CN104680558A - Struck target tracking method using GPU hardware for acceleration - Google Patents

Abstract

The invention discloses a Struck target tracking method using GPU hardware for acceleration, and solves the problem that the tracking performance is poor and the real-time processing capability is inacceptable in the prior art. The method comprises the following steps: 1, acquiring gray images; 2, judging whether the loaded image is a first frame image or not; 3, initializing the position rectangular frame of a tracking target; 4, extracting the features of all test samples of the gray images; 5, determining the position rectangular frame of the tracking target; 6, extracting the features of all training samples of the gray images; 7, initiating the weights and gradients of the training samples; 8, obtaining and updating the weights and gradients of supporting samples; 9, judging whether all the images are loaded or not; 10, ending target tracking. The method can be applied to the real-time tracking of the target in a video on a universal computer.

Description

Use the Struck method for tracking target that GPU hardware is accelerated

Technical field

The invention belongs to field of computer technology, further relate to a kind of structure output Struck method for tracking target based on kernel method using computer graphics processor GPU hardware to accelerate in computer video target following technical field.The present invention can realize accelerating the structure output Struck method for tracking target based on kernel method, and available realization on a general-purpose computer carries out real-time follow-up to the target in video.

Background technology

Have at a high speed, high performance method for tracking target is the core technology of computer vision field.Current method for tracking target is divided into two classes: a class is the tracking of feature based coupling, and the method mainly builds and can represent clarification of objective, is then judged the position of target by the matching degree between feature; Another kind of is the tracking of based target and background separation, and the method uses the method for machine learning to learn one can the sorter of separate targets and background, and learning process is generally on-line training process, judges target location by the sorter learnt.By contrast, the former have calculate simple, but to having illumination, block, the situation of the factors vary such as yardstick can not process well.The latter can solve the problem that the former runs into a certain extent, and has higher robustness, but its computation complexity is higher.

Patented claim " a kind of real-time distribution field method for tracking target based on the global search " (applying date: on June 26th, 2014 that Xibei Univ. of Agricultural & Forest Science & Technology proposes, application number: 201410298728.7, publication number: CN 104036528A) in disclose the tracking of a kind of feature based coupling.The method using the distribution field of target as object module, when calculated candidate areal distribution field and object module correlation matrix, by the good nature of two-dimensional Fourier transform, the search of subrange is expanded to the search of global scope, and ensure that very high processing speed, there is the ability of process in real time.But the deficiency that the method still exists is, have employed a kind of simple linear update method to upgrade target image distribution field model, for target by the situation of temporarily shielding, this update method can lose target information, thus occurs the problem of template drift.

Patented claim " a kind of high-speed target track algorithm of multi-feature extraction the Stepwise Refinement " (applying date: on September 2nd, 2014 that Wuhan University proposes, application number: 201410442527.X, publication number: CN104200216A) in disclose the tracking of a kind of feature based coupling.The various features of target carries out merging as object module by the method, the target location after adopting Mean Shift method to calculate multiple features fusion.Although the method has higher tracking accuracy, avoid the problem of template drift to a certain extent.But the deficiency that the method still exists is, only upgrade object module by initial target model and current candidate object module, therefore, the initialization of the method to target is very sensitive, and the method has the low shortcoming of robustness.

Paper " Struck:Structured Output Tracking with Kernels " (the IEEE International Conference on Computer Vision that Sam Hare, Amir Saffari and Philip H.S.Torr deliver, ICCV 2011, Barcelona, Spain, November 6-13,2011, pages 263 – 270) in openly propose the tracking of a kind of based target and background separation, the method is referred to as Struck method for tracking target.The method adopts structure based supporting vector machine model Structured SVM, learns the sorter that can be distinguished object and background, is judged the position of target by the sorter learnt.Although the method has the high advantage of performance, overcome classic method to a certain extent and can not solve by the problem of blocking the robustness reduction caused with illumination variation.But, the deficiency that the method still exists is, the method adopt the method for serial computing to carry out extracting discriminant value that the weights of sample, gradient and calculating test sample book are supported in the feature of training sample and test sample book, acquisition and renewal, because above each process relates to very intensive, therefore this realization has the slow shortcoming of processing speed.Experiment shows, for a long video sequence, the average treatment speed of the method is approximately 5fps, and obvious the method does not have real-time, thus limits its practical application.

Summary of the invention

The object of the invention is to overcome above-mentioned the deficiencies in the prior art, provide a kind of Struck method for tracking target using GPU hardware to accelerate, can realize carrying out real-time follow-up to the target in video.

Step of the present invention is as follows:

(1) gray level image is obtained:

(1a) two field picture in image sequence to be tracked is loaded in main frame internal memory;

(1b) image being loaded into main frame internal memory is converted into gray level image;

(1c) call the integrogram computing function cvIntegral increased income in the OpenCV of computer vision storehouse, calculate the integrogram of gray level image;

(1d) gray level image is copied in gray level image set;

(2) judge that whether the image loaded is the 1st two field picture in image sequence to be tracked, if so, then perform step (3), otherwise, perform step (4);

(3) the position rectangle frame of initialization tracking target:

In gray level image, choose rectangle frame tracking target be included, using the position rectangle frame of selected rectangle frame as tracking target, perform step (6);

(4) feature of all test sample books of gray level image is extracted:

(4a) to all test sample books of gray level image, full sampling method is adopted respectively, the position rectangle frame of collecting test sample;

(4b) according to the integrogram of gray level image and the position rectangle frame of all test sample books, utilize computer graphics processor GPU, the parallel Lis Hartel extracting all test sample books of gray level image is levied;

(5) the position rectangle frame of tracking target is determined:

Utilize computer graphics processor GPU, the discriminant value of all test sample books of parallel computation gray level image, finds out the test sample book corresponding to maximum discriminant value, using the position rectangle frame of the position rectangle frame of this test sample book as tracking target;

(6) feature of all training samples of gray level image is extracted:

(6a) the position rectangle frame of first of gray level image training sample is set as the position rectangle frame of tracking target;

(6b) to all the other all training samples of gray level image, adopt unique step sampling method respectively, gather the position rectangle frame of training sample;

(6c) by the position rectangle frame of all training samples of gray level image, copy in the rectangle frame set of training sample position;

(6d) according to the integrogram of gray level image and the position rectangle frame of all training samples, utilize computer graphics processor GPU, the parallel Lis Hartel extracting all training samples of gray level image is levied;

(6e) by the feature of all training samples of gray level image, copy in training sample characteristic set;

(7) weights of initialization training sample and gradient:

The weights of the training sample of gray level image and gradient are initialized as 0; The weights of the training sample of gray level image, gradient are joined respectively the set of training sample weights, the set of training sample gradient;

(8) obtain and upgrade weights, the gradient of supporting sample:

(8a) employing minimizes gradient method, from the training sample of gray level image, select two training samples, using two training samples selected as supporting sample A, B; Adopt and maximize gradient difference method, upgrade the weights and gradient of supporting sample A, B; When supporting the sum of sample more than 100, adopting and minimizing disturbance degree method, upgrading weights and the gradient of all support samples;

(8b) adopt maximum-minimize gradient method, from gray level image set, select a gray level image, from the training sample of the gray level image selected, select two training samples, using two training samples selected as supporting sample A, B; Adopt and maximize gradient difference method, upgrade the weights and gradient of supporting sample A, B; If support, the sum of sample is more than 100, then adopt and minimize disturbance degree method, upgrade weights and the gradient of all support samples;

(8c) adopt maximum-minimize gradient method, from gray level image set, select a gray level image, from the training sample of the gray level image selected, select two training samples, using two training samples selected as supporting sample A, B; Adopt and maximize gradient difference method, upgrade the weights and gradient of supporting sample A, B; By this step continued operation 9 times;

(8d) circulation performs step (8b), step (8c) 9 times;

(9) judge whether all two field pictures having loaded image sequence to be tracked, if so, then perform step (10), otherwise, perform step (1);

(10) target end is followed the tracks of.

The present invention compared with prior art, has the following advantages:

The first, the present invention adopts computer image processor GPU, achieves Struck method for tracking target, to overcome in prior art by the problem of blocking the robustness that causes with illumination variation and reducing, makes the present invention have very high robustness.

Second, the present invention adopts computer graphics processor GPU, and the parallel feature extracting training sample and test sample book, overcomes in prior art and extract the slow problem of sample characteristics, make the present invention while the extraction rate improving sample characteristics, substantially increase the speed of tracking target.

3rd, the present invention adopts computer graphics processor GPU, the discriminant value of parallel computation test sample book, overcomes in prior art the problem differentiating that tracking target is slow, make the present invention substantially increase the speed differentiating tracking target, which thereby enhance the speed of tracking target.

4th, the present invention adopts computer graphics processor GPU, parallel weights and the gradient upgrading support sample, overcome in prior art and upgrade the slow problem of tracking target model, the present invention is made to substantially increase the speed upgrading tracking target model, which thereby enhance the speed of tracking target, make the present invention have real-time.

Accompanying drawing explanation

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

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described.

The present invention adopts CUDA language, can realize on the GPU equipment of any a support CUDA framework of NVIDIA.Before enforcement method of the present invention, first should call cudaMalloc function and distribute eight region of memorys on GPU equipment.After use method of the present invention, also should call cudaFree function and discharge this eight sections of region of memorys.

With reference to Fig. 1, the present invention realizes by following steps:

Step 1, obtains gray level image.

Call the image loading function cvLoadImage increased income in the OpenCV of computer vision storehouse, the 1st two field picture in image sequence to be tracked is loaded in main frame internal memory.

Call the Color Channel transfer function cvCvtColor increased income in the OpenCV of computer vision storehouse, the image being loaded into main frame internal memory is converted into gray level image.

Call the integrogram computing function cvIntegral increased income in the OpenCV of computer vision storehouse, calculate the integrogram of gray level image.Call the memory copying function cudaMemcpy of computer graphics processor GPU, by the integrogram of gained from main frame memory copying to computer graphics processor GPU device memory region 1.

Gray level image is copied in gray level image set.

Step 2, judges that whether the image loaded is the 1st two field picture in image sequence to be tracked, if so, then performs step 3, otherwise, perform step 4.

Step 3, the position rectangle frame of initialization tracking target.

In gray level image, choose rectangle frame tracking target be included, using the position rectangle frame of selected rectangle frame as tracking target, perform step 6.

Step 4, extracts the feature of all test sample books of gray level image.

To all test sample books of gray level image, according to the position rectangle frame of the 1st following step and the 2nd step collecting test sample.

1st step, with the rectangle frame center, position of tracking target be the center of circle, 30 pixel wide construct a circle for radius.

2nd step, in circle centered by any point, the wide of the position rectangle frame of tracking target be that the Gao Weigao of the position rectangle frame of wide, tracking target constructs a rectangle frame, using the position rectangle frame of this rectangle frame as a test sample book.

The position rectangle frame data of the test sample book of gained are saved in, calculate in graphic process unit GPU device memory region 7.According to the integrogram data in the position rectangle frame data of test sample book in computer graphics processor GPU device memory region 7 and region of memory 1, utilize computer graphics processor GPU, the Lis Hartel of all test sample books of parallel computation gray level image is levied, and the feature of gained is saved in computer graphics processor GPU device memory region 8.

Step 5, determines the position rectangle frame of tracking target.

According to the data calculated in graphic process unit GPU device memory region 3,4,8, according to the following formula, utilize computer graphics processor GPU, the discriminant value of each test sample book of parallel computation gray level image:

$f=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_i\·\exp (-0.02{\left|\right|x_i-z\left|\right|}^2)$

Wherein, f represents the discriminant value of each test sample book of gray level image; N expresses support for the sum of sample; β _irepresent i-th weights supporting sample; The index operation that it is the end with natural constant e that exp () represents; x _irepresent i-th feature supporting sample; Z represents the feature of each test sample book of gray level image; || || ²represent the square operation asking vector field homoemorphism.

The test sample book with maximum discriminant value is found out, using the position rectangle frame of the position rectangle frame of the test sample book of gained as tracking target from all test sample books of gray level image.

Step 6, extracts the feature of all training samples of gray level image.

The position rectangle frame of first of gray level image training sample is set as the position rectangle frame of tracking target.

To all the other all training samples of gray level image, adopt unique step sampling method respectively, gather the position rectangle frame of training sample.

To all the other all training samples of gray level image, gather the position rectangle frame of training sample according to following 1st step and the 2nd step.

1st step, with the center of the position rectangle frame of tracking target for starting point, the surrounding to this central point disperses 8 line segments equably, and the length of every bar line segment is 60 pixel wide.

2nd step, any line segment is chosen a bit, this point to tracking target rectangle frame center, position between distance be m pixel wide, m ∈ { 12,24,36,48,60}; Centered by selected point, the Gao Weigao of the wide position rectangle frame for wide, tracking target of the position rectangle frame of tracking target constructs a rectangle frame, using the position rectangle frame of this rectangle frame as a training sample.

The position rectangle frame data of all training samples of gray level image are saved in, in computer graphics processor GPU device memory region 2.

According to the integrogram data in the position rectangle frame data of training sample in computer graphics processor GPU device memory region 2 and region of memory 1, utilize computer graphics processor GPU, the Lis Hartel of all training samples of parallel computation gray level image is levied.

The characteristic of all training samples of gray level image is saved in, in computer graphics processor GPU device memory region 3.

Step 7, the weights of initialization training sample and gradient.

The weights of all training samples of gray level image and gradient are initialized as 0, the weights after initialization, gradient are saved in respectively in computer graphics processor GPU device memory region 4,5.

Step 8, obtains and upgrades weights, the gradient of supporting sample.

(8a) according to the following formula, computer graphics processor GPU is utilized, the gradient of each training sample of parallel computation gray level image:

$g=\frac{a}{b}-1-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_i\exp (-0.02{\left|\right|x_i-x\left|\right|}^2)$

Wherein, g represents the gradient of each training sample of gray level image; A represents the rectangle frame region, position of each training sample of gray level image, with the area of rectangle frame region, the position intersecting area of first training sample of gray level image; B represents the rectangle frame region, position of each training sample of gray level image, with the rectangle frame region, position of first training sample of gray level image mutually and the area in region; N expresses support for the sum of sample; β _irepresent i-th weights supporting sample; The index operation that it is the end with natural constant e that exp () represents; x _irepresent i-th feature supporting sample; X represents the feature of each training sample of gray level image; || || ²represent the square operation asking vector field homoemorphism.

The gradient data of all training samples of gained is saved in, in computer graphics processor GPU device memory region 5.

Being copied to by first of gray level image training sample just supports in sample set, using this sample as support sample A.

From the training sample of gray level image, find out the training sample with minimal gradient, the training sample with minimal gradient is copied in negative support sample set, the training sample of minimal gradient will be had as support sample B.

According to the following formula, utilize computer graphics processor GPU, the kernel function value between sample A, B and each support sample is supported in parallel computation respectively:

K(x ₁,x ₂)＝exp(-0.02||x ₁-x ₂|| ²)

Wherein, K (x ₁, x ₂) express support for kernel function value between sample A, B and each support sample; The index operation that it is the end with natural constant e that exp () represents; x ₁express support for the feature of sample A, B; x ₂represent the feature of each support sample; || || ²represent the square operation asking vector field homoemorphism.

The kernel function Value Data of gained is saved in, in computer graphics processor GPU device memory region 6.

According to the following formula, the right value update amount supporting sample is calculated:

$\mathrm{\λ}=\left\{\begin{array}{cc}\max (0,\min (\frac{g_A-g_B}{2-2K(x_A,x_B)},100-{\mathrm{\β}}_A)),& y_A=y\\ \max (0,\min (\frac{g_A-g_B}{2-2K(x_A,x_B)},-{\mathrm{\β}}_A)),& y_A\≠y\end{array}\right.$

Wherein, λ expresses support for the right value update amount of sample; Maxima operation is got in max () expression; Min () represents that getting minimum value operates; g _aexpress support for the gradient of sample A; g _bexpress support for the gradient of sample B; K (x _a, x _b) express support for sample A and support the kernel function value between sample B; β _aexpress support for the weights of sample A; y _aexpress support for the position rectangle frame of sample A; Y expresses support for the position rectangle frame of first training sample of the gray level image belonging to sample A.

To support that the weights of sample A add right value update amount λ, will support that the weights of sample B deduct right value update amount λ.

The weights of support sample A, B after renewal are saved in, in computer graphics processor GPU device memory region 4.

According to the following formula, utilize computer graphics processor GPU, the parallel gradient upgrading each support sample:

g＝g′-λ(K(x,x _A)-K(x,x _B))

Wherein, g represents the gradient upgrading rear each support sample; The gradient of each support sample before g ' expression upgrades; λ represents right value update amount; K (x, x _a) represent each support sample and support the kernel function value between sample A; K (x, x _b) represent each support sample and support the kernel function value between sample B.

The gradient data of all support samples after renewal is saved in, in computer graphics processor GPU device memory region 5.

Judge to support that whether the sum of sample is more than 100, if so, then performs the 1st step below, the 2nd step, the 3rd step, the 4th step and the 5th step, otherwise, perform step (8b).

1st step, according to the following formula, utilizes computer graphics processor GPU, the disturbance degree of each negative support sample of parallel computation:

$e={\mathrm{\β}}_2^2(2-2K(x_1,x_2))$

Wherein, e represents the disturbance degree of each negative support sample; β ₂represent the weights of each negative support sample; K (x ₁, x ₂) represent the just support sample of the gray level image belonging to each negative support sample and the kernel function value between each negative support sample.

2nd step, finds out the negative support sample with minimum influence degree from negative support sample, using having the negative support sample of minimum influence degree as sample to be deleted, is deleted by the negative support sample with minimum influence degree from negative support sample set.

3rd step, according to the following formula, utilizes computer graphics processor GPU, the parallel gradient upgrading each support sample:

g＝g′-β ₂(K(x,x ₁)-K(x,x ₂))

Wherein, g represents the gradient upgrading rear each support sample; β ₂represent the weights of sample to be deleted; K (x, x ₁) represent the kernel function value just supported between sample of each support sample and the gray level image belonging to sample to be deleted; K (x, x ₂) represent kernel function value between each support sample and sample to be deleted.

The gradient data of all support samples after renewal is saved in, in computer graphics processor GPU device memory region 5.

4th step, according to the following formula, upgrades the weights just supporting sample of the gray level image belonging to sample to be deleted:

β ₁＝β ₁′+β ₂′

Wherein, β ₁represent the weights just supporting sample of the gray level image after upgrading belonging to sample to be deleted; β ₁' represent the weights just supporting sample upgrading front gray level image belonging to sample to be deleted; β ₂' represent the weights upgrading front sample to be deleted.

5th step, if what upgrade rear gray level image belonging to sample to be deleted just supports whether the weights of sample equal 0, if, this is just being supported sample is from just supporting to delete sample set, gray level image belonging to sample to be deleted is deleted from gray level image set, otherwise, this is just being supported the weights of sample are saved in computer graphics processor GPU device memory region 4.

(8b) Stochastic choice gray level image from gray level image set, using the gray level image of selection as pending gray level image.

According to the following formula, utilize computer graphics processor GPU, the gradient of each training sample of the pending gray level image of parallel computation:

$g=\frac{a}{b}-1-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_i\exp (-0.02{\left|\right|x_i-x\left|\right|}^2)$

Wherein, g represents the gradient of each training sample of pending gray level image; A represents the area of rectangle frame region, the position intersecting area of the rectangle frame region, position of each training sample of pending gray level image and first training sample of pending gray level image; The rectangle frame region, position that b represents the rectangle frame region, position of each training sample of pending gray level image and first training sample of pending gray level image mutually and the area in region; N expresses support for the sum of sample; β _irepresent i-th weights supporting sample; The index operation that it is the end with natural constant e that exp () represents; x _irepresent i-th feature supporting sample; X represents the feature of each training sample of pending gray level image; || || ²represent the square operation asking vector field homoemorphism.

The gradient data of all training samples of gained is saved in, in computer graphics processor GPU device memory region 5.

Be subordinated to the support sample found out in the support sample of pending gray level image and there is greatest gradient, using the support sample obtained as supporting sample A.

Be subordinated to the training sample found out in the training sample of pending gray level image and there is minimal gradient, using the training sample obtained as supporting sample B.

If support, sample B is not labeled, then support sample B is copied to negative support and, in sample set, according to the following formula, utilize computer graphics processor GPU, the kernel function value between sample B and each support sample is supported in parallel computation:

K(x ₁,x _B)＝exp(-0.02||x ₁-x _B|| ²)

Wherein, K (x ₁, x _b) represent each support sample and support the kernel function value between sample B; The index operation that it is the end with natural constant e that exp () represents; x ₁represent the feature of each support sample; x _bexpress support for the feature of sample B; || || ²represent the square operation asking vector field homoemorphism.

The kernel function Value Data of gained is saved in, in computer graphics processor GPU device memory region 7.

According to the following formula, the right value update amount supporting sample is calculated:

$\mathrm{\λ}=\left\{\begin{array}{cc}\max (0,\min (\frac{g_A-g_B}{2-2K(x_A,x_B)},100-{\mathrm{\β}}_A)),& y_A=y\\ \max (0,\min (\frac{g_A-g_B}{2-2K(x_A,x_B)},-{\mathrm{\β}}_A)),& y_A\≠y\end{array}\right.$

Wherein, λ expresses support for the right value update amount of sample; Maxima operation is got in max () expression; Min () represents that getting minimum value operates; g _aexpress support for the gradient of sample A; g _bexpress support for the gradient of sample B; K (x _a, x _b) express support for sample A and support the kernel function value between sample B; β _aexpress support for the weights of sample A; y _aexpress support for the position rectangle frame of sample A; Y expresses support for the position rectangle frame of first training sample of the gray level image belonging to sample A.

To support that the weights of sample A add right value update amount λ, will support that the weights of sample B deduct right value update amount λ.

The weights of support sample A, B after renewal are saved in, in computer graphics processor GPU device memory region 4.

According to the following formula, utilize computer graphics processor GPU, the parallel gradient upgrading each support sample:

g＝g′-λ(K(x,x _A)-K(x,x _B))

Wherein, g represents the gradient upgrading rear each support sample; The gradient of each support sample before g ' expression upgrades; λ represents right value update amount; K (x, x _a) represent each support sample and support the kernel function value between sample A; K (x, x _b) represent each support sample and support the kernel function value between sample B.

The gradient data of all support samples after renewal is saved in, in computer graphics processor GPU device memory region 5.

Judge to support that whether the sum of sample is more than 100, if so, then performs the 1st step below, the 2nd step, the 3rd step, the 4th step and the 5th step, otherwise, perform step (8c).

1st step, according to the following formula, utilizes computer graphics processor GPU, the disturbance degree of each negative support sample of parallel computation:

$e={\mathrm{\β}}_2^2(2-2K(x_1,x_2))$

Wherein, e represents the disturbance degree of each negative support sample; β ₂represent the weights of each negative support sample; K (x ₁, x ₂) represent the just support sample of the gray level image belonging to each negative support sample and the kernel function value between each negative support sample.

2nd step, finds out the negative support sample with minimum influence degree from negative support sample, using having the negative support sample of minimum influence degree as sample to be deleted, is deleted by the negative support sample with minimum influence degree from negative support sample set.

3rd step, according to the following formula, utilizes computer graphics processor GPU, the parallel gradient upgrading each support sample:

g＝g′-β ₂(K(x,x ₁)-K(x,x ₂))

Wherein, g represents the gradient upgrading rear each support sample; β ₂represent the weights of sample to be deleted; K (x, x ₁) represent the kernel function value just supported between sample of each support sample and the gray level image belonging to sample to be deleted; K (x, x ₂) represent kernel function value between each support sample and sample to be deleted.

The gradient data of all support samples after renewal is saved in, in computer graphics processor GPU device memory region 5.

4th step, according to the following formula, upgrades the weights just supporting sample of the gray level image belonging to sample to be deleted:

β ₁＝β ₁′+β ₂′

Wherein, β ₁represent the weights just supporting sample of the gray level image after upgrading belonging to sample to be deleted; β ₁' represent the weights just supporting sample upgrading front gray level image belonging to sample to be deleted; β ₂' represent the weights upgrading front sample to be deleted.

5th step, if what upgrade rear gray level image belonging to sample to be deleted just supports whether the weights of sample equal 0, if, this is just being supported sample is from just supporting to delete sample set, gray level image belonging to sample to be deleted is deleted from gray level image set, otherwise, this is just being supported the weights of sample are saved in computer graphics processor GPU device memory region 4.

(8c) Stochastic choice gray level image from gray level image set, using the gray level image of selection as pending gray level image.

Be subordinated to the support sample found out in the support sample of pending gray level image and there is greatest gradient, using the support sample obtained as supporting sample A.

Be subordinated to the support sample found out in the support sample of pending gray level image and there is minimal gradient, using the support sample obtained as supporting sample B.

According to the following formula, the right value update amount supporting sample is calculated:

$\mathrm{\λ}=\left\{\begin{array}{cc}\max (0,\min (\frac{g_A-g_B}{2-2K(x_A,x_B)},100-{\mathrm{\β}}_A)),& y_A=y\\ \max (0,\min (\frac{g_A-g_B}{2-2K(x_A,x_B)},-{\mathrm{\β}}_A)),& y_A\≠y\end{array}\right.$

Wherein, λ expresses support for the right value update amount of sample; Maxima operation is got in max () expression; Min () represents that getting minimum value operates; g _aexpress support for the gradient of sample A; g _bexpress support for the gradient of sample B; K (x _a, x _b) express support for sample A and support the kernel function value between sample B; β _aexpress support for the weights of sample A; y _aexpress support for the position rectangle frame of sample A; Y expresses support for the position rectangle frame of first training sample of the gray level image belonging to sample A.

To support that the weights of sample A add right value update amount λ, will support that the weights of sample B deduct right value update amount λ.

The weights of support sample A, B after renewal are saved in, in computer graphics processor GPU device memory region 4.

According to the following formula, utilize computer graphics processor GPU, the parallel gradient upgrading each support sample:

g＝g′-λ(K(x,x _A)-K(x,x _B))

Wherein, g represents the gradient upgrading rear each support sample; The gradient of each support sample before g ' expression upgrades; λ represents right value update amount; K (x, x _a) represent each support sample and support the kernel function value between sample A; K (x, x _b) represent each support sample and support the kernel function value between sample B.

The gradient data of all support samples after renewal is saved in, in computer graphics processor GPU device memory region 5.By step (8c) continued operation 9 times.

(8d) circulation performs step (8b), step (8c) 9 times.

Step 9, judges whether all two field pictures having loaded image sequence to be tracked, if so, then performs step 10, otherwise, perform step 1.

Step 10, target end is followed the tracks of.

Claims (9)

1. the Struck method for tracking target using GPU hardware to accelerate, comprises the steps:

(1) gray level image is obtained:

(1a) two field picture in image sequence to be tracked is loaded in main frame internal memory;

(1b) image being loaded into main frame internal memory is converted into gray level image;

(1c) call the integrogram computing function cvIntegral increased income in the OpenCV of computer vision storehouse, calculate the integrogram of gray level image;

(1d) gray level image is copied in gray level image set;

(2) judge that whether the image loaded is the 1st two field picture in image sequence to be tracked, if so, then perform step (3), otherwise, perform step (4);

(3) the position rectangle frame of initialization tracking target:

In gray level image, choose rectangle frame tracking target be included, using the position rectangle frame of selected rectangle frame as tracking target, perform step (6);

(4) feature of all test sample books of gray level image is extracted:

(4a) to all test sample books of gray level image, full sampling method is adopted respectively, the position rectangle frame of collecting test sample;

(4b) according to the integrogram of gray level image and the position rectangle frame of all test sample books, utilize computer graphics processor GPU, the parallel Lis Hartel extracting all test sample books of gray level image is levied;

(5) the position rectangle frame of tracking target is determined:

Utilize computer graphics processor GPU, the discriminant value of all test sample books of parallel computation gray level image, finds out the test sample book corresponding to maximum discriminant value, using the position rectangle frame of the position rectangle frame of this test sample book as tracking target;

(6) feature of all training samples of gray level image is extracted:

(6a) the position rectangle frame of first of gray level image training sample is set as the position rectangle frame of tracking target;

(6b) to all the other all training samples of gray level image, adopt unique step sampling method respectively, gather the position rectangle frame of training sample;

(6c) by the position rectangle frame of all training samples of gray level image, copy in the rectangle frame set of training sample position;

(6d) according to the integrogram of gray level image and the position rectangle frame of all training samples, utilize computer graphics processor GPU, the parallel Lis Hartel extracting all training samples of gray level image is levied;

(6e) by the feature of all training samples of gray level image, copy in training sample characteristic set;

(7) weights of initialization training sample and gradient:

The weights of the training sample of gray level image and gradient are initialized as 0; The weights of the training sample of gray level image, gradient are joined respectively the set of training sample weights, the set of training sample gradient;

(8) obtain and upgrade weights, the gradient of supporting sample:

(8a) employing minimizes gradient method, from the training sample of gray level image, select two training samples, using two training samples selected as supporting sample A, B; Adopt and maximize gradient difference method, upgrade the weights and gradient of supporting sample A, B; When supporting the sum of sample more than 100, adopting and minimizing disturbance degree method, upgrading weights and the gradient of all support samples;

(8b) adopt maximum-minimize gradient method, from gray level image set, select a gray level image, from the training sample of the gray level image selected, select two training samples, using two training samples selected as supporting sample A, B; Adopt and maximize gradient difference method, upgrade the weights and gradient of supporting sample A, B; If support, the sum of sample is more than 100, then adopt and minimize disturbance degree method, upgrade weights and the gradient of all support samples;

(8c) adopt maximum-minimize gradient method, from gray level image set, select a gray level image, from the training sample of the gray level image selected, select two training samples, using two training samples selected as supporting sample A, B; Adopt and maximize gradient difference method, upgrade the weights and gradient of supporting sample A, B; By this step continued operation 9 times;

(8d) circulation performs step (8b), step (8c) 9 times;

(9) judge whether all two field pictures having loaded image sequence to be tracked, if so, then perform step (10), otherwise, perform step (1);

(10) target end is followed the tracks of.

2. the Struck method for tracking target of use GPU hardware acceleration according to claim 1, it is characterized in that, the concrete steps of the described full sampling method of step (4a) are as follows:

1st step, with the rectangle frame center, position of tracking target be the center of circle, 30 pixel wide construct a circle for radius;

2nd step, in circle centered by any point, the wide of the position rectangle frame of tracking target be that the Gao Weigao of the position rectangle frame of wide, tracking target constructs a rectangle frame, using the position rectangle frame of this rectangle frame as a test sample book.

3. the Struck method for tracking target of use GPU hardware acceleration according to claim 1, is characterized in that, the discriminant value of each test sample book of the gray level image described in step (5), calculates according to the following formula:

$f=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_i\·\exp (-0.02{\left|\right|x_i-z\left|\right|}^2)$

Wherein, f represents the discriminant value of each test sample book of gray level image; N expresses support for the sum of sample; β _irepresent i-th weights supporting sample; The exponential function that it is the end that exp () represents with natural constant e operates; x _irepresent i-th feature supporting sample; Z represents the feature of each test sample book of gray level image; || || ²represent the square operation asking vector field homoemorphism.

4. the Struck method for tracking target of use GPU hardware acceleration according to claim 1, it is characterized in that, the concrete steps of step (6b) described unique step sampling method are as follows:

1st step, with the center of the position rectangle frame of tracking target for starting point, the surrounding to this central point disperses 8 line segments equably, and the length of every bar line segment is 60 pixel wide;

2nd step, any line segment is chosen a bit, this point to tracking target rectangle frame center, position between distance be m pixel wide, m ∈ { 12,24,36,48,60}; Centered by selected point, the Gao Weigao of the wide position rectangle frame for wide, tracking target of the position rectangle frame of tracking target constructs a rectangle frame, using the position rectangle frame of this rectangle frame as a training sample.

5. the Struck method for tracking target of use GPU hardware acceleration according to claim 1, it is characterized in that, the concrete steps minimizing gradient method described in step (8a) are as follows:

1st step, according to the following formula, utilizes computer graphics processor GPU, the gradient of each training sample of parallel computation gray level image:

$g=\frac{a}{b}-1-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_i\exp (-0.02{\left|\right|x_i-x\left|\right|}^2)$

Wherein, g represents the gradient of each training sample of gray level image; A represents the area of rectangle frame region, the position intersecting area of the rectangle frame region, position of each training sample of gray level image and first training sample of gray level image; The rectangle frame region, position that b represents the rectangle frame region, position of each training sample of gray level image and first training sample of gray level image mutually and the area in region; N expresses support for the sum of sample; β _irepresent i-th weights supporting sample; The index operation that it is the end with natural constant e that exp () represents; x _irepresent i-th feature supporting sample; X represents the feature of each training sample of gray level image; || || ²represent the square operation asking vector field homoemorphism;

2nd step, copies to first of gray level image training sample and just supports in sample set, using this sample as support sample A;

3rd step, finds out the training sample with minimal gradient from the training sample of gray level image, is copied to by the training sample mark with minimal gradient in negative support sample set, using this sample as support sample B;

4th step, according to the following formula, utilizes computer graphics processor GPU, and the kernel function value between sample A, B and each support sample is supported in parallel computation respectively:

K(x ₁,x ₂)＝exp(-0.02||x ₁-x ₂|| ²)

Wherein, K (x ₁, x ₂) express support for kernel function value between sample A, B and each support sample; The index operation that it is the end with natural constant e that exp () represents; x ₁express support for the feature of sample A, B; x ₂represent the feature of each support sample; || || ²represent the square operation asking vector field homoemorphism.

6. the Struck method for tracking target of use GPU hardware acceleration according to claim 1, it is characterized in that, the concrete steps maximizing gradient difference method described in step (8a), step (8b) and step (8c) are as follows:

1st step, according to the following formula, calculates the right value update amount supporting sample:

$\mathrm{\λ}=\left\{\begin{array}{cc}\max (0,\min (\frac{g_A-g_B}{2-2K(x_A,x_B)},100-{\mathrm{\β}}_A)),& y_A=y\\ \max (0,\min (\frac{g_A-g_B}{2-2K(x_A,x_B)},-{\mathrm{\β}}_A)),& y_A\≠y\end{array}\right.$

Wherein, λ expresses support for the right value update amount of sample; Maxima operation is got in max () expression; Min () represents that getting minimum value operates; g _aexpress support for the gradient of sample A; g _bexpress support for the gradient of sample B; K (x _a, x _b) express support for sample A and support the kernel function value between sample B; β _aexpress support for the weights of sample A; y _aexpress support for the position rectangle frame of sample A; Y expresses support for the position rectangle frame of first training sample of the gray level image belonging to sample A;

2nd step, will support that the weights of sample A add right value update amount λ, will support that the weights of sample B deduct right value update amount λ;

3rd step, according to the following formula, utilizes computer graphics processor GPU, the parallel gradient upgrading each support sample:

g＝g′-λ(K(x,x _A)-K(x,x _B))

Wherein, g represents the gradient upgrading rear each support sample; The gradient of each support sample before g ' expression upgrades; λ represents right value update amount; K (x, x _a) represent each support sample and support the kernel function value between sample A; K (x, x _b) represent each support sample and support the kernel function value between sample B.

7. the Struck method for tracking target that accelerates of use GPU hardware according to claim 1, is characterized in that, the concrete steps minimizing disturbance degree method described in step (8a) and step (8b) are as follows:

1st step, according to the following formula, utilizes computer graphics processor GPU, the disturbance degree of each negative support sample of parallel computation:

$e={\mathrm{\β}}_2^2(2-2K(x_1,x_2))$

Wherein, e represents the disturbance degree of each negative support sample; β ₂represent the weights of each negative support sample; K (x ₁, x ₂) represent the kernel function value just supported between sample and each negative support sample of the gray level image belonging to each negative support sample;

2nd step, finds out the negative support sample with minimum influence degree from negative support sample, using having the negative support sample of minimum influence degree as sample to be deleted, is deleted by the negative support sample with minimum influence degree from negative support sample set;

3rd step, according to the following formula, utilizes computer graphics processor GPU, the parallel gradient upgrading each support sample:

g＝g′-β ₂(K(x,x ₁)-K(x,x ₂))

Wherein, g represents the gradient upgrading rear each support sample; β ₂represent the weights of sample to be deleted; K (x, x ₁) represent the kernel function value just supported between sample of each support sample and the gray level image belonging to sample to be deleted; K (x, x ₂) represent kernel function value between each support sample and sample to be deleted;

4th step, according to the following formula, upgrades the weights just supporting sample of the gray level image belonging to sample to be deleted:

β ₁＝β ₁′+β ₂′

Wherein, β ₁represent the weights just supporting sample of the gray level image after upgrading belonging to sample to be deleted; β ₁' represent the weights just supporting sample upgrading front gray level image belonging to sample to be deleted; β ₂' represent the weights upgrading front sample to be deleted;

5th step, if after upgrading gray level image belonging to sample to be deleted just support that the weights of sample equal 0, then this is just being supported that sample is from just supporting to delete sample set, deletes the gray level image belonging to sample to be deleted from gray level image set.

8. the Struck method for tracking target of use GPU hardware acceleration according to claim 1, it is characterized in that, the concrete steps of step (8b) described maximum-minimize gradient method are as follows:

1st step, Stochastic choice gray level image from gray level image set, using this gray level image as pending gray level image;

2nd step, according to the following formula, utilizes computer graphics processor GPU, the gradient of each training sample of the pending gray level image of parallel computation:

$g=\frac{a}{b}-1-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_i\exp (-0.02{\left|\right|x_i-x\left|\right|}^2)$

Wherein, g represents the gradient of each training sample of pending gray level image; A represents the area of rectangle frame region, the position intersecting area of the rectangle frame region, position of each training sample of pending gray level image and first training sample of pending gray level image; The rectangle frame region, position that b represents the rectangle frame region, position of each training sample of pending gray level image and first training sample of pending gray level image mutually and the area in region; N expresses support for the sum of sample; β _irepresent i-th weights supporting sample; The index operation that it is the end with natural constant e that exp () represents; x _irepresent i-th feature supporting sample; X represents the feature of each training sample of pending gray level image; || || ²represent the square operation asking vector field homoemorphism;

3rd step, is subordinated in the support sample of pending gray level image, finds out the support sample with greatest gradient, using the support sample obtained as supporting sample A;

4th step, is subordinated in the training sample of pending gray level image, finds out the training sample with minimal gradient, using the training sample obtained as supporting sample B;

5th step, if support, sample B is not labeled, then support sample B is copied to negative support and, in sample set, according to the following formula, utilize computer graphics processor GPU, the kernel function value between sample B and each support sample is supported in parallel computation:

K(x ₁,x _B)＝exp(-0.02||x ₁-x _B|| ²)

Wherein, K (x ₁, x _b) represent each support sample and support the kernel function value between sample B; The index operation that it is the end with natural constant e that exp () represents; x ₁represent the feature of each support sample; x _bexpress support for the feature of sample B; || || ²represent the square operation asking vector field homoemorphism.

9. the Struck method for tracking target of use GPU hardware acceleration according to claim 1, it is characterized in that, the concrete steps of step (8c) described maximum-minimize gradient method are as follows:

1st step, Stochastic choice gray level image from gray level image set, using this gray level image as pending gray level image;

2nd step, is subordinated in the support sample of pending gray level image, finds out the support sample with greatest gradient, using the support sample obtained as supporting sample A;

3rd step, is subordinated in the support sample of pending gray level image, finds out the support sample with minimal gradient, using the support sample obtained as supporting sample B.