CN104050451A - Robust target tracking method based on multi-channel Haar-like characteristics - Google Patents

Abstract

The invention discloses a robust target tracking method based on multi-channel Haar-like characteristics. The robust target tracking method based on the multi-channel Haar-like characteristics comprises the following steps that firstly, the characteristic values of the multi-channel Haar-like characteristics are obtained during target tracking, and then M corresponding candidate weak classifiers are trained according to the multi-channel Haar-like characteristics; secondly, K weak classifiers with the highest response capacities are selected from the M corresponding candidate weak classifiers so that a strong classifier with the response being p(y=1|x) can be constructed, and N weak classifiers with the lowest judgement capacities are replaced; thirdly, an appearance models and the positions of targets are updated according to the weak classifiers, all the steps are executed repeatedly, and then the positions and the appearance models of the targets in all frames are obtained. According to the robust target tracking method based on the multi-channel Haar-like characteristics, tracking of colored images can be achieved, and the method can be adapted to real-time change of the targets and external conditions during tracking.

Description

A kind of robust method for tracking target based on hyperchannel Haar-like feature

Technical field

The invention belongs to pattern-recognition and computer vision field, relate to a kind of robust method for tracking target based on hyperchannel Haar-like feature.

Background technology

As the important research direction of computer vision field, Target Tracking Problem continues to be subject to people's concern in recent years.Motion target tracking technology has wide application prospect in fields such as robot navigation, video content analysis, security monitorings.Although Chinese scholars is through researching and proposing in a large number plurality of target tracking, but due to its intrinsic complicacy, comprise target deformation, rotate, block, illumination variation, motion blur etc., how to construct one accurately, robust and real-time Target Tracking System be a problem that is worth further investigation always.

Display model has critical effect in the track algorithm of a robust.The track algorithm proposing at present can be divided into production model and discriminative model two classes according to the difference of their display model.First production model is learnt a display model and is represented target, then the search region the most similar to this display model in each frame.Owing to not considering background, production model has been lost a lot of Useful Informations in tracing process.Discriminative model i.e. the tracking based on detecting, and it will be followed the tracks of as a binary classification problems, uses sorter that target is separated from background.As the nearest representative algorithm based on detecting tracking, the track algorithm of learning from example has been obtained immense success in target tracking domain more.

Using Haar-like feature to carry out object representation is the key factor that many learn-by-example track algorithms are obtained superperformance.Calendar year 2001, first Viola etc. are incorporated into the AdaBoost algorithm based on Haar-like small echo in face detection.Owing to the thought of integral image being applied in the calculating of Haar-like wavelet character, greatly improve the acquisition speed of feature.Inspired by this, Babenko in 2011 etc., by the method for online many learn-by-example training classifiers, utilize Haar-like feature to realize tracking to a discriminative model of target and background training.Different from traditional histogram object representation based on statistics, Haar-like feature is the feature based on structure, insensitive to the information such as color and texture.Therefore when, original many learn-by-example algorithms are followed the tracks of color video, conventionally adopt a certain single channel information of RGB image or be translated into gray level image and follow the tracks of.Obviously, no matter be adopt single channel information or it is simply fused into gray level image, all can cause information dropout to original color image.In addition, original many learn-by-example algorithms by Haar-like feature the real-time training classifier of response on positive negative sample follow the tracks of.Haar-like feature generates at random in the time of the first frame, and uses in each frame subsequently always.Due to the gradual change of target and background in tracing process, the weak Haar-like feature of some differentiation power can lose efficacy, and in feature selection process, will not used.Haar-like feature " once generate, forever use ", is difficult to meet the real-time change requirement of target itself and external condition in tracing process.

Summary of the invention

The object of the invention is to overcome the shortcoming of above-mentioned prior art, a kind of robust method for tracking target based on hyperchannel Haar-like feature is provided, the method can meet the tracking of coloured image, meets the real-time change of target itself and external condition in tracing process simultaneously.

For achieving the above object, the robust method for tracking target based on hyperchannel Haar-like feature of the present invention comprises the following steps:

1) in the process of target following, user is spotting position in the first frame, then in annular region corresponding to this target location, gather some positive and negative sample block, wherein, corresponding several hyperchannels Haar-like of each positive and negative sample block feature, each hyperchannel Haar-like feature comprises several rectangular blocks, random six parameters that produce each rectangular block, described six parameters are respectively the upper left corner horizontal ordinate of rectangular block, ordinate, width, highly, weight and passage, and adopt the method for integral image to calculate the eigenwert of each rectangular block according to the passage at each rectangular block place, again all eigenwerts of each passage are weighted to summation according to the weight of rectangular block, the result of weighted sum is as the eigenwert of this hyperchannel Haar-like feature, then go out corresponding M candidate's Weak Classifier according to each hyperchannel Haar-like features training, from all M candidate's Weak Classifiers, select again K the Weak Classifier that response is maximum based on Boosting algorithm, then be the strong classifier of p (y=1|x) by K the maximum Weak Classifier tectonic response of response of selecting, wherein, x is the eigenwert of image block, y is binary variable, at moment t, interesting target position is , establish the position that l (x) represents candidate image piece x,

2) in the region of radius s around next frame target, intercept image block collection and calculate x ∈ X ^sthe p (y|x) of all candidate blocks in scope; Then basis more new target location, the more display model of fresh target of while;

3) reject successively N Weak Classifier and respond minimum hyperchannel Haar-like feature, add again N new hyperchannel Haar-like feature of random generation, then according to new M M the Weak Classifier that hyperchannel Haar-like features training makes new advances, from the each Weak Classifier of new M, select again K Weak Classifier of the response maximum making new advances, and build new strong classifier by K Weak Classifier structure of described new response maximum;

4) repeating step 2) and step 3), obtain position and the display model of target in each frame, complete the tracking of target.

Described passage is RGB triple channel.

Each hyperchannel Haar-like feature comprises 2-4 rectangular block.

The present invention has following beneficial effect:

Robust method for tracking target based on hyperchannel Haar-like feature of the present invention adopts the method for integral image to calculate the eigenwert of each rectangular block according to the passage at each rectangular block place, again all eigenwerts of each passage are weighted to summation according to the weight of rectangular block, the result of weighted sum is as the eigenwert of this hyperchannel Haar-like feature, represent the outward appearance of target by more information, thereby meet the tracking of target color; Secondly, in the time producing strong classifier, reject successively N Weak Classifier and respond minimum hyperchannel Haar-like feature, and then newly increase random generation N new Haar-like feature, to adapt to the various variations of target appearance, thereby meet the real-time change of target itself and external condition in tracing process.

Brief description of the drawings

Fig. 1 is the comparison diagram of the present invention and first prior art.

Embodiment

Below in conjunction with accompanying drawing, the present invention is described in further detail:

With reference to figure 1, the robust method for tracking target based on hyperchannel Haar-like feature of the present invention comprises the following steps:

1) in the process of target following, user is spotting position in the first frame, then in annular region corresponding to this target location, gather some positive and negative sample block, wherein, corresponding several hyperchannels Haar-like of each positive and negative sample block feature, each hyperchannel Haar-like feature comprises several rectangular blocks, random six parameters that produce each rectangular block, described six parameters are respectively the upper left corner horizontal ordinate of rectangular block, ordinate, width, highly, weight and passage, and adopt the method for integral image to calculate the eigenwert of each rectangular block according to the passage at each rectangular block place, again all eigenwerts of each passage are weighted to summation according to the weight of rectangular block, the result of weighted sum is as the eigenwert of this hyperchannel Haar-like feature, then go out corresponding M candidate's Weak Classifier according to each hyperchannel Haar-like features training, from all M candidate's Weak Classifiers, select again K the Weak Classifier that response is maximum based on Boosting algorithm, then be the strong classifier of p (y=1|x) by K the maximum Weak Classifier tectonic response of response of selecting, wherein, x is the eigenwert of image block, y is binary variable, at moment t, interesting target position is if l (x) represents the position of candidate image piece x,

2) in the region of radius s around next frame target, intercept image block collection and calculate x ∈ X ^sthe p (y|x) of all candidate blocks in scope; Then basis more new target location, the more display model of fresh target of while;

3) reject successively N Weak Classifier and respond minimum hyperchannel Haar-like feature, add again N new hyperchannel Haar-like feature of random generation, then according to new M M the Weak Classifier that hyperchannel Haar-like features training makes new advances, from the each Weak Classifier of new M, select again K Weak Classifier of the response maximum making new advances, and build new strong classifier by K Weak Classifier structure of described new response maximum;

4) repeating step 2) and step 3), obtain position and the display model of target in each frame, complete the tracking of target.

Described passage is RGB triple channel.

Each hyperchannel Haar-like feature comprises 2-4 rectangular block.

The Organization of Data form of many learn-by-examples is { (X ₁, y ₁) ..., (X _n, y _n), the wherein bag X of multiple example compositions _i={ x _i1..., x _im, y _ifor the label of bag, when at least comprising a positive example in a bag, just think that this bag is for positive closure, otherwise just think that it is negative bag, the label y of example bag _ifor:

$y_i=\underset{j}{\max }\left(y_{\mathrm{ij}}\right)---\left(1\right)$

Wherein, y _ijbe the label of example, the target tracking algorism of learning from example adopts many learn-by-example algorithms and the online training classifier of Boosting method more, and the expression formula of maximized log-likelihood objective function L is:

$L=\underset{i}{\mathrm{\Σ}}\left(\log p\left(y_i\right|X_i)\right)---\left(2\right)$

Wherein, p (y _i| X _i) be the likelihood probability of bag, it is by the likelihood probability p (y of example _i| x _ij) utilize Noisy-OR model to calculate:

$p\left(y_i\right|X_i)=1-\underset{j}{\mathrm{\Π}}(1-p\left(y_i\right|x_{\mathrm{ij}}))---\left(3\right)$

The likelihood probability of each example is calculated by formula (4):

P (y|x)=σ (H (x)) (4) wherein, h (x) is strong classifier, forms by organizing Weak Classifier h (x) cascade more.

In target following, due to cannot all samples of disposable acquisition, therefore keep K candidate's Weak Classifier of M >, adopt the framework of Boosting study to select successively K Weak Classifier and maximize log-likelihood objective function:

$h_k=\underset{h\∈\{h_1,\·\·\·,h_M\}}{\arg \max }L(H_{k-1}+h)---\left(5\right)$

Weak Classifier is by Haar-like feature f _klog-likelihood function than On-line Estimation, wherein Weak Classifier calculates as formula and is:

$h_k\left(x\right)=\log \left[\frac{p_t(y=1|f_k\left(x\right))}{p_t(y=0|f_k\left(x\right))}\right]---\left(6\right)$

Wherein, p _t(y=1|f _k) and p (x) _t(y=0|f _k(x)) be the likelihood function of feature, order

P (y=1)=p (y=0), is obtained by Bayesian formula:

$h_k\left(x\right)=\log \left[\frac{p_t\left(f_k\left(x\right)\right|y=1)}{p_t\left(f_k\left(x\right)\right|y=0)}\right]---\left(7\right)$

Wherein, p _t(f _k(x) | y=1)～N (μ ₁, σ ₁), p _t(f _k(x) | y=0)～N (μ ₀, σ ₀), μ ₁and μ ₀, σ ₁and σ ₀represent respectively average and the standard deviation of positive negative sample Gaussian distribution, when Weak Classifier is received new data { (x ₁, y ₁) ..., (x _n, y _n) time, adopt formula (8) and formula (9) undated parameter:

${\mathrm{\μ}}_1\←{\mathrm{\ρ\μ}}_1+(1-\mathrm{\ρ})\frac{1}{n}\underset{i|y_i=1}{\mathrm{\Σ}}{f}_k\left(x_i\right)---\left(8\right)$

${\mathrm{\σ}}_1\←{\mathrm{\ρ\σ}}_1+(1-\mathrm{\ρ})\sqrt{\frac{1}{n}\underset{i|y_i=1}{\mathrm{\Σ}}{(f_k\left(x_i\right)-{\mathrm{\μ}}_1)}^2}---\left(9\right)$

Wherein 0 < ρ < 1 is learning rate, for μ ₀and σ ₀renewal with similar above.Many learn-by-example algorithms are as algorithm 1.

In the time obtaining next frame target location, according to formula (8) and the more display model of fresh target of formula (9), positive sample packages gathers around target: wherein γ is for gathering radius, and negative sample wraps in an annular region and gathers: wherein, γ is with identical above, and β is a scalar, owing to there being a large amount of negative samples in this region, algorithm in this sample set subset of random choose as negative sample bag.

Claims (3)

1. the robust method for tracking target based on hyperchannel Haar-like feature, is characterized in that, comprises the following steps:

1) in the process of target following, user is spotting position in the first frame, then in annular region corresponding to this target location, gather some positive and negative sample block, wherein, corresponding several hyperchannels Haar-like of each positive and negative sample block feature, each hyperchannel Haar-like feature comprises several rectangular blocks, random six parameters that produce each rectangular block, described six parameters are respectively the upper left corner horizontal ordinate of rectangular block, ordinate, width, highly, weight and passage, and adopt the method for integral image to calculate the eigenwert of each rectangular block according to the passage at each rectangular block place, again all eigenwerts of each passage are weighted to summation according to the weight of rectangular block, the result of weighted sum is as the eigenwert of this hyperchannel Haar-like feature, then go out corresponding M candidate's Weak Classifier according to each hyperchannel Haar-like features training, from all M candidate's Weak Classifiers, select again K the Weak Classifier that response is maximum based on Boosting algorithm, then be the strong classifier of p (y=1|x) by K the maximum Weak Classifier tectonic response of response of selecting, wherein, x is the eigenwert of image block, y is binary variable, at moment t, interesting target position is , establish the position that l (x) represents candidate image piece x,

2) in the region of radius s around next frame target, intercept image block collection and calculate x ∈ X ^sthe p (y|x) of all candidate blocks in scope; Then basis more new target location, the more display model of fresh target of while;

3) reject successively N Weak Classifier and respond minimum hyperchannel Haar-like feature, add again N new hyperchannel Haar-like feature of random generation, then according to new M M the Weak Classifier that hyperchannel Haar-like features training makes new advances, from the each Weak Classifier of new M, select again K Weak Classifier of the response maximum making new advances, and build new strong classifier by K Weak Classifier structure of described new response maximum;

4) repeating step 2) and step 3), obtain position and the display model of target in each frame, complete the tracking of target.

2. the robust method for tracking target based on hyperchannel Haar-like feature according to claim 1, is characterized in that, described passage is RGB triple channel.

3. the robust method for tracking target based on hyperchannel Haar-like feature according to claim 1, is characterized in that, each hyperchannel Haar-like feature comprises 2-4 rectangular block.