CN106503700A - Haar features multiprocessing framework face detection system and detection method based on FPGA - Google Patents
Haar features multiprocessing framework face detection system and detection method based on FPGA Download PDFInfo
- Publication number
- CN106503700A CN106503700A CN201611270462.0A CN201611270462A CN106503700A CN 106503700 A CN106503700 A CN 106503700A CN 201611270462 A CN201611270462 A CN 201611270462A CN 106503700 A CN106503700 A CN 106503700A
- Authority
- CN
- China
- Prior art keywords
- face characteristic
- face
- value
- characteristic value
- threshold value
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/161—Detection; Localisation; Normalisation
- G06V40/166—Detection; Localisation; Normalisation using acquisition arrangements
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/172—Classification, e.g. identification
Abstract
Haar features multiprocessing framework face detection system and detection method based on FPGA, belongs to data facial features localization technical field.Solve existing facial features localization algorithm complicated, the slow problem of processor arithmetic speed.The image storage module of the present invention is used for extracting the haar features in the range detection window of n × n in camera acquisition image, using haar feature extraction face characteristics, and using each pixel point coordinates of face characteristic calculating face characteristic;And face characteristic each pixel point coordinates for obtaining will be calculated be sent to integral image maker, integral image maker calculates face characteristic value according to face characteristic each pixel point coordinates using integrogram method;And the face characteristic value P transmission grader for obtaining will be calculated;Grader is compared with characteristic threshold value A × m using face characteristic value P for receiving, and remains larger than face characteristic value P of characteristic threshold value A × m.The present invention is applied to for Face datection use.
Description
Technical field
The invention belongs to data facial features localization technical field.
Background technology
Face datection such as recognizes that in many applications monitoring plays important work in video conference and camera auto-focusing etc.
With.Due to the extensive work of computer vision field so that Face datection algorithm development is rapidly reaching the degree of real-time detection.
Viola and Jones proposes a kind of method based on machine learning in document, can realize comparing while precision is kept
High frame speed.However, the algorithm is too slow relative to conventional processors (particularly on embedded platform).
Content of the invention
The present invention is that the slow problem of processor arithmetic speed is proposed in order to solve existing facial features localization algorithm complexity
A kind of Haar feature multiprocessing framework face check systems based on FPGA.
Haar feature multiprocessing framework face detection systems based on FPGA of the present invention, it include image storage mould
Block 1, integral image maker 2, grader 3 and detection window Zoom module 4;
Image storage module 1 is used for extracting the haar features in the range detection window of n × n in camera acquisition image,
Using haar feature extraction face characteristics, and using each pixel point coordinates of face characteristic calculating face characteristic;And calculating is obtained
Face characteristic each pixel point coordinates send;Wherein n is positive integer;
Integral image maker 2 is used for receiving face characteristic each pixel point coordinates, according to each pixel of face characteristic
Coordinate calculates face characteristic value using integrogram method;And face characteristic value P for calculating acquisition is sent;
Grader 3 receives and stores face characteristic value information, and using face characteristic value P for receiving and characteristic threshold value A × m
It is compared, remains larger than face characteristic value P of characteristic threshold value A × m, when the number of face characteristic value P is more than B, makes m=m+
1, face characteristic value P is compared with characteristic threshold value A × m, remains larger than face characteristic value P of characteristic threshold value A × m;Until people
When the number of face eigenvalue P is less than B, stops being compared, and face characteristic value P for obtaining more afterwards is exported, complete
Haar feature multiprocessing framework Face datections based on FPGA;Wherein, A, m, P, B are positive integer;
Detection window Zoom module 4 is used for whether the size of the range detection window for judging n × n to cover camera acquisition figure
Picture, if so, completes the detection of face characteristic, and face characteristic value P that features training data device 33 is exported is exported, people is completed
The detection of face feature, otherwise makes n=n × a, sends the scope for extracting n × n in camera acquisition image to image frame buffer 11
Haar character-driven signals in detection window, wherein, a is the integer more than 1.
Further, image storage module 1 includes image frame buffer 11 and address generator 12,
Image frame buffer 11 is used for extracting the haar features in the range detection window of n × n in camera acquisition image,
Using haar feature extraction face characteristics;And the face characteristic information of extraction is sent;
Address generator 12 is used for receiving face characteristic information, and the face characteristic information of reception is stored as two-dimentional battle array
Row, using pixel address=(y*w+x), calculate the coordinate (x, y) of each pixel of face characteristic, and wherein, w is integrogram image
Width, and face characteristic each pixel point coordinates for obtaining will be calculated send.
Further, grader 3 includes feature classifiers 31, stage comparator 32 and features training data device 33;
Feature classifiers 31 are used for receiving and storing face characteristic value information, and using face characteristic value P for receiving and spy
Levy threshold value A × m to be compared, remain larger than face characteristic value P of characteristic threshold value A × m, will be greater than the face of characteristic threshold value A × m
Eigenvalue P sends;Wherein, m=1;
Stage comparator 32 is used for receiving as m=2, face characteristic value P more than characteristic threshold value A × m, and using reception
Face characteristic value P be compared with characteristic threshold value A × m, remain larger than face characteristic value P of characteristic threshold value A × m, will be greater than
Face characteristic value P of characteristic threshold value A × m sends;Wherein, m=2;
Features training data device 33 is used for receiving as m=2, face characteristic value P more than characteristic threshold value A × m, and utilizes
Face characteristic value P of reception is compared with characteristic threshold value A × m, remains larger than face characteristic value P of characteristic threshold value A × m, when
When the number of face characteristic value P is more than B, m=m+1, face characteristic value P is made to be compared with characteristic threshold value A × m, remain larger than
Face characteristic value P of characteristic threshold value A × m;Until when the number of face characteristic value P is less than B, stop being compared, and to comparing
Face characteristic value P for obtaining afterwards is exported, and obtains the face characteristic in detection window, wherein, m > 2;
Detection window Zoom module 4 is used for whether the size of the range detection window for judging n × n to cover camera acquisition figure
Picture, if so, completes the detection of face characteristic, and face characteristic value P that features training data device 33 is exported is exported, people is completed
The detection of face feature, otherwise makes n=n × a, sends the scope for extracting n × n in camera acquisition image to image frame buffer 11
Haar character-driven signals in detection window, wherein, a is the integer more than 1.
Based on the Haar feature multiprocessing framework method for detecting human face of FPGA, the method is concretely comprised the following steps:
The haar features in the range detection window of n × n in camera acquisition image are extracted, using haar feature extraction people
Face feature, and using each pixel point coordinates of face characteristic calculating face characteristic;And face characteristic each picture for obtaining will be calculated
The step of vegetarian refreshments coordinate sends;Wherein n is positive integer;
For receiving face characteristic each pixel point coordinates, according to face characteristic, each pixel point coordinates utilizes integrogram
Method, calculates face characteristic value;And by calculate obtain face characteristic value P send the step of;
Face characteristic value information is received and stored, and is compared with characteristic threshold value A × m using face characteristic value P for receiving
Compared with, remain larger than characteristic threshold value A × m face characteristic value P the step of;In the step, when the number of face characteristic value P is more than B
When, make m=m+1, face characteristic value P be compared with characteristic threshold value A × m, remain larger than the face characteristic of characteristic threshold value A × m
Value P;Until when the number of face characteristic value P is less than B, stopping being compared, and face characteristic value P for obtaining more afterwards being carried out
Output, the step of complete the Haar feature multiprocessing framework Face datections based on FPGA;Wherein, A, m, P, B are positive number;
For judging the step of whether size of the range detection window of n × n covers camera acquisition image, the step
In, face characteristic value P that features training data device 33 is exported is carried out if it is, complete the detection of face characteristic by judged result
Output, completes the detection of face characteristic, otherwise makes n=n × a, sends to image frame buffer 11 and extracts camera acquisition image
Haar character-driven signals in the range detection window of middle n × n.
The present invention proposes a kind of Haar feature multiprocessing framework face detection systems based on FPGA, it is possible to use multiple
Intrinsic concurrency in application specific processor unit algorithm, carries out Multiple detection to a face, will repeat to examine in post-processing stages
The result of survey is combined, and improves the Face datection speed based on Haar features, reaches the purpose of real-time detection.
Description of the drawings
Fig. 1 is the theory diagram of the Haar feature multiprocessing framework face detection systems based on FPGA of the present invention;
Fig. 2 is interior processing unit theory diagram in the face detection system described in specific embodiment one;
Fig. 3 slides for detection window on the entire image and finds the schematic diagram of face;
Fig. 4 is the structural representation of Haar features in a detection window;
Structural representations of the Fig. 5 (a) for edge Haar features;
Structural representations of the Fig. 5 (b) for line Haar features;
Structural representations of the Fig. 5 (c) for specific direction Haar features;
Fig. 6 is region D pixels and schematic diagram.
Specific embodiment
In order that the objects, technical solutions and advantages of the present invention become more apparent, below in conjunction with drawings and Examples, right
The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only in order to explain the present invention, and
It is not used in the restriction present invention.
Specific embodiment one, combine Fig. 1 present embodiment is described, the Haar based on FPGA described in present embodiment is special
Multiprocessing framework face detection system is levied, it includes image storage module 1, integral image maker 2, grader 3 and detection window
Mouth Zoom module 4;
Image storage module 1 is used for extracting the haar features in the range detection window of n × n in camera acquisition image,
Using haar feature extraction face characteristics, and using each pixel point coordinates of face characteristic calculating face characteristic;And calculating is obtained
Face characteristic each pixel point coordinates send;Wherein n is positive integer;
Integral image maker 2 is used for receiving face characteristic each pixel point coordinates, according to each pixel of face characteristic
Coordinate calculates face characteristic value using integrogram method;And face characteristic value P for calculating acquisition is sent;
Grader 3 receives and stores face characteristic value information, and using face characteristic value P for receiving and characteristic threshold value A × m
It is compared, remains larger than face characteristic value P of characteristic threshold value A × m, when the number of face characteristic value P is more than B, makes m=m+
1, face characteristic value P is compared with characteristic threshold value A × m, remains larger than face characteristic value P of characteristic threshold value A × m;Until people
When the number of face eigenvalue P is less than B, stops being compared, and face characteristic value P for obtaining more afterwards is exported, complete
Haar feature multiprocessing framework Face datections based on FPGA;Wherein, A, m, P, B are positive integer;
Detection window Zoom module 4 is used for whether the size of the range detection window for judging n × n to cover camera acquisition figure
Picture, if so, completes the detection of face characteristic, and face characteristic value P that features training data device 33 is exported is exported, people is completed
The detection of face feature, otherwise makes n=n × a, sends the scope for extracting n × n in camera acquisition image to image frame buffer 11
Haar character-driven signals in detection window, wherein, a is the integer more than 1.
The present invention has gone out a kind of cascade classifier framework by parallel processing based on Haar face detection algorithms.Herein in detail
Haar detection-phases are carefully described, while describing pretreatment stage in detail, the picture signal of input A/D conversions is carried out first
And decoding, image trimming is being carried out, each two field picture is being sent to sort module through address generator, finally data is passed through
USB interface transmitter shows in the display.
Image is stored as two-dimensional array using substantial amounts of BRAM on FPGA by frame buffer, conversely, image is deposited
Store up as an one-dimensional vector, pixel (x, y) address is calculated as (y*w+x).The pixel value read from frame buffer is sent to
Calculate the integral image maker module of frame.
The pixel value address that address generator is generated is obtained with the frame buffer of raster fashion surface sweeping framework by reading.
As framework is necessarily drawn to reduce the detection window to run different size size, therefore the address generator is also responsible for basis
Arest neighbors NO emissions reduction algorithm is realized using low complex degree streamline scaling this two field picture, it is possible to achieve per the clock cycle 1
The handling capacity of address is sent to frame buffer, with the pixel value needed for retrieving.
Integration diagram generator is generated in current operation scale and stores the integral image of current frame image.Follow-up
The Haar classifier stage must provide for high bandwidth.Therefore, the integral image of current detection window is stored in depositor, so as to
Multiple values are read simultaneously can.
The method for having several integral images to generate, a kind of is can to calculate the integral image of whole frame and store it in
As pretreatment stage in BRAM.But, the current integration image window in depositor is needed due to us, therefore required letter
Breath must be sent in depositor before Haar classifier stage running;Another kind is direct during Haar detection-phases
Current integration image window is dynamically generated in a register.Architectural framework proposed by the present invention generates product with system dynamics
Partial image combines, to synthesize complete face detection system.Second method has the excellent of minimizing Time & Space Complexity
Point.
Classifier data is divided into several parts based on the cascade classifier of stub, is stored in individually in the way of being grouped
Increase the bandwidth of Face datection in BRAM.The details of each rectangle position in one group of BRAM storage Haar feature.Another
Group BRAM stores the threshold value of each Haar feature.The value on the value on the left side and the right is stored in another group of BRAM.Due to its size
Less, phase threshold is stored in a register.
Stage in due to Haar classifier independently of one another, therefore they can be divided into can be with the N number of of parallel processing
Set, N is positive integer, for example, if number of stages is 20 and has 4 processing units, for first processing units can be given
Give 5 stages to process, second processing unit can give other 5 stages, and remaining 10 stages can distribute to which
Its two processing unit.Feature quantity in each stage is different, and should carry out to the stage between each processing unit
Divide, to obtain equal load distribution and minimum idle time.
Face datection engine and pre- learning value and the position of Haar features of classifier data retrieval threshold, determine current inspection
Survey whether window includes face.The internal structure of the Face datection engine comprising 4 processing units is as shown in Figure 2.The framework can
The higher rate of Face datection frame rate is realized with scaled to include more processing units.Next different sons are described
Module.
Each processing unit (Processing Unit) processes one group of stage, and detection window from classifier data
The all stages for distributing to the processing unit are passed through, have then generated and pass through signal.If detection window failed in any stage, it
It is immediately generated bypass signal.
Fig. 2 is by its position is sent to integral image module from grader, receives feature square from integral image module
Graphic data value.The weighted sum of rectangular characteristic and characteristic threshold value are compared, lvalue or r value are selected according to comparative result.Selected
The value that selects is accumulated by the stage and is compared with phase threshold.If accumulated value is more than threshold value, window passes through the specific rank
Section.If detection window failed in any stage, generate bypass signal and detection window slides into next position.Otherwise,
If the last stage passes through in institute's allocated phase, produce and pass through signal.Due to different processing units have different
Stage, so skip and pass through signal demand synchronously carry out between them.
Synchronization module is made up of four triggers, and each trigger is arranged by corresponding processing unit, and detection window passes through
During all stages for specifying set, corresponding processing unit is generated and passes through signal.All it is set if all of trigger, then anticipates
Taste all stages of grader and is all passed through, and illustrates that detection window includes face.In any one new detection window
Before upper operation grader, trigger can be eliminated.
If skipping synchronization module detection window any one stage failure, corresponding processing unit in processing unit
Being immediately generated bypass signal notifies face detecting and alarm window to fail, and loads next detection window.However, dividing
Before the next detection window of class device operation, it should the current pipeline of flashing processing units.
Specific embodiment two, present embodiment is many to the Haar features based on FPGA described in specific embodiment one
Processing framework face detection system is further illustrated, and image storage module 1 includes image frame buffer 11 and address generator
12,
Image frame buffer 11 is used for extracting the haar features in the range detection window of n × n in camera acquisition image,
Using haar feature extraction face characteristics;And the face characteristic information of extraction is sent;
Address generator 12 is used for receiving face characteristic information, and the face characteristic information of reception is stored as two-dimentional battle array
Row, using pixel address=(y*w+x), calculate the coordinate (x, y) of each pixel of face characteristic, and wherein, w is integrogram image
Width, and face characteristic each pixel point coordinates for obtaining will be calculated send.
Specific embodiment three, present embodiment is many to the Haar features based on FPGA described in specific embodiment one
Processing framework face detection system is further illustrated, and grader 3 includes feature classifiers 31, stage comparator 32 and feature instruction
Practice data device 33;
Feature classifiers 31 are used for receiving and storing face characteristic value information, and using face characteristic value P for receiving and spy
Levy threshold value A × m to be compared, remain larger than face characteristic value P of characteristic threshold value A × m, will be greater than the face of characteristic threshold value A × m
Eigenvalue P sends;Wherein, m=1;
Stage comparator 32 is used for receiving as m=2, face characteristic value P more than characteristic threshold value A × m, and using reception
Face characteristic value P be compared with characteristic threshold value A × m, remain larger than face characteristic value P of characteristic threshold value A × m, will be greater than
Face characteristic value P of characteristic threshold value A × m sends;Wherein, m=2;
Features training data device 33 is used for receiving as m=2, face characteristic value P more than characteristic threshold value A × m, and utilizes
Face characteristic value P of reception is compared with characteristic threshold value A × m, remains larger than face characteristic value P of characteristic threshold value A × m, when
When the number of face characteristic value P is more than B, m=m+1, face characteristic value P is made to be compared with characteristic threshold value A × m, remain larger than
Face characteristic value P of characteristic threshold value A × m;Until when the number of face characteristic value P is less than B, stop being compared, and to comparing
Face characteristic value P for obtaining afterwards is exported, and obtains the face characteristic in detection window, wherein, m > 2;
The Haar feature multiprocessing frameworks Face datection side based on FPGA described in specific embodiment four, present embodiment
Method, the method are concretely comprised the following steps:
The haar features in the range detection window of n × n in camera acquisition image are extracted, using haar feature extraction people
Face feature, and using each pixel point coordinates of face characteristic calculating face characteristic;And face characteristic each picture for obtaining will be calculated
The step of vegetarian refreshments coordinate sends;Wherein n is positive integer;
For receiving face characteristic each pixel point coordinates, according to face characteristic, each pixel point coordinates utilizes integrogram
Method, calculates face characteristic value;And by calculate obtain face characteristic value P send the step of;
Face characteristic value information is received and stored, and is compared with characteristic threshold value A × m using face characteristic value P for receiving
Compared with, remain larger than characteristic threshold value A × m face characteristic value P the step of;In the step, when the number of face characteristic value P is more than B
When, make m=m+1, face characteristic value P be compared with characteristic threshold value A × m, remain larger than the face characteristic of characteristic threshold value A × m
Value P;Until when the number of face characteristic value P is less than B, stopping being compared, and face characteristic value P for obtaining more afterwards being carried out
Output, the step of complete the Haar feature multiprocessing framework Face datections based on FPGA;Wherein, A, m, P, B are positive number;
For judging the step of whether size of the range detection window of n × n covers camera acquisition image, the step
In, face characteristic value P that features training data device 33 is exported is carried out if it is, complete the detection of face characteristic by judged result
Output, completes the detection of face characteristic, otherwise makes n=n × a, sends to image frame buffer 11 and extracts camera acquisition image
Haar character-driven signals in the range detection window of middle n × n.
In the present invention, Face datection algorithm using a square detection window (for example, 20 × 20) scan whole image
To find face characteristic, as shown in Figure 3.
If finding enough features in the ad-hoc location of a window, then the window specification includes a face.For
The various sizes of face of detection, zooms in and out after window surface sweeping whole image and repeats to process.
However, the method is high computational, and it is not suitable for embedded system.In addition, for Embedded Application, will
Image is scaled, while detection window is maintained at 20 × 20 sizes, it need not keep one completely in memory
The many sized image pyramids in ground.Conversely, when needed, closest reduction scheme can be complicated in a low-down hardware
Complete to generate framework by the scale of regulation from original image in the dynamic pipeline of degree.Although accuracy of detection is depending on being used
Scaling algorithm, arest neighbors scheme maintains the accuracy of the Face datection compared favourably with software system.
When using a detection window scanogram, in each position, window has to pass through one Haar features
The cascade classifier of composition.These Haar are characterized in that the simple structure being made up of 2 or 3 rectangles, as shown in Figure 4.
The little characteristic set in different size and position and can have use in the detection windows such as different transverse and longitudinals ratios, from
And thousands of Haar features are generated, such as shown in (a), (b) of Fig. 5, (c).
These Haar features are used for detecting facial characteristics.For example, it is understood that eye areas are colourity than the bridge of the nose and cheek
Darker.
Therefore, an any of the above rectangle is put on human face region, then the pixel in white rectangle region and will be deducted
The value of the pixel sum in black rectangle region is referred to as face characteristic value.This rectangle is put into a non-face region again, then
The eigenvalue for calculating is different from face characteristic value, so the purpose of these rectangles is exactly that face characteristic is quantified, to distinguish people
Face and non-face.Above-mentioned discussion shows, by using two features, can just imply the presence of eyes.In order that detection window bag
Containing whole face, need to calculate thousands of Haar features.More than 6000 Haar features are divided into 38 independences by Viola-Jones
Stage, only when a window is by all stages, it is just classified as a face.Detailed algorithm can be looked in
Arrive.Compared to early stage, later stage has larger complexity and accuracy, makes non-face window as far as possible in early stage quilt
Refusal.
Huge due to Haar-like number of features, cause to calculate pixel and the needs in each rectangle of Haar features
The plenty of time is consumed, detection speed is had a strong impact on.Therefore Viola-Jones proposes integrogram method as pretreatment stage, and this has
Help to calculate in Time constant each feature and.Integrogram is one and picture size identical two-dimensional matrix, integrogram
Coordinate (x, y) comprising the pixel of the coordinate (x, y) in upper left face in original image and.
If integrogram is ii, original graph is i', then the integrated value that position P (x, y) goes out is
Wherein ii (x, y) represents the integration map values at pixel P (x, y) place, and i (x', y') is " original for point (x', y') place
Figure ", represents the gray value of P (x, y) upper left side all pixels point in image.
Rectangular characteristic value is calculated using integrogram, the pixel value of region D can utilize 1,2,3,4 points of integrogram is counting
Calculate, as shown in Figure 6;Region D pixel and it is calculated as:ii(4)+ii(1)-ii(2)-ii(3)
Therefore, the pixel value of ii (1)=region A;
The pixel value of the pixel value of ii (2)=region A+region B;
The pixel value of the pixel value of ii (3)=region A+region C;
The pixel value of the pixel value of the pixel value of the pixel value of ii (4)=region A+region B+region C+region D;
So, the pixel value of region D is ii (4)+ii (1)-ii (2)-ii (3).
According to above-mentioned algorithm can calculate within the section time certain rectangular characteristic pixel and, and rectangle can be drawn
Calculating for eigenvalue is only relevant with the integrogram of this feature end points, unrelated with image coordinate value.Therefore, regardless of rectangular characteristic size
How, the time of eigenvalue calculation consumption is all constant, and simply simple plus and minus calculation, improves detection speed.
Tranining database used in machine learning algorithm is normalized square mean.Therefore, incoming detection window
Needs are normalized, and it is powerful for different lighting conditions that this causes algorithm[14].Can be used to calculate current detection window
The variance of mouth,
Wherein, v is detection window pixel value, ∑ x2The quadratic sum of detection window pixel value is represented, z is to detect the picture in window
Prime number, μ are the average pixel values in detection window, and which can be calculated from integral image window.If in pretreatment rank
The range of summation table of the quadratic sum integral image of pixel value (integrated square image) is generated during section, then can effectively calculate ∑
v2.
Weak Classifier:(each Haar-like feature is equivalent to a Weak Classifier)
Give a series of training sample (x1,y1),(x2,y2),...(xg,yg) wherein yi=0 represents which is that negative sample is (non-
Face), yi=1 is expressed as positive sample (face), and g is training sample total quantity.
Initialization weight w1,i=D (i), D (i) are the inverse of twice positive sample or negative sample, to t=1 ..., T:Normalizing
Change weight:T is sample training data set
Wherein, i is the reference numerals of sample, trains a Weak Classifier h (x, f, p, θ), weak study to calculate to each feature f
Method finds an optimal threshold, makes minimum weight error rate ε of the Weak ClassifierfMinimum, obtaining Weak Classifier is:
εf=∑iqi|h(xi,f,p,θ-yi)| (4)
H (x, f, p, θ)=1pf (x)<p(θ)
H (x, f, p, θ)=0, other (5)
Choose optimal Weak Classifier htX () (has minimal error rate εt);Wherein, x is video in window, and p is for controlling
The designator in the direction of the sign of inequality, θ are in probable range to distinguish face and non-face threshold value;
ht(x)=h (x, ft,pt,θt) (6)
Obtain according to this optimal Weak Classifier adjustment weight:
Wherein ei=0 represents xiCorrectly classified, ei=1 represents xiMistakenly classified,εtFor minimal error
Rate, βtFor vision response test.
Strong classifier:
Show that T optimal Weak Classifier is constituted through T iteration, as a result for:
C (x)=0, other (8)
Wherein,
When the number of Weak Classifier reaches certain amount, the strong classifier of composition just has compared with high detection rate, but examines
Survey than relatively time-consuming, therefore Paul Viola and Michael Jones propose classifiers combination to be got up using the method for cascade, i.e.,
Several strong classifiers are divided into some groups, being then together in series becomes cascade classifier.Be suitable for various parallel, wherein
Some by before work utilize.The grader was made up of several independent stages.The individual stage is special by some Haar
Composition is levied, all detection windows of a framework are required for by this grader.
Non-face for part sample just can be excluded by some small-sized strong classifiers using it, and only face sample
This and the non-face sample of minority enter into the detection of next stage grader.The reason for speed is lifted be in a sub-picture, big portion
It is all non-face sample to divide window to be detected.
The false drop rate of hypothesis whole system is F, and the system is made up of k level graders, then the flase drop of one-level strong classifier
Rate will meet following relation with the false drop rate of whole system:
In the same manner, the relational expression that first-level class device sends out verification and measurement ratio and the verification and measurement ratio of whole system can be obtained:
Wherein, k is the series of grader in system, and wherein, D represents the verification and measurement ratio of whole system, diRepresent one-level to divide by force
The verification and measurement ratio of class device.After every one-level strong classifier verification and measurement ratio and false drop rate is determined, training obtains satisfactory one-level
Strong classifier, so that constitute strong classifier.
The training process of cascade classifier:First have to determine the highest false drop rate F allowed by whole cascade classifiermaxWith
Lowest detection rate Dmin, the cascade classifier determines by how many strong classifiers, i.e. the series of the cascade classifier.Assume that each is strong
The highest false drop rate of grader is fmax, lowest detection rate is dmin, then
There are three kinds of frameworks, be absorbed in and whole system realized within hardware using the concurrency of the face detection algorithm based on Haar
System.Accelerate face detection by the several Haar features in parallel processing moment[8].Document [9] proposes a mixing
The a large amount of parallelizations of solution, wherein starting stage, and later stage phase sequence is executed.Document [10] proposes a kind of framework, its
The parallel running on some different detection windows of middle grader.
Presently preferred embodiments of the present invention is the foregoing is only, not in order to limit the present invention, all in essence of the invention
Any modification, equivalent and improvement that is made within god and principle etc., should be included within the scope of the present invention.
Claims (4)
1. Haar feature multiprocessing framework face detection systems based on FPGA, it is characterised in that it includes image storage module
(1), integral image maker (2), grader (3) and detection window Zoom module (4);
Image storage module (1) is used for extracting the haar features in the range detection window of n × n in camera acquisition image, profit
With haar feature extraction face characteristics, and face characteristic each pixel point coordinates is calculated using face characteristic;And obtain calculating
Face characteristic each pixel point coordinates send;Wherein n is positive integer;
Integral image maker (2) is used for receiving face characteristic each pixel point coordinates, and according to face characteristic, each pixel is sat
Mark calculates face characteristic value using integrogram method;And face characteristic value P for calculating acquisition is sent;
Grader (3) receives and stores face characteristic value information, and is entered with characteristic threshold value A × m using face characteristic value P for receiving
Row compares, and remains larger than face characteristic value P of characteristic threshold value A × m, when the number of face characteristic value P is more than B, makes m=m+1,
Face characteristic value P is compared with characteristic threshold value A × m, remains larger than face characteristic value P of characteristic threshold value A × m;Until face
When the number of eigenvalue P is less than B, stops being compared, and face characteristic value P for obtaining more afterwards is exported, complete base
Haar feature multiprocessing framework Face datections in FPGA;Wherein, A, m, P, B are positive integer;
Detection window Zoom module (4) is used for whether the size of the range detection window for judging n × n to cover camera acquisition figure
Picture, if so, completes the detection of face characteristic, and face characteristic value P that features training data device 33 is exported is exported, people is completed
The detection of face feature, otherwise makes n=n × a, sends the scope for extracting n × n in camera acquisition image to image frame buffer 11
Haar character-driven signals in detection window, wherein, a is the integer more than 1.
2. Haar feature multiprocessing framework face detection systems based on FPGA according to claim 1, it is characterised in that
Image storage module (1) includes image frame buffer (11) and address generator (12),
Image frame buffer (11) is used for extracting the haar features in the range detection window of n × n in camera acquisition image, profit
With haar feature extraction face characteristics;And the face characteristic information of extraction is sent;
Address generator (12) is used for receiving face characteristic information, and the face characteristic information of reception is stored as two-dimensional array,
Using pixel address=(y*w+x), the coordinate (x, y) of each pixel of face characteristic is calculated, wherein, w is integrogram image
Width, and face characteristic each pixel point coordinates for obtaining will be calculated send.
3. Haar feature multiprocessing framework face detection systems based on FPGA according to claim 1, it is characterised in that
Grader (3) includes feature classifiers (31), stage comparator (32) and features training data device (33);
Feature classifiers (31) are used for receiving and storing face characteristic value information, and using face characteristic value P and feature for receiving
Threshold value A × m is compared, and remains larger than face characteristic value P of characteristic threshold value A × m, and the face that will be greater than characteristic threshold value A × m is special
Value indicative P sends;Wherein, m=1;
Stage comparator (32) is used for receiving as m=2, face characteristic value P more than characteristic threshold value A × m, and using reception
Face characteristic value P is compared with characteristic threshold value A × m, is remained larger than face characteristic value P of characteristic threshold value A × m, be will be greater than spy
Face characteristic value P for levying threshold value A × m sends;Wherein, m=2;
Features training data device (33) is used for receiving as m=2, face characteristic value P more than characteristic threshold value A × m, and utilization connects
Face characteristic value P of receipts is compared with characteristic threshold value A × m, is remained larger than face characteristic value P of characteristic threshold value A × m, is worked as people
When the number of face eigenvalue P is more than B, makes m=m+1, face characteristic value P be compared with characteristic threshold value A × m, remain larger than spy
Levy face characteristic value P of threshold value A × m;Until face characteristic value P number be less than B when, stop being compared, and to comparing after
Face characteristic value P of acquisition is exported, and obtains the face characteristic in detection window, wherein, m > 2;
Detection window Zoom module (4) is used for whether the size of the range detection window for judging n × n to cover camera acquisition figure
Picture, if so, completes the detection of face characteristic, face characteristic value P that features training data device (33) is exported is exported, is completed
The detection of face characteristic, otherwise makes n=n × a, sends to image frame buffer (11) and extracts n × n in camera acquisition image
Haar character-driven signals in range detection window, wherein, a is the integer more than 1.
4. the Haar feature multiprocessing framework method for detecting human face based on FPGA, it is characterised in that the method is concretely comprised the following steps:
The haar features in the range detection window of n × n in camera acquisition image are extracted, special using haar feature extractions face
Levy, and using each pixel point coordinates of face characteristic calculating face characteristic;And face characteristic each pixel for obtaining will be calculated
The step of coordinate sends;Wherein n is positive integer;
For receiving face characteristic each pixel point coordinates, according to face characteristic each pixel point coordinates using integrogram method, meter
Calculate face characteristic value;And by calculate obtain face characteristic value P send the step of;
Face characteristic value information is received and stored, and is compared with characteristic threshold value A × m using face characteristic value P for receiving, protected
The step of staying face characteristic value P more than characteristic threshold value A × m;In the step, when the number of face characteristic value P is more than B, m is made
=m+1, face characteristic value P are compared with characteristic threshold value A × m, remain larger than face characteristic value P of characteristic threshold value A × m;Directly
To face eigenvalue P number be less than B when, stop being compared, and face characteristic value P for obtaining more afterwards exported,
The step of completing the Haar feature multiprocessing framework Face datections based on FPGA;Wherein, A, m, P, B are positive number;
For the step of whether size of the range detection window of n × n covers camera acquisition image judged, in the step, sentence
Face characteristic value P that features training data device (33) is exported is carried out defeated if it is, complete the detection of face characteristic by disconnected result
Go out, complete the detection of face characteristic, otherwise make n=n × a, send to image frame buffer (11) and extract camera acquisition image
Haar character-driven signals in the range detection window of middle n × n.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611270462.0A CN106503700A (en) | 2016-12-30 | 2016-12-30 | Haar features multiprocessing framework face detection system and detection method based on FPGA |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611270462.0A CN106503700A (en) | 2016-12-30 | 2016-12-30 | Haar features multiprocessing framework face detection system and detection method based on FPGA |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106503700A true CN106503700A (en) | 2017-03-15 |
Family
ID=58334974
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611270462.0A Pending CN106503700A (en) | 2016-12-30 | 2016-12-30 | Haar features multiprocessing framework face detection system and detection method based on FPGA |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106503700A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107368797A (en) * | 2017-07-06 | 2017-11-21 | 湖南中云飞华信息技术有限公司 | The parallel method for detecting human face of multi-angle, device and terminal device |
CN108549838A (en) * | 2018-03-13 | 2018-09-18 | 苏州奥科德瑞智能科技有限公司 | A kind of back-up surveillance method of view-based access control model system |
CN110110589A (en) * | 2019-03-25 | 2019-08-09 | 电子科技大学 | Face classification method based on FPGA parallel computation |
CN110119678A (en) * | 2019-03-29 | 2019-08-13 | 珠海亿智电子科技有限公司 | A kind of FPGA verifying system and method for recognition of face |
CN111783876A (en) * | 2020-06-30 | 2020-10-16 | 西安全志科技有限公司 | Self-adaptive intelligent detection circuit and image intelligent detection method |
CN111881715A (en) * | 2020-06-03 | 2020-11-03 | 西安电子科技大学 | Face detection hardware acceleration method, system and equipment |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN202351893U (en) * | 2011-10-27 | 2012-07-25 | 上海德致伦电子科技有限公司 | Human face detecting system |
CN202662026U (en) * | 2012-06-05 | 2013-01-09 | 上海锦江电子技术工程有限公司 | Multi-face recognition system |
CN103390152A (en) * | 2013-07-02 | 2013-11-13 | 华南理工大学 | Sight tracking system suitable for human-computer interaction and based on system on programmable chip (SOPC) |
-
2016
- 2016-12-30 CN CN201611270462.0A patent/CN106503700A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN202351893U (en) * | 2011-10-27 | 2012-07-25 | 上海德致伦电子科技有限公司 | Human face detecting system |
CN202662026U (en) * | 2012-06-05 | 2013-01-09 | 上海锦江电子技术工程有限公司 | Multi-face recognition system |
CN103390152A (en) * | 2013-07-02 | 2013-11-13 | 华南理工大学 | Sight tracking system suitable for human-computer interaction and based on system on programmable chip (SOPC) |
Non-Patent Citations (3)
Title |
---|
肖琳: "基于Adaboost的人脸检测算法研究及其FPGA实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
许昀 等: "Adaboost 算法的 FPGA 实现与性能分析", 《微计算机信息》 * |
高金良: "基于Adaboost算法的人脸实时检测及FPGA设计", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107368797A (en) * | 2017-07-06 | 2017-11-21 | 湖南中云飞华信息技术有限公司 | The parallel method for detecting human face of multi-angle, device and terminal device |
CN108549838A (en) * | 2018-03-13 | 2018-09-18 | 苏州奥科德瑞智能科技有限公司 | A kind of back-up surveillance method of view-based access control model system |
CN110110589A (en) * | 2019-03-25 | 2019-08-09 | 电子科技大学 | Face classification method based on FPGA parallel computation |
CN110119678A (en) * | 2019-03-29 | 2019-08-13 | 珠海亿智电子科技有限公司 | A kind of FPGA verifying system and method for recognition of face |
CN111881715A (en) * | 2020-06-03 | 2020-11-03 | 西安电子科技大学 | Face detection hardware acceleration method, system and equipment |
CN111881715B (en) * | 2020-06-03 | 2023-07-28 | 西安电子科技大学 | Face detection hardware acceleration method, system and equipment |
CN111783876A (en) * | 2020-06-30 | 2020-10-16 | 西安全志科技有限公司 | Self-adaptive intelligent detection circuit and image intelligent detection method |
CN111783876B (en) * | 2020-06-30 | 2023-10-20 | 西安全志科技有限公司 | Self-adaptive intelligent detection circuit and image intelligent detection method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106503700A (en) | Haar features multiprocessing framework face detection system and detection method based on FPGA | |
CN101339609B (en) | Image processing apparatus and image processing method | |
CN100361138C (en) | Method and system of real time detecting and continuous tracing human face in video frequency sequence | |
CN105913040B (en) | The real-time pedestrian detecting system of dual camera under the conditions of noctovision | |
TWI439951B (en) | Facial gender identification system and method and computer program products thereof | |
Rekik et al. | A new visual speech recognition approach for RGB-D cameras | |
US20170091950A1 (en) | Method and system for tracking an object | |
CN104573625B (en) | Data processing system and its generating device and type of face detection method | |
CN101211411B (en) | Human body detection process and device | |
EP1973064A2 (en) | Object image detection method and object image detection device | |
US20070058836A1 (en) | Object classification in video data | |
CN109255375A (en) | Panoramic picture method for checking object based on deep learning | |
CN104166841A (en) | Rapid detection identification method for specified pedestrian or vehicle in video monitoring network | |
CN101576953A (en) | Classification method and device of human body posture | |
CN110738160A (en) | human face quality evaluation method combining with human face detection | |
CN104036284A (en) | Adaboost algorithm based multi-scale pedestrian detection method | |
JP2012084140A (en) | Video analysis method and system | |
CN109522883A (en) | A kind of method for detecting human face, system, device and storage medium | |
CN109902576B (en) | Training method and application of head and shoulder image classifier | |
CN108764096A (en) | A kind of pedestrian weight identifying system and method | |
WO2012046426A1 (en) | Object detection device, object detection method, and object detection program | |
Jiang et al. | Lightweight object detection network model suitable for indoor mobile robots | |
US9053354B2 (en) | Fast face detection technique | |
WO2013026205A1 (en) | System and method for detecting and recognizing rectangular traffic signs | |
CN103150575A (en) | Real-time three-dimensional unmarked human body gesture recognition method and system |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | 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: 20170315 |