CN105809709A - Bit plane-based moving target tracking method - Google Patents

Abstract

The invention discloses a bit plane-based moving target tracking method which comprises the following steps: (1) a tracking target is chosen in a first frame of a video and the position of the target is marked manually; (2) a brightness bit plane and a local binary pattern bit plane of a target area image are respectively obtained and then subjected to Gaussian smoothing operation so as to build two appearance models; (3) search areas are determined in a next frame, a smoothed brightness bit plane and a smoothed local binary pattern bit plane of the search areas are respectively obtained, and an area with maximum proximity to the two appearance models is searched and used as the tracking target; (4) according to the built appearance models and a tracking result in a current frame, the appearance models are updated according to a preset updating rate; (5) return to the third step till all video frames are processed. The bit plane-based moving target tracking method is advantaged by tracking precision and robustness, and a problem that a moving target is hard to track under conditions of light condition changes, target position changes, significant appearance changes and the like in the video.

Description

A kind of motion target tracking method based on bit plane

Technical field

The invention belongs to computer vision and area of pattern recognition, be specifically related to a kind of motion target tracking method based on bit plane.

Background technology

Target following is the focus of computer vision research, is obtained for extensive use in fields such as video monitoring, video frequency searching, traffic monitorings.At present, motion target tracking method is broadly divided into two classes: based on the method sentencing method for distinguishing and Model Matching.

Based on sentencing method for distinguishing motion target tracking problem as a classification problem, it is intended to training a grader, separated by moving target from background, this kind of method is also referred to as the tracking based on detection.2004, the support vector machine of off-line training was incorporated into based in the tracking of light stream by Avidan, but when target appearance occurs significantly to change, followed the tracks of and drift occur.In order to solve this problem, it is necessary to grader is carried out online updating.2007, Avidan proposed and some Weak Classifiers is integrated into a strong classifier, and the method for real-time update Weak Classifier, and the method can accurately distinguish that each pixel is target or background.2008, Grabner adopted semi-supervised method training grader, it is only necessary to manually mark sample in first frame of video.But, in these supervision with semi-supervised method, some useful information is lost, if once there is error in the process followed the tracks of, it is easy to make error accumulation, thus causing following the tracks of unsuccessfully.Viola et al. proposes use multi-instance learning method to detect moving target, to overcome the problem of drift in tracking process.In follow-up research, multi-instance learning method is improved by Babenko, Zhang et al. further, the precision of algorithm and real-time be improved significantly.

Model matching method searches the region most like with object module in each frame as following the tracks of result.As far back as 1996, Black et al. proposed the method adopting the display model of off-line learning to carry out target following, but the method does not adapt to the change of target appearance.Subsequently, some adaptive outward appearance modeling methods are gradually introduced to motion target tracking.Along with rarefaction representation is in the successful Application in multiple fields, a lot of scholars begin attempt to the method adopting rarefaction representation to modeling target.From 2012, the concept of distribution field was successively introduced tracking field by Sevilla-Lara, Felsberg et al., and track algorithm has been improved, and achieves good tracking effect.

These algorithms obtain application in specific environment, but complicated and changeable owing to following the tracks of scene, and it tends not to effectively solve illumination change, cosmetic variation, change of shape and block the impact on target following.

Summary of the invention

Present invention aims to the problems such as the change of the illumination condition of existence in video scene, object pose change and outward appearance significantly change, there is provided a kind of based on bit plane, and merge brightness and the motion target tracking method of LBP feature, it is intended to improve the accuracy of motion target tracking under complex environment.

For reaching above-described purpose, the present invention adopts the following technical scheme that:

A kind of motion target tracking method based on bit plane, it is characterised in that comprise the steps:

Step (1) is selected target of following the tracks of in video the first frame, and hand labeled target location；

Step (2), to target area image, asks for its luminance bit plane and local binary patterns bit plane, then carries out Gaussian smoothing, sets up brightness display model and texture appearance model respectively；

Step (3) determines region of search in the next frame, to its ask for respectively smooth after luminance bit plane and local binary patterns bit plane, and search for the brightness display model set up in step (2) and texture appearance model closest to region as tracking target；

Step (4), according to the tracking result in the display model set up and present frame, updates brightness display model and texture appearance model according to renewal rate set in advance；

Step (5) is disposed when all frame of video, then stop calculating；Otherwise, jump procedure (3).

Further, in described step (1), the method for hand labeled target location is: by the selected target of following the tracks of of rectangle frame, and record this rectangle frame upper left corner two-dimensional coordinate (x, y), and the width of rectangle frame and height.

Further, the method setting up brightness display model in described step (2) is:

1. utilize below equation to represent the brightness of each pixel, namely represent the brightness of each pixel by natural binary sequence:

Wherein, (i j) represents the pixel intensity of the i-th row jth row, a to I_i,j,kRepresent the value of kth bit when this pixel intensity natural binary sequence is represented；

2. below equation is utilized to represent for binary gray code Sequence Transformed for natural binary:

Wherein, b_i,j,kRepresent I (i, the value of kth bit when j) representing with binary gray code；

3. utilize below equation that the brightness of each pixel is projected to different bit planes by bit, thus obtaining 8 luminance bit planes:

Wherein, BPGC is luminance bit plane, and i and j represents the row and column of image respectively, and k represents the sequence number of bit plane, k=0,1,2 ..., 7；

4. below equation is utilized to carry out Gaussian smoothing:

Wherein, M₁For brightness display model,Be average it is μ_s, standard deviation is σ_s2 dimension gaussian kernel,Be average it is μ_f, standard deviation is σ_f1 dimension gaussian kernel, * represents convolution algorithm.

Further, the method setting up texture appearance model in described step (2) is:

1. seek the local binary patterns feature of target image, adopt the LBP operator of 3 × 3 traditional sizes, the LBP value centered by all non-edge pixels is calculated for target image；

2. utilize below equation to represent the LBP value of each pixel, namely represent the LBP value of each pixel by natural binary sequence:

Wherein, (i j) represents the LBP value of the pixel of the i-th row jth row, a ' to I '_{I, j, k}Represent the value of kth bit when the LBP value natural binary sequence of this pixel is represented；

3. utilize below equation that LBP value is converted into binary gray code to represent:

Wherein, b '_{I, j, k}Represent the value of kth bit when LBP value binary gray code is represented；

4. utilize below equation that the LBP value of each pixel is projected to different bit planes by bit, thus obtaining 8 texture bits planes:

Wherein, BPGC ' is texture bits plane, and i and j represents the row and column of image respectively, and k represents the sequence number of bit plane, k=0,1,2 ..., 7；

5. below equation is utilized to carry out Gaussian smoothing, thus obtaining texture appearance model M₂:

Wherein, M₂For texture appearance model,Be average it is μ_s, standard deviation is σ_s2 dimension gaussian kernel,Be average it is μ_f, standard deviation is σ_f1 dimension gaussian kernel, * represents convolution algorithm.

Further, the method determining region of search in described step (3) is:

With previous frame target positionCenter be the center of circle, with r for radius, the rectangle in this region is dropped on as candidate target in all centers, adopts below equation to represent:

Wherein,For target position in previous frame, l (x) represents image block x position in present frame, and r is search radius, and X represents the set of all candidate image blocks, | | | | represent Euclidean distance.

Further, described step (3) is tried to achieve candidate region smooth after luminance bit plane C₁Method be:

1. utilize below equation to represent the brightness of each pixel in candidate region, namely represent the brightness of each pixel by natural binary sequence:

Wherein, (i j) represents the pixel intensity of the i-th row jth row, a in candidate region to I_i,j,kRepresent the value of kth bit when this pixel intensity natural binary sequence is represented；

2. below equation is utilized to represent for binary gray code Sequence Transformed for natural binary:

Wherein, b_i,j,kRepresent I (i, the value of kth bit when j) representing with binary gray code；

3. utilize below equation that the brightness of pixel each in candidate region is projected to different bit planes by bit, thus obtaining 8 luminance bit planes:

Wherein, BPGC is luminance bit plane, and i and j represents the row and column of image respectively, and k represents the sequence number of bit plane, k=0,1,2 ..., 7；

4. below equation is utilized to carry out Gaussian smoothing:

Thus obtain candidate region smooth after luminance bit plane C₁。

Further, described step (3) is tried to achieve the local binary patterns bit plane C of candidate region₂Method be:

1. seek the LBP feature of image in candidate region: adopt the LBP operator of 3 × 3 traditional sizes, the LBP value centered by all non-edge pixels is calculated for candidate region image；

2. utilize below equation to represent the LBP value of each pixel in candidate region, namely represent the LBP value of each pixel by natural binary sequence:

Wherein, (i j) represents the LBP value of the i-th row jth row, a ' in candidate region to I '_{I, j, k}Represent the value of kth bit when this value natural binary sequence is represented；

3. utilize below equation that LBP value is converted into binary gray code to represent:

Wherein, b '_{I, j, k}Represent the value of kth bit when LBP value binary gray code is represented；

4. utilize below equation that the LBP value of pixel each in candidate region is projected to different bit planes by bit, thus obtaining 8 texture bits planes:

Wherein,For texture bits plane, i and j represents the row and column of image respectively, and k represents the sequence number of bit plane, k=0,1,2 ..., 7；

5. below equation is utilized to carry out Gaussian smoothing:

Obtain LBP bit plane C₂。

Further, the brightness display model set up in search and step (2) in described step (3) and texture appearance model closest to region as following the tracks of order calibration method be: try to achieve the region x ' making following formula obtain minima:

Wherein,Represent target position in present frame, M₁For brightness display model, M₂For texture appearance model, C₁For smoothed luminance bit plane, C₂For smoothed LBP bit plane, w is weights and 0 < w < 1, Dist () is the distance between two bit planes, tries to achieve as follows:

Wherein, BP₁And BP₂Respectively two bit planes, i, j and k represents line number, row number and bit plane sequence number respectively.

Further, the method updating brightness display model and texture appearance model in described step (4) is:

Wherein, λ is renewal rate and 0 < λ < 1, M₁For brightness display model, M₂For texture appearance model, C₁(x ') luminance bit plane corresponding to target area in present frame, C₂(x ') texture bits plane corresponding to target area in present frame.

Adopt the present invention of as above technical scheme, have the advantages that

Motion target tracking method based on bit plane provided by the invention, make full use of brightness and the LBP textural characteristics of original image, by convolution algorithm by the uncertain introducing tracking process of position, and adopt the method for bit plane that target appearance is set up model, effectively overcome illumination condition change, object pose change and outward appearance significantly change etc. to the harmful effect following the tracks of result, there is good accuracy and robustness.

Accompanying drawing explanation

Fig. 1 is the motion target tracking method flow chart based on bit plane involved in the present invention.

Fig. 2 is gray level bit plane involved in the present invention.

Fig. 3 is texture bits plane involved in the present invention.

Fig. 4-1 is the center error curve diagram in test video sequence involved in the present invention.

Fig. 4-2 is the center error curve diagram in test video sequence involved in the present invention.

Fig. 4-3 is the center error curve diagram in test video sequence involved in the present invention.

Fig. 4-4 is the center error curve diagram in test video sequence involved in the present invention.

Fig. 4-5 is the center error curve diagram in test video sequence involved in the present invention.

Fig. 4-6 is the center error curve diagram in test video sequence involved in the present invention.

Fig. 4-7 is the center error curve diagram in test video sequence involved in the present invention.

Fig. 4-8 is the center error curve diagram in test video sequence involved in the present invention.

Fig. 5 is the tracking comparison diagram in test video sequence involved in the present invention.

Detailed description of the invention

Below in conjunction with drawings and the specific embodiments, the present invention is described in detail, and the explanation of the invention is not limited.

A kind of motion target tracking method based on bit plane, as it is shown in figure 1, comprise the following steps:

(1) selected target of following the tracks of in video the first frame, and hand labeled target location.

In the first frame by the selected target of following the tracks of of rectangle frame, and record this rectangle frame upper left corner two-dimensional coordinate (x, y), and the width of rectangle frame and height.

(2) to target area image, ask for its luminance bit plane and local binary patterns bit plane, then carry out Gaussian smoothing, set up brightness display model and texture appearance model respectively.

1. brightness display model is set up

First, utilize below equation to represent the brightness of each pixel, namely represent the brightness of each pixel by natural binary sequence:

Wherein, (i j) represents the pixel intensity of the i-th row jth row, a to I_i,j,kRepresent the value of kth bit when this pixel intensity natural binary sequence is represented.

Secondly, below equation is utilized to represent for binary gray code Sequence Transformed for natural binary:

Wherein, b_i,j,kRepresent I (i, the value of kth bit when j) representing with binary gray code.

Then, utilize below equation that the brightness of each pixel is projected to different bit planes by bit, thus obtaining 8 luminance bit planes, as shown in Figure 2:

Wherein, BPGC is luminance bit plane, and i and j represents the row and column of image respectively, and k represents the sequence number of bit plane, k=0,1,2 ..., 7.

Finally, below equation is utilized to carry out Gaussian smoothing:

Wherein, M₁For brightness display model,Be average it is μ_s, standard deviation is σ_s2 dimension gaussian kernel,Be average it is μ_f, standard deviation is σ_f1 dimension gaussian kernel, * represents convolution algorithm.

2. texture appearance model is set up

First, local binary patterns (LocalBinaryPattern, the LBP) feature of target image is sought.Adopt the LBP operator of 3 × 3 traditional sizes, the LBP value centered by all non-edge pixels is calculated for target image.

Secondly, utilize below equation to represent the LBP value of each pixel, namely represent the LBP value of each pixel by natural binary sequence:

Wherein, (i j) represents the LBP value of the pixel of the i-th row jth row, a ' to I '_{I, j, k}Represent the value of kth bit when the LBP value natural binary sequence of this pixel is represented.

Then, utilize below equation that LBP value is converted into binary gray code to represent:

Wherein, b '_{I, j, k}Represent the value of kth bit when LBP value binary gray code is represented.

Then, utilize below equation that the LBP value of each pixel is projected to different bit planes by bit, thus obtaining 8 texture bits planes, as shown in Figure 3:

Wherein,For texture bits plane, i and j represents the row and column of image respectively, and k represents the sequence number of bit plane, k=0,1,2 ..., 7.

Finally, below equation is utilized to carry out Gaussian smoothing, thus obtaining texture appearance model M₂:

Wherein, M₂For texture appearance model,Be average it is μ_s, standard deviation is σ_s2 dimension gaussian kernel,Be average it is μ_f, standard deviation is σ_f1 dimension gaussian kernel, * represents convolution algorithm.

(3) determine region of search in the next frame, to its ask for respectively smooth after luminance bit plane and local binary patterns bit plane, and search for two display models closest to region as tracking target.

1. region of search is determined

With previous frame target positionCenter be the center of circle, with r for radius, the rectangle in this region is dropped on as candidate target in all centers, adopts below equation to represent:

Wherein,For target position in previous frame, l (x) represents image block x position in present frame, and r is search radius, and X represents the set of all candidate image blocks, | | | | represent Euclidean distance.

2. the luminance bit plane after asking candidate region smooth

First, utilize below equation to represent the brightness of each pixel in candidate region, namely represent the brightness of each pixel by natural binary sequence:

Wherein, (i j) represents the pixel intensity of the i-th row jth row, a in candidate region to I_i,j,kRepresent the value of kth bit when this pixel intensity natural binary sequence is represented.

Secondly, below equation is utilized to represent for binary gray code Sequence Transformed for natural binary:

Wherein, b_i,j,kRepresent I (i, the value of kth bit when j) representing with binary gray code.

Then, utilize below equation that the brightness of pixel each in candidate region is projected to different bit planes by bit, thus obtaining 8 luminance bit planes:

Wherein, BPGC is luminance bit plane, and i and j represents the row and column of image respectively, and k represents the sequence number of bit plane, k=0,1,2 ..., 7.

Finally, below equation is utilized to carry out Gaussian smoothing:

Thus obtain candidate region smooth after luminance bit plane C₁。

3. the local binary patterns bit plane after asking candidate region smooth

First, the local binary patterns feature of target image is sought.Adopt the LBP operator of 3 × 3 traditional sizes, the LBP value centered by all non-edge pixels is calculated for target image.

Then, utilize below equation to represent the LBP value of each pixel in candidate region, namely represent the LBP value of each pixel by natural binary sequence:

Wherein, (i j) represents the LBP value of the i-th row jth row, a ' in candidate region to I '_{I, j, k}Represent the value of kth bit when this value natural binary sequence is represented.

Then, utilize below equation that LBP value is converted into binary gray code to represent:

Wherein, b '_{I, j, k}Represent the value of kth bit when LBP value binary gray code is represented.

Then, utilize below equation that the LBP value of pixel each in candidate region is projected to different bit planes by bit, thus obtaining 8 texture bits planes:

Wherein,For texture bits plane, i and j represents the row and column of image respectively, and k represents the sequence number of bit plane, k=0,1,2 ..., 7.

Finally, below equation is utilized to carry out Gaussian smoothing:

Obtain LBP bit plane C₂。

According to above step, obtain candidate region smooth after luminance bit plane C₁With LBP bit plane C₂。

4. target location is determined

Find with two display models closest to region as following the tracks of order calibration method be: try to achieve the region x ' making following formula obtain minima:

Wherein,Represent target position in present frame, M₁For brightness display model, M₂For texture appearance model, C₁For smoothed luminance bit plane, C₂For smoothed LBP bit plane, w is weights and 0 < w < 1, Dist () is the distance between two bit planes, tries to achieve as follows:

Wherein, BP₁And BP₂Respectively two bit planes, i, j and k represents line number, row number and bit plane sequence number respectively.

(4) according to the moving target in the display model set up and present frame, brightness display model and texture appearance model are updated according to renewal rate set in advance.

Wherein, λ is renewal rate and 0 < λ < 1, M₁For brightness display model, M₂For texture appearance model, C₁(x ') luminance bit plane corresponding to target area in present frame, C₂(x ') texture bits plane corresponding to target area in present frame.

(5) when all frame of video are disposed, then stop calculating；Otherwise, jump procedure (3).

Moving Target Tracking Algorithm based on bit plane describes as follows:

Input: video sequence V, and the position that target is in the first frame

Output: the target location of each frame in video sequence

Step:

(1) initialized target display model

(2) forf=2to | V | do

(3) determine hunting zone, obtain image collection

(4)Calculate the luminance bit plane of each image block $C_1\left(x\right)={\mathrm{BPGC}}_{Indensity}\left(x\right)*h_{{\mathrm{\&mu;}}_s,{\mathrm{\&sigma;}}_s}*h_{{\mathrm{\&mu;}}_f,{\mathrm{\&sigma;}}_f}$ With LBP bit plane

(5) target location is determined,

(6) the luminance bit plane C of target is calculated₁(x ') and LBP bit plane C₂(x′)

(7) display model, M are updated₁=λ M₁+(1-λ)C₁(x ') and M₂=λ M₂+(1-λ)C₂(x′)

(8)endfor

Embodiment:

In order to assess the performance of the present invention, 8 video sequences provided such as Babenko are tested.These video sequences contain partial occlusion, target deformation, illumination variation, change in size, rapid movement, similar object interference etc., it is respectively compared currently for the good four kinds of algorithms of above-mentioned video sequence tracking effect as a comparison, respectively: the multi-instance learning tracking (WMIL) based on loop structure detecting and tracking method (CSK) of core, online AdaBoost tracking (OAB), multi-instance learning tracking (MIL) and weighting compares from aspects such as mean error, tracking success rates respectively.Method in the present invention is in XP operating system, adopts Matlab7.0.1 programming realization, and allocation of computer is double-core 2.93GHzCPU and 2GB internal memory.

Parameter is arranged:

Code and its parameter provided in article of author's issue are provided with the algorithm compared.Method parameter in the present invention is set to: search radius r=30, model modification speed λ=0.95 (except λ=0.85 in Cliffbar and Dollar), weight w=0.5, and 2 dimension gaussian kernel are sized to 9*9 and its standard deviation sigma_s=1,1 dimension gaussian kernel is sized to 5*1 and its standard deviation sigma_f=0.625.

Quantitative analysis:

Algorithm keeps track position is adopted to compare tracking five kinds different from the off-centring of actual position distance (see table 1) and tracking success rate (see table 2).Wherein, errors of centration reflects the distance of tracing positional and actual position, and the more little explanation tracking error of this value is more little, closer to target actual position, follows the tracks of result more accurate；Otherwise, illustrating to follow the tracks of result deviation actual position more remote, tracking effect is more poor.Tracking success rate is defined as: if following the tracks of the coincidence factor of rectangle frame and actual position rectangle frame > 50%, then it is assumed that follow the tracks of successfully；Otherwise, then it is assumed that failure.The more big explanation tracking effect of this value is more good, otherwise, illustrate that tracking effect is more poor.

Table 1 follows the tracks of result and actual position centre distance mean error (in units of pixel)

Success rate (%) followed the tracks of by table 2

Illustrate: table 1 and in table 2 runic represent best result, italic represents second-best result.

By table 1 and table 2 it can be seen that test video is had and follows the tracks of result preferably by method in the present invention.In five kinds of methods, the average position error of the present invention is minimum and tracking success rate is the highest, and this also reflects the Stability and veracity of method in the present invention.Fig. 4-1 to Fig. 4-8 is the error curve diagram (in units of pixel) between the tracking result of five kinds of methods and target actual position.

Qualitative analysis:

Fig. 5 illustrates the tracking effect contrast of five kinds of method partial frames in 8 video sequences.

This video sequence of Dollar contains the deformation of target and the interference of similar object.Method in the present invention and CSK method and actual position closest to, tracking effect is the most accurate.

There is relatively long time and large-scale partial occlusion in OccludedFace and OccludedFace2 the two video, the target in OccludedFace2 has rotatable head and puts on a hat, and this both increases the difficulty of motion target tracking.In five kinds of trackings, the method tracking effect in the present invention is best.

DavidIndoor contains illumination, target sizes and the change such as attitude, outward appearance.Other four kinds of methods all occur in that drift in various degree, and the method in the present invention shows good stability and accuracy.

Cliffbar contains the interference of the fuzzy and similar background caused because of quick priming.It can be seen that CSK tracking effect is best, the method in the present invention is taken second place, and other three kinds of methods are performed poor.

Occurring in that the rapid movement of target, cosmetic variation, rotation and partial occlusion etc. in CokeCan, both increase the difficulty of tracking, the method in the present invention shows follows the tracks of result preferably.

Twinings contains and rotates because of target 360 degree and move caused cosmetic variation and dimensional variation.In five kinds of methods, the method in the present invention and target actual position closest to.

Girl contain by rotating the deformation and dimensional variation that cause, the interference of other target, and because the outward appearance caused of moving changes completely, this both increases the difficulty of tracking, five kinds of trackings all occur in that error in various degree, and wherein the tracking effect of WMIL is relatively stable.

Generally speaking, illumination condition can be overcome to change based on the track algorithm of bit plane, impact that object pose change and outward appearance significantly change etc. bring, in five kinds of algorithms, show and follow the tracks of Stability and veracity preferably.

Claims (9)

1. the motion target tracking method based on bit plane, it is characterised in that comprise the steps:

Step (1) is selected target of following the tracks of in video the first frame, and hand labeled target location；

Step (2), to target area image, asks for its luminance bit plane and local binary patterns bit plane, then carries out Gaussian smoothing, sets up brightness display model and texture appearance model respectively；

Step (3) determines region of search in the next frame, to its ask for respectively smooth after luminance bit plane and local binary patterns bit plane, and search for the brightness display model set up in step (2) and texture appearance model closest to region as tracking target；

Step (4), according to the tracking result in the display model set up and present frame, updates brightness display model and texture appearance model according to renewal rate set in advance；

Step (5) is disposed when all frame of video, then stop calculating；Otherwise, jump procedure (3).

2. a kind of motion target tracking method based on bit plane according to claim 1, it is characterized in that: in described step (1), the method for hand labeled target location is: follow the tracks of target with rectangle frame is selected, and record the two-dimensional coordinate (x in this rectangle frame upper left corner, y), and the width of rectangle frame and height.

3. a kind of motion target tracking method based on bit plane according to claim 1, it is characterised in that: the method setting up brightness display model in described step (2) is:

1. utilize below equation to represent the brightness of each pixel, namely represent the brightness of each pixel by natural binary sequence:

Wherein, (i j) represents the pixel intensity of the i-th row jth row, a to I_i,j,kRepresent the value of kth bit when this pixel intensity natural binary sequence is represented；

2. below equation is utilized to represent for binary gray code Sequence Transformed for natural binary:

Wherein, b_i,j,kRepresent I (i, the value of kth bit when j) representing with binary gray code；

3. utilize below equation that the brightness of each pixel is projected to different bit planes by bit, thus obtaining 8 luminance bit planes:

Wherein, BPGC is luminance bit plane, and i and j represents the row and column of image respectively, and k represents the sequence number of bit plane, k=0,1,2 ..., 7；

4. below equation is utilized to carry out Gaussian smoothing:

Wherein, M₁For brightness display model,Be average it is μ_s, standard deviation is σ_s2 dimension gaussian kernel,Be average it is μ_f, standard deviation is σ_f1 dimension gaussian kernel, * represents convolution algorithm.

4. a kind of motion target tracking method based on bit plane according to claim 1, it is characterised in that: the method setting up texture appearance model in described step (2) is:

1. seek the local binary patterns feature of target image, adopt the LBP operator of 3 × 3 traditional sizes, the LBP value centered by all non-edge pixels is calculated for target image；

2. utilize below equation to represent the LBP value of each pixel, namely represent the LBP value of each pixel by natural binary sequence:

Wherein, I ' (i, j) represents the LBP value of the pixel of the i-th row jth row,Represent the value of kth bit when the LBP value natural binary sequence of this pixel is represented；

3. utilize below equation that LBP value is converted into binary gray code to represent:

Wherein, b '_{I, j, k}Represent the value of kth bit when LBP value binary gray code is represented；

4. utilize below equation that the LBP value of each pixel is projected to different bit planes by bit, thus obtaining 8 texture bits planes:

Wherein, BPGC ' is texture bits plane, and i and j represents the row and column of image respectively, and k represents the sequence number of bit plane, k=0,1,2 ..., 7；

5. below equation is utilized to carry out Gaussian smoothing, thus obtaining texture appearance model M₂:

Wherein, M₂For texture appearance model,Be average it is μ_s, standard deviation is σ_s2 dimension gaussian kernel,Be average it is μ_f, standard deviation is σ_f1 dimension gaussian kernel, * represents convolution algorithm.

5. a kind of motion target tracking method based on bit plane according to claim 1, it is characterised in that: the method determining region of search in described step (3) is:

With previous frame target positionCenter be the center of circle, with r for radius, the rectangle in this region is dropped on as candidate target in all centers, adopts below equation to represent:

Wherein,For target position in previous frame, l (x) represents image block x position in present frame, and r is search radius, and X represents the set of all candidate image blocks, | | | | represent Euclidean distance.

6. a kind of motion target tracking method based on bit plane according to claim 1, it is characterised in that: described step (3) is tried to achieve candidate region smooth after luminance bit plane C₁Method be:

1. utilize below equation to represent the brightness of each pixel in candidate region, namely represent the brightness of each pixel by natural binary sequence:

Wherein, (i j) represents the pixel intensity of the i-th row jth row, a in candidate region to I_i,j,kRepresent the value of kth bit when this pixel intensity natural binary sequence is represented；

2. below equation is utilized to represent for binary gray code Sequence Transformed for natural binary:

Wherein, b_i,j,kRepresent I (i, the value of kth bit when j) representing with binary gray code；

3. utilize below equation that the brightness of pixel each in candidate region is projected to different bit planes by bit, thus obtaining 8 luminance bit planes:

Wherein, BPGC is luminance bit plane, and i and j represents the row and column of image respectively, and k represents the sequence number of bit plane, k=0,1,2 ..., 7；

4. below equation is utilized to carry out Gaussian smoothing:

Thus obtain candidate region smooth after luminance bit plane C₁。

7. a kind of motion target tracking method based on bit plane according to claim 1, it is characterised in that: described step (3) is tried to achieve the local binary patterns bit plane C of candidate region₂Method be:

1. seek the LBP feature of image in candidate region: adopt the LBP operator of 3 × 3 traditional sizes, the LBP value centered by all non-edge pixels is calculated for candidate region image；

2. utilize below equation to represent the LBP value of each pixel in candidate region, namely represent the LBP value of each pixel by natural binary sequence:

Wherein, (i j) represents the LBP value of the i-th row jth row, a ' in candidate region to I '_{I, j, k}Represent the value of kth bit when this value natural binary sequence is represented；

3. utilize below equation that LBP value is converted into binary gray code to represent:

Wherein, b '_{I, j, k}Represent the value of kth bit when LBP value binary gray code is represented；

4. utilize below equation that the LBP value of pixel each in candidate region is projected to different bit planes by bit, thus obtaining 8 texture bits planes:

Wherein,For texture bits plane, i and j represents the row and column of image respectively, and k represents the sequence number of bit plane, k=0,1,2 ..., 7；

5. below equation is utilized to carry out Gaussian smoothing:

Obtain LBP bit plane C₂。

8. a kind of motion target tracking method based on bit plane according to claim 1, it is characterised in that: the brightness display model set up in search and step (2) in described step (3) and texture appearance model closest to region as following the tracks of order calibration method be: try to achieve the region x ' making following formula obtain minima:

Wherein,Represent target position in present frame, M₁For brightness display model, M₂For texture appearance model, C₁For smoothed luminance bit plane, C₂For smoothed LBP bit plane, w is weights and 0 < w < 1, Dist () is the distance between two bit planes, tries to achieve as follows:

Wherein, BP₁And BP₂Respectively two bit planes, i, j and k represents line number, row number and bit plane sequence number respectively.

9. a kind of motion target tracking method based on bit plane according to claim 1, it is characterised in that: the method updating brightness display model and texture appearance model in described step (4) is:

Wherein, λ is renewal rate and 0 < λ < 1, M₁For brightness display model, M₂For texture appearance model, C₁(x ') luminance bit plane corresponding to target area in present frame, C₂(x ') texture bits plane corresponding to target area in present frame.