CN104123570B - Human hand grader and training and detection method based on the combination of shared Weak Classifier - Google Patents
Human hand grader and training and detection method based on the combination of shared Weak Classifier Download PDFInfo
- Publication number
- CN104123570B CN104123570B CN201410351088.1A CN201410351088A CN104123570B CN 104123570 B CN104123570 B CN 104123570B CN 201410351088 A CN201410351088 A CN 201410351088A CN 104123570 B CN104123570 B CN 104123570B
- Authority
- CN
- China
- Prior art keywords
- weak classifier
- classifier
- combination
- human hand
- training
- 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.)
- Active
Links
Landscapes
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of human hand graders and its detection method based on the combination of shared Weak Classifier, and multi-pose hand can be detected from mixed and disorderly background.The each round of boosting during classifier training all chooses one group of sharing feature, and building a corresponding shared Weak Classifier using this group of sharing feature combines.Meanwhile build corresponding multi-threshold using sharing feature combination and share Weak Classifier combination, improve the classification capacity of grader.
Description
Technical field
The invention belongs to computer vision and machine learning field, suitable for having the multifarious target of stronger appearance
One strong classifier of training, and the target is detected in complicated background using the grader.
Background technology
With the fast development of computer science and technology, variation that the mode of human-computer interaction becomes, interactive experience also has aobvious
The raising of work, the gesture manipulation based on computer vision is to apply a kind of more mode in recent years.At present, many smart televisions
It is provided with the function of gesture manipulation.Using gesture carry out equipment manipulation, core technology for gesture identification and hand with
Track, and the two core technologies premise is that accurately detecting the position of appearance in one's hands from the image that camera captures.
However, since the variation diversity of hand is very strong, the effect that the hand of many attitude is detected from mixed and disorderly background is unsatisfactory.
The method that opponent is detected at present can generally be divided into two major classes, taken the photograph merely with common monocular optical respectively
The image captured as head carries out the method for the detection of hand and receives dress using camera and other ancillary equipments such as infrared emission
It the obtained information such as puts and carrys out hand in detection image.The latter obtains the letter on target more horn of plenty using the equipment of high cost
Breath, accurate testing result is obtained so as to which simple software algorithm is used only.And the former equipment cost is relatively low, easily
It obtains, but the information captured is less, and so algorithm is complex, and detection result is a bit weaker.It is taken the photograph using common monocular optical
The image captured as head carries out the detection of hand, and specific method can be divided into three classes again, respectively the method based on colouring information, base
Method in movable information and the method based on machine learning.
(1) method based on colouring information
This method establishes complexion model according to the color of human hand skin, by complexion model and the image captured progress
Match somebody with somebody, the region of successful match is area of skin color, you can the region being considered where hand.Other assistant strips can also further be passed through
Part further screens area of skin color, to obtain the position of more accurately hand.This method is simple and practicable, calculates complicated
Low, detection efficiency height is spent, but due to Color-sensitive, being easily subject to ambient lighting variation and the shadow of Different Individual colour of skin otherness
It rings, so as to cause the reduction of accuracy in detection.
(2) method based on movable information
The sequence of frames of video that this method is continuously captured according to camera determines the body part moved in image, and by its
It splits.Increase restrictive condition again on the basis of splitting herein, further determine that the position where hand.The method avoids
Method ambient light based on color is according to variation and influence of the colour of skin difference to detection result, but a disadvantage is that cannot detect static
Target, and it must be fixed placement to require camera.
(3) method based on machine learning
Carry out opponent by extracting feature, training grader from image based on the method for machine learning and be detected.It is this
Method is complex compared with first two method, and it is one than relatively time-consuming process to train grader by machine learning.Such as
What effectively extracts feature from image, and trains a hand to many attitude all with relatively strong point using the less time
The problem of grader of class ability and very fast detection speed is most critical.CVPRs of the Antonio Torralba et al. in 2004
The paper delivered in (IEEE Conference of Computer Vision and Pattern Recognition) meeting
《Sharing features:efficient boosting procedures for multiclass object
detection》The JointBoost algorithms of middle proposition.The algorithm has used the sharing feature between inhomogeneity target to train classification
Device, but the classification not covered for sharing feature cannot effectively be classified, and some shared Weak Classifier selection is
It is no suitable, the selection of subsequent characteristics and the structure of grader can be seriously affected, causes to influence entire classifier performance.
The content of the invention
It is an object of the invention to provide a kind of human hand graders and its implementation based on the combination of shared class device.Make
With the feature extracted from sample set image, a grader with higher robustness, the i.e. hand to many attitude are trained
It can effectively detect, while detection speed is faster compared with the grader that other methods are trained for the grader.
Based on this purpose, human hand grader of the invention by multistage strong classifier cascade form, per level-one strong classifier by
Multipair shared Weak Classifier combination is formed;Target image to be checked must be judged as " comprising human hand " by every level-one strong classifier, whole
A cascade classifier just can finally be judged as " comprising human hand ", otherwise be judged as " not including human hand ".
The training method of human hand grader of the present invention is:Assuming that a shared C class sample class, N number of training sample,
Each training sample viWeight for classification c isShare C weight;Nf Weak Classifier is combined as one, and sample
This collection is divided for nf subset, then, k-th of Weak Classifier is:
Wherein, m is the wheel number for being currently located the training of level-one strong classifier,For the kth of i-th of sample in training sample
The value of a feature,Correspond to the threshold value of classification c, S for k-th of Weak Classifierk(nf) k-th of the subset concentrated for nf son;
This group of Weak Classifier combination is expressed as:
Wherein, in the set of the combination, subset mutual exclusion two-by-two, the union of whole subsets is whole sample class;
The loss function of this group of Weak Classifier combination is:
The number of this group of Weak Classifier is average often to reduce 1, and the increase of the error in classification summation of this group of Weak Classifier combination is:
Wherein, JCWhen for the number of Weak Classifier being C, the sum of the deviations of this group of Weak Classifier combination, JnfFor Weak Classifier
Number when being nf, the sum of the deviations of this group of Weak Classifier combination,
In above-mentioned formula (4), withMinimum target chooses Weak Classifier combination;
During training is per level-one strong classifier, often training obtains a Weak Classifier combination, calculates current state
The false drop rate that lower strong classifier classifies to training sample, if the false drop rate is higher than the false drop rate threshold value of this grade of strong classifier,
Continue to train, otherwise training terminates.
Human hand detection method of the present invention comprises the following steps:
(1) image to be detected is gathered;
(2) feature extraction is carried out to image to be detected;
(3) feature of step (2) extraction is obtained into the defeated of Weak Classifier compared with the threshold value of each Weak Classifier
Go out;
(4) output valve of each Weak Classifier under strong classifier is added up to be added summation, then by the summation and strong classification
The threshold value comparison of device, if summation is judged as human hand, carries out the judgement of next stage, only often more than the threshold value of strong classifier
Level-one is judged as human hand, and final human hand grader is just judged as human hand, otherwise terminates.
The feature extracted in the step 2 includes:Haar features, HOG features and Variance feature.
Compared with prior art, the present invention passes through one group of side for sharing Weak Classifier of training in each round of boosting
Method and the method that the corresponding multiple threshold values of target classification are used in each Weak Classifier, avoid JointBoost algorithms
In the target classification covered for not being shared feature that is likely to occur the problem of can not effectively classifying.And due to multiple classifications
Sharing feature between target is using what the Weak Classifier that the sharing feature is trained was detected the target of these classifications
When, Weak Classifier can also be shared, and overall judgement number during so as to reduce detection accelerates detection speed.
Description of the drawings
Explanation and specific embodiment are described in further detail the present invention below in conjunction with the accompanying drawings.
The human hand grader overall structure block diagram that Fig. 1 is combined based on shared Weak Classifier;
Fig. 2 is per level-one strong classifier training flow chart;
Fig. 3 rises grader number schematic diagram in criterion selection Weak Classifier combination using most slow error;
Specific embodiment
Human hand detection method of the present invention is divided into image reading to be checked, box counting algorithm, human hand object judgement, with
And testing result exports this four part composition.Specifically include foundation, the feature extraction side of sample of human hand training set and test set
The training method of method and grader.For the training method of grader, in order to achieve the object of the present invention, the present invention exists
It is made that many places are improved using following methods on the basis of JointBoost algorithms:
First, during training grader, in each round of boosting, the group of several sharing features is obtained
It closes.For the set mutual exclusion two-by-two that the sample class of each feature in shared combination is formed, and the union of entire set is
Whole sample class.According to the combination of such a sharing feature, one group of corresponding Weak Classifier is trained.In order to determine
The number of feature in sharing feature combination defines the criterion of one " minimal error rising ".
Secondly, for each Weak Classifier, the output valve of function is value and the classification of some feature according to sample
The threshold value of device is compared what the result drawn determined.It is every in of the invention due to being related to the detection of multi-pose or multi-class targets
A Weak Classifier sets each classification one threshold value being more suitable for, to strengthen grader respectively according to the difference of classification
For the classification capacity between different classes of target and background.
First, the overall structure and the course of work of grader are described.Fig. 1 is the present invention is based on shared Weak Classifiers
The overall structure block diagram of the human hand grader of combination.After image window to be detected is read into, it is necessary first to image be carried out special
Sign extraction.The grader of the present invention according to the Haar features, HOG features and Variance feature of target image to be checked judge
It is no to include human hand, it is therefore desirable to the target image to be checked be calculated, extract corresponding three kinds of features.
After obtaining Haar, HOG and Variance feature of target image to be checked, the value of these three features is input to trained
To grader in.After compared with a series of threshold value, comprehensive all corresponding scoring events of feature, grader
Final judgement is provided, whether which includes human hand.Grader is using the cascade structure of multistage strong classifier, mesh to be checked
Logo image must be judged as " comprising human hand ", entire cascade classifier just can finally judge by every level-one strong classifier
For " including human hand ", otherwise, arbitrary level-one strong classifier is judged as that " not including human hand " can all cause target image to be checked at this
Grade exits, and is cascaded grader and is finally judged as " not including human hand ".The use of cascade classifier significantly reduces flase drop
Generation.
The training method of the human hand grader to being combined the present invention is based on shared Weak Classifier is described below.
Entire grader is made of the cascade of multistage strong classifier, and per level-one strong classifier by several to sharing Weak Classifier
Combination collectively forms.Acquisition for every level-one strong classifier sets a false drop rate threshold value first, and each group is obtained in training
After shared Weak Classifier combination, training sample is once sentenced in all comprehensive all Weak Classifiers combination obtained at present
It is disconnected, if false drop rate is higher than default threshold value, continue the training of the Weak Classifier combination of next round;If on the contrary, less than default
Threshold value, then no longer carry out next round, this grade of strong classifier training finishes.The shared Weak Classifier that all training are obtained, which combines, to be closed
It is this grade of strong classifier together.
It is described below for the combination of shared Weak Classifier and training method.
Assuming that a shared C classes, one shares N number of training sample, each training sample viWeight for classification c is
And share C weight.Assuming that have nf Weak Classifier as a combination, and sample set is divided for nf subset, then kth
A Weak Classifier is expressed as:
WhereinCorrespond to the threshold value of classification c, so each Weak Classifier h for the gradermFor different classifications
Just have multiple and different threshold values.This group of Weak Classifier combination can be expressed as:
WhereinAnd
The form of loss function Weak Classifier at this time can be expressed as:
If representing loss function using Weak Classifier combination, for:
It as nf=1, represents that the feature chosen is shared by all sample packets, therefore only selects a shared spy
Sign, and build the Weak Classifier that all sample packets are shared;And as nf=C, that is, represent by sample set according to sample from
The classification of body is grouped, and every one kind is individually trained, therefore without using sharing feature.As nf=C, letter is lost
Several values reaches minimum, but grader quantity reaches maximum, and it is also most slow to cause detection speed.And when nf is gradually decreased
When, the value of loss function can be increased, but grader quantity is reduced, and detection speed is speeded.
Suitable nf values in order to obtain, the criterion risen using most slow error choose the value of nf.First, I
This C classes sample is each trained respectively using exclusive feature, then have C unshared weak points in the combination of the Weak Classifier that obtains
Class device, and assume that their sum of the deviations is JC.Then, reduce in classifiers combination that (one weak to C-1 for the number of Weak Classifier
Grader becomes the shared Weak Classifier of two class samples), sum of the deviations is J at this timeC-1.So, when Weak Classifier number reduces 1
When, error in classification summation adds Δ JC-1=JC-1-JC.This process is repeated, when Weak Classifier number subtracts in classifiers combination
(some Weak Classifiers may be shared between a few class samples, and it is respective to be likely present some remaining a few class samples during as little as nf
Exclusive Weak Classifier), error in classification summation adds Δ Jnf=Jnf-JC, the number of Weak Classifier is average often to reduce 1, weak typing
The increase of the error in classification summation of device combination so represents:
Formula (5) is illustrated combined with Weak Classifier in grader number reduce, the increased speed of error in classification summation misses
The concept of poor rising gradient.Error increment (molecule) is bigger, and the decrement (denominator) of Weak Classifier number is smaller in combination, thenValue it is bigger, when representing to choose the combination of nf Weak Classifier, counter productive that error in classification increase embodies compared with
The good effect that the reduction of grader number embodies is more apparent;On the contrary, molecule is smaller, denominator is bigger, thenIt is worth smaller, expression
It is more apparent to reduce the good effect that grader number is brought, and unobvious are got in the increase of error in classification.Therefore, as shown in Figure 1, I
Select so thatValue reach minimum Weak Classifier combination.Abscissa is the number of Weak Classifier in classifiers combination,
Ordinate is the error in classification summation of a classifiers, and line segment 1 and line segment 2 are represented respectively when Weak Classifier number in grader group
When being reduced to 1 and nf from C, the increased trend of error in classification summation.Line segment 1,2 and the angle of axis of abscissas are respectively α1, α2,
And tan α1With tan α2Value show respectively the gradient of line segment 1 and line segment 2, i.e., error in classification summation is with Weak Classifier in combination
It reduces and increased speed degree.Tan α in figure1> tan α2, i.e., 2 error increase is slower compared with 1, so 2 be one compared with 1
A better Weak Classifier assembled scheme.
Claims (3)
1. a kind of training method of the human hand grader of shared Weak Classifier combination, the human hand grader is by multistage strong classifier grade
Connection is formed, and is combined and formed by multipair shared Weak Classifier per level-one strong classifier;Target image to be checked must by force be divided by every level-one
Class device is judged as " comprising human hand " that entire cascade classifier just can finally be judged as " comprising human hand ", otherwise be judged as " not including
Human hand ", it is characterised in that:
Assuming that a shared C class sample class, N number of training sample, each training sample viWeight for classification c is WiC, altogether
There is C weight;Nf Weak Classifier is as a combination, and sample set is divided for nf subset, then, k-th of weak typing
Device is:
Wherein, m is the wheel number of training,For the characteristic value of k-th of sample in training sample,Correspond to for human hand grader
The threshold value of classification c, Sk(nf) k-th of the subset concentrated for nf son;
This group of Weak Classifier combination is expressed as:
Wherein, in the set of the combination, subset mutual exclusion two-by-two, the union of whole subsets is whole sample class;
The loss function of this group of Weak Classifier combination is:
The number of this group of Weak Classifier is average often to reduce 1, and the increase of the error in classification summation of this group of Weak Classifier combination is:
Wherein, JCWhen for the number of Weak Classifier being C, the sum of the deviations of this group of Weak Classifier combination, JnfFor of Weak Classifier
When number is nf, the sum of the deviations of this group of Weak Classifier combination,
In above-mentioned formula (4), withMinimum target chooses Weak Classifier combination;
During training is per level-one strong classifier, often training obtains a Weak Classifier combination, calculates strong under current state
The false drop rate that grader classifies to training sample, if the false drop rate continues higher than the false drop rate threshold value of this grade of strong classifier
Training, otherwise training terminate.
2. a kind of human hand detection method based on training method described in claim 1, it is characterised in that:Comprise the following steps:
(1) image to be detected is gathered;
(2) feature extraction is carried out to image to be detected;
(3) feature of step (2) extraction is obtained into the output of Weak Classifier compared with the threshold value of each Weak Classifier;
(4) output valve of each Weak Classifier under strong classifier is added up to be added summation, then by the summation and strong classifier
Threshold value comparison, if summation is judged as human hand, carries out the judgement of next stage, only per level-one more than the threshold value of strong classifier
It is judged as human hand, final human hand grader is just judged as human hand, otherwise terminates.
3. detection method according to claim 2, it is characterised in that:The feature extracted in the step 2 includes:Haar is special
Sign, HOG features and Variance feature.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410351088.1A CN104123570B (en) | 2014-07-22 | 2014-07-22 | Human hand grader and training and detection method based on the combination of shared Weak Classifier |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410351088.1A CN104123570B (en) | 2014-07-22 | 2014-07-22 | Human hand grader and training and detection method based on the combination of shared Weak Classifier |
Publications (2)
Publication Number | Publication Date |
---|---|
CN104123570A CN104123570A (en) | 2014-10-29 |
CN104123570B true CN104123570B (en) | 2018-06-05 |
Family
ID=51768975
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201410351088.1A Active CN104123570B (en) | 2014-07-22 | 2014-07-22 | Human hand grader and training and detection method based on the combination of shared Weak Classifier |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN104123570B (en) |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101763515A (en) * | 2009-09-23 | 2010-06-30 | 中国科学院自动化研究所 | Real-time gesture interaction method based on computer vision |
WO2013091370A1 (en) * | 2011-12-22 | 2013-06-27 | 中国科学院自动化研究所 | Human body part detection method based on parallel statistics learning of 3d depth image information |
CN103793056A (en) * | 2014-01-26 | 2014-05-14 | 华南理工大学 | Mid-air gesture roaming control method based on distance vector |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8648959B2 (en) * | 2010-11-11 | 2014-02-11 | DigitalOptics Corporation Europe Limited | Rapid auto-focus using classifier chains, MEMS and/or multiple object focusing |
US8995772B2 (en) * | 2012-11-09 | 2015-03-31 | Microsoft Technology Licensing, Llc | Real-time face detection using pixel pairs |
-
2014
- 2014-07-22 CN CN201410351088.1A patent/CN104123570B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101763515A (en) * | 2009-09-23 | 2010-06-30 | 中国科学院自动化研究所 | Real-time gesture interaction method based on computer vision |
WO2013091370A1 (en) * | 2011-12-22 | 2013-06-27 | 中国科学院自动化研究所 | Human body part detection method based on parallel statistics learning of 3d depth image information |
CN103793056A (en) * | 2014-01-26 | 2014-05-14 | 华南理工大学 | Mid-air gesture roaming control method based on distance vector |
Non-Patent Citations (2)
Title |
---|
一种人手检测系统的设计与实现;李丹立等;《计算机应用与软件》;20110930;第28卷(第9期);56-59 * |
基于肤色和Real AdaBoost的人手检测方法研究;张玲玲;《中国优秀硕士学位论文全文数据库 信息科技辑》;20130715(第7期);I138-1294 * |
Also Published As
Publication number | Publication date |
---|---|
CN104123570A (en) | 2014-10-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106650806B (en) | A kind of cooperating type depth net model methodology for pedestrian detection | |
CN105160317B (en) | One kind being based on area dividing pedestrian gender identification method | |
CN105139039B (en) | The recognition methods of the micro- expression of human face in video frequency sequence | |
CN104517104B (en) | A kind of face identification method and system based under monitoring scene | |
CN111523462B (en) | Video sequence expression recognition system and method based on self-attention enhanced CNN | |
CN106778796B (en) | Human body action recognition method and system based on hybrid cooperative training | |
CN111931758B (en) | Face recognition method and device combining facial veins | |
CN107463920A (en) | A kind of face identification method for eliminating partial occlusion thing and influenceing | |
CN104504362A (en) | Face detection method based on convolutional neural network | |
CN107463892A (en) | Pedestrian detection method in a kind of image of combination contextual information and multi-stage characteristics | |
CN102521616B (en) | Pedestrian detection method on basis of sparse representation | |
CN107944431A (en) | A kind of intelligent identification Method based on motion change | |
CN102194108A (en) | Smiley face expression recognition method based on clustering linear discriminant analysis of feature selection | |
Wei et al. | Deep spatial pyramid ensemble for cultural event recognition | |
CN102163281A (en) | Real-time human body detection method based on AdaBoost frame and colour of head | |
CN105956570B (en) | Smiling face's recognition methods based on lip feature and deep learning | |
WO2013075295A1 (en) | Clothing identification method and system for low-resolution video | |
Guo et al. | Real-time hand detection based on multi-stage HOG-SVM classifier | |
CN110599463A (en) | Tongue image detection and positioning algorithm based on lightweight cascade neural network | |
CN109190456A (en) | Pedestrian detection method is overlooked based on the multiple features fusion of converging channels feature and gray level co-occurrence matrixes | |
CN105608443B (en) | A kind of face identification method of multiple features description and local decision weighting | |
CN108090513A (en) | Multi-biological characteristic blending algorithm based on particle cluster algorithm and typical correlation fractal dimension | |
CN106529441A (en) | Fuzzy boundary fragmentation-based depth motion map human body action recognition method | |
Zhang et al. | Iris image classification based on color information | |
Pang et al. | F-DR Net: Face detection and recognition in One Net |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |