CN106910203A - The method for quick of moving target in a kind of video surveillance - Google Patents

Abstract

The invention discloses a kind of method for quick of moving target in video surveillance, the quick detection of moving target in realizing video surveillance by way of background model, detection moving region are set up to video to be measured, target to be measured is determined and carries out specific tracking etc. to target.The moving target that the method for quick of moving target can be in quick detection video in the video surveillance that the present invention is provided, the calculating and detection of characteristic goal are carried out according to the target data Sample Storehouse for prestoring, and target missing inspection and the problem picked up again are avoided well.

Description

The method for quick of moving target in a kind of video surveillance

Technical field

The invention belongs to field of video detection, and in particular to the method for quick of moving target in a kind of video surveillance.

Background technology

With full-scale digital, the development of the video monitoring system of networking, the effect of video monitoring becomes more substantially, The exploration of its height, integration and flexibility, for the development of whole security protection industry provides more wide development space, intelligence The features such as energy video monitoring is due to assigning more intellectuality, activeization, validity, as video monitoring trend of new generation.

The demand of intelligent video monitoring system mostlys come from those occasions sensitive to safety requirements, such as army, public security, Bank, road, parking lot etc..When theft occur or it is abnormal occur when, such system can actively to guard promptly and accurately Send alarm, enable staff to make full use of video surveillance network to implement alarm linkage and emergency command and dispose so that Avoid the generation of crime.Decrease the input for employing large quantities of monitoring personnel simultaneously.

Important content and basis as intelligent monitoring technology, to target (especially vehicle and pedestrian target) detect and with Continuing to optimize for track is the constantly progressive necessary process of intelligent monitoring technology.The following problem that current target following detection is present： 1) moving target monitoring directly affects the effect of tracking as the basis of target following.But due to holding during video acquisition It is vulnerable to the interference of extraneous factor, such as DE Camera Shake, illumination variation, target shadow interference, target occlusion, homochromy ambient interferences Deng these reality factors all can bring sizable difficulty to target monitoring.Such as the variation targets shade of lighting angle exists Target prospect can be detected as by different degrees of in monitoring process, target global shape is produced a very large impact；Target occlusion makes Detection process is difficult to obtain the overall shape of target, loses target relevant information, is unfavorable for follow-up tracking；Illumination variation, take the photograph Camera shake, ambient interferences can regard noise as, the accuracy of same influence target detection.2) occlusion issue is current target The Important Problems of tracking, present most of Target Tracking Systems not can effectively solve the problem that and receive background or other mesh with target Mutual occlusion issue between mark.Target occlusion is random, uncertain problem during tracking.Asked for such The simple background modeling that relies on of topic realizes that target detection or track algorithm are insecure, it is necessary to set up preferably object module or Feature templates, and solved with the accurate match of model or feature using those visible target parts.Therefore, target inspection is improved The efficiency and accuracy rate of survey are vital contents in intelligent video monitoring system.

Target detection in monitor video, is broadly divided into two classes：(1) moving object detection；(2) based on image recognition Target detection.Both respectively have advantage and disadvantage, and moving object detection can fast and effeciently detect the moving target in monitor video, but Target to being still in scene cannot be detected, and also seem that comparing is powerless in the separation of adhesion target；Based on image The target detection of identification, to full figure in all targets detect, either moving target or static target, due to relying on Judge in feature, these are excellent for influence of this kind of method situations such as obtaining accuracy rate higher, recall rate and being less subject to adhesion While gesture, it generally require that more run times.

The content of the invention

Move mesh in a kind of video surveillance it is an object of the invention to be provided to overcome above the deficiencies in the prior art Target method for quick.

Technical scheme is as follows：

The method for quick of moving target, comprises the following steps in a kind of video surveillance：

Step one, video V={ I to be measured₀,I₁,I₂,…,I_k, wherein I_kIt is the kth frame image in video V, by video 0th two field picture is set as initial back-ground model D₀, i.e. D₀=I₀；

Step 2, by calculating F_k=I_k-D_k-1Extraction prospect, and in F_kThe upper set S for extracting potential moving region_kIf, Determine S_kThe initial value O of the set of middle target_k={ } is empty set；It is the image I that currently will analyzing wherein to extract prospect_kWith the back of the body Scape model D_k-1Point-to-point to subtract each other, difference exceedes the point of constant 10 as foreground point, otherwise as background dot；It is point-to-point to have subtracted each other Morphology open and close operation is used after, noise filtering is crossed and is made UNICOM region relatively more regular, then obtained using region growing method Obtain wherein as all UNICOM regions of prospect, close region is merged, obtain the set S of potential moving region_k；

Step 3, for step 2 S_iIn each potential moving region s, detection s in target to be checked that may be present, institute There is the target for detecting to add set O_kIn；

Step 4, to O_k-1In all targets, be tracked in kth frame, gained target to be checked is also added to O_k In, and continue to appear in present frame to the last time of lost target in all short time forward and be tracked, gained target It is also added to current goal set O_kIn；

Step 5, for set O_kThe target that middle certain hour is occurred without makees disappearance treatment, by the target from set O_kIn delete Remove；Step 6, updates background model；

Step 7, for next two field picture, repeats the above steps two to step 6, until detecting video last frame figure As after, testing result is obtained.

Further, in described video surveillance moving target method for quick, in step 3 detect s in may deposit Mesh calibration method to be checked be multiple dimensioned, floating window and target identification based on HoG feature calculations method.

Further, in described video surveillance moving target method for quick, based on HoG feature calculations Target identification, in addition it is also necessary to set a decision mechanism, first demarcates a target data Sample Storehouse, data sample manually headed by the mechanism This storehouse is made up of the picture of a large amount of same sizes, and all having demarcated it per pictures whether there is particular detection target, per pictures In the HoG features of each pixel be acquired；HoG features and then shape that decision mechanism passes through each pixel in the every pictures of acquisition In pairs in the identification model data of detection target；During particular detection, first by window to be measured be adjusted to in decision mechanism Picture identical size, then by whether there is particular detection target in identification model data judging window to be measured.

Further, in described video surveillance moving target method for quick, in step 3 detect s in may deposit Mesh calibration method to be checked can also be carried out according to following：

1) using the first two field picture as initial background, i.e. C (x, y, 1)=T (x, y, k), wherein x, y are the seat of pixel Mark, k is frame number number；

2) to present frame T (x, y, k), the calculating of target identification matrix D (x, y, k) is carried out, computing formula is：

Wherein C (x, y, k) is current background image, and the span of F (k) is 20-40；

3) each pixel A is counted_i,j,kAnd A_i+n,j+m,k+fProbability within the time and space field of pixel to be measured is close Degree, calculates the value of information gap and test point pixel, and wherein information gap M (x, y, k) is：

In formula, g (A_i,j,k) and g (A_i+n,j+m,k+f) it is respectively pixel A_i,j,kAnd A_i+n,j+m,k+fIn time-domain and spatial domain Interior probability density function；Value Z (the A of pixel_i,j,k) be：Z(A_i,j,k)=g (A_i,j,k)+M (x, y, k)/26, (k >=2)；

4) by the value Z (A of pixel_i,j,kThe region of) >=0.02 marks as prospect, labeled as Q^q；The value Z of pixel (A_i,j,kThe region of)≤0.02 as context marker, labeled as Q^b；

5) the multiscale morphological gradient image g of present frame is calculated using Multiscale Morphological method；

6) the mark Q more than^qWith Q^bOptimize：G ,=imimposemin (g, Q^q|Q^b), then calculated using watershed Method obtains target image：Contour=watershed (g)；

7) to next frame image update and set up new background, repeat step 2) -6), until last frame image, is examined Survey result.

Further, in described video surveillance moving target method for quick, in step 4 track method such as Under：

1) it is initial retrieval position p to initialize first in frame to be tracked with target location identical position in tracking source frame；

2) centered on retrieving position p in frame to be tracked, size and the target equal-sized rectangle frame in source frame is tracked As potential tracking result, the target in it and tracking source is sought into mean square deviation, if this value is less than threshold value 2, then it is assumed that with Track is completed, and have found tracking target, and algorithm terminates；

3) mean square deviation that previous step is obtained is unsatisfactory for if condition, is eight positions diamond shaped positions around p：It is tiltedly right Four adjoint points at angle+up and down each every the position of any each calculated respectively as rectangle frame center with tracking target with Mean square deviation in track source frame, thinks to be tied in the absence of tracking if this eight values exceed the mean square deviation that previous step is obtained Really, algorithm terminates；Otherwise, p is replaced with the point for producing lowest mean square difference；

4) last time to the target in a range of frame before present frame occurs, all using in above step Method is tracked.

Further, in described video surveillance moving target method for quick, background model is updated in step 6 Specially it is the point of (x, y) to position, such as the brightness value in current background model is d, and the brightness value in present image It is p, then by present image, the value of background model is updated to D_k=(1- α) × d+ α × p, if the point that position is (x, y) exists It is judged to background dot in step 2, then α values take 0.1, if being judged to foreground point in step 2, α values take 0.01.

Further, in described video surveillance moving target method for quick, step 4) in before present frame A range of frame refers to preceding 25 frame of present frame.

The motion that the method for quick of moving target can be in quick detection video in the video surveillance that the present invention is provided Target, the calculating and detection of characteristic goal are carried out according to the target data Sample Storehouse for prestoring, and target leakage is avoided well Inspection and the problem picked up again.

Brief description of the drawings

Fig. 1 is the flow chart of the moving target method for quick described in the embodiment of the present invention 1；

Fig. 2 is the picture of the before processing described in the embodiment of the present invention 1；

Fig. 3 is the differentiated picture of background described in the embodiment of the present invention 1；

Fig. 4 is through the picture after morphological operation described in the embodiment of the present invention 1；

Fig. 5 is the picture of the final detection structure described in the embodiment of the present invention 1.

Specific embodiment

Embodiment 1

To make the object, technical solutions and advantages of the present invention become more apparent, below in conjunction with specific embodiment, and reference Accompanying drawing, the present invention is described in more detail.

Fig. 1 is the flow chart of the moving target method for quick that the present embodiment is provided, and is comprised the following steps：

Step M1, to video sequence in each two field picture, detect all potential moving regions therein, and obtain The position in these regions；

In the present embodiment, we choose the HD video segment that one section of CCTV camera shoots, and resolution ratio is 1920x1080, video scene is the road traffic that both sides are greenbelts, the point pixel of the picture in video sequence, its brightness value Between 0~255, video is designated as V={ I₀,I₁,I₂,…,I_k, wherein I_kIt is the kth frame image in video V, ours is follow-up Analysis is all based on brightness value to be carried out, and Fig. 2 is the pictures for picking up from the video.

Background model D_kIt is the data needed to use in step M1, it represents the prediction to the background of current scene. Under initial situation, D_kIt is set equal to the first pictures, i.e. D in video sequence₀=I₀, during subsequent execution, D₀Carry out Real-time update；

Step M1 is further comprising the steps：

Step M11, the current picture analyzed in video sequence and background model D₀It is point-to-point to subtract each other extraction prospect, That is F_k=I_k-D_k-1, now k=1 (Fig. 3 shows the result that picture subtracts each other with background model in Fig. 2), wherein difference exceedes certain The point of constant (such as, 10) is considered as foreground point, otherwise it is assumed that being background dot.By the step, we have known current figure The preliminary judgement of the foreground/background of the every bit as in；

Step M12, the preliminary judgement of the foreground/background obtained according to previous step, plucks out the potential fortune in present image Dynamic region S_k.Setting S_kThe initial value O of the set of middle target_k={ } is empty set, and then order uses morphology open and close operation, mistake Noise filtering simultaneously makes UNICOM region relatively more regular, on this basis, is obtained wherein as the institute of prospect using region growing method There is UNICOM region, each UNICOM region is exactly a potential moving region, to these UNICOM regions, is once closed on area Domain merges：If the minimum in Liang Ge UNICOMs region is no more than certain threshold value (such as, 3) comprising the distance between rectangle, then by this Liang Ge UNICOMs region merging technique.Fig. 3 is last all these by UNICOM's region merging technique by such as Fig. 4 of the result after morphological operation Highlight regions form the potential moving region of only one, and (needing to carry out also can be as seen from the figure the reason for adjacent area merges Come, object vehicle therein has been partitioned into several adjoining UNICOM regions, if do not merged, it is possible to draw automobile It is divided into several fractions)；

Step M13, background model D is updated with present image₀.It is the point of (x, y) to position, it is assumed that in current background model In value be d, and the value in present image is p, then by present image, the value of background model is updated to D₁=(1- α) × d + α × p, if the point of the position is judged as background in M11, uses a larger α value (such as, 0.1), if should The point of position is judged as prospect in M11, then use a less α value (such as, 0.01)；

Step M2, each potential moving region to being exported in step M12 carries out target identification, wherein real to find out Target in meaning, step M2 is further included：

Step M21, determines division of the multiple dimensioned floating window to region, and the size of floating window is according to the actual possibility of target Size is set.For certain specific target, it is assumed that minimum dimension of such target in video image be s × t (for Target pedestrian, this value is used such as：S=8, t=16；For target vehicle, this value is used such as：S=24, T=24), then it is s × t window size to be defined on first yardstick, and in the upper left position in region, then the window floats first The window's position float from left to right, from the top down successively until the lower right corner in region, the displacement floated every time is window size 1/2 (i.e. the span of displacement from left to right be s/2, from the top down the span of displacement be t/2), when the location free procedure terminates, i.e., The location free procedure of first yardstick terminates；Into second floating of yardstick, in this second floating of yardstick, floating frame It is wide, high while be adjusted to the constant times (such as, 1.05) of certain wide, high of floating frame in a yardstick more than 1, so After be similar to and equally floated in first yardstick, next one yardstick is entered after terminating until window size has exceeded target Full-size untill (such as：Two times of wide, height of minimum dimension).All stop places of each of the above yardstick floating window, shape Into the region division to the region, we can be judged in each of which position with the presence or absence of the specific mesh in step below Mark.So it is exactly to form this region division (region has been partitioned into many window's positions, between these positions that M21 is actual May mutually overlap)；

Step M22, to region in calculate HoG features a little, the calculating of HoG features can directly use OpenCV Respective function complete；

Step M23, to M21 region divisions each the window's position out, is carried out based on HoG features in the window Target identification (carries out pedestrian/vehicle identification) respectively, and used as the result of identification, each window of region division is judged In the window there is target in Yes/No.

Based on the target identification of HoG features, this needs to use step M4 and carries out machine learning and obtains decision mechanism (i.e. Identification model, that is, model data D3).

In order to step M4 carries out machine learning, a target data Sample Storehouse D2 has been demarcated manually first (for pedestrian, car It is each to have such set of data samples by oneself, illustrated just for a certain kind therein now), data sample storehouse D2 (such as 32 × 64) are constituted by the picture of the same size of a certain amount of (such as, 1000 width), each picture has been demarcated and has wherein been Or it is no there is specific objective (i.e. pedestrian/vehicle), the HoG features of each pixel are obtained and (obtained special in each picture Levy)；

Step M4, the identification model data that machine learning obtains the target are carried out based on D2.Due to the size of picture in D2 Fixed, the byte number of the HoG features of each pixel is also fixed in picture, therefore it is exactly to regard one as that each pictures are actual The individual high dimension vector formed by the HoG features arranged in sequence of all pixels, and whether include specific mesh in this pictures Mark (0,1) is classification, and this is a very typical classification problem, and we complete this learning process using SVM algorithm (SVM algorithm is directly carried out using the function in OpenCV), so as to obtain the identification model data D3 of the target；

In M23, we are first adjusted to window size the size of picture in D2, are called then or with SVM algorithm Model data D3, completes identification, and Yes/No has the specific objective in judging the window；

Step M24, arranges recognition result.Due to there is overlap largely between those windows of M21 region divisions, Repeatedly recognized it is therefore possible to simple target, it is therefore desirable to which it is weight which the simple target for judging to be recognized in M23 wherein has Multiple, by arranging, the final all targets arrived in the region detection are obtained, and these targets are added to set O_kIn (Fig. 5 It is the final recognition result of Fig. 4, it is seen that come it and two targets of the adhesion on moving region are separated)；

For potential moving region S_kIn target truly detection, we also carried out in research process with Lower detection mode：

1) using the first two field picture as initial background, i.e. C (x, y, 1)=T (x, y, k), wherein x, y are the seat of pixel Mark, k is frame number number；

2) to present frame T (x, y, k), the calculating of target identification matrix D (x, y, k) is carried out, computing formula is：

Wherein C (x, y, k) is current background image, and the span of F (k) is 20-40；

3) each pixel A is counted_i,j,kAnd A_i+n,j+m,k+fProbability within the time and space field of pixel to be measured is close Degree, calculates the value of information gap and test point pixel, and wherein information gap M (x, y, k) is：

In formula, g (A_i,j,k) and g (A_i+n,j+m,k+f) it is respectively pixel A_i,j,kAnd A_i+n,j+m,k+fIn time-domain and spatial domain Interior probability density function；Value Z (the A of pixel_i,j,k) be：Z(A_i,j,k)=g (A_i,j,k)+M (x, y, k)/26, (k >=2)；

4) by the value Z (A of pixel_i,j,kThe region of) >=0.02 marks as prospect, labeled as Q^q；The value Z of pixel (A_i,j,kThe region of)≤0.02 as context marker, labeled as Q^b；

5) the multiscale morphological gradient image g of present frame is calculated using Multiscale Morphological method；

6) the mark Q more than^qWith Q^bOptimize：G ,=imimposemin (g, Q^q|Q^b), then calculated using watershed Method obtains target image：Contour=watershed (g)；

7) new background, repeat step 2 are set up to next frame image update and according to above step M13) -6), until finally One two field picture.

By above detecting step, finally give and step M24 identical particular detection target vehicles and pedestrian.

Step M3, carries out target following.Step M2 have detected target to all potential moving region of present image, him Integrate all targets just obtained in present image, it is known that these targets be not it is isolated exist, it can regarded Persistently exist in a period of time in frequency sequence, we string together the target between different images, so as to obtain by step 3 Obtain the motion track of target；

Step M31, screens out some substantially undesirable targets, and this is a reservation step, in a variety of causes Be likely to require and do a preliminary screening, such as boundary, such as due to wrong report for bringing of recognition accuracy deficiency etc.；

Step M32, to present image I_kSet O before_k-1In target, to tracking before present frame k is carried out, tracking Method is specific as follows：

1) it is initial retrieval position p to initialize first in frame to be tracked with target location identical position in tracking source frame；

2) centered on retrieving position p in frame to be tracked, size and the target equal-sized rectangle frame in source frame is tracked As potential tracking result, the target in it and tracking source is sought into mean square deviation, if this value is less than threshold value 2, then it is assumed that with Track is completed, and have found tracking target, and algorithm terminates；

3) mean square deviation that previous step is obtained is unsatisfactory for if condition, is eight positions diamond shaped positions around p：It is tiltedly right Four adjoint points at angle+up and down each every the position of any each calculated respectively as rectangle frame center with tracking target with Mean square deviation in track source frame, thinks to be tied in the absence of tracking if this eight values exceed the mean square deviation that previous step is obtained Really, algorithm terminates；Otherwise, p is replaced with the point for producing lowest mean square difference；

4) last time to the target in a range of frame before present frame (such as, 25 frame) occurs, and all uses Method in above step is tracked, and sees whether it occurs, if there is, judge whether be among the testing result of M24, If then both are connected by identifying identical ID, if not then it is added to as a newfound target working as Preceding goal set O_kIn；

Step M33, it would be desirable to which disappearance treatment is not made in the target for occurring to some for a long time；Once make disappearance treatment, Then system will forget the target, and any target for occurring later will not may make any association with it again, in our realization In, the target not occurred in some nearest frames (such as, 600 frame) is simply made disappearance treatment by we.

Thus, the quick detection to moving target in monitor video is completed.

Particular embodiments described above, has been carried out further in detail to the purpose of the present invention, technical scheme and beneficial effect Describe in detail bright, should be understood that and the foregoing is only specific embodiment of the invention, be not intended to limit the invention, it is all Within the spirit and principles in the present invention, any modification, equivalent substitution and improvements done etc., should be included in guarantor of the invention Within the scope of shield.

Claims (7)

1. in a kind of video surveillance moving target method for quick, it is characterised in that comprise the following steps：

Step one, video V={ I to be measured₀,I₁,I₂,…,I_k, wherein I_kIt is the kth frame image in video V, by the 0th of video the Two field picture is set as initial back-ground model D₀, i.e. D₀=I₀；

Step 2, by calculating F_k=I_k-D_k-1Extraction prospect, and in F_kThe upper set S for extracting potential moving region_k, set S_kIn The initial value O of the set of target_k={ } is empty set；It is the image I that currently will analyzing wherein to extract prospect_kWith background model D_k-1Point-to-point to subtract each other, difference exceedes the point of constant 10 as foreground point, otherwise as background dot；It is point-to-point subtract each other completion after Using morphology open and close operation, cross noise filtering and make UNICOM region relatively more regular, then obtained wherein using region growing method As all UNICOM regions of prospect, close region is merged, obtain the set S of potential moving region_k；

Step 3, for step 2 S_iIn each potential moving region s, detection s in target to be checked that may be present, Suo Youjian The target for measuring adds set O_kIn；

Step 4, to O_k-1In all targets, be tracked in kth frame, gained target to be checked is also added to O_kIn, and after Continuous to appear in present frame to the last time of lost target in all short time forward and be tracked, gained target is equally added To current goal set O_kIn；

Step 5, for set O_kThe target that middle certain hour is occurred without makees disappearance treatment, by the target from set O_kMiddle deletion；

Step 6, updates background model；

Step 7, for next two field picture, repeats the above steps two to step 6, until detecting video last frame image Afterwards, testing result is obtained.

2. in video surveillance according to claim 1 moving target method for quick, it is characterised in that in step 3 Mesh calibration method to be checked that may be present is multiple dimensioned, floating window and the target identification based on HoG feature calculations in detection s Method.

3. in video surveillance according to claim 2 moving target method for quick, it is characterised in that based on HoG The target identification of feature calculation, in addition it is also necessary to set a decision mechanism, first demarcates a target data manually headed by the mechanism Sample Storehouse, data sample storehouse is made up of the picture of a large amount of same sizes, and all having demarcated it per pictures whether there is particular detection Target, the HoG features of each pixel are acquired in every pictures；Decision mechanism is by each pixel in the every pictures of acquisition HoG features so formed for detect target identification model data；During particular detection, window to be measured is adjusted to first With picture identical size in decision mechanism, then by identification model data judging window to be measured whether there is particular detection Target.

4. in video surveillance according to claim 1 moving target method for quick, it is characterised in that in step 3 Mesh calibration method to be checked that may be present can also be carried out according to following in detection s：

1) using the first two field picture as initial background, i.e. C (x, y, 1)=T (x, y, k), wherein x, y are the coordinate of pixel, and k is Frame number number；

2) to present frame T (x, y, k), the calculating of target identification matrix D (x, y, k) is carried out, computing formula is：

$D(x,y,k)=\left\{\begin{array}{c}1,|T(x,y,k)-C(x,y,k)|\≥F\left(k\right)\\ 0,|T(x,y,k)-C(x,y,k)|\≤F\left(k\right)\end{array}\right.$

Wherein C (x, y, k) is current background image, and the span of F (k) is 20-40；

3) each pixel A is counted_i,j,kAnd A_i+n,j+m,k+fProbability density within the time and space field of pixel to be measured, calculates The value of information gap and test point pixel, wherein information gap M (x, y, k) are：

$M(x,y,k)=\underset{n=-1}{\overset{1}{\mathrm{\Σ}}}\underset{m=-1}{\overset{1}{\mathrm{\Σ}}}\underset{f=-2}{\overset{0}{\mathrm{\Σ}}}log\frac{g\left(A_{i,j,k}\right)}{g\left(A_{i+n,j+m,k+f}\right)},(k\≥2)$

In formula, g (A_i,j,k) and g (A_i+n,j+m,k+f) it is respectively pixel A_i,j,kAnd A_i+n,j+m,k+fIn time-domain and spatial domain Probability density function；Value Z (the A of pixel_i,j,k) be：Z(A_i,j,k)=g (A_i,j,k)+M (x, y, k)/26, (k >=2)；

4) by the value Z (A of pixel_i,j,kThe region of) >=0.02 marks as prospect, labeled as Q^q；The value Z of pixel (A_i,j,kThe region of)≤0.02 as context marker, labeled as Q^b；

5) the multiscale morphological gradient image g of present frame is calculated using Multiscale Morphological method；

6) the mark Q more than^qWith Q^bOptimize：G ,=imimposemin (g, Q^q|Q^b), then obtained using watershed algorithm To target image：Contour=watershed (g)；

7) to next frame image update and set up new background, repeat step 2) -6), until last frame image, obtain detection knot Really.

5. in video surveillance according to claim 1 moving target method for quick, it is characterised in that in step 4 The method of tracking is as follows：

1) it is initial retrieval position p to initialize first in frame to be tracked with target location identical position in tracking source frame；

2) using centered on retrieving position p in frame to be tracked, size and target in source frame is tracked equal-sized rectangle frame as Potential tracking result, mean square deviation is sought by the target in it and tracking source, if this value is less than threshold value 2, then it is assumed that tracked Into, tracking target is have found, algorithm terminates；

3) mean square deviation that previous step is obtained is unsatisfactory for if condition, is eight positions diamond shaped positions around p：Diagonally opposing corner Four adjoint points+each each calculated respectively as rectangle frame center with tracking target in tracking source every the position of any up and down Mean square deviation in frame, thinks, in the absence of tracking result, to calculate if this eight values exceed the mean square deviation that previous step is obtained Method terminates；Otherwise, p is replaced with the point for producing lowest mean square difference；

4) last time to the target in a range of frame before present frame occurs, all using the method in above step It is tracked.

6. in video surveillance according to claim 1 moving target method for quick, it is characterised in that in step 6 Update background model be specially to position be (x, y) point, such as the brightness value in current background model is d, and is schemed currently Brightness value as in is p, then by present image, the value of background model is updated to D_k=(1- α) × d+ α × p, if position is The point of (x, y) is judged to background dot in step 2, then α values take 0.1, if being judged to foreground point in step 2, α values take 0.01。

7. in video surveillance according to claim 5 moving target method for quick, it is characterised in that step 4) in A range of frame refers to preceding 25 frame of present frame before present frame.