CN107844739A - Robustness target tracking method based on adaptive rarefaction representation simultaneously - Google Patents

Abstract

The robustness target tracking method based on adaptive rarefaction representation simultaneously of the present invention, comprises the following steps：S1, the size according to Laplacian noise energy, adaptive foundation while sparse tracing model；S2, the tracing model established is solved；S3, template is updated.Method for tracing track identification effect provided by the invention is good, and anti-interference is stronger, can realize that more accurate and real-time target tracking, method for tracing are relatively stable.

Description

Robust Object Tracking Method Based on Adaptive Simultaneous Sparse Representation

technical field

The invention belongs to the technical field of computer image processing and relates to a target tracking method, in particular to a robust target tracking method based on adaptive simultaneous sparse representation.

Background technique

Object tracking occupies an important position in the field of computer vision. With the use of high-quality computers and cameras and the need for automatic video analysis, people are interested in object tracking. The main tasks of target tracking include: detection of moving targets of interest, continuous tracking between video frames and behavior analysis of tracked targets. Currently, related applications of object tracking include: motion recognition, video retrieval, human-computer interaction, traffic monitoring, vehicle navigation, etc.

At present, although many tracking algorithms have been proposed, the target tracking technology still faces many challenges. In the actual tracking process, video noise pollution is often encountered, such as target pose changes, lighting changes, background clutter, partial occlusion and complete occlusion, etc. These problems often lead to target tracking failure (tracking drift). In particular, the impact of long-term occlusion on target tracking is catastrophic. Illumination changes depend on the environment in which the target is located. The greater the illumination change, the greater the impact on the tracking effect. Background clutter and target pose changes will affect the accuracy of tracking. Therefore, it will also lead to tracking drift, video noise pollution will make tracking errors accumulate, and eventually cause tracking failure.

For target tracking, it is a basic and challenging problem to solve the target shape change. The target shape change can be divided into internal shape change and external shape change. Posture change is an internal shape change, while lighting changes, background clutter, partial occlusion, and full occlusion are external shape changes. Addressing these appearance changes requires an adaptive tracking method, namely an online learning method. At present, online learning methods can be roughly divided into two categories, namely Generative Approaches and Discriminative Approaches. The generation method (GA) is a method of searching for the region most similar to the tracking target, and the discriminant method (DA) can be regarded as a binary classification problem, the main purpose of which is to use known training samples to train a classifier. Used to distinguish objects and backgrounds. Although the discriminant method and the generative method have achieved reliable tracking to a certain extent, they also have their own shortcomings. First, the discriminant method has high requirements for feature extraction, so it is sensitive to noise in the actual tracking process, and it may appear for noisy targets. Tracking fails, and the generation method cannot accurately find areas similar to the target in the mixed background, so it is prone to tracking failure; second, the discriminant method needs sufficient training sample sets, good samples can improve the performance of the classifier, and bad samples It will weaken the performance of the classifier. If bad samples are introduced into the classifier, the tracking effect will be affected, and the generation method is more sensitive to the template. Once the occluded target is introduced into the template by mistake, tracking failure may occur. Therefore, the two methods are in the real scene. are not sufficiently robust in tracking.

Contents of the invention

The purpose of the present invention is to solve the above problems and provide a method that can cope with various challenges such as illumination changes, scale changes, occlusion, deformation, motion blur, fast motion, rotation, background clutter, and low resolution of the target object in the video. Robust target tracking method for continuous and accurate tracking of target objects.

The object of the present invention is achieved in this way, based on the robust target tracking method of adaptive simultaneous sparse representation, comprising the following steps:

S1. According to the size of the Laplacian noise energy, adaptively establish a simultaneous sparse tracking model;

S2, solving the established tracking model;

S3. The template is updated.

Further, the execution method of the step S1 is: comparing the Laplacian mean noise ||S|| ₂ with the size of the given noise energy threshold τ, and adaptively establishing a simultaneous sparse tracking model according to the comparison result:

When ||S|| ₂ ≤τ, the simultaneous sparse tracking model is:

When ||S|| ₂ >τ, the sparse tracking model is:

Among them, D is the tracking template, expressed as D=[T, I], T is the target template,

T=[T ₁ ,T ₂ ,…,T _n ]∈R ^d×n (d>>n), where d represents the dimension of the image, n represents the number of template base vectors, and each column of T is passed The zero-meanized vector D represents the tracking template, Y is the candidate target set, λ ₁ and λ ₂ are the regular parameters of the model, X is the sparse coefficient, S is the Laplacian noise, is the fidelity item of the sparse model, ||X|| _1,1 is the coefficient matrix, and ||S|| _1,1 represents the Laplacian noise energy.

Further, the step S2 includes:

S21. Calculate the similarity between the tracking target y _t and the mean value of the template T, denoted as sim;

S22. Determine the size of the angle threshold α between sim and cosine, and adaptively select a model for tracking:

When sim<α, the tracking model of the above formula (1) is adopted, and the solution is solved by the alternating direction multiplier (ADMM) method to obtain the sparse coefficient X;

When sim≥α, the tracking model of the above formula (2) is adopted, and the method of alternating direction multiplier (ADMM) is used to solve it and obtain the sparse coefficient X and Laplacian noise S.

Further, in the step S21, the similarity between the tracking target and the template mean value is calculated according to the following formula:

Among them, c is the mean value of the template T, and y is the tracked target, then ||τ|| ₂ can be equivalent to the arccosine of the mean value of the target and the template, that is, the cosine angle.

Further, when ||s|| ₂ ≤τ, the template T is updated.

Further, the template update method includes:

S31. Perform singular value decomposition on the current template T and the tracked target y respectively as follows:

S32. Use singular vectors u, s, v to incrementally update U, S, V, thereby obtaining new singular vectors U _* , S _* , V _* , then the new template is expressed as:

T _* ＝U _* S _* V _* ^T (5)

Further, step S33 is also included: using the unsupervised learning K-means method to train templates, given that the number of initial classes is k, then

Among them, i represents the i-th sample, when When it belongs to class k, then r _ik =1, otherwise r _ik =0, u _k is the average value of all samples belonging to class k, and J represents the sum of distances from sample points to the average value of sample points of this class; through formula (6) Then the average value of all samples of the kth class can be obtained, which reduces the dimension of the original template; thus, the updated template is:

T _new ＝[u ₁ ,u ₂ ,…,u _k ] (7)

The present invention also provides a fast target tracking method for adaptive sparse representation, comprising the following steps:

A. Establish a target tracking model:

B. Use the alternate iteration method to obtain the optimal tracking target;

C. Update the target template T and obtain a new template T ^* ;

D. Return the best tracking target and target template set, and continue the target tracking of the next frame.

Preferably, said step C comprises:

C1. Calculate the similarity between the tracking target and the mean value of the template, and record the similarity as sim;

C2. Compare the similarity sim with the cosine angle threshold α, β. If sim<α, execute step C3 to continue updating the template. If α≤sim≤β, then y←m, execute step C3 to update the template. If sim> β, at this time, the tracking target is basically not similar to the template, indicating that the target suffers from serious noise pollution, and the template is not updated;

C3. Perform singular value decomposition on the template T to obtain T=U∑V ^T , incrementally update the left singular vector U of the template T and obtain a new singular vector U ^* , combine formulas (5) and (6), and calculate new template

Compared with the prior art, the beneficial effects of the present invention are reflected in:

(1) The present invention proposes a new template selection and update method under the framework of sparse representation. This method places more emphasis on the real-time update of the template. At the same time, since the real-time update is the left singular vector of the template, the error introduced by the update can be controlled. It is very low. Compared with removing noise from the target and then introducing a new target template, K-means technology is also introduced in the template update of the present invention, which reduces the dimension of the template, effectively reduces redundant template vectors, and improves tracking performance. real-time, and reduce the impact of noise. However, the traditional template update method is easy to introduce a large noise error, which causes a lot of uncertainty to the tracking of the target in the next frame;

(2) fully consider the influence of Gaussian noise and Laplacian noise in the target tracking model of the present invention, according to the regular item of the adaptive selection model of the size of Laplacian noise energy, this not only improves the precision of tracking, It also improves the real-time performance of tracking;

(3) The present invention combines the ADMM algorithm into the solution of the tracking model, and makes the solution of the model more stable by controlling the parameters of the regular term.

Detailed ways

The robust target tracking method based on adaptive simultaneous sparse representation of the present invention will be further described below in conjunction with specific embodiments.

A robust target tracking method based on adaptive simultaneous sparse representation, including the following steps:

S1. According to the size of the Laplacian noise energy, adaptively establish a simultaneous sparse tracking model

Compare the Laplacian mean noise ||S|| ₂ with the given noise energy threshold τ, and adaptively establish a simultaneous sparse tracking model based on the comparison results:

When ||S|| ₂ ≤τ, the simultaneous sparse tracking model is:

When ||S|| ₂ >τ, the sparse tracking model is:

Among them, the definition D=[T,I] represents the tracking template, I represents the trivial template, and the image set of the given target template

T=[T ₁ ,T ₂ ,...,T _n ]∈R ^d×n (d>>n),，

d represents the dimension of the image, n represents the number of template base vectors, and each column of T is a vector after zero-meanization. Among them, Y=[y ₁ ,y ₂ ,…,y _m ], representing all candidate targets, assuming that the noise obeys the Gaussian Laplace distribution, then:

Y＝TZ+S+E (9)

S means Laplacian noise, E means Gaussian noise, X is sparse coefficient, λ ₁ and λ ₂ are regular parameters of the model, is the fidelity item of the sparse model, and the model has this deformation after considering the Laplacian noise, ||X|| _1,1 is the coefficient matrix, which can better extract the similarity between particles and effectively remove the redundant information of the template,| |S|| _1,1 characterizes Laplace noise energy.

Although the existing sparse representation target tracking algorithms can solve the effects of partial occlusion, illumination changes, pose changes and background clutter to a certain extent, the models are all based on the noise obeying the Gaussian distribution, and the noise distribution is considered too simply. situation, so tracking failure may occur in the face of some complex noise distribution situations. In the target tracking model of the present invention, the influence of Gaussian noise and Laplacian noise is fully considered, and the regular term of the selection model is adaptively selected according to the size of Laplacian noise energy, which not only improves the accuracy of tracking, but also improves the real-time tracking.

S2. Solve the established tracking model

Preferably, said step S2 includes:

S21. Calculate the similarity between the tracking target y _t and the mean value of the template, denoted as sim;

Preferably, the similarity between the tracking target and the template mean is calculated according to the following formula:

S22. Determine the size of the angle threshold α between sim and cosine, and adaptively select a model for tracking:

When sim<α, the tracking model of the above formula (1) is used to solve it through the Alternating Direction Multiplier (ADMM) method and obtain the sparse coefficient X;

When sim≥α, the tracking model of the above formula (2) is used to solve it by the Alternating Direction Multiplier (ADMM) method and obtain the sparse coefficient X and noise S.

In order to explain the solution scheme of the established tracking model more clearly, the following example illustrates the solution idea of the above formula (2):

The objective function of the above formulas (1) and (2) is a convex optimization problem, so an unconstrained optimization method can be used to solve the objective function. Alternating direction multiplier method (ADMM) is a classical method for unconstrained solution. With the advantages of stable solution and fast convergence speed, the ADMM method is used to solve the optimization problem (2) as follows:

First, transform the constrained problem into an unconstrained problem as

in is an indicator function ( _xi represents the i-th row of X, and if x _i is non-negative, then τ ₊ ( _xi ) is equal to 0; otherwise τ ₊ ( _xi ) is equal to +∞). Therefore, the optimization problem (2) has the following equivalent form:

Among them, V ₁ , V ₂ , and V ₃ are dual variables, and the formula (11) is further optimized as

here,

The augmented Lagrange function of formula (11) is

Where β represents the Lagrange multiplier, U=[U ₁ ,U ₂ ,U ₃ ] ^T .

Formula (15) can be decomposed into three sub-optimization problems, which are the sub-problems of X, the sub-problems of S and the sub-problems of V. The following sub-optimization problems are solved as follows:

Therefore, according to the extreme value principle, only the first derivative of the above sub-problem needs to be calculated, and the optimal solution of equation (15) can be obtained as follows:

V _1* ＝[β(TX−U ₁ )+(YS)]/(1+β)

V _2* ＝shrink(XU ₂ ,λ ₁ /β)

V _3* ＝max(0,XU ₃ )

S _* ＝shrink(YV ₁ ,λ ₂ )

X _* ＝(T ^T T+2I) ^-1 [T ^T (V ₁ +U ₁ )+V ₂ +U ₂ +V ₃ +U ₃ ]

where shrink is a shrinkage operator, that is, for a non-negative vector p, there is

shrinkI(x,p)=sgn(x)ο max{|x|-p,0}.

Similarly, the ADMM method can still be used to solve the above formula (1), and the following solution can be obtained:

V _1* ＝[β(TX−U ₁ )+Y]/(1+β)

V ₂ *＝shrink(XU ₂ ,λ ₁ /β)

V ₃ *＝max(0,XU ₃ )

X _* ＝(T ^T T+2I) ^-1 [T ^T (V ₁ +U ₁ )+V ₂ +U ₂ +V ₃ +U ₃ ]

In this way, through the analysis and solution of the sub-problems, the general forms of the solutions of formulas (1) and (2) are obtained. If a given Lagrangian multiplier β is input, the optimal numerical solutions of equations (1) and (2) can be obtained by iterating in alternate directions, where Table 1 shows the ADMM solution algorithm process of equation (1), and Table 2 It is the ADMM solution algorithm process of formula (2).

Table 1

Table 2

S3, update the template

As a preferred solution, when ||s|| ₂ ≤τ, the template T is updated.

Further, the template update method includes:

S31. Singular value decomposition is performed on the current template T and the tracked target y respectively as follows:

Down:

T = USV ^T

y=usv ^T (4)

S32. Use singular vectors u, s, v to incrementally update U, S, V, thereby obtaining new singular vectors U*, S*, V*, then the new template is expressed as:

T _* ＝U _* S _* V _* ^T (5)

Preferably, step S33 is also included: using the unsupervised learning K-means method to train templates, given that the number of initial classes is k, the K-means learning method is as follows:

where i represents the i-th sample, when When it belongs to class k, then r _ik =1, otherwise r _ik =0, u _k is the average value of all samples belonging to class k, and J represents the sum of distances from sample points to the average value of sample points of this class; through formula (6) Then the average value of all samples of the kth class can be obtained, which reduces the dimension of the original template. Thus, the updated template is:

T _new ＝[u ₁ ,u ₂ ,…,u _k ] (7)

The template update method provided by the present invention is more robust to occlusion and illumination changes. This method is different from the traditional template update. It emphasizes the selection of templates that have important contributions to target tracking and avoids the use of trivial templates. , and the K-means algorithm is used to perform unsupervised training on the template, which greatly eliminates the redundant information of the template, thereby improving the real-time performance of tracking.

As shown in Table 3, it is the target tracking algorithm process of the adaptive simultaneous sparse representation of the embodiment of the present invention:

table 3

Below, will compare the target tracking method (Ours) provided by the present invention with other five kinds of existing methods with good tracking performance by experiment, these five kinds of tracking algorithms are respectively the circular matrix tracking (CSK) of kernel skill, Accelerated Dual Gradient Pursuit (L1APG), Multi-Task Tracking (MTT), Sparse Prototype Tracking (SPT) and Sparse Joint Tracking (SCM). The following experiments are all based on Matlab 2012a, the computer memory is 2GB, and the CPU is Intel(R) Core(TM) i3 platform.

Data and Experiment Description:

The experiment selected 14 different tracking-challenging videos, including occlusion, lighting changes, background clutter, pose changes, low resolution and fast motion, and other factors that affect the tracking results. The profile of the video sequence attributes is shown in Table 4. The video in Table 4 contains different noises, where OV means target loss, BC means background clutter, OCC means complete occlusion, OCP means partial occlusion, OPR means rotate out of plane, LR means low resolution, FM means fast motion, SV means Size changes. There are three evaluation methods used in the experiment of the embodiment of the present invention, and each evaluation method can explain the quality of the tracking performance to a certain extent, which are respectively Center Local Error (Center Local Error), Overlap Ratio (Overlap Ratio) and Curve Area Under Curve. Given the real target frame R _g (ground truth) and the tracked target frame R _t (tracked target bounding) of the frame, we might as well set their center positions as: p _g ＝(x _g , y _g ) and p _t ＝(x _t ,y _t ), then the local center error is CLE＝||p _g -p _t || ₂ , and the overlap rate

area(·) represents all the pixels in the area, and the value of each point of the area curve (AUC) represents the success rate of the video tracking when the overlap rate is greater than a given threshold η. In particular, we set η=0.5, and when the overlap ratio OR>0.5, it is considered that the tracking of the frame is successful. The relevant tracking results are shown in Tables 4, 5, and 6.

Table 4

video sequence number of frames noise(s) video sequence number of frames noise(s) Walking2 495 SV,OCP,LR suv 945 OCC,OV,BC Car4 659 IV, SV CarDark 393 IV,BC,LR Car2 913 IV, SV, BC Deer 71 FM,LR,BC girl 500 OPR, OCC, LR Singr2 366 IV, OPR, BC FaceOcc2 812 OCC,OPR,IV Skater2 435 SV,OPR Football 362 OCC,OPR,BC Dudek 1145 OCC,BC,OV FaceOcc1 892 OCC Subway 175 OCC,BC

Table 5 is the comparison data of various algorithm performances based on the average overlap rate, where AOR represents the total average overlap rate, wherein the larger the average overlap rate, the better the tracking performance; the parameters are set as follows in the experiment: regularization parameter λ ₁ = 0.1, λ ₂ = 0.1, penalty factor β = 0.1, the minimum cosine angle threshold is α _min = 20, the maximum is α _max = 35, the maximum number of template basis vectors is 15, the number of particle samples is 600, the size of the image block is 25×25, the maximum number of iterations in the experiment Loop=20, and the convergence error tol=0.001. In the experiment, the parameters λ ₁ and λ ₂ are obtained through the cross-validation method, and the adjustment of the λ ₂ parameter satisfies the following rules. change), at this time the value of λ ₂ should be smaller, and vice versa. As can be seen from the experimental comparison data in Table 5, the tracking method (Ours) provided by the present invention is obviously higher than other tracking methods no matter from the average overlap rate of different video categories or the total average overlap rate, that is, the tracking method of the present invention has achieved The best tracking performance results.

table 5

Table 6 shows the comparison of the performance of various algorithms based on the average local center error, where ACLE represents the total average center error, and the smaller the average center error, the better the tracking performance. It is obvious from the data in Table 6 that the tracking method (Ours) provided by the present invention is significantly lower than other tracking methods no matter from the average center error of different video categories or the total average center error, that is, the tracking method of the present invention has achieved the best results. Excellent final performance effect.

Table 6

As shown in Table 7, it is a comparison of the performance of various algorithms based on the average success rate, where ASR represents the overall average success rate;

Table 7

From the comparative data in Table 7, it can be known that the tracking method (Ours) provided by the present invention is evaluated by other tracking methods no matter the average success rate of different video categories or the overall average success rate.

In order to further understand the tracking algorithm proposed by the present invention, the specific influence of the Laplacian noise mentioned in the model and the template update criterion on the tracking effect will be explained below.

The traditional template update method is directly updated by tracking the similarity between the target and the template. If the similarity is greater than a given threshold, it is considered that the target has encountered large noise pollution, so it is necessary to replace the tracking target with a smaller original weight. template vector, this replacement is actually relatively rough, because a large noise error is introduced, which causes a lot of uncertainty for the tracking of the target in the next frame, and the new template update method proposed by the present invention weakens the Noise effect. The specific performance is as follows:

(1) The new template update method effectively weighs the weight between the original template vector and the new tracking target, and realizes template update through the forgetting factor;

(2) The new template update method introduces the K-means method, which can effectively reduce redundant template vectors and improve the real-time performance of tracking, and the calculation of the class center is obtained by weighted average, so it can also effectively reduce noise.

The specific experiments are given below to compare the influence of template update and Laplacian on the experimental results, and the selection of experimental objects: MTT algorithm, ASSAT algorithm (only Laplacian), ASSAT (only template update), ASSAT (Laplace + template update), experimental data selection sequence Skater2, Dudek, SUV, Walking2, Subway, Deer, etc.

Table 8 shows the comparison of the effect of Laplacian on the experimental results. It can be seen from Table 8 that except for the Walking2 sequence, the tracking effect of adding Laplacian noise is better than that of the MTT algorithm, but the original template update method limits its performance. tracking performance, and the proposed new template update method improves the tracking performance of the ASSAT algorithm.

Table 8

Table 9 compares the impact of different template updates on the experimental results. From Table 9, it can be seen that the tracking effect of the ASSAT method and the IVT method using only template updates is similar, and there is not much improvement. For Skater2 and Subway sequence, the two methods have the same effect. Not good, the reason is that these two sequences contain large occlusions, and the ASSAT algorithm that only considers template updates without considering Laplacian noise cannot effectively track the target, and the same is true for IVT. In fact, for this kind of situation with large occlusion, if Laplacian noise is not considered, it can be attributed to the fact that the noise factor affects the sparse structure of the solution X in formula (1).

Table 9

The impact of different noise selections on the solution of the target in the occlusion situation is also manifested in that the solution obtained when considering Laplacian noise is sparse, and the solution is optimal at this time, while the solution obtained when Laplacian noise is not considered The obtained solution is a dense non-optimal solution, so maintaining the sparse structure of the solution directly affects the tracking performance of the algorithm.

The present invention also provides a fast target tracking method for adaptive sparse representation, comprising the following steps:

A. Establish a target tracking model:

B. Use the alternate iteration method to obtain the optimal tracking target;

C. Update the target template T and get a new template T ^*

D. Return the best tracking target and target template set, and continue the target tracking of the next frame.

As preferably, said step C comprises:

C1. Calculate the similarity between the tracking target and the mean value of the template, and record the similarity as sim;

C2. Compare the similarity sim with the cosine angle threshold α, β. If sim<α, execute step C3 to continue updating the template. If α≤sim≤β, then y←m, execute step C3 to update the template. If sim> β, at this time, the tracking target is basically not similar to the template, indicating that the target suffers from serious noise pollution, and the template is not updated;

C3. Perform singular value decomposition on the template T to obtain T=U∑V ^T , incrementally update the left singular vector U of the template T and obtain a new singular vector U ^* , combine formulas (5) and (6), and calculate new template

Although the embodiments of the present invention have been shown and described, those skilled in the art can understand that various changes, modifications and substitutions can be made to these embodiments without departing from the principle and spirit of the present invention. and modifications, the scope of the invention is defined by the appended claims and their equivalents.

Claims (8)

1. A robust target tracking method based on adaptive simultaneous sparse representation, comprising the following steps: S1. According to the size of the Laplacian noise energy, adaptively establish a simultaneous sparse tracking model; S2, solving the established tracking model; S3. The template is updated. 2. The robust target tracking method based on adaptive simultaneous sparse representation according to claim 1, characterized in that, the execution method of the step S1 is: comparing Laplacian mean noise ||S|| ₂ with The size of the noise energy threshold τ is given, and the simultaneous sparse tracking model is established adaptively according to the comparison results: When ||S|| ₂ ≤τ, the simultaneous sparse tracking model is: When ||S|| ₂ >τ, the sparse tracking model is: Among them, D represents the tracking template, Y is the candidate target set, λ ₁ and λ ₂ are the regular parameters of the model, X is the sparse coefficient, S is the Laplacian noise, is the fidelity item of the sparse model, ||X|| _1,1 is the coefficient matrix, and ||S|| _1,1 represents the Laplacian noise energy. 3. The robust target tracking method based on adaptive simultaneous sparse representation according to claim 2, wherein the tracking template D is expressed as D=[T, I], where T is a target template, and T ＝[T ₁ ,T ₂ ,…,T _n ]∈R ^d×n (d>>n), where d represents the dimension of the image, n represents the number of template basis vectors, each column of T is passed through zero Averaged vector. 4. The robust target tracking method based on adaptive simultaneous sparse representation according to claim 3, wherein said step S2 comprises: S21. Calculate the similarity between the tracking target y _t and the mean value of the template T, denoted as sim; S22. Determine the size of the angle threshold α between sim and cosine, and adaptively select a model for tracking: When sim<α, the tracking model of the above formula (1) is used to solve it by the method of alternating direction multiplier (ADMM) and obtain the sparse coefficient X; When sim≥α, the tracking model of the above formula (2) is used to solve it by the Alternating Direction Multiplier (ADMM) method and obtain the sparse coefficient X and Laplacian noise S. 5. The robust target tracking method based on adaptive simultaneous sparse representation according to claim 4, characterized in that, in the step S21, the similarity between the tracking target and the template T mean value is calculated according to the following formula: Among them, c is the mean value of the template T, and y is the tracked target, then ||τ|| ₂ can be equivalent to the arccosine of the mean value of the target and the template, that is, the cosine angle. 6. The robust target tracking method based on adaptive simultaneous sparse representation according to any one of claims 1-5, wherein when ||s|| ₂ ≤τ, the template T is updated. 7. the robust target tracking method based on adaptive simultaneous sparse representation according to claim 6, is characterized in that, described template update method comprises: S31. Perform singular value decomposition on the current template T and the tracked target y respectively as follows: S32. Use the singular vectors u, s, v to incrementally update U, S, V, thereby obtaining new singular vectors U _* , S _* , V _* , and the new template is expressed as: T _* ＝U _* S _* V _* ^T (5) 8. The robust target tracking method based on adaptive simultaneous sparse representation according to claim 7, further comprising step S33: using the unsupervised learning K-means method to train templates, given the number of initial classes is k, then Among them, i represents the i-th sample, when When it belongs to class k, then r _ik =1, otherwise r _ik =0, u _k is the average value of all samples belonging to class k, and J represents the sum of distances from sample points to the average value of sample points of this class; by formula (6) Then the average value of all samples of the kth class can be obtained, thus, the updated template is: T _new = [u ₁ , u ₂ , . . . , u _k ] (7).