CN105139426A - Video moving object detection method based on non-down-sampling wavelet transformation and LBP - Google Patents

Abstract

The invention provides a video moving object detection method based on non-down-sampling wavelet transformation and an LBP. The method comprises the steps of: establishing an initial background image; carrying out non-down-sampling wavelet transformation to obtain wavelet transformation sub-images; calculating an LBP binary characteristic vector of each pixel in each wavelet transformation sub-image; dividing each wavelet transformation sub-image into a non-overlap pixel block with N*N pixels, and calculating texture differences of corresponding pixel blocks of the background image and a current frame image; according to a set threshold, judging whether each non-overlap pixel block is a foreground block or a background block; and using the pixels marked as the background block for background updating so as to generate a background image for the moving object detection of a next frame. According to the invention, the non-down-sampling wavelet transformation and the LBP are combined, so that the translation change occurs in the wavelet transformation process is effectively avoided, the calculation is simple, the operation speed is high, and the application prospect is relatively good.

Description

A kind of video moving object detection method based on undecimated wavelet transform and LBP

Technical field

The present invention relates to a kind of detection method of video frequency motion target, especially a kind of detection method of the video frequency motion target combined based on undecimated wavelet transform and LBP.

Background technology

So-called video frequency motion target detects and refers to from the video sequence image of Still Camera shooting, is split by Motion-changed Region and extracts, namely detect foreground moving object from the background of video frame images.Video moving object subdivision is the key link of many computer vision systems, is widely used in many computer vision systems, as the three-dimensionalreconstruction, complex behavior identification etc. of moving object tracking, free viewpoint video.Therefore, many researchers are devoted to the research of video foreground moving object detection, and most of algorithm uses the color of pixel and strength information to detect foreground object, and this algorithm has the video of Similar color for foreground object and background, seems helpless.Therefore, some researcher's texture feature extractions (as LBP (LocalBinaryPattern)) are to improve the detection perform of foreground object.But, when foreground object and the background area residing for it all obey be uniformly distributed time, LBP can not detect prospect well, and LBP can not extract tiny smooth textural characteristics.Although existing algorithm uses the information such as color, intensity, texture, gradient, edge to detect moving target.But these algorithms accurately cannot detect with video background color similarity and have the foreground object of tiny smooth texture, because only utilize the color of original image and texture (as LBP) feature can not distinguish foreground object and background.

LBP texture description algorithm based on wavelet transformation is applied to the classification of radioscopic image, recognition of face and personal recognition.These algorithms are all extract LBP histogram as textural characteristics to the frequency subgraph of down-sampling wavelet transformation, because down-sampling wavelet transformation does not have translation invariance, therefore, cause being permitted great wavelet coefficient by multiple wavelet transformation layer.

Summary of the invention

The technical problem to be solved in the present invention is that existing video foreground moving object detection adopts down-sampling wavelet transformation to there will be translation change, and computation process is complicated, travelling speed is slow.

In order to solve the problems of the technologies described above, the invention provides 1, a kind of video moving object detection method based on undecimated wavelet transform and LBP, it is characterized in that, comprise the steps:

Step 1, the image utilizing initial a few frame of video or tens frames not to comprise foreground moving object sets up initial background image;

Step 2, carries out non-lower sampling L layer wavelet transformation respectively to initial background image and present image, obtains the wavelet transformation subgraph of background image and the wavelet transformation subgraph of present image respectively;

Step 3, calculates the LBP vector of binary features of each pixel in each wavelet transformation subgraph;

Step 4, is divided into the non-overlapped block of pixels of N × N pixel by each wavelet transformation subgraph, calculate the texture difference of the respective pixel block of background image and current frame image, concrete steps are:

Step 4.1, for background image and current frame image, calculates the background pixel of same coordinate position of the equidirectional frequency subgraph of same wavelet decomposition layer and the Hamming distances of the LBP vector of binary features of current frame pixel;

Step 4.2, for the pixel of each coordinate position, at the corresponding Hamming distances of each frequency subgraph, calculate the Hamming distances of Hamming distances average as the pixel of this coordinate position of M frequency subgraph, thus obtain the matrix that is element with pixel Hamming distances;

Step 4.3, matrix trace inequality step 4.2 obtained becomes the non-overlapped block of pixels of N × N pixel, and in compute matrix, the average of all elements is as the Hamming distances of this non-overlapped block of pixels;

Step 5, judge whether each non-overlapped block of pixels meets foreground blocks Rule of judgment, if the Hamming distances of current non-overlapped block of pixels is more than or equal to the distance threshold of setting and the number of pixels of current background pixel is greater than the quantity threshold of setting, then marking this non-overlapped block of pixels is foreground blocks, otherwise to mark this non-overlapped block of pixels be background block;

Step 6, is used for context update by the pixel being labeled as background block, to be the moving object detection generation background image of next frame.

Adopt undecimated wavelet transform to replace existing down-sampling wavelet transformation, there is the advantage of the translation invariance of conversion, be more suitable for the accurate detection of Local textural feature; LBP scale-of-two textural characteristics describing method is adopted to replace existing LBP histogram textural characteristics describing method, utilize the binary number of LBP as textural characteristics, and the tiny texture difference in local of Pixel-level is measured with Hamming distances, divide block of pixels again, the textural characteristics of statistical pixel block level, measured the difference of background texture and prospect texture by fine granularity to coarseness step by step, not only reduce the complexity of calculating, and substantially increase arithmetic speed and precision.

As further restriction scheme of the present utility model, background image method for building up initial in step 1 is the pixel value of pixel grey scale average as initial background image of the initial a few frame of calculating or tens frames, thus sets up initial background image.Adopt gray average as the pixel value of initial background image, fast and easy can calculate the suitable background image of acquisition.

As further restriction scheme of the present utility model, in step 2, non-lower sampling L layer wavelet transformation is carried out respectively to initial background image and present image, obtain the wavelet transformation subgraph of background image and the wavelet transformation subgraph of present image respectively, wherein, if I _tbe t current frame image, B _tfor its background image, represent current frame image I _tthe low channel subgraph of undecimated wavelet transform and hf channel subgraph; S=1,2, K, S, represent the number of plies of wavelet transformation, in video moving object detection method, make S=3; O=1,2,3,4, represent the level of low channel and hf channel, vertical and direction, three, diagonal angle respectively successively.

As further restriction scheme of the present utility model, in described step 2, the span of L is 2 ~ 4; In step 4.2, the span of M is 7 ~ 13; When L is 2, M is 7; When L is 3, M is 10; When L is 4, M is 13.Adopt the Hamming distances average of 7 ~ 13 frequency subgraphs as the Hamming distances of the pixel of this coordinate position, not only increase computational accuracy, ensure that operation efficiency, and the generation of incident can be avoided, there is higher reliability.

As further restriction scheme of the present utility model, in step 4, the value of N is 16, each wavelet transformation subgraph is divided into the non-overlapped block of pixels of 16 × 16 pixels, calculates the texture difference of the respective pixel block of background image and current frame image:

Step 4.1, for background image and current frame image, calculates the background pixel of same coordinate position of the equidirectional frequency subgraph of same wavelet decomposition layer and the Hamming distances of the LBP vector of binary features of current frame pixel;

In step 4.2, the value of M is 10, for the pixel of each coordinate position, at the corresponding Hamming distances of each frequency subgraph, calculate the Hamming distances of Hamming distances average as the pixel of this coordinate position of 10 frequency subgraphs, thus obtain the matrix that is element with pixel Hamming distances, if LBP _m(i, j, I _t) be current frame image I _tthe coordinate of m wavelet transformation subgraph be the LBP proper vector of the pixel of (i, j), LBP _m(i, j, B _t) be current background image B _tthe coordinate of m wavelet transformation subgraph be the LBP proper vector of the pixel of (i, j), LBP calculates 8 neighborhoods adopting pixel, and therefore LBP vector of binary features is 8 dimensions, order ${\mathrm{LBP}}_m(i,j,I_t)=\{{\mathrm{LBP}}_m^I\left(1\right),{\mathrm{LBP}}_m^I\left(2\right),L{\mathrm{LBP}}_m^I\left(8\right)\},$ With season ${\mathrm{LBP}}_m(i,j,B_t)=\{{\mathrm{LBP}}_m^B\left(1\right),{\mathrm{LBP}}_m^B\left(2\right),L{\mathrm{LBP}}_m^B\left(8\right)\},$ Then current frame image I _twith background image B _tlBP vector of binary features LBP (i, j, I _t) and LBP (i, j, B _t) Hamming distances be:

$HM(i,j,I_t)=\underset{m-1}{\overset{10}{\Σ}}\underset{k=1}{\overset{8}{\Σ}}\mathrm{\δ}\left(k\right)---\left(1\right)$

In formula,

$\mathrm{\δ}\left(k\right)=\left\{\begin{array}{cc}1& {\mathrm{LBP}}_m^I\left(k\right)\≠{\mathrm{LBP}}_m^B\left(k\right)\\ 0& {\mathrm{LBP}}_m^I\left(k\right)={\mathrm{LBP}}_m^B\left(k\right)\end{array}\right.,k=1,2,L8---\left(2\right)$

Therefore, the matrix H M (t) that is element with pixel Hamming distances is obtained;

Step 4.3, the matrix H M (t) step 4.2 obtained is divided into the non-overlapped block of pixels of 16 × 16 pixels, and in compute matrix, the average of all elements is as the Hamming distances of this non-overlapped block of pixels:

$HM(I_t,n)=\frac{1}{256}\underset{j=1}{\overset{16}{\Σ}}\underset{i=1}{\overset{16}{\Σ}}HM(i,j,I_t)---\left(5\right)$

In formula, which non-overlapped block of pixels n represents.

As further restriction scheme of the present utility model, in step 5, judge whether each non-overlapped block of pixels meets foreground blocks Rule of judgment, if the Hamming distances of current non-overlapped block of pixels is more than or equal to the distance threshold of setting and the number of pixels of current background pixel is greater than the quantity threshold λ of setting ₂, then marking this non-overlapped block of pixels is foreground blocks, otherwise to mark this non-overlapped block of pixels be background block, for each element HM (i, j, the I of matrix H M (t) _t), the average with following formula coordinates computed position (i, j):

$\stackrel{\‾}{HM}(i,j)=\frac{1}{t-1}\underset{k=1}{\overset{t-1}{\Σ}}HM(i,j,I_k)---\left(3\right)$

If pixel (i, j) meets following condition, then this pixel is judged as background pixel:

$\begin{array}{ccc}HM(i,j,I_t)\≥{\mathrm{\α}}_1\stackrel{\‾}{HM}(i,j)& and& HM(i,j,I_t)>{\mathrm{\λ}}_1\end{array}---\left(4\right)$

In formula, α ₁and λ ₁for adjustment parameter, be set to 1.5≤α ₁≤ 2.5 and 14≤λ ₁≤ 20;

For current frame image I _thamming distance matrix HM (t) each 16 × 16 block of pixels, calculate the average of t-1 block of pixels before same position with following formula:

$\stackrel{\‾}{HM}(I_t,n)=\frac{1}{t-1}\underset{k=1}{\overset{t-1}{\Σ}}HM(I_k,n)---\left(6\right)$

If current block meets following condition, then this block of pixels is judged as background block:

$\begin{array}{ccc}HM(I_t,n)\≥{\mathrm{\α}}_2\stackrel{\‾}{HM}(I_t,n)& and& NP\_B\left(n\right)\≥{\mathrm{\λ}}_2\end{array}---\left(7\right)$

In formula, NP_B (n) is for being judged as the number of pixels of background, α in current block ₂and λ ₂for adjustment parameter, be set to 1.1≤α ₂≤ 1.3 and 60≤λ ₂≤ 90.

Beneficial effect of the present invention is: (1) adopts undecimated wavelet transform to replace existing down-sampling wavelet transformation, has the advantage of the translation invariance of conversion, is more suitable for the accurate detection of Local textural feature; (2) LBP scale-of-two textural characteristics describing method is adopted to replace existing LBP histogram textural characteristics describing method, utilize the binary number of LBP as textural characteristics, and the tiny texture difference in local of Pixel-level is measured with Hamming distances, divide block of pixels again, the textural characteristics of statistical pixel block level, measured the difference of background texture and prospect texture by fine granularity to coarseness step by step, not only reduce the complexity of calculating, and substantially increase arithmetic speed and precision.

Accompanying drawing explanation

Fig. 1 is method flow diagram of the present invention;

Fig. 2 is the structural drawing of undecimated wavelet transform of the present invention.

Embodiment

As shown in Figure 1, a kind of video moving object detection method based on undecimated wavelet transform and LBP of the present invention, its video is compressed video or the non-compression video of using camera collection, initial a few frame or tens two field pictures of this video should be the background frames not comprising foreground moving object, in order to set up initial background image, comprise the steps:

Step 1, the image utilizing initial a few frame of video or tens frames not to comprise foreground moving object sets up initial background image, initial background image method for building up is the pixel value of pixel grey scale average as initial background image of the initial a few frame of calculating or tens frames, thus sets up initial background image;

Step 2, carries out non-lower sampling L layer wavelet transformation respectively to initial background image and present image, obtains the wavelet transformation subgraph of background image and the wavelet transformation subgraph of present image respectively, wherein, if I _tbe t current frame image, B _tfor its background image, represent current frame image I _tthe low channel subgraph of undecimated wavelet transform and hf channel subgraph; S=1,2, K, S, represent the number of plies of wavelet transformation, in the present invention, make S=3; O=1,2,3,4, represent the level (o=2) of low channel (o=1) and hf channel, vertical (o=3) and diagonal angle (o=4) three directions respectively successively, the structural drawing of undecimated wavelet transform as shown in Figure 2;

Step 3, calculates each wavelet transformation subgraph in LBP (LocalBinaryPattern) vector of binary features of each pixel;

Step 4, is divided into the non-overlapped block of pixels of 16 × 16 pixels by each wavelet transformation subgraph, calculate the texture difference of the respective pixel block of background image and current frame image, concrete steps are:

Step 4.1, for background image and current frame image, calculates the background pixel of same coordinate position of the equidirectional frequency subgraph of same wavelet decomposition layer and the Hamming distances of the LBP vector of binary features of current frame pixel;

Step 4.2, for the pixel of each coordinate position, at the corresponding Hamming distances of each frequency subgraph, the value of M is 10, calculate the Hamming distances of Hamming distances average as the pixel of this coordinate position of 10 frequency subgraphs, thus obtain the matrix that is element with pixel Hamming distances, setting LBP _m(i, j, I _t) be current frame image I _tthe coordinate of m wavelet transformation subgraph be the LBP proper vector of the pixel of (i, j), LBP _m(i, j, B _t) be current background image B _tthe coordinate of m wavelet transformation subgraph be the LBP proper vector of the pixel of (i, j), LBP calculates 8 neighborhoods adopting pixel, and therefore LBP vector of binary features is 8 dimensions, order ${\mathrm{LBP}}_m(i,j,I_t)=\{{\mathrm{LBP}}_m^I\left(1\right),{\mathrm{LBP}}_m^I\left(2\right),L{\mathrm{LBP}}_m^I\left(8\right)\},$ With season ${\mathrm{LBP}}_m(i,j,B_t)=\{{\mathrm{LBP}}_m^B\left(1\right),{\mathrm{LBP}}_m^B\left(2\right),L{\mathrm{LBP}}_m^B\left(8\right)\},$ Then current frame image I _twith background image B _tlBP vector of binary features LBP (i, j, I _t) and LBP (i, j, B _t) Hamming distances be:

$HM(i,j,I_t)=\underset{m-1}{\overset{10}{\Σ}}\underset{k=1}{\overset{8}{\Σ}}\mathrm{\δ}\left(k\right)---\left(1\right)$

In formula,

$\mathrm{\δ}\left(k\right)=\left\{\begin{array}{cc}1& {\mathrm{LBP}}_m^I\left(k\right)\≠{\mathrm{LBP}}_m^B\left(k\right)\\ 0& {\mathrm{LBP}}_m^I\left(k\right)={\mathrm{LBP}}_m^B\left(k\right)\end{array}\right.,k=1,2,L8---\left(2\right)$

Therefore, the matrix H M (t) that is element with pixel Hamming distances is obtained;

Step 4.3, the matrix H M (t) step 4.2 obtained is divided into the non-overlapped block of pixels of 16 × 16 pixels, in compute matrix, the average of all elements is as the Hamming distances of this non-overlapped block of pixels, with the Hamming distances of formula (5) calculating the n-th non-overlapped block of pixels is:

$HM(I_t,n)=\frac{1}{256}\underset{j=1}{\overset{16}{\Σ}}\underset{i=1}{\overset{16}{\Σ}}HM(i,j,I_t)---\left(5\right)$

Step 5, judges whether each non-overlapped block of pixels meets foreground blocks Rule of judgment, if the Hamming distances of current non-overlapped block of pixels is more than or equal to the distance threshold of setting and the number of pixels of current background pixel is greater than the quantity threshold λ of setting ₂, then marking this non-overlapped block of pixels is foreground blocks, otherwise to mark this non-overlapped block of pixels be background block, for each element HM (i, j, the I of matrix H M (t) _t), the average with following formula coordinates computed position (i, j):

$\stackrel{\‾}{HM}(i,j)=\frac{1}{t-1}\underset{k=1}{\overset{t-1}{\Σ}}HM(i,j,I_k)---\left(3\right)$

If pixel (i, j) meets following condition, then this pixel is judged as background pixel:

$\begin{array}{ccc}HM(i,j,I_t)\≥{\mathrm{\α}}_1\stackrel{\‾}{HM}(i,j)& and& HM(i,j,I_t)>{\mathrm{\λ}}_1\end{array}---\left(4\right)$

In formula, α ₁and λ ₁for adjustment parameter, be set to 1.5≤α ₁≤ 2.5 and 14≤λ ₁≤ 20;

For current frame image I _thamming distance matrix HM (t) each 16 × 16 block of pixels, calculate the average of t-1 block of pixels before same position with following formula:

$\stackrel{\‾}{HM}(I_t,n)=\frac{1}{t-1}\underset{k=1}{\overset{t-1}{\Σ}}HM(I_k,n)---\left(6\right)$

If current block meets following condition, then this block of pixels is judged as background block:

$\begin{array}{ccc}HM(I_t,n)\≥{\mathrm{\α}}_2\stackrel{\‾}{HM}(I_t,n)& and& NP\_B\left(n\right)\≥{\mathrm{\λ}}_2\end{array}---\left(7\right)$

In formula, NP_B (n) is for being judged as the number of pixels of background, α in current block ₂and λ ₂for the adjustment parameter of setting, be set to 1.1≤α ₂≤ 1.3 and 60≤λ ₂≤ 90;

Step 6, the pixel being labeled as background block is used for upgrading background image, to be the moving object detection generation background image of next frame, if a block of pixels is judged as background block, pixel then in this block of pixels just thinks background pixel, therefore the pixel in this block of pixels is used as context update, and more new formula is:

B _t+1(x)＝β·B _t(x)+(1-β)I _t(x)(8)

In formula, β is context update parameter, and establishes 0.0001≤β≤0.2.

Claims (6)

1., based on a video moving object detection method of undecimated wavelet transform and LBP, it is characterized in that, comprise the steps:

Step 1, the image utilizing initial a few frame of video or tens frames not to comprise foreground moving object sets up initial background image;

Step 2, carries out non-lower sampling L layer wavelet transformation respectively to initial background image and present image, obtains the wavelet transformation subgraph of background image and the wavelet transformation subgraph of present image respectively;

Step 3, calculates the LBP vector of binary features of each pixel in each wavelet transformation subgraph;

Step 4, is divided into the non-overlapped block of pixels of N × N pixel by each wavelet transformation subgraph, calculate the texture difference of the respective pixel block of background image and current frame image, concrete steps are:

Step 4.1, for background image and current frame image, calculates the background pixel of same coordinate position of the equidirectional frequency subgraph of same wavelet decomposition layer and the Hamming distances of the LBP vector of binary features of current frame pixel;

Step 4.2, for the pixel of each coordinate position, at the corresponding Hamming distances of each frequency subgraph, calculate the Hamming distances of Hamming distances average as the pixel of this coordinate position of M frequency subgraph, thus obtain the matrix that is element with pixel Hamming distances;

Step 4.3, matrix trace inequality step 4.2 obtained becomes the non-overlapped block of pixels of N × N pixel, and in compute matrix, the average of all elements is as the Hamming distances of this non-overlapped block of pixels;

Step 5, judge whether each non-overlapped block of pixels meets foreground blocks Rule of judgment, if the Hamming distances of current non-overlapped block of pixels is more than or equal to the distance threshold of setting and the number of pixels of current background pixel is greater than the quantity threshold of setting, then marking this non-overlapped block of pixels is foreground blocks, otherwise to mark this non-overlapped block of pixels be background block;

Step 6, is used for context update by the pixel being labeled as background block, to be the moving object detection generation background image of next frame.

2. the video moving object detection method based on undecimated wavelet transform and LBP according to claim 1, it is characterized in that, background image method for building up initial in described step 1 is the pixel value of pixel grey scale average as initial background image of the initial a few frame of calculating or tens frames, thus sets up initial background image.

3. the video moving object detection method based on undecimated wavelet transform and LBP according to claim 1, it is characterized in that, in described step 2, non-lower sampling L layer wavelet transformation is carried out respectively to initial background image and present image, obtain the wavelet transformation subgraph of background image and the wavelet transformation subgraph of present image respectively, wherein, if I _tbe t current frame image, B _tfor its background image, represent current frame image I _tthe low channel subgraph of undecimated wavelet transform and hf channel subgraph; S=1,2, K, S, represent the number of plies of wavelet transformation, in video moving object detection method, make S=3; O=1,2,3,4, represent the level of low channel and hf channel, vertical and direction, three, diagonal angle respectively successively.

4. the video moving object detection method based on undecimated wavelet transform and LBP according to claim 1 and 2, it is characterized in that, in described step 2, the span of L is 2 ~ 4; In step 4.2, the span of M is 7 ~ 13; When L is 2, M is 7; When L is 3, M is 10; When L is 4, M is 13.

5. the video moving object detection method based on undecimated wavelet transform and LBP according to claim 1 and 2, it is characterized in that, in described step 2, L is 3; In described step 4, the value of N is 16, each wavelet transformation subgraph is divided into the non-overlapped block of pixels of 16 × 16 pixels, calculates the texture difference of the respective pixel block of background image and current frame image:

Step 4.1, for background image and current frame image, calculates the background pixel of same coordinate position of the equidirectional frequency subgraph of same wavelet decomposition layer and the Hamming distances of the LBP vector of binary features of current frame pixel;

In step 4.2, the value of M is 10, for the pixel of each coordinate position, at the corresponding Hamming distances of each frequency subgraph, calculate the Hamming distances of Hamming distances average as the pixel of this coordinate position of 10 frequency subgraphs, thus obtain the matrix that is element with pixel Hamming distances, if LBP _m(i, j, I _t) be current frame image I _tthe coordinate of m wavelet transformation subgraph be the LBP proper vector of the pixel of (i, j), LBP _m(i, j, B _t) be current background image B _tthe coordinate of m wavelet transformation subgraph be the LBP proper vector of the pixel of (i, j), LBP calculates 8 neighborhoods adopting pixel, and therefore LBP vector of binary features is 8 dimensions, order ${\mathrm{LBP}}_m(i,j,I_t)=\{{\mathrm{LBP}}_m^I\left(1\right),{\mathrm{LBP}}_m^I\left(2\right),L{\mathrm{LBP}}_m^I\left(8\right)\},$ With season ${\mathrm{LBP}}_m(i,j,B_t)=\{{\mathrm{LBP}}_m^B\left(1\right),{\mathrm{LBP}}_m^B\left(2\right),L{\mathrm{LBP}}_m^B\left(8\right)\},$ Then current frame image I _twith background image B _tlBP vector of binary features LBP (i, j, I _t) and LBP (i, j, B _t) Hamming distances be:

$HM(i,j,I_t)=\underset{m-1}{\overset{10}{\Σ}}\underset{k=1}{\overset{8}{\Σ}}\mathrm{\δ}\left(k\right)---\left(1\right)$

In formula,

$\mathrm{\δ}\left(k\right)=\left\{\begin{array}{cc}1& {\mathrm{LBP}}_m^I\left(k\right)\≠{\mathrm{LBP}}_m^B\left(k\right)\\ 0& {\mathrm{LBP}}_m^I\left(k\right)={\mathrm{LBP}}_m^B\left(k\right)\end{array}\right.,k=1,2,L8---\left(2\right)$

Therefore, the matrix H M (t) that is element with pixel Hamming distances is obtained;

Step 4.3, the matrix H M (t) step 4.2 obtained is divided into the non-overlapped block of pixels of 16 × 16 pixels, and in compute matrix, the average of all elements is as the Hamming distances of this non-overlapped block of pixels:

$HM(I_t,n)=\frac{1}{256}\underset{j=1}{\overset{16}{\Σ}}\underset{i=1}{\overset{16}{\Σ}}HM(i,j,I_t)---\left(5\right)$

In formula, which non-overlapped block of pixels n represents.

6. the video moving object detection method based on undecimated wavelet transform and LBP according to claim 5, it is characterized in that, in described step 5, judge whether each non-overlapped block of pixels meets foreground blocks Rule of judgment, if the Hamming distances of current non-overlapped block of pixels is more than or equal to the distance threshold of setting and the number of pixels of current background pixel is greater than the quantity threshold λ of setting ₂, then marking this non-overlapped block of pixels is foreground blocks, otherwise to mark this non-overlapped block of pixels be background block, for each element HM (i, j, the I of matrix H M (t) _t), the average with following formula coordinates computed position (i, j):

$\stackrel{\‾}{HM}(i,j)=\frac{1}{t-1}\underset{k=1}{\overset{t-1}{\Σ}}HM(i,j,I_k)---\left(3\right)$

If pixel (i, j) meets following condition, then this pixel is judged as background pixel:

$\begin{array}{ccc}HM(i,j,I_t)\≥{\mathrm{\α}}_1\stackrel{\‾}{HM}(i,j)& and& HM(i,j,I_t)>{\mathrm{\λ}}_1\end{array}---\left(4\right)$

In formula, α ₁and λ ₁for adjustment parameter, be set to 1.5≤α ₁≤ 2.5 and 14≤λ ₁≤ 20;

For current frame image I _thamming distance matrix HM (t) each 16 × 16 block of pixels, calculate the average of t-1 block of pixels before same position with following formula:

$\stackrel{\‾}{HM}(I_t,n)=\frac{1}{t-1}\underset{k=1}{\overset{t-1}{\Σ}}HM(I_k,n)---\left(6\right)$

If current block meets following condition, then this block of pixels is judged as background block:

$\begin{array}{ccc}HM(I_t,n)\≥{\mathrm{\α}}_2\stackrel{\‾}{HM}(I_t,n)& and& NP\_B\left(n\right)\≥{\mathrm{\λ}}_2\end{array}---\left(7\right)$

In formula, NP_B (n) is for being judged as the number of pixels of background, α in current block ₂and λ ₂for adjustment parameter, be set to 1.1≤α ₂≤ 1.3 and 60≤λ ₂≤ 90.