CN101620673A - Robust face detecting and tracking method - Google Patents

Abstract

The invention relates to a robust face detecting and tracking method comprising a robust face detecting method and a robust face tracking method. The robust face detecting method comprises the following steps: firstly, detecting a face based on similar-Haar small-wave characteristic extraction and a weight trimming AdaBoost algorithm; and then, verifying the face by using obvious detection. The robust face tracking method comprises the following steps: firstly, ensuring a face to be tracked based on face detection and skin color verification; and then, tracking the face based on Camshift track and skin color verification. When the face is lost, a face to be tracked is supplied according to the face detection and the skin color verification.

Description

A kind of people's face of robust detects and tracking

Technical field

The present invention relates to a kind of method of technical field of image processing, especially people's face of robust detects and tracking.

Background technology

It is to determine in a given image whether people's face is arranged that people's face detects, and exports the position of everyone face.The method that people's face detects roughly is divided into method based on knowledge, based on the method for feature, based on the method for template, based on the method for outward appearance etc.Comparatively successful fast face detecting method is based on the sensation target detection method of integral image and AdaBoost algorithm at present.Because the variability of people's face and the complicacy of actual application environment, situation detects with people's face of realizing robust need to examine yardstick, position, quantity, direction, expression, illumination, block etc.

Face tracking is meant follows the tracks of one or more people's faces in continuous sequence of video images.Face tracking in the video sequence can be applied to supervisory system, also can be used as the pre-service that recognition of face, facial pose identification and Expression Recognition contour level are used.Face tracking method is broadly divided into method based on feature, based on the method for model, based on the method for the colour of skin with based on method of shape etc.Have the characteristics of quick and low calculated amount based on the method for the colour of skin, but the background color of the variation of illumination and the similar colour of skin may cause the inefficacy of face tracking.Be not subjected to the influence of background color and illumination variation based on the method for shape, but highly in disorder background may have influence on tracking effect.The influence of factor such as the face tracking of robust need consider illumination, block also will reach the requirement of real-time simultaneously.

For people's face of realizing robust under the complex background detects and follows the tracks of, people's face that the present invention proposes a kind of robust detects and tracking, wherein the method for detecting human face of robust can be realized higher detection rate and lower false drop rate, and the face tracking method of robust can be issued to higher tracking accuracy in the various situations of blocking.

Summary of the invention

The objective of the invention is to detect and follow the tracks of applied environment at people's face of complexity, the people's face that proposes robust detects and face tracking method.The method for detecting human face of robust at first extracts based on class Haar wavelet character and weights pruning AdaBoost algorithm carries out the detection of people's face, utilizes conspicuousness to detect then people's face is verified.The face tracking method of robust at first detects based on people's face and people's face to be tracked is determined in colour of skin checking, carry out face tracking based on Camshift tracking and colour of skin checking then, under the situation that people's face is lost, provide people's face to be tracked by detection of people's face and colour of skin checking again.

People's face of a kind of robust of the present invention detects and tracking, comprises the method for detecting human face of robust and the face tracking method of robust.

About the method for detecting human face of robust, its step is as follows:

Step 1: utilize weights to prune the AdaBoost algorithm and carry out the training of people's face sorter, its process is as follows:

Step 1.1: given training image (x ₁, y ₁) ..., (x _n, y _n) y wherein _i=0,1 corresponds respectively to negative sample and positive sample.

Step 1.2: initialization weights ${\mathrm{\ω}}_{1,i}=\frac{1}{2m},\frac{1}{2l}$ Correspond respectively to y _i=0,1, wherein m and l are respectively positive sample and negative sample number.

Step 1.3: for t=1 ..., T:

Step 1.3.1: normalization weights ${\mathrm{\ω}}_{t,i}\←\frac{{\mathrm{\ω}}_{t,i}}{{\mathrm{\Σ}}_{j=1}^n{\mathrm{\ω}}_{t,j}}$ Make ω _{T, i}Set ω _tIt is a probability distribution.

Step 1.3.2: as t＞T ₀, determine threshold value θ _t, make $\underset{{\mathrm{\&omega;}}_{t,i}<{\mathrm{\&theta;}}_t}{\mathrm{\&Sigma;}}{\mathrm{\&omega;}}_{t,i}<0.01\&le;\underset{{\mathrm{\&omega;}}_{t,i}\&le;{\mathrm{\&theta;}}_t}{\mathrm{\&Sigma;}}{\mathrm{\&omega;}}_{t,i}.$ With weights less than θ _tSample get rid of, be not used in the training of Weak Classifier.

Step 1.3.3:, train a sorter h for each feature j _jMake its corresponding feature.Calculating is for ω _tError, ε _j=∑ _iω _i| h _j(x _i)-y _i|.

Step 1.3.4: selection sort device h _t, make it have least error ε _t

Step 1.3.5: refreshing weight: ${\mathrm{\&omega;}}_{t+1,i}={\mathrm{\&omega;}}_{t,i}{\mathrm{\&beta;}}_t^{1-e_i}$ If sample x wherein _iBy the then e that correctly classifies _i=0, otherwise e _i=1, ${\mathrm{\&beta;}}_t=\frac{{\mathrm{\&epsiv;}}_t}{1-{\mathrm{\&epsiv;}}_t}.$

Step 1.4: final strong classifier is:

Wherein ${\mathrm{\&alpha;}}_t=\log \frac{1}{{\mathrm{\&beta;}}_t}---\left(1\right)$

Step 2: input picture is carried out window scanning, each window is extracted the class Haar wavelet character (the paper An extended set of Haar-like features for rapid object detection that list of references: Lienhart people such as (sharp grace Harts) delivered on International Conference on Image Processing (Flame Image Process international conference) in 2002 (fast target based on the class Haar feature of expanding detects)) of one group of expansion, as shown in Figure 2;

Step 3: the feature of extracting is sent into the AdaBoost sorter shown in the formula (1) carry out the detection of people's face, if h (x)=1, then this scanning window is people's face window to be verified; If h (x)=0, then this scanning window is not people's face window.

Step 4: people's face window to be verified is carried out the conspicuousness checking, as the formula (2):

$H_D=-\underset{i}{\mathrm{\&Sigma;}}{P}_D\left(d_i\right){\log }_2{P}_D\left(d_i\right)---\left(2\right)$

Wherein be P (d _i) image value d _iProbability.

If step 5: selected threshold θ is H _D〉=θ, people's face window then to be verified is by checking, i.e. this window area behaviour face; If H _DBy checking, promptly this window area is not people's face for＜θ, people's face window then to be verified.

About the face tracking method of robust, its step is as follows:

Step 1: use method for detecting human face of the present invention to detect people's face,, carry out step 2 if detect people's face; If do not detect people's face, next frame carry out step 1.

Step 2: detected people's face is carried out colour of skin checking, if detected people's face verify by the colour of skin, carry out step 3; If detected people's face is disallowable with this people's face not by colour of skin checking, next frame carry out step 1.

Step 3: people's face window that will be by colour of skin checking is as object to be tracked.

Step 4: utilize Camshift algorithm (the paper Computer vision face tracking for use in a perceptualuser interface (being applied to the computer vision face tracking of a perception user interface) that list of references: Bradski (this base of cloth rad) delivered on Intel TechnologyJournal (Intel technical journal) in 1998) that object to be tracked is followed the tracks of, obtain the central point and the size of target window.

Step 5: the people's face that traces into is carried out colour of skin checking, if the people's face that traces into verify by the colour of skin, carry out step 6; If by colour of skin checking, then this people's face is not disallowable, for next frame, jumps to step 1 for the people's face that traces into.

Step 6: for next frame, people's face window of step 3 is carried out colour of skin checking,, the center of this people's face and size as target window, are jumped to step 4 if by colour of skin checking; If by colour of skin checking, people's face center of step 5 and size as target window, are jumped to step 4.

Good effect of the present invention and advantage are:

1. method of robust human face detection of the present invention has stronger robustness, not only has the higher detection rate, and lower false drop rate, can reject the chaff interference of the similar people's face that does not possess the conspicuousness characteristics;

2. robust human face tracking of the present invention has robustness to various partial occlusions, can avoid blocking the people's face that causes tracking and lose.

Description of drawings

Fig. 1 method step block scheme.

The class Haar wavelet character of one group of expansion of Fig. 2.

Embodiment

People's face of robust of the present invention detects and tracking;

One, the method for detecting human face of robust, its step is as follows:

Step 1: utilize weights to prune the AdaBoost algorithm and carry out the training of people's face sorter, its process is as follows:

Step 1.1: given training image (x ₁, y ₁) ..., (x _n, y _n) y wherein _i=0,1 corresponds respectively to negative sample and positive sample.

Step 1.2: initialization weights ${\mathrm{\&omega;}}_{1,i}=\frac{1}{2m},\frac{1}{2l}$ Correspond respectively to y _i=0,1, wherein m and l are respectively positive sample and negative sample number.

Step 1.3: for t=1 ..., T:

Step 1.3.1: normalization weights ${\mathrm{\&omega;}}_{t,i}\&LeftArrow;\frac{{\mathrm{\&omega;}}_{t,i}}{{\mathrm{\&Sigma;}}_{j=1}^n{\mathrm{\&omega;}}_{t,j}}$ Make ω _{T, i}Set ω _tIt is a probability distribution.

Step 1.3.2: as t＞T ₀, determine threshold value θ _t, make $\underset{{\mathrm{\&omega;}}_{t,i}<{\mathrm{\&theta;}}_t}{\mathrm{\&Sigma;}}{\mathrm{\&omega;}}_{t,i}<0.01\&le;\underset{{\mathrm{\&omega;}}_{t,i}\&le;{\mathrm{\&theta;}}_t}{\mathrm{\&Sigma;}}{\mathrm{\&omega;}}_{t,i}.$ With weights less than θ _tSample get rid of, be not used in the training of Weak Classifier.

Threshold value θ wherein _tDefinite method be, at first with all sample weights ω _{T, i}Sort, then according to from small to large the order weights ω that adds up one by one _{T, i}, when it with just greater than 0.01 the time, then selected current ω _{T, i}Be θ _tStep 1.3.3:, train a sorter h for each feature j _jMake its corresponding feature.Calculating is for ω _tError, ε _j=∑ _iω _i| h _j(x _j)-y _i|.

Step 1.3.4: selection sort device h _t, make it have least error ε _t

Step 1.3.5: refreshing weight: ${\mathrm{\&omega;}}_{t+1,i}={\mathrm{\&omega;}}_{t,i}{\mathrm{\&beta;}}_t^{1-e_i}$ If sample x wherein _iBy the then e that correctly classifies _i=0, otherwise e _i=1, ${\mathrm{\&beta;}}_t=\frac{{\mathrm{\&epsiv;}}_t}{1-{\mathrm{\&epsiv;}}_t}.$

Step 1.4: final strong classifier is:

Wherein ${\mathrm{\&alpha;}}_t=\log \frac{1}{{\mathrm{\&beta;}}_t}---\left(1\right)$

Step 2: to input picture carry out from left to right, the scanning of from top to bottom window, window is of a size of h * h, wherein h is since 24 and be that step-length increases gradually with 2, but is no more than the length of input picture or wide.Each window is extracted the class Haar wavelet character (the paper An extended set of Haar-like features forrapid object detection that list of references: Lienhart people such as (sharp grace Harts) delivered on International Conferenceon Image Processing (Flame Image Process international conference) in 2002 (fast target based on the class Haar feature of expanding detects)) of one group of expansion, as shown in Figure 2;

Step 3: the feature of extracting is sent into the AdaBoost sorter shown in the formula (1) carry out the detection of people's face, if h (x)=1, then this scanning window is people's face window to be verified; If h (x)=0, then this scanning window is not people's face window.

Step 4: people's face window to be verified is carried out the conspicuousness checking, as the formula (2):

$H=-\underset{i}{\mathrm{\&Sigma;}}P\left(d_i\right){\log }_{2}P\left(d_i\right)---\left(2\right)$

Wherein be P (d _i) image value d _iProbability.

If step 5: selected threshold θ is H _D〉=θ, people's face window then to be verified is by checking, i.e. this window area behaviour face; If H _DBy checking, promptly this window area is not people's face for＜θ, people's face window then to be verified.

Two, the face tracking method of robust, its step is as follows:

Step 1: use method for detecting human face of the present invention to detect people's face,, carry out step 2 if detect people's face; If do not detect people's face, next frame carry out step 1.

Step 2: detected people's face is carried out colour of skin checking, if detected people's face verify by the colour of skin, carry out step 3; If detected people's face is disallowable with this people's face not by colour of skin checking, next frame carry out step 1.

Wherein colour of skin checking is provided by formula (3), and when R (red), G (green), B (indigo plant) pixel value of a pixel satisfies this three subformulas simultaneously, this pixel just is classified as the colour of skin.When carrying out colour of skin checking, given human face region to be verified if the ratio of skin pixel number that is detected and regional total pixel number surpasses a threshold value P ∈ [0.5,1], judges that then this zone is people's face.

L _R/G＜R/G＜U _R/G

L _R/B＜R/B＜U _R/B (3)

L _G/B＜G/B＜U _G/B

L wherein _R/G, L _R/B, L _G/BAnd U _R/G, U _R/B, U _G/BBe respectively low threshold value and high threshold, the present invention adopts L _R/G=1.08, L _R/B=1.14, L _G/B=0.96, U _R/G=1.86, U _R/B=3.92, U _G/B=2.15.

Step 3: people's face window that will be by colour of skin checking is as object to be tracked.

Step 4: utilize Camshift algorithm (the paper Computer vision face tracking for use in a perceptualuser interface (being applied to the computer vision face tracking of a perception user interface) that list of references: Bradski (this base of cloth rad) delivered on Intel TechnologyJournal (Intel technical journal) in 1998) that object to be tracked is followed the tracks of, obtain the central point and the size of target window.

Step 5: the people's face that traces into is carried out colour of skin checking, if the people's face that traces into verify by the colour of skin, carry out step 6; If by colour of skin checking, then this people's face is not disallowable, for next frame, jumps to step 1 for the people's face that traces into.

Step 6: for next frame, people's face window of step 3 is carried out colour of skin checking,, the center of this people's face and size as target window, are jumped to step 4 if by colour of skin checking; If by colour of skin checking, people's face center of step 5 and size as target window, are jumped to step 4.

Claims (2)

1, a kind of method for detecting human face of robust is characterized in that: this method for detecting human face step is as follows:

Step 1: utilize weights to prune the AdaBoost algorithm and carry out the training of people's face sorter, its process is as follows:

Step 1.1: given training image (x ₁, y ₁) ..., (x _n, y _n) y wherein _i=0,1 corresponds respectively to negative sample and positive sample;

Step 1.2: initialization weights ${\mathrm{\&omega;}}_{1,i}=\frac{1}{2m},\frac{1}{2l}$ Correspond respectively to y _i=0,1, wherein m and l are respectively positive sample and negative sample number;

Step 1.3: for t=1 ..., T:

Step 1.3.1: normalization weights ${\mathrm{\&omega;}}_{t,i}\&LeftArrow;\frac{{\mathrm{\&omega;}}_{t,i}}{{\mathrm{\&Sigma;}}_{j=1}^n{\mathrm{\&omega;}}_{t,j}}$ Make ω _{T, i}Set ω _tIt is a probability distribution;

Step 1.3.2: as t＞T ₀, determine threshold value θ _t, make $\underset{{\mathrm{\&omega;}}_{t,i}<{\mathrm{\&theta;}}_t}{\mathrm{\&Sigma;}}{\mathrm{\&omega;}}_{t,i}<0.01\&le;\underset{{\mathrm{\&omega;}}_{t,i}\&le;{\mathrm{\&theta;}}_t}{\mathrm{\&Sigma;}}{\mathrm{\&omega;}}_{t,i}.$ With weights less than θ _tSample get rid of, be not used in the training of Weak Classifier;

Step 1.3.3:, train a sorter h for each feature j _jMake its corresponding feature.Calculating is for ω _tError, ε _j=∑ _iω _i| h _j(x _i)-y _i|;

Step 1.3.4: selection sort device h _t, make it have least error ε _t

Step 1.3.5: refreshing weight: ${\mathrm{\&omega;}}_{t+1,i}={\mathrm{\&omega;}}_{t,i}{\mathrm{\&beta;}}_t^{1-e_i}$ If sample x wherein _iBy the then e that correctly classifies _i=0, otherwise

e _i＝1， ${\mathrm{\&beta;}}_t=\frac{{\mathrm{\&epsiv;}}_t}{1-{\mathrm{\&epsiv;}}_t};$

Step 1.4: final strong classifier is:

Wherein $a_t=\log \frac{1}{{\mathrm{\&beta;}}_t}---\left(1\right)$

Step 2: input picture is carried out window scanning, each window is extracted the class Haar wavelet character of one group of expansion;

Step 3: the feature of extracting is sent into the AdaBoost sorter shown in the formula (1) carry out the detection of people's face, if h (x)=1, then this scanning window is people's face window to be verified; If h (x)=0, then this scanning window is not people's face window;

Step 4: people's face window to be verified is carried out the conspicuousness checking, as the formula (2):

$H_D=-\underset{i}{\mathrm{\&Sigma;}}{P}_D\left(d_i\right){\log }_2{P}_D\left(d_i\right)---\left(2\right)$

Wherein be P (d _i) image value d _iProbability;

If step 5: selected threshold θ is H _D〉=θ, people's face window then to be verified is by checking, i.e. this window area behaviour face; If H _DBy checking, promptly this window area is not people's face for＜θ, people's face window then to be verified.

2, a kind of face tracking method of robust is characterized in that: this face tracking method step is as follows:

Step 1: use method for detecting human face of the present invention to detect people's face,, carry out step 2 if detect people's face; If do not detect people's face, next frame carry out step 1;

Step 2: detected people's face is carried out colour of skin checking, if detected people's face verify by the colour of skin, carry out step 3; If detected people's face is disallowable with this people's face not by colour of skin checking, next frame carry out step 1;

Step 3: people's face window that will be by colour of skin checking is as object to be tracked;

Step 4: utilize the Camshift algorithm that object to be tracked is followed the tracks of, obtain the central point and the size of target window;

Step 5: the people's face that traces into is carried out colour of skin checking, if the people's face that traces into verify by the colour of skin, carry out step 6; If by colour of skin checking, then this people's face is not disallowable, for next frame, jumps to step 1 for the people's face that traces into;

Step 6: for next frame, people's face window of step 3 is carried out colour of skin checking,, the center of this people's face and size as target window, are jumped to step 4 if by colour of skin checking; If by colour of skin checking, people's face center of step 5 and size as target window, are jumped to step 4.

Images