CN104751448A - Online video tracking method based on PCA (Principal Component Analysis) and noise separation - Google Patents

Abstract

The invention discloses an online video tracking method based on PCA (Principal Component Analysis) and noise separation. The method comprises the following steps: step 1, acquiring a first frame image, and initializing a transformation parameter; step 2, carrying out PCA spatial modeling and initializing a base vector and a mean value; step 3, acquiring a next frame of image; step 4: extracting images according to the transformation parameter; step 5, calculating representing coefficients and abnormal noise according to the PCA vector sum and the mean value; step 6, calculating a new transformation parameter by an Lucas-Kanade algorithm; step 7, judging whether the loop iteration is restrained, updating and outputting the transformation parameter if yes, updating the PCA vector sum and the mean value, and returning to the step 3; if yes, returning to the step 4.

Description

The Online Video tracking of Based PC A and noise separation

Technical field

The present invention relates to the method for tracking target in a kind of computer vision field, especially a Based PC A(Principal components analysis for computational accuracy and speed can relatively be improved, principal component analysis (PCA)) and the Online Video tracking of noise separation.

Background technology

At present, the target following based on video has been applied to the fields such as modernization military affairs, intelligent video monitoring, intelligent transportation, intelligent vision navigation, man-machine interaction.Tradition tracking is confined to follow the tracks of specific objective under controlled condition more, and the online vision that existing method for tracking target is mostly the arbitrary target in reality scene is followed the tracks of is followed the tracks of.The practicality of online visual tracking method is stronger, but due to the complicacy of reality scene and tracked target, as the change of the inherence such as targeted attitude, shape and illumination, motion blur, block, the external factor such as background is mixed and disorderly, realize online vision and follow the tracks of and there is lot of challenges.Existing Online Video tracking generally can be divided into three major types according to ultimate principle and technology: the track algorithm based on state estimation, the track algorithm based on sorter and the track algorithm based on template matches.

With the track algorithm that the immediate prior art of the present invention is based on template matches, such algorithm general feature histogram, gray-scale pixels template and proper subspace are as To Template, in tracing process, the image-region that such algorithm is found in candidate target region Yu To Template similarity is maximum or cluster is minimum.The people such as Lucas and Kanade propose a kind of iterative image registration Algorithm Lucas-Kanade algorithm in paper An iterative image registration technique with an application to stereo vision, this algorithm is the basis of optical flow tracking algorithm, its advantage is fast to rigidity target processing speed in illumination-constant situation, shortcoming be can not the illumination of processing target well, attitude and the change of blocking.The people such as Li propose the Lucas-Kanade algorithm based on rarefaction representation in paper Robust Registration-based tracking by sparse representation with model update, attempt to utilize sparse representation model to portray the change of target in tracing process, but because this algorithm does not carry out explicitly analysisanddiscusion to extraordinary noise, when process such as to block at the challenge, effect is still undesirable.

In a lot of problems of computer vision and pattern-recognition, usual observation noise can be assumed to be the single noise of obeying certain distribution.The Multi-dimensional Gaussian distribution of the variances such as the observation noise obedience of such as principal component analysis (PCA) (PCA) algorithm hypothesis sample data, can maximize the major component of projecting direction as sample of variance by selection.The people such as Ross propose a kind of increment subspace tracking algorithm in paper Incremental learning for robust visual tracking, can on-line study PCA subspace adapt to target light according to and the change of attitude, this algorithm can processing target rotation well, yardstick and illumination variation.But when there is the extraordinary noise such as partial occlusion or damage, single noise just cannot be utilized to consider target normal variation and extraordinary noise simultaneously.Recently, in the estimation problem of people in paper Decomposing and regularizing sparse/non-sparse components for motion field estimation such as Ayvaci, matching error is decomposed into two parts: Gaussian noise and sparse noise.Subsequently, similar fractionation is done to matching error and regularization term in the estimation problem of the people such as Chen in paper Occlusion detection and motion estimation with convex optimization simultaneously, achieved good effect.But, up to now also less than the relevant report of the Online Video tracking about Based PC A and noise separation.

Summary of the invention

The present invention is the above-mentioned technical matters in order to solve existing for prior art, provides a kind of and relatively can improve the Based PC A of computational accuracy and speed and the Online Video tracking of noise separation.

Technical solution of the present invention is: the Online Video tracking of a kind of Based PC A and noise separation, it is characterized in that carrying out in accordance with the following steps successively:

Step 1: gather the first two field picture, initialization conversion parameter;

Step 2:PCA spatial modeling, initialization base vector and average;

Step 3: gather next frame image;

Step 4: take image according to conversion parameter;

Step 5: represent coefficient and extraordinary noise according to PCA base vector and mean value computation;

Step 6:Lucas-Kanade algorithm calculates new conversion parameter;

Step 7: judge that loop iteration is restrained? be upgrade and output transform parameter, upgrade PCA base vector and average, return step 3; No, be back to step 4.

Described step 2 takes target image in the first two field picture, around target image, collect sample, utilizes PCA subspace Modling model , wherein, for sample to be observed, for mean vector, for the base vector of PCA, for representing coefficient.

Described step 5 is according to objective function solve and represent coefficient x and extraordinary noise, in formula, e is sparse extraordinary noise, , for presetting regularization parameter;

Given optimum , optimum obtained by map operation ; Given optimum , optimum acquisition is operated, namely by soft-threshold , described in for soft-threshold function , wherein for sign function.

Described step 6 obtains according to Lucas-Kanade algorithm gauss-Newton iterative gradient decline form: ,

In formula:

for conversion parametric variable; it is extra large gloomy matrix; it is image gradient; be with corresponding image block; i(w (z, p)) be with y (image block that w (z, p) is corresponding, wherein represent affined transformation, z is image coordinate before conversion, and p is the conversion parameter stored.

The present invention utilizes PCA base vector to describe tracked target, the normal variation utilizing the thought processing target of burbling noise and the extraordinary noise that may occur, PCA is represented being embedded into registration with the thought of burbling noise follows the tracks of in framework, proposes corresponding objective function.This objective function can obtain the conversion parameter of the expression coefficient of target, extraordinary noise value and target by iterative, thus the tracing process of realize target.Experimental result shows that the tracking accuracy of the inventive method on challenging video database has certain advantage with speed compared with other trackings, especially in the situation such as target occlusion, illumination variation, attitudes vibration, background be mixed and disorderly, has stronger robustness.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the embodiment of the present invention.

Fig. 2 is the present invention, LK algorithm, the SLK algorithm qualitative test design sketch of (the serious circumstance of occlusion of target) on FaceOcc and Caviar video.

Fig. 3 is the present invention, LK algorithm, the qualitative test design sketch of SLK algorithm on DavidIn and Car video situations such as () illumination variation, attitudes vibration, background are mixed and disorderly.

Embodiment

As shown in Figure 1: carry out in accordance with the following steps successively:

Step 1: gather the first two field picture, initialization conversion parameter;

Step 2:PCA spatial modeling, initialization base vector and average:

In the first two field picture, take target image, around target image, collect sample, utilize PCA subspace Modling model , wherein, for sample to be observed, for mean vector, for the base vector of PCA, for representing coefficient.

Step 3: gather next frame image;

Step 4: image is taken to this current gathered collection image according to the conversion parameter stored;

Step 5: represent coefficient and extraordinary noise according to PCA base vector and mean value computation:

Suppose that observation noise comprises intensive Gaussian noise and sparse extraordinary noise two parts, use represent, wherein n is the intensive Gaussian noise of independent identically distributed little variance, and e is sparse extraordinary noise.Noise e is regarded as an extra variable, as known e, just obey independent identically distributed multivariate Gaussian variable.Theoretical according to maximal possibility estimation, (namely Gaussian noise item n can to absorb in L2 norm imply obey as independent identically distributed multivariate Gaussian variable).

According to objective function solve and represent coefficient x and extraordinary noise, in formula for sparse extraordinary noise, , for presetting regularization parameter;

Objective function changeable type is , form by two: (a) is about the least mean-square error item of variable x and e; B () is to the L1 norm regularization item of variable e.Therefore optimization problem is a convex optimization problem, and local minimum is equivalent to global minimum.

Objective function utilizes following two character iteration to carry out iterative: character 1: given optimum , optimum can be obtained by map operation (consider the orthogonality of PCA base, ).Character 2: given optimum , optimum acquisition can be operated, namely by soft-threshold , for soft-threshold function , wherein for sign function.

Step 6:Lucas-Kanade algorithm calculates new conversion parameter:

PCA subspace is represented being embedded into registration with noise separation thought follows the tracks of in framework, obtains following optimization problem: , there is not closed solution in this optimization problem, therefore adopts iterative algorithm to solve.Obtain optimum solution .Then, fixing x and e, namely optimization problem is converted into standard registration tracking equations: , wherein, be with corresponding image block, be with corresponding image block. there is clearer and more definite physical significance, with corresponding composition can regard the modeling to object variations during noise without exception as, and the composition corresponding with e can regard the process to extraordinary noise as.Obtain according to Lucas-Kanade algorithm gauss-Newton iterative gradient decline form: .In formula:

for conversion parametric variable; it is extra large gloomy matrix; it is image gradient; be with corresponding image block; i(w (z, p)) be with y (image block that w (z, p) is corresponding, wherein represent affined transformation, z is image coordinate before conversion, and p is the conversion parameter stored.

Step 7: judge that loop iteration is restrained? be upgrade and store and export new conversion parameter, upgrade PCA base vector and average, return step 3; No, be back to step 4.

Experiment effect:

The present invention realizes under Matlab 2009B software platform, be configured to Intel i7-2.66GHz double-core CPU, in save as speed that the computing machine of 4GB runs for average 20 frames per second.Design parameter arranges as follows: the dimension of PCA subspace is set to , regularization parameter is set to .Zoom to unified for the image taken pixel, Inner eycle iterations is set to 10, and outer circulation iterations is set to 30.The present invention utilizes 4 challenging video sequences to test and comparison algorithm, and they are FaceOcc, Caviar, DavidIn and Car respectively.These video test sequence are open test data, comprise the challenge factors such as illumination variation, partial occlusion, dimensional variation, attitudes vibration, background be mixed and disorderly.

The inventive method compares qualitatively with two maximally related algorithms and analyzes, these two algorithms are the template track algorithm of LK(based on Lucas-Kanade respectively, paper An iterative image registration technique with an application to stereo vision) and SLK(track algorithm that rarefaction representation and Lucas-Kanade algorithm are combined, paper Robust Registration-based tracking by sparse representation with model update).

Fig. 2 is that the present invention, LK algorithm, SLK algorithm are at FaceOcc(a) and Caviar(b) the qualitative test design sketch of (the serious circumstance of occlusion of target) on video.The present invention in Fig. 2 (solid line), LK algorithm (dotted line), SLK algorithm (dot-and-dash line).As can be seen from Figure 2, the tracking that the present invention proposes can process circumstance of occlusion well, and its basic reason is that the noise separation model adopted herein can be followed the tracks of in framework at registration and considers and process extraordinary noise situation.

Fig. 3 is that the present invention, LK algorithm, SLK algorithm are at DavidIn(a) and qualitative test design sketch Car(b) on video situations such as () illumination variation, attitudes vibration, background are mixed and disorderly.The present invention in Fig. 3 (solid line), LK algorithm (dotted line), SLK algorithm (dot-and-dash line).As can be seen from Figure 3, the present invention also can obtain good effect when these are challenged, this give the credit to the present invention adopt PCA represent can portray well target illumination and among a small circle attitudes vibration time outward appearance change.

Utilize the frame number (FPS:frame per second) of central point error evaluation criterion and process p.s. to come tracking accuracy and the speed of appraisal procedure, and compare (as IVT, L1 with more track algorithm, OSPT, LK, SLK etc.), result is as shown in table 1.As can be seen from Table 1, the present invention all achieves good effect in precision He in speed.Although precision of the present invention is a little less than OSPT algorithm, tracking velocity is more faster than OSPT algorithm.In addition, although speed of the present invention is lower than LK algorithm, compared with LK algorithm, the precision improvement of the inventive method is a lot.So, consider tracking accuracy and speed, the inventive method achieve remarkable must be progressive.

Table 1

。

Claims (4)

1. an Online Video tracking for Based PC A and noise separation, is characterized in that carrying out in accordance with the following steps successively:

Step 1: gather the first two field picture, initialization conversion parameter;

Step 2:PCA spatial modeling, initialization base vector and average;

Step 3: gather next frame image;

Step 4: take image according to conversion parameter;

Step 5: represent coefficient and extraordinary noise according to PCA base vector and mean value computation;

Step 6:Lucas-Kanade algorithm calculates new conversion parameter;

Step 7: judge that loop iteration is restrained? be upgrade and output transform parameter, upgrade PCA base vector and average, return step 3; No, be back to step 4.

2. the Online Video tracking of Based PC A and noise separation according to claim 1, is characterized in that described step 2 takes target image in the first two field picture, collects sample, utilize PCA subspace Modling model around target image , wherein, for sample to be observed, for mean vector, for the base vector of PCA, for representing coefficient.

3. the Online Video tracking of Based PC A and noise separation according to claim 2, is characterized in that described step 5 is according to objective function solve and represent coefficient x and extraordinary noise, in formula, e is sparse extraordinary noise, , for presetting regularization parameter;

Given optimum , optimum obtained by map operation ; Given optimum , optimum acquisition is operated, namely by soft-threshold , described in for soft-threshold function , wherein for sign function.

4. the Online Video tracking of Based PC A and noise separation according to claim 3, is characterized in that described step 6 obtains according to Lucas-Kanade algorithm gauss-Newton iterative gradient decline form: ,

In formula:

for conversion parametric variable; it is extra large gloomy matrix; It is image gradient; be with corresponding image block; i(w (z, p)) be with y (image block that w (z, p) is corresponding, wherein represent affined transformation, z is image coordinate before conversion, and p is the conversion parameter stored.