CN109255339A  Classification method based on adaptive depth forest body gait energy diagram  Google Patents
Classification method based on adaptive depth forest body gait energy diagram Download PDFInfo
 Publication number
 CN109255339A CN109255339A CN201811222012.3A CN201811222012A CN109255339A CN 109255339 A CN109255339 A CN 109255339A CN 201811222012 A CN201811222012 A CN 201811222012A CN 109255339 A CN109255339 A CN 109255339A
 Authority
 CN
 China
 Prior art keywords
 forest
 sample
 adaptive
 current iteration
 depth
 Prior art date
 Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
 Granted
Links
 230000003044 adaptive Effects 0.000 title claims abstract description 163
 230000005021 gait Effects 0.000 title claims abstract description 111
 238000010586 diagram Methods 0.000 title claims abstract description 68
 238000007637 random forest analysis Methods 0.000 claims description 48
 238000005070 sampling Methods 0.000 claims description 16
 238000004364 calculation methods Methods 0.000 claims description 8
 230000014509 gene expression Effects 0.000 claims description 8
 230000000007 visual effect Effects 0.000 claims description 8
 230000002708 enhancing Effects 0.000 claims description 7
 210000004940 Nucleus Anatomy 0.000 claims description 6
 230000000875 corresponding Effects 0.000 claims description 6
 239000000284 extracts Substances 0.000 claims description 6
 238000000605 extraction Methods 0.000 claims description 4
 281000017677 Target Video companies 0.000 claims description 2
 239000012141 concentrates Substances 0.000 claims description 2
 238000000528 statistical tests Methods 0.000 claims description 2
 238000005457 optimization Methods 0.000 abstract 1
 238000000034 methods Methods 0.000 description 10
 230000001537 neural Effects 0.000 description 10
 230000006399 behavior Effects 0.000 description 5
 230000003542 behavioural Effects 0.000 description 3
 230000000694 effects Effects 0.000 description 3
 230000001133 acceleration Effects 0.000 description 2
 239000000203 mixtures Substances 0.000 description 2
 280000814325 Academia Sinica companies 0.000 description 1
 281000019761 Intel, Corp. companies 0.000 description 1
 230000015572 biosynthetic process Effects 0.000 description 1
 238000003062 neural network model Methods 0.000 description 1
 230000000750 progressive Effects 0.000 description 1
 238000003786 synthesis reactions Methods 0.000 description 1
 230000002194 synthesizing Effects 0.000 description 1
 239000011901 water Substances 0.000 description 1
Classifications

 G—PHYSICS
 G06—COMPUTING; CALCULATING; COUNTING
 G06K—RECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
 G06K9/00—Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
 G06K9/00335—Recognising movements or behaviour, e.g. recognition of gestures, dynamic facial expressions; Lipreading
 G06K9/00342—Recognition of whole body movements, e.g. for sport training
 G06K9/00348—Recognition of walking or running movements, e.g. gait recognition

 G—PHYSICS
 G06—COMPUTING; CALCULATING; COUNTING
 G06N—COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
 G06N3/00—Computer systems based on biological models
 G06N3/004—Artificial life, i.e. computers simulating life
 G06N3/006—Artificial life, i.e. computers simulating life based on simulated virtual individual or collective life forms, e.g. single "avatar", social simulations, virtual worlds or particle swarm optimisation
Abstract
The invention discloses a kind of classification methods based on adaptive depth forest body gait energy diagram, the steps include: (1) building training set and test set；(2) adaptive depth forest model is constructed；(3) nicety of grading of adaptive depth forest model is calculated；(4) judge whether adaptive depth forest model reaches requirement；(5) if training set accuracy rate and depth not up to require, extend adaptive depth forest model, it solves customized secondary convex optimization problem and obtains adaptive depth forest model in the Enhanced feature of training set, the training set of adaptive depth forest after being expanded by the Enhanced feature on training set；(6) test set classification accuracy rate is calculated.The present invention reduces the complexities of model and model to need the weight of training, while reducing gait energy diagram sample size required for training pattern.
Description
Technical field
The invention belongs to technical field of image processing, further relate to one of image identification technical field and are based on certainly
Adapt to the body gait energy diagram classification method of depth forest.The present invention can be used for body gait energy diagram GEI (Gait
Energy Image) in gait feature classify.
Background technique
Gait Recognition technology is that a kind of biotechnology of identification is carried out according to the posture that people in video sequence walks.By
In Gait Recognition have the characteristics that noninfringement property, remote identity and be difficult to it is hiding, therefore Gait Recognition country it is public
The fields such as safety, financial security, authentication, video monitoring have a wide range of applications.
A kind of patent document " gait knowledge based on deep learning of scientific and technological (Beijing) Co., Ltd of milky way water droplet in its application
It is proposed in other method " (number of patent application: 201410587758X, application publication number: CN104299012A) a kind of based on depth
The gait recognition method of study.This method describes gait sequence using gait energy diagram, is instructed by depth convolutional neural networks
Practice Matching Model, to match the identity of Gait Recognition people.The training process of this method are as follows: to marked good identity be related to it is more
The training gait video sequence at a visual angle extracts gait energy diagram, repeats to choose any two of them to based on convolutional neural networks
Matching Model be trained until model restrain；The identification process of this method are as follows: to singleview gait video to be identified and
Registration gait video sequence extracts gait energy diagram respectively, utilizes based on convolutional neural networks trained in training process
The gait energy diagram and the registered each gait energy diagram of gait video sequence of singleview gait video to be identified are calculated with model
Similarity, the size according to similarity carries out identity prediction, and exports recognition result.This method is to the gait across large viewing
Variation has very strong robustness, and it is not high to efficiently solve existing gait Recognition technology precision when handling across visual angle Gait Recognition
The problem of.But the shortcoming that this method still has is, since this method is the classification based on depth convolutional neural networks
Method, and a large amount of training sample is needed based on the classification method of depth convolutional neural networks, it is fewer in training sample amount
In the case of, the classification method of depth convolutional neural networks cannot sufficiently learn the gait feature in body gait energy diagram, thus
The ability in feature extraction of depth convolutional neural networks is reduced, this is not suitable for the classification method of depth convolutional neural networks
The data of small sample, and the classification performance based on depth convolutional neural networks depends critically upon parameter testing process, trains
It needs to consume biggish computing resource in journey.
Inst. of Computing Techn. Academia Sinica is (special in the patent document " gait identification method and system " of its application
Sharp application number: CN201710136803.3, application publication number: CN107016346A) in a kind of gait recognition method is provided.The party
Method extracts behavioural characteristic, position feature and gait feature from the acceleration information that the intelligent terminal carried by user obtains；Benefit
User's current behavior is identified according to the behavioural characteristic with preparatory trained Activity recognition model；Using in advance it is trained
Position identification model identifies the present bit of the intelligent terminal according to the position feature and the user's current behavior identified
It sets；Using preparatory trained Gait Recognition model according to the gait feature, the user's current behavior and the intelligence that are identified
Can the current location of terminal identify the identity of user.This method improves the knowledge of gait identity by being layered progressive identification method
Other accuracy rate and robustness, and do not need to limit the position and direction that related sensor is put, it is very flexibly and square
Just it uses.But the shortcoming that this method still has is, adds since this method needs user to carry intelligent terminal to obtain
Speed data extracts behavioural characteristic, position feature and the gait feature of user using the acceleration information of acquisition, this is weakened
The noninfringement of Approach for Gait Classification, and this method is by preparatory trained Activity recognition model, position identification model and gait
Feature extraction, can not all be dissolved into a model, cannot achieve and divide end to end by three model compositions of identification model
Class.
Summary of the invention
It is a kind of based on adaptive depth forest people it is an object of the invention in view of the above shortcomings of the prior art, propose
The classification method of body gait energy diagram.For the present invention compared with other existing Approach for Gait Classification, adaptive depth forest model is logical
The weight for solving each tree is crossed, the Enhanced feature of adaptive depth forest model is obtained, it is adaptive with the Enhanced feature training of acquisition
Depth forest model is answered, keeps weight required for training less, training speed is fast.In the training process according to test set and training
The accuracy of collection is continuously increased the depth of adaptive depth forest, make the complexity of adaptive depth forest with training process from
It adapts to adjust, reduces the complexity of model, take full advantage of computing resource.
The thinking for realizing the object of the invention is, first to all gait Video Composition gait energy diagrams, to utilize random combine
The method of sampling obtains training set and test set.Then an initial adaptive depth forest model is constructed, and to model
Parameter is initialized.The nicety of grading of the adaptive depth forest model of iteration before calculating.Judge the adaptive of current iteration
Whether the nicety of grading on training set of depth forest model meets the requirements, the deconditioning if meeting the requirements, this model is
The weight of each tree in depth forest is calculated, the adaptive depth of current iteration is obtained if being unsatisfactory for requiring for required model
Enhanced feature of the forest model on test set is spent, if the accuracy rate on test set reaches requirement, then with regard to deconditioning, this
Model is required model, is otherwise extended to current depth forest, and the training set Enhanced feature of depth forest is obtained,
The training set of current depth forest is updated, the depth of depth forest is increased, and continues to be trained current depth forest, until
Model required for obtaining.
The specific steps of the present invention are as follows:
(1) sample set is constructed:
(1a) carries out background modeling and background realtime update to 124 target person videos of input, obtains each target
The human body contour outline foreground image of personage；
(1b) extracts body gait figure from the human body contour outline foreground image of each target person video；
(1c) is respectively by the gray value of all human body contour images of each target person obtained in each gait cycle
After addition, then the sum of the human body contour outline image divided by each target person, obtain gait energy corresponding with the gait cycle
Figure；
(1d) takes each target person under the conditions of normal walking and visual angle is 90 degree of six gait energy diagrams, and totally 744
It opens gait energy diagram and constitutes sample set；
(2) training set and test set are obtained:
(2a) carries out stochastical sampling to sample set, obtains 992 positive samples and negative sample using random combine sampling splicing method
This, forms training set for all positive samples and negative sample；
(2b) carries out stochastical sampling to sample set, obtains 992 positive samples and negative sample using random combine sampling splicing method
This, forms test set for all positive samples and negative sample；
(3) adaptive depth forest model is constructed:
(3a) builds the adaptive depth forest model being made of two random forests and two completely random forests；
(3b) by adaptive depth forest model two random forests and two completely random forests in the number set
Be disposed as 400, the depth capacity of tree be disposed as 8, bootstrapping sample mark is disposed as True, and the nucleus number of parallel processing is all provided with
It is set to nucleus number in the maximum of server, sets t for the depth of adaptive depth forest model；
(4) nicety of grading of the adaptive depth forest model of current iteration is calculated:
The training set of the adaptive depth forest model of current iteration is input to the adaptive depth of current iteration by (4a)
In forest model, the corresponding class probability of each sample in training set is exported；
(4b) utilizes training set nicety of grading calculation formula, calculates adaptive depth forest of the training set Jing Guo current iteration
The nicety of grading of adaptive depth forest model after model training；
(5) judge whether the nicety of grading of the adaptive depth forest model of current iteration is more than or equal to 95%；If so, holding
Row step (12) otherwise executes step (6)；
(6) weight of each tree in the adaptive depth forest model of current iteration is obtained:
(6a) utilizes following customized secondary convex optimizing expressions, obtains the adaptive depth forest model of current iteration
The weight of each tree in each random forest:
ξ_{i} ^{k}>=0, i ∈ K
Wherein,ξ when minimizing target function type is sought in expression_{i},w^{k}Operation, i indicates in training set under sample
Mark, K indicate that the sum of training set sample, ∑ indicate sum operation, ξ_{i} ^{k}Indicate in training set ith sample current iteration from
The basic loss in depth forest model in kth of random forest is adapted to, λ indicates regularization coefficient,     indicate 2 norms behaviour
Make, w^{k}Indicate total weight of kth of forest in the adaptive depth forest model of current iteration, s.t. indicates constraint condition symbol
Number, t indicates the subscript set in kth of random forest in the adaptive depth forest model of current iteration, T_{k}Indicate current iteration
Adaptive depth forest model in the sum set in kth of random forest, P_{i} ^{(t,k)}Indicate ith of sample in current iteration
Basic loss in adaptive depth forest model in tth of tree of kth of random forest, w^{(t,k)}Indicate current iteration from
Adapt to the t several weight of kth of random forest in depth forest model, τ presentation class interval；
(6b) forms the weight of each tree in the adaptive each random forest of depth forest model of current iteration current
The weight of the adaptive depth forest model of iteration；
(7) Enhanced feature of the adaptive depth forest model of current iteration on test set is obtained:
(7a) passes through the adaptive of current iteration according to the weight of each tree in the adaptive depth forest model of current iteration
The Enhanced feature solution formula of depth forest model is answered, obtaining each sample in test set respectively is positive sample and negative sample
Enhanced feature；
(7b) by test set the positive sample of all samples and the Enhanced feature of negative sample constitute the adaptive of current iteration
Answer the Enhanced feature of depth forest model；
(8) accuracy of test set is calculated:
According to Enhanced feature of the adaptive depth forest model of current iteration on test set, using current iteration from
Adapt to the accuracy that depth forest model test set accuracy calculation method calculates test set；
(9) judge whether the adaptive depth forest model test set accuracy of current iteration is being greater than the maximum of test set just
True rate otherwise, executes step (12) if so, thening follow the steps (10)；
(10) Enhanced feature of the adaptive depth forest model of current iteration on training set is obtained:
(10a) according to the weight of each tree in the adaptive depth forest model of current iteration, by current iteration from
The Enhanced feature solution formula of depth forest model is adapted to, obtaining each sample in training set respectively is positive sample and negative sample
Enhanced feature；
(10b) by training set the positive sample of all samples and the Enhanced feature of negative sample constitute the adaptive of current iteration
Answer the Enhanced feature of depth forest model；
(11) the adaptive depth forest model of current iteration is extended:
(11a) is by the adaptive depth forest of the training set of the adaptive depth forest model of current iteration and current iteration
The Enhanced feature head and the tail of model splice, and obtain the training set of the adaptive depth forest of current iteration；
(11b) is using the depth t+1 of the adaptive depth forest model of current iteration as the adaptive depth of current iteration
The depth t of forest model；
(12) judge whether the depth t of the adaptive depth forest model of current iteration is greater than 100, if so, being instructed
Step (13) are executed after the adaptive depth forest model perfected, otherwise, are executed step (4)；
(13) classify to gait energy diagram all in test set:
The test set of the adaptive depth forest model of current iteration is input to trained adaptive depth forest mould
In type, classify to gait energy diagram in test set.
The present invention compared with prior art, has the advantage that
First, since the present invention is by the weight of solution each tree, the Enhanced feature of adaptive depth forest model is obtained,
Using the adaptive depth forest model of Enhanced feature training, overcomes the prior art and need to train during training neural network
A large amount of weight, the not applicable problem of data set of the neural network model to small sample, so that the present invention has skilled weight
The advantages of quantity is few, and training speed is fast, is suitable for small data set.
Second, since the present invention is continuously increased adaptive depth according to the accuracy of test set and training set in the training process
The depth for spending forest makes the complexity of adaptive depth forest with training process automatic adjusument, overcomes prior art mind
The problem of model complexity immobilizes in the training process through network, so that The present invention reduces the complexities of model, sufficiently
Computing resource is utilized.
Detailed description of the invention
Fig. 1 is flow chart of the invention.
Specific embodiment
With reference to the accompanying drawing 1, the specific steps realized to the present invention are further described.
Step 1, sample set is constructed.
Background modeling and background realtime update are carried out to 124 target person videos of input, obtain each target person
Human body contour outline foreground image.
The background modeling refers to background realtime update, and the previous frame of human target video is subtracted each other with background frames, is obtained
To the human body contour outline foreground image of target person video；The power function in vision library is called, it will be in human body contour outline foreground image
Existing cavity and the removal of scattered white point.
Body gait figure is extracted from the human body contour outline foreground image of each target person video.
The step of described extraction body gait figure is, by the point on each human body contour outline most left side and the point on the most right side away from
From the width as human body contour outline, height of the uppermost point at a distance from nethermost point as human body contour outline；Utilize human body
Profile width the characteristic sexually revised synchronizing cycle occurs at any time, gait is divided by the change width signal of human body contour outline
Period；Human body contour outline image is deducted by the width and height of human body contour outline and keeps depthwidth ratio constant, by all human body contour outlines
Image scaling is to same size.
The gray value of all human body contour images of each target person obtained in each gait cycle is added respectively
Afterwards, the sum of the human body contour outline image then divided by each target person obtains gait energy diagram corresponding with the gait cycle.
It takes each target person under the conditions of normal walking and visual angle is 90 degree of six gait energy diagrams, totally 744 steps
State energy diagram constitutes sample set.
Step 2, training set and test set are obtained.
Using random combine sampling splicing method, stochastical sampling is carried out to sample set, obtains 992 positive samples and negative sample,
All positive samples and negative sample are formed into training set.
Using random combine sampling splicing method, stochastical sampling is carried out to sample set, obtains 992 positive samples and negative sample,
All positive samples and negative sample are formed into test set.
The step of random combine sampling splicing method, is as follows:
The first step randomly selects 6 gaits of the human target of a unselected mistake from 124 human targets of sample set
Energy diagram is as currently processed object.
Second step randomly selects four pairs of gait energy diagrams from currently processed object, each pair of gait energy diagram is carried out horizontal
To splicing, constitutes a positive sample and set 0 for its label, the positive sample quantity of each target person is 4.
Third step randomly selects a gait energy diagram from currently processed object.
4th step randomly selects 4 different target personages from sample set in addition to current goal personage respectively, from institute
In respective 6 gait energy diagrams of the target person of selection, a gait energy diagram is randomly selected respectively, totally 4 gait energy
Figure.
5th step, the 4 state energy diagrams obtained with the 4th step respectively with the gait energy diagram that third step obtains are twobytwo
Pairing, constitutes 4 pairs of gait energy diagrams, each pair of gait energy diagram is carried out horizontallyspliced, constitutes a negative sample and by its label
It is set as 1, the negative sample quantity of each human target is 4.
Step 6: being processed by currently processed object tag.
Step 7: to 124 target persons, if 6 gait energy diagrams of each target person be collectively labeled as it is processed, that
496 positive samples and 496 negative samples are obtained, i.e. 992 samples altogether, to 124 target persons, target person if it exists
6 gait energy diagrams of object are untreated to be finished, and the first step is executed.
Step 3, adaptive depth forest model is constructed.
Build the adaptive depth forest model being made of two random forests and two completely random forests.
By in adaptive depth forest model two random forests and two completely random forests in the number set be all provided with
Be set to 400, the depth capacity of tree be disposed as 8, bootstrapping sample mark is disposed as True, and the nucleus number of parallel processing is disposed as
The depth of adaptive depth forest model is set t by nucleus number in the maximum of server.
Step 4, the nicety of grading of the adaptive depth forest model of current iteration is calculated.
The training set of the adaptive depth forest model of current iteration is input to the adaptive depth forest of current iteration
In model, the corresponding class probability of each sample in training set is exported.
Using training set nicety of grading calculation formula, adaptive depth forest model of the training set Jing Guo current iteration is calculated
The nicety of grading of adaptive depth forest model after training.
Step 5, judge the nicety of grading of adaptive depth forest model of current iteration whether less than 95%；If so, holding
Otherwise row step 6 executes step 12.
Step 6, the weight of each tree in the adaptive depth forest model of current iteration is obtained.
(6.1) following customized secondary convex optimizing expressions is utilized, the adaptive depth forest mould of current iteration is obtained
The weight of each tree in each random forest of type:
ξ_{i} ^{k}>=0, i ∈ K
Wherein,ξ when minimizing target function type is sought in expression_{i},w^{k}Operation, i indicate training set in sample mark
Number, K indicates that the sum of training set sample, ∑ indicate sum operation, ξ_{i} ^{k}Indicate in training set ith sample current iteration from
The basic loss in depth forest model in kth of random forest is adapted to, λ indicates regularization coefficient,     indicate 2 norms behaviour
Make, w^{k}Indicate total weight of kth of forest in the adaptive depth forest model of current iteration, s.t. indicates constraint condition symbol
Number, t indicates the subscript set in kth of random forest in the adaptive depth forest model of current iteration, T_{k}Indicate current iteration
Adaptive depth forest model in the sum set in kth of random forest, P_{i} ^{(t,k)}Indicate ith of sample in current iteration
Basic loss in adaptive depth forest model in tth of tree of kth of random forest, w^{(t,k)}Indicate current iteration from
Adapt to the t several weight of kth of random forest in depth forest model, τ presentation class interval；
Ith of sample is in the adaptive depth forest model of current iteration in tth of tree of kth of random forest
Basic loss expression formula are as follows:
P_{i} ^{(t,k)}=Z_{i}*(P_{i,0} ^{(t,k)}P_{i,1} ^{(t,k)})
Wherein, Z_{i}Indicator variable is indicated, if the label of ith of sample is 0, Z in training set_{i}=1, if in training set
The label of i sample is 1, then Z_{i}=1, * indicate multiplication operation, P_{i,0} ^{(t,k)}, P_{i,1} ^{(t,k)}Indicate ith of sample in current iteration
Class probability in tth of tree in kth of random forest in adaptive depth forest model.
(6.2) weight of each tree in the adaptive each random forest of depth forest model of current iteration is formed current
The weight of the adaptive depth forest model of iteration；
Step 7, Enhanced feature of the adaptive depth forest model of current iteration on test set is obtained.
According to the weight of each tree in the adaptive depth forest model of current iteration, pass through the adaptive depth of current iteration
The Enhanced feature solution formula of forest model is spent, obtains the enhancing that each sample in test set is positive sample and negative sample respectively
Feature.
By in test set the positive sample of all samples and the Enhanced feature of negative sample constitute the adaptive depth of current iteration
Spend the Enhanced feature of forest model.
The Enhanced feature solution formula of the adaptive depth forest model of the current iteration is as follows:
Wherein, v_{i,0}Indicate enhancing probability of ith of sample for positive sample, p in training set_{i,0} ^{(t,k)}It indicates ith in training set
The probability of a sample positive sample in tth of tree of kth of random forest in the adaptive depth forest model of current iteration,
v_{i,1}Indicate enhancing probability of ith of sample for negative sample, p in training set_{i,1} ^{(t,k)}Indicate that ith of sample is current in training set
In the adaptive depth forest model of iteration in tth of tree of kth of random forest negative sample probability.
Step 8, the accuracy of test set is calculated.
According to Enhanced feature of the adaptive depth forest model of current iteration on test set, using current iteration from
Adapt to the accuracy that depth forest model test set accuracy calculation method calculates test set.
The adaptive depth forest model test set accuracy calculation method of the current iteration is as follows:
The first step, to each sample in test set, the Enhanced feature numerical value of positive sample is greater than the Enhanced feature of negative sample
Value, and the label of sample is 0, then and sample classification is correct, otherwise sample classification mistake.
Second step, to each sample in test set, the Enhanced feature numerical value of positive sample is less than the Enhanced feature of negative sample
Value, and the label of sample is 1, then and sample classification is correct, otherwise sample classification mistake.
Third step, statistical test concentrate the sum for the sample correctly classified.
4th step obtains adaptive with the sum for the sample correctly classified in test set divided by the sum of test set sample
Answer the test set accuracy of depth forest model.
Step 9, judge whether the adaptive depth forest model test set accuracy of current iteration is greater than test set most
Otherwise big accuracy, executes step 12 if so, thening follow the steps 10.
The maximum accuracy of the test set refer to test set depth be 1,2 ..., the adaptive depth forest model of t
The maximum value of upper accuracy.
Step 10, Enhanced feature of the adaptive depth forest model of current iteration on training set is obtained.
According to the weight of each tree in the adaptive depth forest model of current iteration, pass through the adaptive depth of current iteration
The Enhanced feature solution formula of forest model is spent, obtains the enhancing that each sample in training set is positive sample and negative sample respectively
Feature.
By in training set the positive sample of all samples and the Enhanced feature of negative sample constitute the adaptive depth of current iteration
Spend the Enhanced feature of forest model.
Step 11, the adaptive depth forest model of current iteration is extended.
By the adaptive depth forest model of the training set of the adaptive depth forest model of current iteration and current iteration
Enhanced feature head and the tail splice, obtain the training set of the adaptive depth forest model of current iteration.
Using the depth t+1 of the adaptive depth forest model of current iteration as adaptive depth forest when current iteration
The depth t of model.
Step 12, judge whether the depth t of the adaptive depth forest model of current iteration is greater than 100, if so, obtaining
Step 13 is executed after trained adaptive depth forest model, otherwise, executes step 4.
Step 13, classify to gait energy diagram all in test set.
The test set of the adaptive depth forest model of current iteration is input to trained adaptive depth forest mould
In type, classify to gait energy diagram in test set.
Effect of the invention is described further below with reference to emulation experiment:
1, emulation experiment condition:
The hardware platform of emulation experiment of the invention are as follows: Intel Core i78700K@3.70GHz CPU, 64GB
RAM、NVIDIA Geforce GTX1080Ti GPU。
The software platform of emulation experiment of the invention are as follows: python2.7 and TensorFlow1.4.
Data used in emulation experiment of the invention are disclosed Gait Recognition data set Dataset B, including 124 targets
The profile of personage's gait of human body under three kinds of walking conditions (usual terms wears overcoat, carries package condition) and 11 visual angles
Sketch figure, this method synthesize the gait energy diagram of 124 target persons, and choose each target person at normal walking and visual angle
For six gait energy diagrams under conditions of 90 degree, totally 744 gait energy diagrams constitute sample set, are sampled and are spliced with random combine
Method carries out stochastical sampling, composing training collection and test set to sample set.
2, emulation content and interpretation of result:
In emulation experiment of the invention, first according to each of method synthesis Dataset B data library of the invention
The gait energy diagram at different conditions of target person, and each target person is chosen under the conditions of normal walking and visual angle is
90 degree of six gait energy diagrams, totally 744 gait energy diagrams constitute sample set, are sampled splicing method with random combine, to sample
Collection carries out stochastical sampling, forms training set and test set.Using the adaptive depth forest model of training set training, then with use
Test sample is tested in trained adaptive depth forest model, obtains test set nicety of grading of the invention.
Emulation experiment of the invention is under identical data set, using the method for the present invention and prior art random forest
The test set that (paper " Random Forests " 2001 " Mach.Learning " of Leo Breiman) constructs the present invention
In gait energy diagram classify.
In order to verify the classifying quality of method and prior art random forest of the invention, using following formula, this is calculated separately
The nicety of grading of invention and random forest, and compare.
The sample correctly classified in nicety of grading=test set/test set total sample number
1. two methods nicety of grading list of table
Emulation experiment method  Random forest  This method 
Nicety of grading  75.43%  92.93% 
The present invention and existing random forest technology classification accuracy computation are listed in table 1 respectively as a result, as seen from Table 1,
Nicety of grading of the invention be 92.93%, this index be much higher than random forest classification method, it was demonstrated that the present invention can obtain than
The nicety of grading of the higher gait energy diagram of random forest method.
Claims (8)
1. a kind of classification method based on adaptive depth forest body gait energy diagram, which is characterized in that solve adaptive deep
The weight for spending each tree in forest model, obtains the Enhanced feature of adaptive depth forest model, according to depth forest model
Depth adaptive depthkeeping degree forest model complexity really utilizes the Enhanced feature training depth forest for obtaining depth forest model
Model；The specific steps of this method include the following:
(1) sample set is constructed:
(1a) carries out background modeling and background realtime update to 124 target person videos of input, obtains each target person
Human body contour outline foreground image；
(1b) extracts body gait figure from the human body contour outline foreground image of each target person video；
The gray value of all human body contour images of each target person obtained in each gait cycle is added by (1c) respectively
Afterwards, the sum of the human body contour outline image then divided by each target person obtains gait energy diagram corresponding with the gait cycle；
Six gait energy diagrams that (1d) takes each target person under the conditions of normal walking and visual angle is 90 degree, totally 744 walk
State energy diagram constitutes sample set；
(2) training set and test set are obtained:
(2a) carries out stochastical sampling to sample set, obtains 992 positive samples and negative sample using random combine sampling splicing method,
All positive samples and negative sample are formed into training set；
(2b) carries out stochastical sampling to sample set, obtains 992 positive samples and negative sample using random combine sampling splicing method,
All positive samples and negative sample are formed into test set；
(3) adaptive depth forest model is constructed:
(3a) builds the adaptive depth forest model being made of two random forests and two completely random forests；
(3b) by adaptive depth forest model two random forests and two completely random forests in the number set be all provided with
Be set to 400, the depth capacity of tree be disposed as 8, bootstrapping sample mark is disposed as True, and the nucleus number of parallel processing is disposed as
The depth of adaptive depth forest model is set t by nucleus number in the maximum of server；
(4) nicety of grading of the adaptive depth forest model of current iteration is calculated:
The training set of the adaptive depth forest model of current iteration is input to the adaptive depth forest of current iteration by (4a)
In model, the corresponding class probability of each sample in training set is exported；
(4b) utilizes training set nicety of grading calculation formula, calculates adaptive depth forest model of the training set Jing Guo current iteration
The nicety of grading of depth forest model after training；
(5) judge the nicety of grading of adaptive depth forest model of current iteration whether less than 95%；If so, executing step
(6), step (12) otherwise, are executed；
(6) weight of each tree in the adaptive depth forest model of current iteration is obtained:
(6a) utilizes following customized secondary convex optimizing expressions, and the adaptive depth forest model for obtaining current iteration is each
The weight of each tree in random forest:
ξ_{i} ^{k}>=0, i ∈ K
Wherein,ξ when minimizing target function type is sought in expression_{i},w^{k}Operation, i indicate training set in sample label, K table
Show that the sum of training set sample, ∑ indicate sum operation, ξ_{i} ^{k}Indicate that ith of sample is in the adaptive depth of current iteration in training set
The basic loss in forest model in kth of random forest is spent, λ indicates regularization coefficient,     indicate the operation of 2 norms, w^{k}
Indicate total weight of kth of forest in the adaptive depth forest model of current iteration, s.t. indicates constraint condition symbol, t table
Show the subscript set in kth of random forest in the adaptive depth forest model of current iteration, T_{k}Indicate the adaptive of current iteration
Answer the sum set in kth of random forest in depth forest model, P_{i} ^{(t,k)}Indicate ith of sample in the adaptive of current iteration
Basic loss in depth forest model in tth of tree of kth of random forest, w^{(t,k)}Indicate the adaptive depth in current iteration
Spend the t several weight of kth of random forest in forest model, τ presentation class interval；
The weight of each tree in the adaptive each random forest of depth forest model of current iteration is formed current iteration by (6b)
Adaptive depth forest model weight；
(7) Enhanced feature of the adaptive depth forest model of current iteration on test set is obtained:
(7a) passes through the adaptive depth of current iteration according to the weight of each tree in the adaptive depth forest model of current iteration
The Enhanced feature solution formula of forest model is spent, obtains the enhancing that each sample in test set is positive sample and negative sample respectively
Feature；
(7b) by test set the positive sample of all samples and the Enhanced feature of negative sample constitute the adaptive depth of current iteration
Spend the Enhanced feature of model；
(8) accuracy of test set is calculated:
According to Enhanced feature of the adaptive depth forest model of current iteration on test set, the adaptive of current iteration is utilized
The accuracy of depth forest model test set accuracy calculation method calculating test set；
(9) judge whether the adaptive depth forest model test set accuracy of current iteration is greater than the maximum of test set correctly
Rate otherwise, executes step (12) if so, thening follow the steps (10)；
(10) Enhanced feature of the adaptive depth forest model of current iteration on training set is obtained:
(10a) passes through the adaptive of current iteration according to the weight of each tree in the adaptive depth forest model of current iteration
The Enhanced feature solution formula of depth forest model obtains the increasing that each sample in training set is positive sample and negative sample respectively
Strong feature；
(10b) by training set the positive sample of all samples and the Enhanced feature of negative sample constitute the adaptive depth of current iteration
Spend the Enhanced feature of forest model；
(11) the adaptive depth forest model of current iteration is extended:
(11a) is by the increasing of the training set of the adaptive depth forest model of current iteration and the adaptive depth model of current iteration
Strong feature head and the tail splice, and obtain the training set of the adaptive depth forest model of current iteration；
(11b) is gloomy as adaptive depth when current iteration using the depth t+1 of the adaptive depth forest model of current iteration
The depth t of woods model；
(12) judge whether the depth t of the adaptive depth forest model of current iteration is greater than 100, if so, being trained
Depth forest model after execute step (13), otherwise, execute step (4)；
(13) classify to gait energy diagram all in test set:
The test set of the adaptive depth forest model of current iteration is input in trained depth forest model, to test
Gait energy diagram is concentrated to classify.
2. the classification method according to claim 1 based on adaptive depth forest body gait energy diagram, feature exist
In: background modeling described in step (1a) refers to background realtime update, by the previous frame of human target video and background frames phase
Subtract, obtains the human body contour outline foreground image of target person video；The power function in vision library is called, by human body contour outline foreground picture
The cavity as present in and scattered white point remove.
3. the classification method according to claim 1 based on adaptive depth forest body gait energy diagram, feature exist
In: the step of extraction body gait figure described in step (1b), is, by the point on each human body contour outline most left side and the most right side
Width of the distance of point as human body contour outline, height of the uppermost point at a distance from nethermost point as human body contour outline；Benefit
The characteristic sexually revised synchronizing cycle occurs at any time with the profile width of human body, is drawn by the change width signal of human body contour outline
Divide gait cycle；Human body contour outline image is deducted by the width and height of human body contour outline and keeps depthwidth ratio constant, by owner
Body contour images zoom to same size.
4. the classification method according to claim 1 based on adaptive depth forest body gait energy diagram, feature exist
It is as follows in the step of: random combine described in step (2) samples splicing method:
The first step randomly selects 6 gait energy of the human target of a unselected mistake from 124 human targets of sample set
Figure is used as currently processed object；
Second step randomly selects four pairs of gait energy diagrams from currently processed object, and each pair of gait energy diagram is carried out lateral spelling
It connects, constitute a positive sample and sets 0 for its label, the positive sample quantity of each target person is 4；
Third step randomly selects a gait energy diagram from currently processed object；
4th step randomly selects 4 different target personages from sample set in addition to current goal personage respectively, from selected
Respective 6 gait energy diagrams of target person in, randomly select a gait energy diagram respectively, totally 4 gait energy diagrams；
5th step is matched twobytwo with 4 state energy diagrams that the gait energy diagram that third step obtains is obtained with the 4th step respectively,
4 pairs of gait energy diagrams are constituted, the progress of each pair of gait energy diagram is horizontallyspliced, it constitutes a negative sample and sets its label to
1, the negative sample quantity of each human target is 4；
Step 6: being processed by currently processed object tag；
Step 7: to 124 target persons, if 6 gait energy diagrams of each target person be collectively labeled as it is processed, then total
496 positive samples and 496 negative samples are obtained, i.e. 992 samples altogether, to 124 target persons, target person if it exists
6 gait energy diagrams are untreated to be finished, and the first step is executed.
5. the classification method according to claim 1 based on adaptive depth forest body gait energy diagram, feature exist
In: the t of ith of sample described in step (6a), kth of random forest in the adaptive depth forest model of current iteration
Basic loss expression formula in a tree are as follows:
P_{i} ^{(t,k)}=Z_{i}*(P_{i,0} ^{(t,k)}P_{i,1} ^{(t,k)})
Wherein, Z_{i}Indicator variable is indicated, if the label of ith of sample is 0, Z in training set_{i}=1, if ith in training set
The label of sample is 1, then Z_{i}=1, * indicate multiplication operation, P_{i,0} ^{(t,k)}, P_{i,1} ^{(t,k)}Indicate ith sample current iteration from
Adapt in depth forest model in kth of random forest class probability in tth of tree.
6. the classification method according to claim 1 based on adaptive depth forest body gait energy diagram, feature exist
In: the Enhanced feature solution formula of the adaptive depth forest model of current iteration described in step (7a) is as follows:
Wherein, v_{i,0}Indicate enhancing probability of ith of sample for positive sample, p in training set_{i,0} ^{(t,k)}Indicate ith of sample in training set
The probability of this positive sample in tth of tree of kth of random forest in the adaptive depth forest model of current iteration, v_{i,1}Table
Show enhancing probability of ith of sample for negative sample, p in training set_{i,1} ^{(t,k)}Indicate that ith of sample is in current iteration in training set
In adaptive depth forest model in tth of tree of kth of random forest negative sample probability.
7. the classification method according to claim 1 based on adaptive depth forest body gait energy diagram, feature exist
In: the adaptive depth forest model test set accuracy calculation method of current iteration described in step (8) is as follows:
The first step, to each sample in test set, the Enhanced feature numerical value of positive sample is greater than the Enhanced feature value of negative sample, and
And the label of sample is 0, then sample classification is correct, otherwise sample classification mistake；
Second step, to each sample in test set, the Enhanced feature numerical value of positive sample is less than the Enhanced feature value of negative sample, and
And the label of sample is 1, then sample classification is correct, otherwise sample classification mistake；
Third step, statistical test concentrate the sum for the sample correctly classified；
4th step obtains depth forest with the sum for the sample correctly classified in test set divided by the sum of test set sample
The test set accuracy of model.
8. the classification method according to claim 1 based on adaptive depth forest body gait energy diagram, feature exist
Refer to that test set in depth is 1,2 in: the maximum accuracy of test set described in step (9) ..., the adaptive depth forest of t
The maximum value of accuracy on model.
Priority Applications (1)
Application Number  Priority Date  Filing Date  Title 

CN201811222012.3A CN109255339B (en)  20181019  20181019  Classification method based on selfadaptive deep forest human gait energy map 
Applications Claiming Priority (1)
Application Number  Priority Date  Filing Date  Title 

CN201811222012.3A CN109255339B (en)  20181019  20181019  Classification method based on selfadaptive deep forest human gait energy map 
Publications (2)
Publication Number  Publication Date 

CN109255339A true CN109255339A (en)  20190122 
CN109255339B CN109255339B (en)  20210406 
Family
ID=65045465
Family Applications (1)
Application Number  Title  Priority Date  Filing Date 

CN201811222012.3A Active CN109255339B (en)  20181019  20181019  Classification method based on selfadaptive deep forest human gait energy map 
Country Status (1)
Country  Link 

CN (1)  CN109255339B (en) 
Cited By (1)
Publication number  Priority date  Publication date  Assignee  Title 

CN110706738A (en) *  20191030  20200117  腾讯科技（深圳）有限公司  Method, device, equipment and storage medium for predicting structure information of protein 
Citations (9)
Publication number  Priority date  Publication date  Assignee  Title 

WO2001039655A2 (en) *  19991206  20010607  Trustees Of Boston University  Inshoe remote telemetry gait analysis system 
CN101241546A (en) *  20071120  20080813  西安电子科技大学  Method for compensating for gait binary value distortion 
EP2439492A1 (en) *  20101007  20120411  Honeywell International, Inc.  System and method for waveletbased gait classification 
CN103473539A (en) *  20130923  20131225  智慧城市系统服务（中国）有限公司  Gait recognition method and device 
CN104200200A (en) *  20140828  20141210  公安部第三研究所  System and method for realizing gait recognition by virtue of fusion of depth information and grayscale information 
CN104299012A (en) *  20141028  20150121  中国科学院自动化研究所  Gait recognition method based on deep learning 
CN105574510A (en) *  20151218  20160511  北京邮电大学  Gait identification method and device 
CN107212890A (en) *  20170527  20170929  中南大学  A kind of motion identification and fatigue detection method and system based on gait information 
US9811720B2 (en) *  20131022  20171107  Bae Systems Information And Electronic Systems Integration Inc.  Mobile device based gait biometrics 

2018
 20181019 CN CN201811222012.3A patent/CN109255339B/en active Active
Patent Citations (9)
Publication number  Priority date  Publication date  Assignee  Title 

WO2001039655A2 (en) *  19991206  20010607  Trustees Of Boston University  Inshoe remote telemetry gait analysis system 
CN101241546A (en) *  20071120  20080813  西安电子科技大学  Method for compensating for gait binary value distortion 
EP2439492A1 (en) *  20101007  20120411  Honeywell International, Inc.  System and method for waveletbased gait classification 
CN103473539A (en) *  20130923  20131225  智慧城市系统服务（中国）有限公司  Gait recognition method and device 
US9811720B2 (en) *  20131022  20171107  Bae Systems Information And Electronic Systems Integration Inc.  Mobile device based gait biometrics 
CN104200200A (en) *  20140828  20141210  公安部第三研究所  System and method for realizing gait recognition by virtue of fusion of depth information and grayscale information 
CN104299012A (en) *  20141028  20150121  中国科学院自动化研究所  Gait recognition method based on deep learning 
CN105574510A (en) *  20151218  20160511  北京邮电大学  Gait identification method and device 
CN107212890A (en) *  20170527  20170929  中南大学  A kind of motion identification and fatigue detection method and system based on gait information 
NonPatent Citations (1)
Title 

ZHIHUA ZHOU 等: "Deep Forest: Towards An Alternative to Deep Neural Networks", 《ARXIV:1702.08835V2》 * 
Cited By (1)
Publication number  Priority date  Publication date  Assignee  Title 

CN110706738A (en) *  20191030  20200117  腾讯科技（深圳）有限公司  Method, device, equipment and storage medium for predicting structure information of protein 
Also Published As
Publication number  Publication date 

CN109255339B (en)  20210406 
Similar Documents
Publication  Publication Date  Title 

Yi et al.  Age estimation by multiscale convolutional network  
CN110032926B (en)  Video classification method and device based on deep learning  
Huo et al.  Deep age distribution learning for apparent age estimation  
CN106570521B (en)  Multilingual scene character recognition method and recognition system  
Fanelli et al.  Hough forestbased facial expression recognition from video sequences  
CN107122375A (en)  The recognition methods of image subject based on characteristics of image  
Rao et al.  Multipose facial expression recognition based on SURF boosting  
CN106203356B (en)  A kind of face identification method based on convolutional network feature extraction  
CN104036255A (en)  Facial expression recognition method  
CN104268593A (en)  Multiplesparserepresentation face recognition method for solving small sample size problem  
CN102013011B (en)  Frontfacecompensationoperatorbased multipose human face recognition method  
CN106909938B (en)  Visual angle independence behavior identification method based on deep learning network  
CN107808113B (en)  Facial expression recognition method and system based on differential depth features  
CN105893947A (en)  Bivisualangle face identification method based on multilocal correlation characteristic learning  
CN110427867B (en)  Facial expression recognition method and system based on residual attention mechanism  
CN108154133A (en)  Human face portrait based on asymmetric combination learningphoto array method  
CN106503616A (en)  A kind of Mental imagery Method of EEG signals classification of the learning machine that transfinited based on layering  
CN109255289B (en)  Crossaging face recognition method based on unified generation model  
CN108537120A (en)  A kind of face identification method and system based on deep learning  
CN108509833A (en)  A kind of face identification method, device and equipment based on structured analysis dictionary  
Chanti et al.  Improving bagofvisualwords towards effective facial expressive image classification  
Pratiwi  The use of self organizing map method and feature selection in image database classification system  
CN103714340A (en)  Selfadaptation feature extracting method based on image partitioning  
CN110598603A (en)  Face recognition model acquisition method, device, equipment and medium  
CN109255339A (en)  Classification method based on adaptive depth forest body gait energy diagram 
Legal Events
Date  Code  Title  Description 

PB01  Publication  
PB01  Publication  
SE01  Entry into force of request for substantive examination  
SE01  Entry into force of request for substantive examination  
GR01  Patent grant  
GR01  Patent grant 