CN109934162A - Facial image identification and video clip intercept method based on Struck track algorithm - Google Patents
Facial image identification and video clip intercept method based on Struck track algorithm Download PDFInfo
- Publication number
- CN109934162A CN109934162A CN201910183195.0A CN201910183195A CN109934162A CN 109934162 A CN109934162 A CN 109934162A CN 201910183195 A CN201910183195 A CN 201910183195A CN 109934162 A CN109934162 A CN 109934162A
- Authority
- CN
- China
- Prior art keywords
- sequence
- struck
- feature
- image
- facial image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Image Analysis (AREA)
Abstract
Facial image identification and video clip intercept method based on Struck track algorithm, belong to field of face identification.The problem of current most area remains in manpower to the retrospect of day net engineering video data and searches the stage, monitoring data treatment effeciency low problem.A kind of facial image identification and video clip intercept method based on Struck track algorithm.Struck tracker is automatically initialized after detecting face using AdaBoost method combination Haar-like feature, tracker truncated picture frame sequence is pre-processed, using One-classSVM as searching classification device after the feature of increment PCA method abstraction sequence, the target that video clip associated in video content is exported after inputting specified image is realized.Video retrieval method based on facial image of the invention at the appointed time can judge whether target occurs in section, and export the relevant video segments of retrieval.
Description
Technical field
The present invention relates to a kind of facial image recognition methods.In particular to a kind of face figure based on Struck track algorithm
As recognition methods.
Background technique
Due to the development of day net engineering, public security department has grasped most monitoring camera.These cameras are every
The order of magnitude that it generates video data is very big.Although the video data of magnanimity is an important leverage of public safety,
Want to find specific target in video information using manpower, there is that time-consuming, subjective judgement is strong, and serious forgiveness is low etc. asks
Topic.
Therefore, it is necessary to the automatic methods that one kind can accurately and rapidly find target.
Summary of the invention
The purpose of the present invention is to solve current most areas to remain in the retrospect of day net engineering video data
The problem of manpower is searched the stage, monitoring data treatment effeciency low problem, and propose a kind of based on Struck track algorithm
Facial image identification and video clip intercept method.
It is a kind of based on Struck track algorithm facial image identification with video clip intercept method, the method by with
Lower step
It is rapid to realize:
Step 1: using human face data in AdaBoost method combination Haar-like feature detection video；
Step 2: automatically initialize Struck tracker using the human face data detected, using tracker carry out with
Track obtains human face image sequence；
Step 3: being pre-processed to the facial image frame sequence of tracker interception；
Step 4: establishing feature space for facial image and being classified with One-classSVM；It realizes and inputs specified image
Sequence
The target of video clip associated in video content is exported after the feature of column.
The invention has the benefit that
The present invention is the video retrieval method based on facial image, at the appointed time can judge whether target goes out in section
It is existing, and export the relevant video segments of retrieval.Video frequency searching mode based on facial image, using Struck track algorithm.
Struck algorithm be it is a kind of can on-line study computer vision track algorithm.It can be with auto-initiation in conjunction with AdaBoost algorithm
Tracker, furthermore AdaBoost is run simultaneously with tracker, and tracking accuracy can be improved.
If tracker loses current goal, next target can be also detected in time, improve the efficiency of segment interception.
Specific embodiment
Specific embodiment 1:
The facial image based on Struck track algorithm of present embodiment identifies and video clip intercept method, the side
Method is logical
Cross following steps realization:
Step 1: using human face data in AdaBoost method combination Haar-like feature detection video；
Step 2: automatically initialize Struck tracker using the human face data detected, using tracker carry out with
Track obtains human face image sequence；
Step 3: being pre-processed to the facial image frame sequence of tracker interception；
Step 4: establishing feature space for facial image and being classified with One-classSVM；It realizes and inputs specified image
The target of video clip associated in video content is exported after the feature of sequence.
Specific embodiment 2:
Unlike specific embodiment one, the facial image identification based on Struck track algorithm of present embodiment
With video clip intercept method, in the step two, Struck is automatically initialized using the human face data detected and is tracked
Device is tracked to obtain the process of human face image sequence using tracker are as follows:
The step of face tracking includes detection and tracking, herein to the image frame sequence in video using cascade Ada
Boost algorithm detects face.Ada Boost is a kind of iterative algorithm, and the polymerization of several Weak Classifiers is become one strong classification
Device can effectively improve the accuracy of learning method.
It is the grey scale change of image that Haar characteristic value is corresponding.Certain features of face substantially can be simple by rectangular characteristic
Ground description, such as the opposite relationship between light and dark of face: the color contrast of eye color and cheek, the bridge of the nose, mouth and ambient color
Comparison etc..
But there are apparent weakness by Haar-like: rectangular characteristic is only to some simple graphic structures, and single features are only
It can describe indicate boundary, line segment, corner angle etc. by the structure for the particular orientation that small area rectangle indicates.
Step 2 one, on the basis of polymerizeing multiple features by Haar-like according to certain rule using Ada Boost algorithm
Classifier, the classifier after polymerizeing in this way are enough to judge whether a certain image-region is face；Cascade Ada Boost will judge respectively
The classifier polymerization of kind face characteristic realizes the identification from eyes, nose to whole face to obtain the higher classifier of effectiveness.
Step 2 two after being handled using first frame of the Ada Boost method to each sequence, will be detected in the frame
Human face data for initializing tracker；
The structural trained track algorithm of step 2 three, Struck full name:
Firstly, getting tracking target in first frame in given Target Photo sequence, tracked with this object initialization
The parameter of device；
Then, second the position that following object is likely to occur is estimated using the forecasting mechanism of own；
Finally, handling emerging tracking target sample, optimal segmentation plane is solved, it is anti-that threshold decision is added
Only supporting vector crosses growth, traces into target object and updates supporting vector collection.
Specific embodiment 3:
Unlike specific embodiment two, the facial image identification based on Struck track algorithm of present embodiment
With video clip intercept method, in the step three, pretreated mistake is carried out to the facial image frame sequence of tracker interception
Journey are as follows: the influence to reduce light to characteristics of image is pre-processed to each frame in image sequence using histogram equalization,
It is normalized again；Normalization is that posture, sample image not of uniform size are obtained one group of characteristic point pair by affine maps
Together, image of the same size reuses the center that Gravity-center Template finds face；Improving image quality, unified image gray scale
Value and size eliminate the noise useless to feature extraction.
Specific embodiment 4:
Unlike specific embodiment three, the facial image identification based on Struck track algorithm of present embodiment
With video clip intercept method, in the step four, establishes feature space for facial image and classified with One-classSVM；
The process of the target of video clip associated in video content, tool are exported after the feature of the sequence of the specified image of realization input
Body are as follows:
Increment PCA method extracts the feature of pretreated sequence；Using One-classSVM as searching classification device, realize
The target of video clip associated in video content is exported after the feature of the sequence of the specified image of input；PCA establishes feature
Space, and incremental update frame by frame, while updating One-classSVM.
Specific embodiment 5:
Unlike specific embodiment four, the facial image identification based on Struck track algorithm of present embodiment
The process for the feature for extracting pretreated sequence with video clip intercept method, the increment PCA method is,
The projecting direction that can utmostly indicate sample distribution, i.e. unit vector are found under lowest mean square meaning, it is
A kind of common feature extracting method.The high dimensional feature linear transformation of initial data is one group of each dimension linear independence by PCA
Vector indicates, is widely used in the extraction of data main feature component, is the common method of high dimensional data dimensionality reduction.PCA algorithm
Core concept be to be mapped to high dimensional feature in the space dimension for being less than former data dimension, the relatively low-dimensional that this neotectonics comes out is
Completely new orthogonal characteristic, referred to as principal component.
Feature vector Y of the piece image X in lower dimensional space is obtained by formula (1) projection pattern
Y=MTX(1)
X is that the vector of normalized picture frame indicates as pretreated data, X ∈ Rm, Y ∈ Rn, m > n, in above formula
M represent projection matrix, Y is the new feature extracted.Enable m dimensional vector Xi(i=1,2,3 ..., l) represents the face after normalization
Image, X represent XiThe mean value of vector, the covariance matrix of these samples:
The column vector of the projection matrix M of principal component analysis by covariance matrix the preceding maximum feature vector of m characteristic value
Composition, sorts from large to small λ for characteristic value1≥λ2≥λ1, and the corresponding feature vector of these characteristic values is set as μi, each in this way
Facial image can project to μ1,μ2,…,μ1The feature vector of l dimension, the dimension of l are obtained in the feature space constituted
It is very big, it is not necessary that be fully retained, m feature vector makes before choosing:
In above formula, the α close to 1 can be chosen as far as possible, this illustrates that energy of the sample on preceding m axis can represent substantially
Entire energy.
Although kind of algorithm is common but serious forgiveness is smaller, because its requirement to training sample set is very high, face is usually trained
Conditions are consistent as far as possible can be only achieved best matching effect with the illumination of target face, background etc.；And the face usually in video
Image change fluctuation is very big, is not restrained by above-mentioned extraneous factor.
Using increment PCA learning algorithm, increment PCA can be gradually reduced during increment on the basis of PCA because
The divergence for extracting feature is effectively reduced in the extraneous factors such as background, light, angle difference caused by image sequence different frame；
One training sequence is a fixed personage, feeds back result to subspace and single class supporting vector in identification process
To classify on machine One-class SVM, dynamic adjusts the unit vector direction of subspace and the classifying face of One-classSVM,
The influence of the factors such as illumination, background is reduced in this way.One-Class SVM classifier is suitable for two classification problems, is
Or it is not.Because exporting there are two results, training data only needs enough positive samples, it is ensured that classifier can be accurate
Judge one of result, meets application scenarios.SVM different inner product kernel functions will obtain reflecting for different higher dimensional spaces
It penetrates, the kernel function used herein is radial basis function and multinomial kernel function:
(1) radial basis function (RBF):
(2) multinomial kernel:
Multinomial kernel and RBF core are the common two kinds of kernel functions of SVM, and RBF kernel function is most widely used, it can be by one
Sample is mapped to the space of a more higher-dimension, and RBF network can approach arbitrary nonlinear function, be difficult to parse in processing system
Rule, generalization ability is good, fast convergence rate.
The present invention can also have other various embodiments, without deviating from the spirit and substance of the present invention, this field
Technical staff makes various corresponding changes and modifications in accordance with the present invention, but these corresponding changes and modifications all should belong to
The protection scope of the appended claims of the present invention.
Claims (5)
1. facial image identification and video clip intercept method based on Struck track algorithm, it is characterised in that: the method
It is realized by following steps:
Step 1: using human face data in AdaBoost method combination Haar-like feature detection video；
Step 2: automatically initializing Struck tracker using the human face data detected, track using tracker
To human face image sequence；
Step 3: being pre-processed to the facial image frame sequence of tracker interception；
Step 4: establishing feature space for facial image and being classified with One-classSVM；Realize the sequence for inputting specified image
Feature after export the target of associated video clip in video content.
2. the facial image identification and video clip intercept method according to claim 1 based on Struck track algorithm,
It is characterized by: automatically initialize Struck tracker using the human face data detected in the step two, using with
Track device is tracked to obtain the process of human face image sequence are as follows:
Step 2 one, polymerize multiple features by Haar-like according to certain rule using Ada Boost algorithm on the basis of classification
Device, to judge whether a certain image-region is face；Cascade Ada Boost will judge various face characteristics classifier polymerization with
The higher classifier of effectiveness is obtained, realizes the identification from eyes, nose to whole face.
Step 2 two, after being handled using first frame of the Ada Boost method to each sequence, the people that will be detected in the frame
Face data are for initializing tracker；
The structural trained track algorithm of step 2 three, Struck full name:
Firstly, tracking target is got in first frame, with this object initialization tracker in given Target Photo sequence
Parameter；
Then, second the position that following object is likely to occur is estimated using the forecasting mechanism of own；
Finally, handling emerging tracking target sample, optimal segmentation plane is solved, threshold decision, which is added, to be prevented from propping up
It holds vector and crosses growth, trace into target object and update supporting vector collection.
3. the facial image identification and video clip intercept method according to claim 2 based on Struck track algorithm,
It is characterized by: carrying out pretreated process to the facial image frame sequence of tracker interception are as follows: use in the step three
Histogram equalization pre-processes the influence to reduce light to characteristics of image to each frame in image sequence, then carries out normalizing
Change；By posture, sample image not of uniform size by affine maps, one group of feature point alignment, image of the same size are obtained, then
The center of face is found using Gravity-center Template；Improving image quality, unified image gray value and size, elimination propose feature
Take useless noise.
4. the facial image identification and video clip intercept method according to claim 3 based on Struck track algorithm,
It is characterized by: establishing feature space for facial image in the step four and being classified with One-classSVM；Realize input
The process of the target of video clip associated in video content is exported after the feature of the sequence of specified image, specifically:
Increment PCA method extracts the feature of pretreated sequence；Using One-classSVM as searching classification device, input is realized
The target of video clip associated in video content is exported after the feature of the sequence of specified image；PCA establishes feature space,
And incremental update frame by frame, while updating One-classSVM.
5. the facial image identification and video clip intercept method according to claim 4 based on Struck track algorithm,
It is characterized by: the process for the feature that the increment PCA method extracts pretreated sequence is,
Using increment PCA learning algorithm, increment PCA can be gradually reduced during increment on the basis of PCA because background,
The divergence for extracting feature is effectively reduced in the extraneous factors such as light, angle difference caused by image sequence different frame；One training
Sequence is a fixed personage, feeds back result to subspace and one-class support vector machines One-class SVM in identification process
On classify, dynamic adjust subspace unit vector direction and One-classSVM classifying face, the kernel function used for
Radial basis function and multinomial kernel function:
(1) radial basis function (RBF):
(2) multinomial kernel:
The most widely used space that a sample is mapped to a more higher-dimension of RBF kernel function, RBF network can approach arbitrary non-
Linear function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910183195.0A CN109934162A (en) | 2019-03-12 | 2019-03-12 | Facial image identification and video clip intercept method based on Struck track algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910183195.0A CN109934162A (en) | 2019-03-12 | 2019-03-12 | Facial image identification and video clip intercept method based on Struck track algorithm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109934162A true CN109934162A (en) | 2019-06-25 |
Family
ID=66987071
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910183195.0A Pending CN109934162A (en) | 2019-03-12 | 2019-03-12 | Facial image identification and video clip intercept method based on Struck track algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109934162A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110879970A (en) * | 2019-10-21 | 2020-03-13 | 武汉兴图新科电子股份有限公司 | Video interest area face abstraction method and device based on deep learning and storage device thereof |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130004018A1 (en) * | 2011-06-29 | 2013-01-03 | Postech Academy - Industry Foundation | Method and apparatus for detecting object using volumetric feature vector and 3d haar-like filters |
| CN104616319A (en) * | 2015-01-28 | 2015-05-13 | 南京信息工程大学 | Multi-feature selection target tracking method based on support vector machine |
| CN105279773A (en) * | 2015-10-27 | 2016-01-27 | 杭州电子科技大学 | TLD framework based modified video tracking optimization method |
| CN106778484A (en) * | 2016-11-16 | 2017-05-31 | 南宁市浩发科技有限公司 | Moving vehicle tracking under traffic scene |
-
2019
- 2019-03-12 CN CN201910183195.0A patent/CN109934162A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130004018A1 (en) * | 2011-06-29 | 2013-01-03 | Postech Academy - Industry Foundation | Method and apparatus for detecting object using volumetric feature vector and 3d haar-like filters |
| CN104616319A (en) * | 2015-01-28 | 2015-05-13 | 南京信息工程大学 | Multi-feature selection target tracking method based on support vector machine |
| CN105279773A (en) * | 2015-10-27 | 2016-01-27 | 杭州电子科技大学 | TLD framework based modified video tracking optimization method |
| CN106778484A (en) * | 2016-11-16 | 2017-05-31 | 南宁市浩发科技有限公司 | Moving vehicle tracking under traffic scene |
Non-Patent Citations (1)
| Title |
|---|
| 王楠: "基于 Struck 跟踪算法的人脸图像识别与截取研究", 《科技创新与应用》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110879970A (en) * | 2019-10-21 | 2020-03-13 | 武汉兴图新科电子股份有限公司 | Video interest area face abstraction method and device based on deep learning and storage device thereof |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104517104B (en) | A kind of face identification method and system based under monitoring scene | |
| CN107316036B (en) | Insect pest identification method based on cascade classifier | |
| Ismael et al. | Face recognition using Viola-Jones depending on Python | |
| US7773781B2 (en) | Face detection method and apparatus and security system employing the same | |
| US20070031041A1 (en) | Apparatus and method for detecting a face | |
| CN109145742A (en) | A kind of pedestrian recognition method and system | |
| CN105760858A (en) | Pedestrian detection method and apparatus based on Haar-like intermediate layer filtering features | |
| CN109376717A (en) | Personal identification method, device, electronic equipment and the storage medium of face comparison | |
| CN103530648A (en) | Face recognition method based on multi-frame images | |
| Verma et al. | Local invariant feature-based gender recognition from facial images | |
| CN102314598A (en) | Retinex theory-based method for detecting human eyes under complex illumination | |
| Araar et al. | Traffic sign recognition using a synthetic data training approach | |
| Dhinesh et al. | Detection of leaf disease using principal component analysis and linear support vector machine | |
| CN110909678B (en) | Face recognition method and system based on width learning network feature extraction | |
| CN109934162A (en) | Facial image identification and video clip intercept method based on Struck track algorithm | |
| CN113468925A (en) | Shielded face recognition method, intelligent terminal and storage medium | |
| CN107657201A (en) | NEXT series of products characteristics of image identifying systems and its recognition methods | |
| Li et al. | Real-time traffic sign detection algorithm based on dynamic threshold segmentation and SVM | |
| Li et al. | Face detection technology based on combining skin color model with improved adaboost algorithm | |
| Reddy et al. | Comparison of HOG and fisherfaces based face recognition system using MATLAB | |
| Vivekanandam et al. | Face recognition from video frames using hidden markov model classification model based on modified random feature extraction | |
| Hbali et al. | Object detection based on HOG features: Faces and dual-eyes augmented reality | |
| Micheal et al. | Multi-view face detection using Normalized Pixel Difference feature | |
| Zhang et al. | Face detection system based on video stream | |
| Hassanpour et al. | Sensing Image Regions for Enhancing Accuracy in People Re-identification |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190625 |