—Registering or indicating performance data other than driving, working, idle, or waiting time, with or without registering driving, working, idle or waiting time
- G07C5/0841—Registering performance data
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
A kind of method for being learnt driving style based on own coding regularization network proposed in the present invention, its main contents are included:Gps data conversion, own coding regularization network, object function and approximate, run-length encoding framework, driver's quantity survey, its process is, in one group of unknown stroke, it is input into the gps data of automobile, it is used as network inputs by building statistical nature matrix, unsupervised autocoder is introduced using the label of limited training set as priori, rebuild hidden layer RNN features, the bottleneck layer of regularization own coding structure is extracted as final driving style character representation layer, so as to estimate the quantity of driver in stroke.The present invention breaches the limitation that existing method is difficult to describe unknown driver's driving style, an own coding regularization network is devised directly from the driving habit of gps data learner driver, the accuracy to different drivers identification classification is improve, for auxiliary, the design of automated driving system provide safer accurate method.
Description
Technical field
The present invention relates to drive identification field, wind is driven based on the study of own coding regularization network more particularly, to a kind of
The method of lattice.
Background technology
Drive identification to be usually used in the automatic Pilot of car insurance industry and automobile and aid in the fields such as driving, identification drives
Member's identity and estimation driver's number.Specifically, various fields particularly feel emerging to the driving style information from gps data
Interest, because good identification can be answered such as, and whether accident moment driver's driver behavior is proper and how many people shares this
The problems such as car etc., reliable risk assessment can be done accordingly, directly affect the payable declaration form expense of insurance company.Except this it
Outward, good driving style is represented helps preferably to model and understand the behavior that people drive, and helps to improve auxiliary, automatic
The design of control loop.Although existing method can well describe unknown stroke known to driver, when driver is
In the case that new person fails to be trained to collection cognition, the effect for being showed even is not so good as people's will, on the other hand, it is also difficult to collect foot
Enough Driver datas, and sufficiently large stroke training set is difficult to ensure that for each driver, as number increases,
The difficulty of the training of grader is again corresponding to be improved.
The present invention proposes a kind of method for learning driving style based on own coding regularization network, using own coding canonical
Change deep neural network (AutoReNet) and run-length encoding framework, supervision and unsupervised feature learning are combined in a framework
In, directly from the driving behavior of GPS recording learning drivers.Specifically, in one group of unknown stroke, the GPS numbers of automobile are input into
According to, be used as network inputs by building statistical nature matrix, using the label of limited training set as priori introduce it is unsupervised from
Dynamic encoder, rebuilds hidden layer RNN features, and the bottleneck layer of regularization own coding structure is extracted as into final driving style mark sheet
Show layer, so as to estimate the quantity of driver in this section of stroke.The present invention breaches existing method and is difficult to describe unknown driver to drive
The limitation of sailing lattice, devises an own coding regularization network directly from the driving habit of gps data learner driver, improves
To the accuracy of different drivers identification classification, for auxiliary, the design of automated driving system provide safer accurate side
Method.
The content of the invention
It is difficult to describe the problem of unknown driver's driving style for existing method, it is an object of the invention to provide one kind
The method for learning driving style based on own coding regularization network, using own coding regularization deep neural network
(AutoReNet) and run-length encoding framework, supervision and unsupervised feature learning are combined in a framework, is directly remembered from GPS
The driving behavior of learner driver is recorded, the accuracy to different drivers identification classification is improve.
To solve the above problems, the present invention provides a kind of method for learning driving style based on own coding regularization network,
Its main contents includes:
(1) gps data conversion;
(2) own coding regularization network;
(3) object function and approximate;
(4) run-length encoding framework;
(5) driver's quantity survey.
Wherein, described gps data conversion, stroke (i.e. GPS is defined by the tuple sequence (u, v, t) of length change
Track), wherein (u, v) represents geographical position, t represents the time, and neutral net input, a stroke are built from raw GPS data
It is first divided into regular length LsAnd by LsThe window of/2 displacements, each section of coding derived from adjacent gps data five it is instantaneous
Motor racing feature, i.e. essential characteristic:Speed norm, the difference of speed norm, acceleration norm, the difference of acceleration norm and angle
Speed, then each section further apply length Lf(Lf< Ls) and by LfThe sliding window of/2 displacements, each window is produced
One frame includes seven essential characteristic statistical values:Average value, minimum value, maximum, 25%, 50% and 75% quartile and standard
Deviation, can then obtain 5 × 7=35 rows and 2 × L from given strokes/LfOne group of statistical nature matrix of row.
Further, described statistical nature matrix, the eigenmatrix for describing a trip segment is defined as being input to god
Through a sample of network, for example, the gps data such as given sampling per second once, uses Ls=256s, Lf=4s, defines size
It is 35 × 128 eigenmatrix as network inputs, the stroke belonging to the tag inheritance from correspondent section of sample.
Wherein, described own coding regularization network (AutoRNet), own coding regularization network (AutoRNet) is directly
From gps data learning, in AutoReNet, the output of supervised learning by the use of Recognition with Recurrent Neural Network (RNN) is hidden as shared
Layer, is rebuild with unsupervised combinations of features, and supervision message is incorporated into unsupervised feature learning:The label of limited training sample is
Unsupervised autocoder brings priori into so that the essential characteristic of study is more meaningful and more ability to see things in their true light,
AutoReNet frameworks include three parts:Stack Recognition with Recurrent Neural Network (stack RNN), for the autocoder and use rebuild
Returned in the Softmax of classification
Further, described stack Recognition with Recurrent Neural Network (stack RNN), 35 × 128 input, i.e., one are represented with x
Trip segment, stack RNN reads x to extract higher level another characteristic, because driving style is typically the ageing of driver behavior,
So sequences of the x for length 128, each element therein is 35 dimensional vectors, herein using the gating cycle unit of 2 layer stacks
(GRU) framework carrys out application order dependence, and first GRU layers (gru1) reads the input x that size is for 35 × 128, along time shaft
Launch 128 steps, and export the sequence that identical length is 128, wherein each element is a vector, the size of vector (is tieed up
Degree) quantity of hidden unit in gru1 is equal to, second GRU layers (gru2) is affixed to gru1, and it also launches 128 steps, but not
It is to export a sequence and simply a vector, the size of vector is equal to the quantity of hidden unit in gru2.One discarding layer quilt
Gru2 is applied to reduce over-fitting, the layer of discarding makes up supervision and unsupervised learning, order as hiding characteristic layer is sharedTable
Show the output of given input x.
Further, described autocoder, carries out feature reconstruction, using autocoding using 3 layers of autocoder
Think highly of and buildRather than x, be so conducive to preferably learning driving style, bottleneck layer (fc1) being fully connected is used for study
Compression expression s, in fc1 using amendment linear unit (ReLU) non-linear f (z)=max (0, z) to ensure s non-negative, will be by
Used in trip2vec codings (see formula (6)), by l1Sparse regularization is applied to s (see formula (1)), and full articulamentum (fc2) is
The output layer of autocoder, wherein f (z)=tanh (z) are used for approximateFor rebuilding.
Further, described Softmax is returned, and Softmax recurrence is the extension that logic (Logistic) is returned, logic
Return for solving two class regression problems, and it is to solve the problems, such as that (multiple drives multiple centrifugal pumps to return purpose using Softmax
Member), the layer (fc3) that is fully connected that will be attached to abandon layer is used to classify, and dividing for class label is produced using Softmax recurrence
Cloth, the quantity (being represented by c) of class is equal to the driver's quantity in training set.
Wherein, described object function and approximate, given training set { xi, yi, its i ∈ { 1 ..., n }, yi∈ 1 ...,
C }, entry function is defined as rebuilding the combination with class object, rebuilds loss and is defined as:
WhereinIt is the output of RNN discarding layers,It is k vectorial " traversal ",It is correspondence's
" code vector ",First term be reconstruction error, it is intended to find traversalWith new expression siTo rebuild the study from RNN
Featurel1Regularization is used for sparse si, formula (1) is sparse code target.Classification Loss is defined as standard cross entropy:
Wherein 1 { } is target function, θ={ θ1..., θcIt is Softmax regression parameters, sum up, composite object
Function is defined as:
If excluding autocoder layer (fc1+fc2), network becomes a stack RNN, andCan serve as x's
Character representation, by this way,Study only instructed by the supervision message of stroke label (i.e. driver ID), abandoning layer can be with
Help reduces over-fitting, but the quantity of unknown driver again may be very big, and training data is limited, and study still can cross plan
Close the known driver in training set, it is difficult to represent unknown driver well, therefore by minimumOne tool of habit of attending a school by taking daily trips
There is the expression of cluster feature, the K-means based on Classic Clustering Algorithms:
It is intended to find out cluster barycenter μkSo that the distance between data point and nearest barycenter minimum, similarly formula (4),
K-means can be regarded as rebuilding xiA kind of method:
Compare formula (1) and formula (5), their reconstruction targets to same type are optimized, only difference is that public
Formula (1) is allowed in each siIt is middle to there is more than one nonzero term, enabling more accurately to represent each xi, therefore, minimizeThe expression of study can be made has Clustering features and more compact, in order to minimizeIn AutoReNet, layer is used
Dropout+fc1+fc2 goes approximately to obtain a unified structure and is easier training, using the sparse reconstruction of formula (1) study andEfficient coding, share sparse autocoder, dropout+fc1+fc2 is typical own coding structure, and fc1's
L is used in output1Regularization, it is sparseCoding.
Wherein, described run-length encoding framework, once trained AutoReNet, just can be used layer gru1+gru2+
Dropout+fc1 as stroke segment scrambler, but, it be from section level run-length data in extract driving style information, that
Just can still be influenceed by the place such as road shape and transportation condition factor, therefore be proposed a run-length encoding frame
Frame, based on training stroke segment scrambler and use bag of words (BoW) latent structure strategy, it is popular for i.e. can be random length
The stroke of degree is as one " article ", and each section regards one " paragraph " as, and whole " theme " of article just can be by polymerization
Paragraph level information is derived, and similarly, based on section level driving style, is determined using the normalization and value of all Utterance level features vector
Adopted stroke level driving style is represented, it is assumed that a stroke trQ sections is divided into, and coding section is characterized inThen stroke level driving style character representation is defined as:
WhereinIt is vector sum,Represent that its jth is tieed up
Wherein, described driver's quantity survey, in one group of unknown stroke, is input into the gps data of automobile, by building
Statistical nature matrix introduces unsupervised autocoder, weight as network inputs using the label of limited training set as priori
Hidden layer RNN features are built, the bottleneck layer of regularization own coding structure final driving style character representation layer is extracted as, so as to estimate
Count the quantity of driver in this section of stroke.
Brief description of the drawings
Fig. 1 is a kind of system flow chart of the method based on own coding regularization network study driving style of the present invention.
Fig. 2 is a kind of work of the GLMB wave filters of the method based on own coding regularization network study driving style of the present invention
Make functional-block diagram.
Fig. 3 is a kind of multiple target tracking effect of the method based on own coding regularization network study driving style of the present invention
Figure.
Fig. 4 is a kind of multiple target tracking flow of the method based on own coding regularization network study driving style of the present invention
Figure.
Specific embodiment
It should be noted that in the case where not conflicting, the feature in embodiment and embodiment in the application can phase
Mutually combine, the present invention is described in further detail with specific embodiment below in conjunction with the accompanying drawings.
Fig. 1 is a kind of system flow chart of the method based on own coding regularization network study driving style of the present invention.It is main
To include that gps data conversion, own coding regularization network, object function and approximate, run-length encoding framework, driver's quantity are estimated
Meter.
Wherein, described gps data conversion, stroke (i.e. GPS is defined by the tuple sequence (u, v, t) of length change
Track), wherein (u, v) represents geographical position, t represents the time, and neutral net input, a stroke are built from raw GPS data
It is first divided into regular length LsAnd by LsThe window of/2 displacements, each section of coding derived from adjacent gps data five it is instantaneous
Motor racing feature, i.e. essential characteristic:Speed norm, the difference of speed norm, acceleration norm, the difference of acceleration norm and angle
Speed, then each section further apply length Lf(Lf< Ls) and by LfThe sliding window of/2 displacements, each window is produced
One frame includes seven essential characteristic statistical values:Average value, minimum value, maximum, 25%, 50% and 75% quartile and standard
Deviation, can then obtain 5 × 7=35 rows and 2 × L from given strokes/LfOne group of statistical nature matrix of row, describes one
The eigenmatrix of trip segment is defined as being input to a sample of neutral net, for example, the GPS such as given sampling per second once
Data, use Ls=256s, Lf=4s, it is 35 × 128 eigenmatrix as network inputs to define size, the label of sample after
Hold the stroke belonging to correspondent section.
Wherein, described own coding regularization network (AutoReNet), own coding regularization network (AutoReNet) is straight
Connect from gps data learning, in AutoReNet, the output of supervised learning by the use of Recognition with Recurrent Neural Network (RNN) is hidden as sharing
Layer is hidden, is rebuild with unsupervised combinations of features, supervision message is incorporated into unsupervised feature learning:The label of limited training sample
For unsupervised autocoder brings priori into so that the essential characteristic of study is more meaningful and more ability to see things in their true light,
AutoReNet frameworks include three parts:Stack Recognition with Recurrent Neural Network (stack RNN), for the autocoder and use rebuild
Returned in the Softmax of classification
(1) stack Recognition with Recurrent Neural Network (stack RNN), the trip segment of 35 × 128 input, i.e., stack are represented with x
RNN reads x to extract higher level another characteristic, because driving style is typically the ageing of driver behavior, x is length
128 sequence, each element therein is 35 dimensional vectors, is answered using gating cycle unit (GRU) framework of 2 layer stacks herein
With schedule dependence, first GRU layers (gru1) reads the input x that size is for 35 × 128,128 steps is launched along time shaft, and defeated
Go out the sequence that identical length is 128, wherein each element is a vector, the size (i.e. dimension) of vector is hidden equal in gru1
Hide unit quantity, second GRU layer (gru2) is affixed to gru1, and it also launches 128 steps, but be not export a sequence and
A simply vector, the size of vector is equal to the quantity of hidden unit in gru2.One abandons layer and is applied to gru2 to reduce
Over-fitting, the layer of discarding makes up supervision and unsupervised learning, order as hiding characteristic layer is sharedRepresent that given input x's is defeated
Go out.
(2) autocoder, feature reconstruction is carried out using 3 layers of autocoder, is rebuild using autocoderRather than
X, is so conducive to preferably learning driving style, and bottleneck layer (fc1) being fully connected is used for studyCompression expression s,
In fc1 using amendment linear unit (ReLU) non-linear f (z)=max (0, z) to ensure s non-negative, will be used in trip2vec volume
Code (see formula (6)), by l1Sparse regularization is applied to s (see formula (1)), and full articulamentum (fc2) is the defeated of autocoder
Go out layer, wherein f (z)=tanh (z) is used for approximateFor rebuilding.
(3) Softmax is returned, and Softmax recurrence is the extension that logic (Logistic) is returned, and logistic regression is used to solve
Two class regression problems, and it is to solve the problems, such as multiple centrifugal pumps (multiple drivers) to return purpose using Softmax, will be attached to lose
That abandons layer is fully connected layer (fc3) for classifying, and the distribution of class label is produced using Softmax recurrence, and the quantity of class is (by c
Represent) it is equal to the driver's quantity in training set.
Wherein, described object function and approximate, given training set { xi, yi, its i ∈ { 1 ..., n }, yi∈ 1 ...,
C }, entry function is defined as rebuilding the combination with class object, rebuilds loss and is defined as:
WhereinIt is the output of RNN discarding layers,It is k vectorial " traversal ",It is correspondence's
" code vector ",First term be reconstruction error, it is intended to find traversalWith new expression siTo rebuild the study from RNN
Featurel1Regularization is used for sparse si, formula (1) is sparse code target.Classification Loss is defined as standard cross entropy:
Wherein 1 { } is target function, θ={ θ1..., θcIt is Softmax regression parameters, sum up, composite object
Function is defined as:
If excluding autocoder layer (fc1+fc2), network becomes a stack RNN, andCan serve as x's
Character representation, by this way,Study only instructed by the supervision message of stroke label (i.e. driver ID), abandoning layer can be with
Help reduces over-fitting, but the quantity of unknown driver again may be very big, and training data is limited, and study still can cross plan
Close the known driver in training set, it is difficult to represent unknown driver well, therefore by minimumOne tool of habit of attending a school by taking daily trips
There is the expression of cluster feature, the K-means based on Classic Clustering Algorithms:
It is intended to find out cluster barycenter μkSo that the distance between data point and nearest barycenter minimum, similarly formula (4),
K-means can be regarded as rebuilding xiA kind of method:
Compare formula (1) and formula (5), their reconstruction targets to same type are optimized, only difference is that public
Formula (1) is allowed in each siIt is middle to there is more than one nonzero term, enabling more accurately to represent each xi, therefore, minimizeThe expression of study can be made has Clustering features and more compact, in order to minimizeIn AutoReNet, layer is used
Dropout+fc1+fc2 goes approximately to obtain a unified structure and is easier training, using the sparse reconstruction of formula (1) study andEfficient coding, share sparse autocoder, dropout+fc1+fc2 is typical own coding structure, and fc1's
L is used in output1Regularization, it is sparseCoding.
Wherein, described run-length encoding framework, once trained AutoReNet, just can be used layer gru1+gru2+
Dropout+fc1 as stroke segment scrambler, but, it be from section level run-length data in extract driving style information, that
Just can still be influenceed by the place such as road shape and transportation condition factor, therefore be proposed a run-length encoding frame
Frame, based on training stroke segment scrambler and use bag of words (BoW) latent structure strategy, it is popular for i.e. can be random length
The stroke of degree is as one " article ", and each section regards one " paragraph " as, and whole " theme " of article just can be by polymerization
Paragraph level information is derived, and similarly, based on section level driving style, is determined using the normalization and value of all Utterance level features vector
Adopted stroke level driving style is represented, it is assumed that a stroke trQ sections is divided into, and coding section is characterized inThen stroke level driving style character representation is defined as:
WhereinIt is vector sum,Represent that its jth is tieed up
Wherein, described driver's quantity survey, in one group of unknown stroke, is input into the gps data of automobile, by building
Statistical nature matrix introduces unsupervised autocoder, weight as network inputs using the label of limited training set as priori
Hidden layer RNN features are built, the bottleneck layer of regularization own coding structure final driving style character representation layer is extracted as, so as to estimate
Count the quantity of driver in this section of stroke.
Fig. 2 is a kind of AutoReNet frameworks of the method based on own coding regularization network study driving style of the present invention
Figure.Directly from gps data learning, in AutoReNet, supervised learning is utilized own coding regularization network (AutoReNet)
The output of Recognition with Recurrent Neural Network (RNN) is rebuild as shared hidden layer with unsupervised combinations of features, and supervision message is incorporated into nothing
The feature learning of supervision:The label of limited training sample is that unsupervised autocoder brings priori into so that study
Essential characteristic is more meaningful and more ability to see things in their true light, and AutoReNet frameworks include three parts:Stack Recognition with Recurrent Neural Network (stack
RNN), for the autocoder rebuild and the Softmax recurrence for classification.
Wherein, stack Recognition with Recurrent Neural Network (stack RNN), the trip segment of 35 × 128 input, i.e., stack are represented with x
RNN reads x to extract higher level another characteristic, because driving style is typically the ageing of driver behavior, x is length
128 sequence, each element therein is 35 dimensional vectors, is answered using gating cycle unit (GRU) framework of 2 layer stacks herein
With schedule dependence, first GRU layers (gru1) reads the input x that size is for 35 × 128,128 steps is launched along time shaft, and defeated
Go out the sequence that identical length is 128, wherein each element is a vector, the size (i.e. dimension) of vector is hidden equal in gru1
Hide unit quantity, second GRU layer (gru2) is affixed to gru1, and it also launches 128 steps, but be not export a sequence and
A simply vector, the size of vector is equal to the quantity of hidden unit in gru2.One abandons layer and is applied to gru2 to reduce
Over-fitting, the layer of discarding makes up supervision and unsupervised learning, order as hiding characteristic layer is sharedRepresent that given input x's is defeated
Go out.
Wherein, autocoder, feature reconstruction is carried out using 3 layers of autocoder, is rebuild using autocoderWithout
It is x, is so conducive to learning more preferable broad sense driving style representing, bottleneck layer (fc1) being fully connected is used for studyPressure
Contracting represent s, in fc1 using amendment linear unit (ReLU) non-linear f (z)=max (0, z) to ensure s non-negative, will be used in
Trip2vec is encoded (see formula (6)), by l1Sparse regularization is applied to s (see formula (1)), and full articulamentum (fc2) is automatic
The output layer of encoder, wherein f (z)=tanh (z) are used for approximateFor rebuilding.
Wherein, Softmax is returned, and Softmax recurrence is the extension that logic (Logistic) is returned, and logistic regression is used to solve
Certainly two class regression problem, and it is solves the problems, such as multiple centrifugal pumps (multiple drivers) to return purpose using Softmax, be will be attached to
Abandon layer is fully connected layer (fc3) for classifying, and produces the distribution of class label using Softmax recurrence, the quantity of class (by
C is represented) it is equal to the driver's quantity in training set.
Fig. 3 is a kind of run-length encoding framework of the method based on own coding regularization network study driving style of the present invention
Figure.Once trained AutoReNet, layer gru1+gru2+dropout+fc1 just can be used as stroke segment scrambler, based on instruction
Experienced stroke segment scrambler and using the strategy of bag of words (BoW) latent structure, it is popular for i.e. can be the stroke of indefinite length
As one " article ", and each section regards one " paragraph " as, and whole " theme " of article can just be believed by the paragraph level that is polymerized
Cease to derive, similarly, based on section level driving style, stroke level is defined using the normalization and value of all Utterance level features vector
Driving style is represented, it is assumed that a stroke trQ sections is divided into, and coding section is characterized inThen stroke
Level driving style character representation is defined as formula (6) whereinIt is vector sum,Represent that its jth is tieed up
Fig. 4 is a kind of multiple target tracking flow of the method based on own coding regularization network study driving style of the present invention
Figure.In one group of unknown stroke, by being input into the gps data of automobile, statistical nature matrix is built as network inputs, will be limited
The label of training set introduces unsupervised autocoder as priori, hidden layer RNN features is rebuild, by regularization own coding knot
The bottleneck layer of structure is extracted as final driving style character representation layer, so as to estimate the quantity of driver in this section of stroke.
For those skilled in the art, the present invention is not restricted to the details of above-described embodiment, without departing substantially from essence of the invention
In the case of god and scope, the present invention can be realized with other concrete forms.Additionally, those skilled in the art can be to this hair
Bright to carry out various changes and modification without departing from the spirit and scope of the present invention, these improvement also should be regarded as of the invention with modification
Protection domain.Therefore, appended claims are intended to be construed to include preferred embodiment and fall into all changes of the scope of the invention
More and modification.
Claims (10)
1. a kind of method for learning driving style based on own coding regularization network, it is characterised in that main to turn including gps data
Change (one);Own coding regularization network (two);Object function and approximate (three);Run-length encoding framework (four);Driver's quantity is estimated
Meter (five).
2. based on gps data conversion () described in claims 1, it is characterised in that by the tuple sequence of length change
(u, v, t) defines stroke (i.e. GPS track), wherein (u, v) represents geographical position, t represents the time, from raw GPS data structure
Neutral net input is built, a stroke is first divided into regular length LsAnd by LsThe window of/2 displacements, each section coding is from adjacent
Five derived from gps data instantaneous motor racing features, i.e. essential characteristic:Speed norm, the difference of speed norm, acceleration model
Number, the difference of acceleration norm and angular speed, then each section further apply length Lf(Lf< Ls) and by Lf/ 2 displacements
Sliding window, each window produce a frame include seven essential characteristic statistical values:Average value, minimum value, maximum, 25%,
50% and 75% quartile and standard deviation, can then obtain 5 × 7=35 rows and 2 × L from given strokes/LfRow
One group of statistical nature matrix.
3. based on the statistical nature matrix described in claims 2, it is characterised in that one eigenmatrix quilt of trip segment of description
It is defined as being input to a sample of neutral net, for example, the gps data such as given sampling per second once, uses Ls=256s,
Lf=4s, definition size is 35 × 128 eigenmatrix as network inputs, the row belonging to the tag inheritance from correspondent section of sample
Journey.
4. based on the own coding regularization network (AutoReNet) (two) described in claims 1, it is characterised in that own coding
Directly from gps data learning, in AutoReNet, supervised learning is utilized and circulates nerve net regularization network (AutoReNet)
The output of network (RNN) is rebuild as shared hidden layer with unsupervised combinations of features, and supervision message is incorporated into unsupervised feature
Study:The label of limited training sample is that unsupervised autocoder brings priori into so that the essential characteristic of study is more
Meaningful and more ability to see things in their true light, AutoReNet frameworks include three parts:Stack Recognition with Recurrent Neural Network (stack RNN), for weight
The autocoder and the Softmax for classifying built are returned.
5. based on stack Recognition with Recurrent Neural Network (stack RNN) described in claims 4, it is characterised in that 35 are represented with x ×
The trip segment of 128 input, i.e., one, stack RNN reads x to extract higher level another characteristic, because driving style is typically to drive
The ageing of action is sailed, so sequences of the x for length 128, each element therein is 35 dimensional vectors, herein using 2 layers of heap
Gating cycle unit (GRU) framework of stack carrys out application order dependence, and it is 35 × 128 that first GRU layers (gru1) reads size
Input x, launch 128 steps along time shaft, and export the sequence that identical length is 128, wherein each element is a vector,
The size (i.e. dimension) of vector is equal to the quantity of hidden unit in gru1, and second GRU layers (gru2) is affixed to gru1, it
Launch 128 steps, but be not to export a sequence and simply a vector, the size of vector is equal to the number of hidden unit in gru2
Amount, one abandons layer and is applied to gru2 to reduce over-fitting, and the layer of discarding makes up supervision and nothing as hiding characteristic layer is shared
Supervised learning, orderRepresent the output of given input x.
6. based on the autocoder described in claims 4, it is characterised in that carry out feature weight using 3 layers of autocoder
Build, rebuild using autocoderRather than x, be so conducive to preferably learning driving style, the bottleneck layer being fully connected
(fc1) it is used for studyCompression expression s, in fc1 using amendment linear unit (ReLU) non-linear f (z)=max (0, z)
To ensure s non-negative, trip2vec codings will be used in (see formula (6)), by l1Sparse regularization is applied to s (see formula (1)),
Full articulamentum (fc2) is the output layer of autocoder, and wherein f (z)=tanh (z) is used for approximateFor rebuilding.
7. returned based on the Softmax described in claims 4, it is characterised in that it is logic (Logistic) that Softmax is returned
The extension of recurrence, logistic regression is used to solve two class regression problems, and it is to solve multiple centrifugal pumps to return purpose using Softmax
Problem (multiple drivers), the layer (fc3) that is fully connected that will be attached to abandon layer is used to classify, and is produced using Softmax recurrence
The distribution of raw class label, the quantity (being represented by c) of class is equal to the driver's quantity in training set.
8. based on the object function described in claims 1 and approximately (three), it is characterised in that given training set { xi, yi, its i
∈ { 1 ..., n }, yi∈ { 1 ..., c }, entry function is defined as rebuilding the combination with class object, rebuilds loss definition
For:
WhereinIt is the output of RNN discarding layers,It is k vectorial " traversal ",It is correspondence" code
Vector ",First term be reconstruction error, it is intended to find traversalWith new expression siTo rebuild the learning characteristic from RNNl1Regularization is used for sparse si, formula (1) is sparse code target, and Classification Loss is defined as standard cross entropy:
Wherein 1 { } is target function, θ={ θ1..., θcIt is Softmax regression parameters, sum up, composite object function
It is defined as:
If excluding autocoder layer (fc1+fc2), network becomes a stack RNN, andCan serve as the feature of x
Represent, by this way,Study only by stroke label (i.e. driver ID) supervision message instruct, abandon layer can help
Over-fitting is reduced, but the quantity of unknown driver again may be very big, and training data is limited, study still can over-fitting instruction
Practice the known driver for concentrating, it is difficult to represent unknown driver well, therefore by minimumHabit of attending a school by taking daily trips one has poly-
The expression of category feature, the K-means based on Classic Clustering Algorithms:
It is intended to find out cluster barycenter μkSo that the distance between data point and nearest barycenter minimum, similarly formula (4), K-
Means can be regarded as rebuilding xiA kind of method:
Compare formula (1) and formula (5), their reconstruction targets to same type are optimized, only difference is that formula (1)
Allow in each siIt is middle to there is more than one nonzero term, enabling more accurately to represent each xi, therefore, minimizeCan
So that the expression of study has Clustering features and more compact, in order to minimizeIn AutoReNet, layer dropout is used
+ fc1+fc2 goes approximately to obtain a unified structure and is easier training, using the sparse reconstruction of formula (1) study andHave
Effect coding, shares sparse autocoder, and dropout+fc1+fc2 is typical own coding structure, and in the output of fc1
Use l1Regularization, it is sparseCoding.
9. based on the run-length encoding framework (four) described in claims 1, it is characterised in that once AutoReNet is trained, just
Layer gru1+gru2+dropout+fc1 can be used as stroke segment scrambler, but, it is to be carried from the run-length data of section level
Take driving style information, then just can still be influenceed by the place such as road shape and transportation condition factor, therefore proposed
One run-length encoding framework, the stroke segment scrambler based on training and using the strategy of bag of words (BoW) latent structure is popular
Saying i.e. can be the stroke of indefinite length as one " article ", and each section regards one " paragraph " as, the whole " master of article
Topic " can just be derived by being polymerized paragraph level information, similarly, based on section level driving style, using all Utterance level features to
The normalization of amount and value are represented defining stroke level driving style, it is assumed that a stroke trQ sections is divided into, and coding section is special
Levying isThen stroke level driving style character representation is defined as:
WhereinIt is vector sum,Represent that its jth is tieed up
10., based on the driver's quantity survey (five) described in claims 1, it is characterised in that in one group of unknown stroke, lead to
The gps data of input automobile is crossed, statistical nature matrix is built as network inputs, the label of limited training set is drawn as priori
Enter unsupervised autocoder, rebuild hidden layer RNN features, the bottleneck layer of regularization own coding structure is extracted as finally driving
Sailing lattice character representation layer, so as to estimate the quantity of driver in this section of stroke.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710124624.8A CN106875511A (en) | 2017-03-03 | 2017-03-03 | A kind of method for learning driving style based on own coding regularization network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710124624.8A CN106875511A (en) | 2017-03-03 | 2017-03-03 | A kind of method for learning driving style based on own coding regularization network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106875511A true CN106875511A (en) | 2017-06-20 |
Family
ID=59170427
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710124624.8A Pending CN106875511A (en) | 2017-03-03 | 2017-03-03 | A kind of method for learning driving style based on own coding regularization network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106875511A (en) |
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108282262A (en) * | 2018-04-16 | 2018-07-13 | 西安电子科技大学 | Intelligent clock signal sorting technique based on gating cycle unit depth network |
CN108320051A (en) * | 2018-01-17 | 2018-07-24 | 哈尔滨工程大学 | A kind of mobile robot dynamic collision-free planning method based on GRU network models |
CN108319980A (en) * | 2018-02-05 | 2018-07-24 | 哈工大机器人(合肥)国际创新研究院 | A kind of recurrent neural network multi-tag learning method based on GRU |
CN109117945A (en) * | 2017-06-22 | 2019-01-01 | 上海寒武纪信息科技有限公司 | Processor and its processing method, chip, chip-packaging structure and electronic device |
CN109670597A (en) * | 2017-09-20 | 2019-04-23 | 顾泽苍 | A kind of more purpose control methods of the machine learning of automatic Pilot |
CN110097008A (en) * | 2019-04-30 | 2019-08-06 | 苏州大学 | A kind of human motion recognition method |
CN110471411A (en) * | 2019-07-26 | 2019-11-19 | 华为技术有限公司 | Automatic Pilot method and servomechanism |
CN111104953A (en) * | 2018-10-25 | 2020-05-05 | 北京嘀嘀无限科技发展有限公司 | Driving behavior feature detection method and device, electronic equipment and computer-readable storage medium |
WO2020119363A1 (en) * | 2018-12-13 | 2020-06-18 | 华为技术有限公司 | Automatic driving method, training method and related apparatuses |
CN111413957A (en) * | 2018-12-18 | 2020-07-14 | 北京航迹科技有限公司 | System and method for determining driving actions in autonomous driving |
CN111443701A (en) * | 2018-12-29 | 2020-07-24 | 南京理工大学 | Unmanned vehicle/robot behavior planning method based on heterogeneous deep learning |
CN111882114A (en) * | 2020-07-01 | 2020-11-03 | 长安大学 | Short-term traffic flow prediction model construction method and prediction method |
CN112533136A (en) * | 2020-11-26 | 2021-03-19 | 南京工业大学 | WLAN fingerprint positioning method based on deep learning |
CN112559968A (en) * | 2020-12-09 | 2021-03-26 | 深圳大学 | Driving style representation learning method based on multi-situation data |
CN114332520A (en) * | 2022-03-14 | 2022-04-12 | 中汽信息科技(天津)有限公司 | Abnormal driving behavior recognition model construction method based on deep learning |
US20220292316A1 (en) * | 2021-03-10 | 2022-09-15 | GM Global Technology Operations LLC | Shape-biased image classification using deep convolutional networks |
WO2023173987A1 (en) * | 2022-03-16 | 2023-09-21 | International Business Machines Corporation | Prediction and operational efficiency for system-wide optimization of an industrial processing system |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101633358A (en) * | 2008-07-24 | 2010-01-27 | 通用汽车环球科技运作公司 | Adaptive vehicle control system with integrated driving style recognition |
CN102320301A (en) * | 2010-04-07 | 2012-01-18 | 通用汽车环球科技运作有限责任公司 | Be used to make the ride characteristic of vehicle to adapt to the method for chaufeur conversion |
WO2015074798A1 (en) * | 2013-11-25 | 2015-05-28 | Robert Bosch Gmbh | Method for evaluating the behaviour of a driver in a vehicle |
CN104900063A (en) * | 2015-06-19 | 2015-09-09 | 中国科学院自动化研究所 | Short distance driving time prediction method |
CN105426638A (en) * | 2015-12-24 | 2016-03-23 | 吉林大学 | Driver behavior characteristic identification device |
CN105930625A (en) * | 2016-06-13 | 2016-09-07 | 天津工业大学 | Design method of Q-learning and neural network combined smart driving behavior decision making system |
CN106127586A (en) * | 2016-06-17 | 2016-11-16 | 上海经达信息科技股份有限公司 | Vehicle insurance rate aid decision-making system under big data age |
CN106203626A (en) * | 2016-06-30 | 2016-12-07 | 北京奇虎科技有限公司 | Car steering behavioral value method and device, automobile |
-
2017
- 2017-03-03 CN CN201710124624.8A patent/CN106875511A/en active Pending
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101633358A (en) * | 2008-07-24 | 2010-01-27 | 通用汽车环球科技运作公司 | Adaptive vehicle control system with integrated driving style recognition |
CN102320301A (en) * | 2010-04-07 | 2012-01-18 | 通用汽车环球科技运作有限责任公司 | Be used to make the ride characteristic of vehicle to adapt to the method for chaufeur conversion |
WO2015074798A1 (en) * | 2013-11-25 | 2015-05-28 | Robert Bosch Gmbh | Method for evaluating the behaviour of a driver in a vehicle |
CN104900063A (en) * | 2015-06-19 | 2015-09-09 | 中国科学院自动化研究所 | Short distance driving time prediction method |
CN105426638A (en) * | 2015-12-24 | 2016-03-23 | 吉林大学 | Driver behavior characteristic identification device |
CN105930625A (en) * | 2016-06-13 | 2016-09-07 | 天津工业大学 | Design method of Q-learning and neural network combined smart driving behavior decision making system |
CN106127586A (en) * | 2016-06-17 | 2016-11-16 | 上海经达信息科技股份有限公司 | Vehicle insurance rate aid decision-making system under big data age |
CN106203626A (en) * | 2016-06-30 | 2016-12-07 | 北京奇虎科技有限公司 | Car steering behavioral value method and device, automobile |
Non-Patent Citations (2)
Title |
---|
WEISHAN DONG: "Autoencoder Regularized Network For Driving Style Representation Learning", 《HTTP://ARXIV.ORG》 * |
张磊: "基于驾驶员特性自学习方法的车辆纵向驾驶辅助系统", 《中国博士学位论文全文数据库工程科技Ⅱ辑》 * |
Cited By (27)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109117945A (en) * | 2017-06-22 | 2019-01-01 | 上海寒武纪信息科技有限公司 | Processor and its processing method, chip, chip-packaging structure and electronic device |
CN109670597A (en) * | 2017-09-20 | 2019-04-23 | 顾泽苍 | A kind of more purpose control methods of the machine learning of automatic Pilot |
CN108320051B (en) * | 2018-01-17 | 2021-11-23 | 哈尔滨工程大学 | Mobile robot dynamic collision avoidance planning method based on GRU network model |
CN108320051A (en) * | 2018-01-17 | 2018-07-24 | 哈尔滨工程大学 | A kind of mobile robot dynamic collision-free planning method based on GRU network models |
CN108319980A (en) * | 2018-02-05 | 2018-07-24 | 哈工大机器人(合肥)国际创新研究院 | A kind of recurrent neural network multi-tag learning method based on GRU |
CN108282262A (en) * | 2018-04-16 | 2018-07-13 | 西安电子科技大学 | Intelligent clock signal sorting technique based on gating cycle unit depth network |
CN111104953A (en) * | 2018-10-25 | 2020-05-05 | 北京嘀嘀无限科技发展有限公司 | Driving behavior feature detection method and device, electronic equipment and computer-readable storage medium |
CN111104953B (en) * | 2018-10-25 | 2024-06-07 | 北京嘀嘀无限科技发展有限公司 | Driving behavior feature detection method, driving behavior feature detection device, electronic device and computer-readable storage medium |
WO2020119363A1 (en) * | 2018-12-13 | 2020-06-18 | 华为技术有限公司 | Automatic driving method, training method and related apparatuses |
US11155264B2 (en) | 2018-12-18 | 2021-10-26 | Beijing Voyager Technology Co., Ltd. | Systems and methods for determining driving action in autonomous driving |
CN111413957A (en) * | 2018-12-18 | 2020-07-14 | 北京航迹科技有限公司 | System and method for determining driving actions in autonomous driving |
CN111413957B (en) * | 2018-12-18 | 2021-11-02 | 北京航迹科技有限公司 | System and method for determining driving actions in autonomous driving |
CN111443701A (en) * | 2018-12-29 | 2020-07-24 | 南京理工大学 | Unmanned vehicle/robot behavior planning method based on heterogeneous deep learning |
CN110097008B (en) * | 2019-04-30 | 2021-02-19 | 苏州大学 | Human body action recognition method |
CN110097008A (en) * | 2019-04-30 | 2019-08-06 | 苏州大学 | A kind of human motion recognition method |
CN110471411A (en) * | 2019-07-26 | 2019-11-19 | 华为技术有限公司 | Automatic Pilot method and servomechanism |
CN111882114B (en) * | 2020-07-01 | 2023-10-31 | 长安大学 | Short-time traffic flow prediction model construction method and prediction method |
CN111882114A (en) * | 2020-07-01 | 2020-11-03 | 长安大学 | Short-term traffic flow prediction model construction method and prediction method |
CN112533136A (en) * | 2020-11-26 | 2021-03-19 | 南京工业大学 | WLAN fingerprint positioning method based on deep learning |
CN112559968A (en) * | 2020-12-09 | 2021-03-26 | 深圳大学 | Driving style representation learning method based on multi-situation data |
CN112559968B (en) * | 2020-12-09 | 2023-01-13 | 深圳大学 | Driving style representation learning method based on multi-situation data |
US20220292316A1 (en) * | 2021-03-10 | 2022-09-15 | GM Global Technology Operations LLC | Shape-biased image classification using deep convolutional networks |
US11893086B2 (en) * | 2021-03-10 | 2024-02-06 | GM Global Technology Operations LLC | Shape-biased image classification using deep convolutional networks |
CN114332520A (en) * | 2022-03-14 | 2022-04-12 | 中汽信息科技(天津)有限公司 | Abnormal driving behavior recognition model construction method based on deep learning |
CN114332520B (en) * | 2022-03-14 | 2022-06-17 | 中汽信息科技(天津)有限公司 | Abnormal driving behavior recognition model construction method based on deep learning |
WO2023173987A1 (en) * | 2022-03-16 | 2023-09-21 | International Business Machines Corporation | Prediction and operational efficiency for system-wide optimization of an industrial processing system |
US12066813B2 (en) | 2022-03-16 | 2024-08-20 | International Business Machines Corporation | Prediction and operational efficiency for system-wide optimization of an industrial processing system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106875511A (en) | A kind of method for learning driving style based on own coding regularization network | |
CN105956560B (en) | A kind of model recognizing method based on the multiple dimensioned depth convolution feature of pondization | |
CN104978580B (en) | A kind of insulator recognition methods for unmanned plane inspection transmission line of electricity | |
CN104537393B (en) | A kind of traffic sign recognition method based on multiresolution convolutional neural networks | |
CN104850890B (en) | Instance-based learning and the convolutional neural networks parameter regulation means of Sadowsky distributions | |
CN109242251A (en) | Vehicular behavior safety detecting method, device, equipment and storage medium | |
CN106845541A (en) | A kind of image-recognizing method based on biological vision and precision pulse driving neutral net | |
CN107766850A (en) | Based on the face identification method for combining face character information | |
CN103996056A (en) | Tattoo image classification method based on deep learning | |
CN106971194A (en) | A kind of driving intention recognition methods based on the double-deck algorithms of improvement HMM and SVM | |
CN107730904A (en) | Multitask vehicle driving in reverse vision detection system based on depth convolutional neural networks | |
CN107239730A (en) | The quaternary number deep neural network model method of intelligent automobile Traffic Sign Recognition | |
CN110321862B (en) | Pedestrian re-identification method based on compact ternary loss | |
CN108280397A (en) | Human body image hair detection method based on depth convolutional neural networks | |
CN106778796A (en) | Human motion recognition method and system based on hybrid cooperative model training | |
CN110097029A (en) | Identity identifying method based on Highway network multi-angle of view Gait Recognition | |
CN112046489A (en) | Driving style identification algorithm based on factor analysis and machine learning | |
CN106326873B (en) | The manipulation Intention Anticipation method of CACC driver's limbs electromyography signal characterization | |
CN105718955A (en) | Visual terrain classification method based on multiple encoding and feature fusion | |
Li et al. | Cluster naturalistic driving encounters using deep unsupervised learning | |
CN111797936B (en) | Image emotion classification method and device based on saliency detection and multi-level feature fusion | |
CN110378397A (en) | A kind of driving style recognition methods and device | |
CN105404858A (en) | Vehicle type recognition method based on deep Fisher network | |
Tran et al. | Hyperparameter optimization for improving recognition efficiency of an adaptive learning system | |
CN114241458A (en) | Driver behavior recognition method based on attitude estimation feature fusion |
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: 20170620 |