CN103886607B - A kind of detection for disturbance target and suppressing method - Google Patents

Abstract

The invention discloses a kind of detection for disturbance target and suppressing method, comprise the following steps: background photograph album generates, background photograph album updates, background photograph album is replaced, space and time continuous attribute is set up, target trajectory is set up and disturbance object judgement.The method can effectively suppress the interference that random noise in dynamic scene, shadow change rapidly, background object slight perturbations brings, effectively detect moving target, and by moving target being set up initial motion track in time series, strengthen foreground target and enter the condition of follow-up identification process, reduce algorithm complex, add efficiency and the precision of target identification module.

Description

A kind of detection for disturbance target and suppressing method

Technical field

Image processing method in the present invention relates to moving object detection and identifying, a kind of for disturbance target Detection and suppressing method.

Background technology

It is partitioned into moving target, then method moving target being identified by recognizer based on foreground detection algorithm It it is the main path of motion estimate under Still camera.The Candidate Motion target providing foreground detection algorithm is sieved The mechanism of choosing, have impact on efficiency and the precision of follow-up differentiation process to a great extent.At intelligent video monitoring, road traffic prison In many application such as control, railway pedestrian's monitoring, random noise, scene shadow change rapidly, object slightly changes disturbing of bringing Moving-target brings the biggest difficulty to identification process.These a few class disturbance targets have a feature that three below is important: 1. from time Seeing between, target location also exists slight change.2., from spatially, target appearance also exists slight change.3. exist from target See in the outward appearance sequence formed in time series, the fluctuation of target gray level various degrees.Former just because of these three Cause, can inevitably bring this kind of target to meet the foreground detection algorithm of certain verification and measurement ratio, thus cause follow-up identification During need to identify the huge number of target, and because the randomness of this kind of disturbance target appearance and recognizer exist one The wrong identification determining error and cause.Therefore, differentiation and the elimination of this type of disturbance target are the most meaningful.

At present, being primarily present two kinds of thinkings in the problem of disturbance suppression target, the first gets down to the inspection of improvement prospect Method of determining and calculating, is inherently eliminated disturbance target.This kind of method is by optimizing the model of scene described in background modeling, it is considered to image Local message to single pixel design conditions probability, be partitioned in image the moving target being substantially distinguished from background.Based on The algorithm of time-space domain model has taken into full account the space-time consistency that color of image is distributed, and combines modeling at space-time and carries out moving target Detection, shows preferable performance during background perturbation in processing dynamic scene.Algorithm based on space-time modeling is due to needs Processing substantial amounts of time-space domain data, computation complexity is high, and memory requirements is big, algorithm poor real, owing to isolated noise disturbs shadow Ringing, final testing result also needs to carry out the post processing such as morphologic filtering and image segmentation just can preferably be detected knot Really.The second thinking gets down to follow-up corrigendum mechanism, and each target produced by foreground detection algorithm is carried out certain journey The suppression of degree, according to target between front and back two frame in attribute difference such as hue, saturation, intensities, determines target before and after suppression Whether feature reaches requirement, reaches requirement and is then detected as foreground target.This kind of algorithm can preferably remove those in appearance There is no a target of large change, but the robustness changed for rapid shadow is poor, such as, it is impossible to eliminate cloud and sail and bring Shadow effect.

Along with video monitoring system was developed to cybertimes by the epoch of simulating, video camera also develops towards intelligent direction, Increasing intelligent video Processing Algorithm includes moving object detection algorithm, needs to transplant, at video camera to intelligent camera On carry out embedded realization.But, the existing video frequency motion target detection algorithm that can process ambient noise in dynamic scene, Not only computation complexity is high, and memory requirements is the biggest, it is difficult to apply on embedded intelligence Camera Platform.

Summary of the invention

Apply in especially intelligent video monitoring system the moving object detection towards dynamic scene for computation vision, hold Be vulnerable to random noise, shadow changes rapidly, background object slight perturbations brings interference problem, it is desirable to provide a kind of For detection and the suppressing method of disturbance target, to improve efficiency and the accuracy of detection of identification process such that it is able to detect rapidly Identify specific moving target.

For achieving the above object, the present invention is by the following technical solutions:

A kind of detection for disturbance target and suppressing method, comprise the following steps:

Background photograph album generation step: set up the image set that capacity is N and be on N number of random node in preserving a period of time Background image, in order to preserve background change at random information, the N width image preserved is all gray level image, N >=2；

Background photograph album updates step: often reads in a two field picture, is converted into background image to be selected by greyscale transformation, and Based on a default update probability P_updateJudge whether to add in described background photograph album this frame background image；

Background photograph album replacement step: in the case of current background image is judged as joining in background photograph album, will the back of the body A certain image in scape photograph album replaces with current background image；

Space and time continuous attribute establishment step: be directed to each prospect that foreground detection algorithm detects in present image Target, characterizes this target by extracting the boundary rectangle of its connected domain, standard rectangle parameter Rect (x, y, w, h) on add Adding a time parameter t, wherein (x, y) represents target transverse and longitudinal coordinate on 2d respectively, w and h represents the width of rectangle And high, these four parameter basis add a time parameter t, sets up new structure TRect{x, y, w, h, t}, It contains target position spatially and size information, and continuation degree in time, is to have time and space continuity mesh Mark rectangle frame；

Target trajectory establishment step: Utilization prospects detection algorithm obtains target sequence, specifies with minimum distance criterion Inheritance between target in frame front and back, and updated described time parameter t by the inheritance between target；

Non-disturbance target determines step: by mating of target and background photograph album, it is determined that target when in former frame whether There is disturbance, and update the t value of target in present frame according to judged result, then, by the t value of target and threshold value set in advance Compare, if the t value of target exceedes threshold value set in advance, be then judged to non-disturbance target, and deliver at subsequent process Reason.

Preferably, it is thus achieved that after a new target, according to its position, (x, y, w h) obtain same position in background photograph album Coupling target carry out matching operation, by definition matching factor, it is determined that target is when whether there is disturbance, root in former frame According to disturbance result of determination, in conjunction with maximum succession threshold value T set₁Update present frame target.For the target sequence of present frame, Described threshold value set in advance is exceeded, it is determined that it is non-disturbance target, and delivers to follow-up identification if there is mesh target value Journey is identified.

Preferably, in described background photograph album updates step, take random number rand with the equiprobability condition of [0-1], As rand >=P_updateTime, it is judged that current background image is joined in background photograph album.

Preferably, in described background photograph album replacement step, current background image replaces a certain in background photograph album Image, the mechanism of used equiprobability random replacement.

Preferably, in described target trajectory establishment step, the distance criterion used is two target rectangle centers Euclidean distance.

Preferably:

In described space and time continuous attribute establishment step:

Detect that the target sequence that some connected domains are constituted is { Obj by foreground detection algorithm_1,T,Obj_2,T,…, Obj_m,T, wherein Obj_i,TRepresent at moment T(i.e. T frame) target sequence in i-th target, each target is the most permissible By a TRect, (x, y, w, h t) represent；

In described target trajectory establishment step:

For target sequence { Obj_1,T,Obj_2,T,…,Obj_m,TThe t value of initialized target is Obj_i,T.t=1, by Neighbour inherits criterion, and the t value of former frame target is passed to present frame target,

Wherein, for former frame target sequence { Obj_1,T-1,Obj_2,T-1,…,Obj_n,T-1Target Obj_j,T-1If, Relative to target Obj in present frame target sequence_j,t-1Preset range D in there is one or more target, then in these targets In find and target Obj_j,T-1Closest target $\hat{i}=\underset{i}{\arg \min }\mathrm{dist}({\mathrm{Obj}}_{j,T-1},{\mathrm{Obj}}_{i,T}),$ Europe between two targets Family name's distance is:

$\mathrm{dist}({\mathrm{Obj}}_i,{\mathrm{Obj}}_j)=\sqrt{{[({\mathrm{Obj}}_i.x+0.5\×{\mathrm{Obj}}_i.w)-({\mathrm{Obj}}_j.x+0.5\×{\mathrm{Obj}}_j.w)]}^2+{[({\mathrm{Obj}}_i.y+0.5\×{\mathrm{Obj}}_i.h)-({\mathrm{Obj}}_j.y+0.5\×{\mathrm{Obj}}_j.h)]}^2}$

If there is such targetThen will be according to following condition Obj_j,T-1T value pass toIf Obj_j,T-1.t ＜ T₁, thenOtherwiseWherein T₁For default maximum succession threshold Value；

Determine in step in described non-disturbance target:

At current goal sequence { Obj_1,T,Obj_2,T,…,Obj_m,TTarget t value is selected more than or equal to T in }₁Sub-sequence Row:For each target Obj in subsequence_i,T, according to its position, (x, y, w, h), in background Photograph album obtains N number of coupling target of same position, if Obj_i,tCoupling target sequence be { paObj₁,paObj₂..., paObj_N, wherein paObj_iBe expressed as in background photograph album in i-th background pictures position for (x, y, w, coupling target h), will Current goal Obj_TWith mate target { paObj₁,paObj₂..., paObj_NBe normalized to set pixel size,

It is its matching factor r that definition current goal mates the normalizated correlation coefficient between target with i-th_i:

$r_i=\frac{E({\mathrm{Obj}}_T\·{\mathrm{paObj}}_i)-E\left({\mathrm{Obj}}_T\right)\·E\left(\mathrm{pa}{\mathrm{Obj}}_i\right)}{\sqrt{D\left({\mathrm{Obj}}_T\right)}\·\sqrt{D\left(\mathrm{pa}{\mathrm{Obj}}_i\right)}}$

Wherein, E represents the expectation of variable, and D represents the variance of variable,

DefinitionFor Obj_T(x, y, w, h, t) with the matching factor of background photograph album, whenMore than certain Threshold value Thr_paTime, it is determined that target, when there is disturbance in former frame, is otherwise judged to the most non-disturbance, if currently Obj in frame_i,TIt is judged as non-disturbance, then increases Obj according to following condition_i,TT value:

If Obj_i,T.t<T₂, then Obj_i,T.t=Obj_i,T.t+1, otherwise Obj_i,T.t=T₂, wherein T₂For predetermined threshold value, T₂＞ T₁；

Target sequence { Obj for present frame_1,T,Obj₂,_T,…,Obj_m,T, the t value if there is target is T₂, the most really Determining it is non-disturbance target.

Preferably, D is 5.

Preferably, T₁It is 4.

Preferably, T₂It is 10.

Preferably, Thr_paIt is 0.987

The Advantageous Effects of the present invention:

Disturbance target suppressing method based on background photograph album and movement locus proposed by the invention mainly has the advantage that

N two field picture in historical frames sequence is retained, from the time, remains the average degree of disturbance target travel With cosmetic variation details.

With revisable update probability P_updateUpdate background photograph album so that the time registered depth of background photograph album can It is adjusted, meets the needs of different scene.

Set up the movement locus of target, be not only able to, for single goal record object movable information on space-time, more can Enough for multiple target to distinguish different target so that interference-free each other, the most multiobject detection and tracking purpose.

Use simple match algorithm based on image block to contrast the difference between target and background photograph album, reduce algorithm multiple Miscellaneous degree, anticipates background photograph album image by histogram equalization, highlights object variations details, improves object matching essence Degree.

Combining target movement locus and background photograph album, judge the disturbance of target with succession and the matching result of successive frame Degree, and unconventional judgement based on single-frame analysis result, while not affecting detection response, can improve disturbance again The degree of accuracy of target discrimination.

Sum it up, the present invention not only achieves the suppression under dynamic scene to disturbance target, have in real time at self Property, while the advantage that memory requirements is little, also overcome tradition moving object detection based on dynamic scene existing for pair time Information utilization between complexity low, spatially is high, memory requirements is big, be not easy to the shortcomings such as embedded realization.Calculating Moving object detection can be carried out on machine platform in real time, and be suitable to the target detection application of embedded platform.

Accompanying drawing explanation

Fig. 1 is detection and the flow chart of a kind of embodiment of suppressing method that the present invention is directed to disturbance target；

Fig. 2 is the display result of " background photograph album ", and wherein the ash of unduplicated history frame of video opened by background photograph album by random N Degree image construction；

Fig. 3 is the connected domain extracted by foreground detection algorithm, wherein the connected domain Sequence composition target sequence of present frame, Interframe target sequence constitutes again the track of target continuously；

Fig. 4 is the pedestrian target detected after non-disturbance target carries out svm classifier judgement.

Detailed description of the invention

Below in conjunction with accompanying drawing, embodiments of the invention are elaborated.It is emphasized that the description below is only shown Example rather than in order to limit the scope of the present invention and application thereof.

Embodiments of the invention refer to Fig. 1 to Fig. 4, and detection and suppressing method for disturbance target include following step Rapid:

(1) generation of background photograph album:

Relatively big in view of background intensity of variation in dynamic scene, its Changing Pattern within a period of time with certain with Machine, at the different nodes of continuous print sequence of frames of video, in background, the position range of disturbance target, outward appearance, gray level all have Having a certain degree of fluctuation, there is randomness in difference.This is the notable feature of ambient interferences target.Set up the image that capacity is N Collection { Img₁,Img₂,…,Img_NPreserve a period of time in be in the background image on N number of random node, preserve these with Machine change information.Fig. 2 shows a background photograph album, as can be seen from Figure the capacity N=6 of this photograph album.

(2) renewal of background photograph album:

By a update probability P_updateUpdate background photograph album.Often read in a frame video image Img_t, become by gray scale Change and be converted into background gray level image to be selected.Take random number rand with the equiprobability condition of [0-1], when rand >= P_updateTime, this frame background gray level image will be added in background photograph album.

(3) replacement of background photograph album:

If by the strategy of step (2), background image Img_tIt is judged as joining in background photograph album, will be with equiprobability Random replacement falls { Img₁,Img₂,…,Img_NAn image in }.The background photograph album shown by Fig. 2 can be seen that, background photograph album with Machine saves N random in historical frames sequence and opens unduplicated background gray level image, and this N opens the average gray phase not to the utmost of photo Together, show that background photograph album have recorded background average light photograph situation of change in a period of time, and the outward appearance saving disturbance target is thin Microvariations.

(4) acquisition of target sequence:

It is directed to each foreground target that foreground detection algorithm detects in present image, by extracting its connected domain Boundary rectangle characterize this target.As seen from Figure 3, by foreground detection algorithm and Morphological scale-space, comprise disturbance target Represented by connected domain one by one at interior moving-target with true pedestrian target.The rectangle of one standard may be characterized as: Rect (x, y, w, h), wherein (x, y) represents target transverse and longitudinal coordinate on 2d respectively, w and h represents width and the height of rectangle, Adding a time parameter t on these four parameter basis, set up new structure TRect{x, y, w, h, t}, it contains mesh It is marked on position spatially and size information, and continuation degree in time, is that there is time and space continuity target rectangle frame. In order to set up the continuity between target on space-time, with new structure TRect{x, y, w, h, t} characterize with space-time even The target rectangle frame of continuous property.So aerial at consecutive hours, will there is continuity thus constitute the fortune of target in target on frame sequence Dynamic track.The target sequence that the some connected domains detected by foreground detection algorithm are constituted is { Obj_1,T,Obj_2,T,…, Obj_m,T, wherein Obj_i,TRepresent at moment T(i.e. T frame) target sequence in i-th target, each target is the most permissible By a TRect, (x, y, w, h t) represent.

(5) foundation of movement locus:

After step (4), for target sequence { Obj_1,T,Obj_2,T,…,Obj_m,TThe t value of initialized target is Obj_i,T=1.Pass through Arest neighbors inherits criterion, and the t value of former frame target passes to present frame target, and rule is as follows: for former frame target sequence { Obj_1,T-1, Obj_2,T-1,…,Obj_n,T-1Target Obj_j,T-1, in present frame target sequence in certain limit D, if there is one or more mesh Mark, then find the target closest with it in these targets

Distance is defined with Euclidean distance between two targets:

$\mathrm{dist}({\mathrm{Obj}}_i,{\mathrm{Obj}}_j)=\sqrt{{[({\mathrm{Obj}}_i.x+0.5\&times;{\mathrm{Obj}}_i.w)-({\mathrm{Obj}}_j.x+0.5\&times;{\mathrm{Obj}}_j.w)]}^2+{[({\mathrm{Obj}}_i.y+0.5\&times;{\mathrm{Obj}}_i.h)-({\mathrm{Obj}}_j.y+0.5\&times;{\mathrm{Obj}}_j.h)]}^2}$

If there is such targetThen by Obj_j,T-1T value pass toDelivery rules is as follows: if Obj_j,T-1.t ＜ T₁, thenOtherwise

Specific to choosing different D values in different application scene, D is a threshold value that can regulate.Selection rule For: the ultimate range of motion it is marked in before and after two frame more than certain classification needing detection.This value can be by training sample This statistics gets, and such as, if desired for detecting pedestrian in railway scene, and video sample shows that pedestrian is at this scene video frame In image sequence, the move distance of the most every two interframe is not over 3 pixels, then can set D as the value more than 3.D is more Greatly, it is possible to the object of tolerance rapid movement, but more likely introducing noise targets, and D is the least, the movement locus of foundation is the most smart Really, the most more likely track is caused to rupture because of target rapid movement.Preferably D=5, meets most application scenarios.

T₁It is the threshold value that can regulate, in principle a T₁As long as more than 1, typically can be by statistics video sample In flash the frame number of noise continued presence and determine.T₁The biggest, then the most just filter the ability of the target that cannot set up movement locus The strongest, detect real motion target simultaneously required for frame number the most, i.e. detection the response time the longest.Preferably, T₁= 4。

(6) non-disturbance target is determined:

With mating of background photograph album: after step (5), in current goal sequence according to { Obj_1,T,Obj_2,T,…, Obj_m,TTarget t value is selected more than or equal to T in }₁Subsequence:As seen from Figure 3, continuously In frame, the disturbance target of a part is unable to reach T due to its path length₁And filtered, it is impossible to enter next step matching process. For each target Obj in subsequence_i,T, according to its position, (x, y, w h) obtain the N of same position in background photograph album Individual coupling target, if Obj_i,TCoupling target sequence be { paObj₁,paObj₂..., paObj_N, wherein paObj_iIt is expressed as the back of the body In scape photograph album, in i-th background pictures, position is (x, y, w, coupling target h).By current goal Obj_TWith mate target {paObj₁,paObj₂..., paObj_NIt is normalized to 15 × 15 pixel sizes.Definition current goal mate with i-th target it Between normalizated correlation coefficient be its matching factor: $r_i=\frac{E({\mathrm{Obj}}_T\&CenterDot;{\mathrm{paObj}}_i)-E\left({\mathrm{Obj}}_T\right)\&CenterDot;E\left(\mathrm{pa}{\mathrm{Obj}}_i\right)}{\sqrt{D\left({\mathrm{Obj}}_T\right)}\&CenterDot;\sqrt{D\left(\mathrm{pa}{\mathrm{Obj}}_i\right)}}$ (E represents variable Expectation, D represents the variance of variable).DefinitionWith background photograph album mate system Number.WhenMore than certain threshold value Thr_paTime, it is determined that target, when there is disturbance in former frame, is otherwise judged to the most non- Disturbance.If Obj in the current frame_i,_TIt is judged as non-disturbance, then increases Obj_i,TT value, increase rule be: if Obj_{i, T}.t<T₂, then Obj_i,T.t=Obj_i,T.t+1, otherwise Obj_i,T.t=T₂, wherein T₂For predetermined threshold value, T₂＞ T₁。

Thr_paIt is a threshold value that can regulate, can add up according to the video sample in concrete application scenarios and get, should It is worth the least, it is possible to the level of disruption of suppression is the biggest, the most also can suppress the human body target that movement degree is less.Preferably, Thr_pa=0.987。

T₂It is a threshold value that can regulate, as long as comparing T in principle₁Big.T₂Compare T₁Big must be the most, then disturbance suppression The ability of noise is the strongest, and the detection response time is the longest simultaneously.Preferably, T₂=10。

Non-disturbance object judgement: disturbance target in this process can be filtered, non-disturbance target then can enter into follow-up Identification process.Target sequence { Obj for present frame_1,T,Obj_2,T,…,Obj_m,T, the t value if there is target is T₂, then Determine that it is non-disturbance target, and deliver to the process such as subsequent process is identified, and t value is less than T₂Target then by as disturbance Target is inhibited, it is impossible to enter into follow-up identification process.Found out by Fig. 3 and Fig. 4, in the numerous targets in target sequence, Only minority target can enter SVM(SVM is Support Vector Machine, the i.e. abbreviation of SVMs) classify and sentence Other process, thus detect pedestrian.

Above-mentioned determine non-disturbance target during contain twice filtration:

Filter for the first time and be present in 1 T₁Between, the disturbance target of the most previously described part is due to its path length It is unable to reach T₁And filtered, for filtering out the target that can not set up movement locus.In this renewal process, target according to Arest neighbors is inherited criterion and is set up movement locus more fresh target t value, and the t value of target is less than T₁, its t value is only possible to+1, as Really the t value of target is more than or equal to T₁, then in this renewal process, its t value is constant.

Second time filters and is present in T₁—T₂Between, it is used for filtering out disturbance target.Second time filters and background photograph album Matching process, in this renewal process, target according to mate with background photograph album the t value that updates t value, only target more than or Equal to T1, its t value is only possible to+1, and otherwise in this renewal process, its t value is constant.Threshold value T₂The filtration played, Less than threshold value T₂Target filtered and be cannot be introduced into final SVM detection, thus pedestrian target cannot be detected as.

The present invention is towards the actual application of intelligent video monitoring system, for the foreground detection in dynamic scene easily by environment Noise jamming problem, it is proposed that the suppression removing method of disturbance target, the method in a kind of dynamic scene based on background photograph album Can effectively suppress the interference that random noise in dynamic scene, shadow change rapidly, background object slight perturbations brings, effectively Detect moving target, and by moving target being set up initial motion track in time series, strengthening foreground target and entering Enter the condition of follow-up identification process, reduce algorithm complex, add efficiency and the precision of target identification module so that based on The suppressing method of jamming target in the dynamic scene of background photograph album, is not only able to realize dynamic scene motion on existing PC platform Target is monitored in real time, is further adapted for the application of embedded intelligence Camera Platform.

Above content is to combine concrete preferred embodiment further description made for the present invention, it is impossible to assert Being embodied as of the present invention is confined to these explanations.For general technical staff of the technical field of the invention, On the premise of present inventive concept, it is also possible to make some simple deduction or replace, all should be considered as belonging to the present invention's Protection domain.

Claims (9)

1. the detection for disturbance target and suppressing method, it is characterised in that comprise the following steps:

Background photograph album generation step: set up the image set that capacity is N and be in the back of the body on N number of random node in preserving a period of time Scape image, in order to preserve background change at random information, the N width image preserved is all gray level image, N >=2；

Background photograph album updates step: often read in a two field picture, is converted into background image to be selected by greyscale transformation, and based on One default update probability P_updateJudge whether to add in described background photograph album this frame background image；

Background photograph album replacement step: in the case of current background image is judged as joining in background photograph album, by background phase A certain image in Ce replaces with current background image；

Space and time continuous attribute establishment step: be directed to each prospect mesh that foreground detection algorithm detects in present image Mark, characterize this target by extracting the boundary rectangle of its connected domain, the target of a standard can by rectangle Rect (x, y, w, H) describe, wherein (x, y) represents target transverse and longitudinal coordinate on 2d respectively, w and h represents width and the height of rectangle, Adding a time parameter t on these four parameter basis, set up new structure TRect{x, y, w, h, t}, it contains mesh It is marked on position spatially and size information, and continuation degree in time, is that there is time and space continuity target rectangle frame；

Target trajectory establishment step: Utilization prospects detection algorithm obtains target sequence, before and after specifying with minimum distance criterion Inheritance between target in frame, and updated described time parameter t by the inheritance between target；

Non-disturbance target determines step: by mating of target and background photograph album, it is determined that whether target is when existing in former frame Disturbance, and update the t value of target in present frame according to judged result, then, the t value of target is carried out with threshold value set in advance Relatively, if the t value of target exceedes threshold value set in advance, then it is judged to non-disturbance target, and delivers to subsequent process process.

2. the method for claim 1, it is characterised in that in described background photograph album updates step, the grade with [0-1] is general Rate condition takes random number rand, as rand >=P_updateTime, it is judged that current background image is joined in background photograph album.

3. the method for claim 1, it is characterised in that in described background photograph album replacement step, current background image Replace a certain image in background photograph album, the mechanism of used equiprobability random replacement.

4. the method for claim 1, it is characterised in that in described target trajectory establishment step, used Distance criterion is the Euclidean distance at two target rectangle centers.

5. the method as described in any one of Claims 1-4, it is characterised in that

In described space and time continuous attribute establishment step:

Detect that the target sequence that some connected domains are constituted is { Obj by foreground detection algorithm_1,T,Obj_2,T,…,Obj_m,T, its Middle Obj_i,TRepresenting the i-th target in the moment T i.e. target sequence of T frame, each target specifically can be by a TRect (x, y, w, h t) represent, m is the number of the target detected；

In described target trajectory establishment step:

For target sequence { Obj_1,T,Obj_2,T,…,Obj_m,TThe t value of initialized target is Obj_i,T.t=1, arest neighbors is passed through Inherit criterion, the t value of former frame target passed to present frame target,

Wherein, for former frame target sequence { Obj_1,T-1,Obj_2,T-1,…,Obj_n,T-1Target Obj_j,T-1If, currently Relative to target Obj in frame target sequence_j,T-1Preset range D in there is one or more target, D for allow a target exist The ultimate range that between consecutive frame, position changes, then find and target Obj in these targets_j,t-1Closest target Euclidean distance between two targets is:

$dist\left({\mathrm{Obj}}_i,{\mathrm{Obj}}_j\right)=\sqrt{{\&lsqb;\left({\mathrm{Obj}}_i.x+0.5\&times;{\mathrm{Obj}}_i.w\right)-\left({\mathrm{Obj}}_j.x+0.5\&times;{\mathrm{Obj}}_j.w\right)\&rsqb;}^2+{\&lsqb;\left({\mathrm{Obj}}_i.y+0.5\&times;{\mathrm{Obj}}_i.h\right)-\left({\mathrm{Obj}}_j.y+0.5\&times;{\mathrm{Obj}}_j.h\right)\&rsqb;}^2}$

If there is such targetThen will be according to following condition Obj_j,T-1T value pass toIf Obj_{j,T- 1}.t ＜ T₁, thenOtherwiseWherein T₁For default maximum succession threshold value, T₁> 1；

Determine in step in described non-disturbance target:

At current goal sequence { Obj_1,T,Obj_2,T,…,Obj_m,TTarget t value is selected more than or equal to T in }₁Subsequence:For each target Obj in subsequence_i,T, according to its position, (x, y, w, h), in background phase N number of coupling target of same position is obtained in Ce, if Obj_i,TCoupling target sequence be { paObj₁,paObj₂..., paObj_N, wherein paObj_iBe expressed as in background photograph album in i-th background pictures position for (x, y, w, coupling target h), will Current goal Obj_TWith mate target { paObj₁,paObj₂..., paObj_NBe normalized to set pixel size,

It is its matching factor r that definition current goal mates the normalizated correlation coefficient between target with i-th_i:

$r_i=\frac{E({\mathrm{Obj}}_T\&CenterDot;{\mathrm{paObj}}_i)-E\left({\mathrm{Obj}}_T\right)\&CenterDot;E\left({\mathrm{paObj}}_i\right)}{\sqrt{D\left({\mathrm{Obj}}_T\right)}\&CenterDot;\sqrt{D\left({\mathrm{paObj}}_i\right)}}$

Wherein, E represents the expectation of variable, and D represents the variance of variable,

DefinitionFor Obj_T(x, y, w, h, t) with the matching factor of background photograph album, whenMore than certain threshold value Thr_paTime, it is determined that target, when there is disturbance in former frame, is otherwise judged to the most non-disturbance, if in the current frame Obj_i,TIt is judged as non-disturbance, then increases Obj according to following condition_i,TT value:

If Obj_i,T.t<T₂, then Obj_i,T.t=Obj_i,T.t+1, otherwise Obj_i,T.t=T₂, wherein T₂For predetermined threshold value, T₂>T₁；

Target sequence { Obj for present frame_1,T,Obj_2,T,…,Obj_m,T, the t value if there is target is T₂, it is determined that it For non-disturbance target.

6. method as claimed in claim 5, it is characterised in that D is 5.

7. method as claimed in claim 5, it is characterised in that T₁It is 4.

8. method as claimed in claim 5, it is characterised in that T₂It is 10.

9. method as claimed in claim 5, it is characterised in that Thr_paIt is 0.987.