CN101017573A - Method for detecting and identifying moving target based on video monitoring - Google Patents

Abstract

This invention relates to one test and identification method for visual monitor system move aim, which is based on block frame different and background deduction and comprises the following steps: firstly processing blocks sorting on visual images; then processing difference computation by frame difference method; according to difference result to realize move image rough cut to separate background and prospect area; the background area is to form and update background module; accordingly by use of background images to process difference of rough cut move area and the difference image grey value meets certain valve value; then polymer the aim pixel points for shadow process in the HSV to get visual image for accurate move aim.

Description

A kind of moving object detection and recognition methods based on video monitoring

One, technical field

What the present invention relates to is motion target detection and recognition methods in a kind of video monitoring system, the image processing method that especially by video technique people or vehicle etc. is carried out intelligent monitoring in traffic, safety-security area.

Two, technical background

Intelligent video monitoring is importance in the computer vision field, and its groundwork is exactly to detect, discern and follow the tracks of interested moving target from the video image of dynamic scene, and wishes can analyze, understand and describe the behavior of monitored object.Traditional supervisory system generally is to obtain the video image of monitoring scene by multiple-camera, finishes supervision by human eye; Perhaps record a video to monitoring the scene, after incident took place, video image was as evidence.Fact proved that the monitoring efficiency of this mode is very low.Intelligentized video monitoring not only can substitute people's eyes, and can finish the task of various monitoring automatically.At present, intelligent video monitoring system is a focus of a computer vision research.

In intelligentized video monitoring system, monitor that the motion target detection in the scene is very crucial technology with identification.But from present achievement in research, motion detection is quite important but is the problem of comparison difficulty fast and accurately.This is because the image of catching in the dynamic environment is subjected to many-sided influence, confusion such as the variation of the variation of weather, illumination condition, background is disturbed, blocking even the motion of video camera etc. between shadow, object and the environment of moving target or between object and the object, and these have brought difficulty all for accurately and effectively motion segmentation.Just the shadow with moving target is an example, and it may link to each other with detected target, also may separate with target.Under former instance, shadow has twisted the shape of target, thereby no longer reliable based on the recognition methods of shape after making; Under latter instance, shadow might be mistaken as and be full of prunes target in the scene.At present image motion is cut apart and is mainly utilized the background subtraction method, is still the quite problem of difficulty but how to set up the background model that dynamic change for any complex environment all has adaptivity.

In intelligentized video monitoring system, monitor that the ability of moving target identification is to judge the important indicator of system intelligent degree in the scene, be the key that can complete successfully automatic monitor task.In the video monitoring process, the monitoring environment complexity, video camera difference is set, the interference that causes the video image of collection to be subjected under many circumstances is very big.For example, video camera is from the distance of monitoring scene, cause the target area size of detection to change, and the distance of different target video cameras area on the contrary can be big also may cause the little object image-forming of target the time, this just causes with the target area is that the Feature Recognition method lost efficacy.The main mathematical method that moving target identification is at present adopted has statistical method, neural network, support vector machine etc.Along with the application scenario of video monitoring constantly enlarges, to the accuracy rate of identification, the real-time of identification also requires more and more higher.It is good how to design a recognition effect, applied range, and adaptable moving target recognition methods is a challenge.

At present, the method for target detection in the video image is mainly contained optical flow method, method of difference and background subtraction method, every kind of method all have relative merits separately.Trend is merged several method now, to obtain better to detect effect.Aspect Target Recognition, the design of feature selecting and sorter is the emphasis of research.The application that neural net method is extensively sent out, current support vector machine more and more are subjected to paying attention to widely.Can be referring to [1] Maurin B, MasoudO, Papanikolopoulos N P.Tracking all traffic:computer vision algorithms for monitoringvehicles, individuals, and crowds.IEEE Robotics﹠amp; Automation Magazine, 2005,12 (1): 29～36.[2] Weiming Hu, Tieniu Tan, Liang Wang, and Steve Maybank, A survey on visualsurveillance of object motion and behaviors, IEEE Transactions on Systems, Man andCybemetics, Part C:Applications and Reviews, Vol.34, No.3,2004, pp.334-352.[3] Wang Liang, Hu Weiming, the pool car pusher. the visual analysis summary of people's motion. Chinese journal of computers, 2002,25 (3): 225-237.[4] Pan Feng etc., based on the human detection under the complex background of support vector machine. Chinese image graphics journal, 2005,10 (2): 181-186.[5] Li Yuancheng, Fang Tingjian. based on the Support Vector Machine Study on Forecasting Method .2003 Vol.18 No.2 of rough set theory P.199-203.[6] Lipton A, Fujiyoshi H and Patil R.Moving target classification and tracking from real-timevideo.In:Proc IEEE Workshop on Applications of Computer Vision, Princeton, NJ, 1998,8-14.[7] Platt J C, Cristianini N, Shawe-Taylor J.Large Margin DAGs for MulticlassClassification.In Advances in Neural Information Processing Systems, MITPress, 2000:12:547-553.[8] Cucch iara R, Grana C, PiccardiM et al.Improving shadowsuppression in moving object detection with HSV color information[A] .In:Proceedings ofIEEE Intelligent T ranspo rtation Systems Conference[C], Oakland, CA, USA, 2001:334～339.

The applicant's CN200510094306.9 video frequency motion target is cut apart and tracking: adopt the background subtraction method to cut apart moving target, carry out the extraction and the feature identification of target signature again, realize the tenacious tracking to target; Wherein, the model of background is the model of a real-time update, adopts based on the scene partitioning algorithm of s self-organizing feature map SOFM network and the merging of zonule mark and the mode that video motion information combines, and self-adaptation generates first frame background; During feature representation, adopt two kinds of expression-forms: parameter characteristic and Pixel-level spatial feature; During target following, adopt progressively meticulous matching strategy from coarse to fine,, realize identification of targets and tracking by thick coupling and thin coupling;

The mode that scene partitioning algorithm that described s self-organizing feature map (SOFM) neural network and zonule mark merge and video motion information combine is: bigger to network settings one earlier output classification number, it is trained, its image to input is carried out over-segmentation, adopting many tolerance Fuzzy Criteria of a kind of zonule mark to carry out iteration to each zone that splits then merges, obtains final segmentation result; The SOFM network is made up of two-layer, be respectively input layer and competition layer, the input layer number is n, the competition layer neuron is m, there are variable weights to be connected between each input node and output node, the input pattern of tieing up arbitrarily is mapped to one dimension or two-dimensional discrete figure in competition layer output, and keeps its topological structure constant;

With scene image data the SOFM network is trained, the a certain specific neuron that makes SOFM is to certain gray feature sensitivity, when image to be split is transfused to network, just can carry out self-organizing clustering according to the gray feature of data to data, finish image is carried out Region Segmentation; Zone many tolerance blurred picture merges: criterion

The image of exporting after SOFM neural network cluster is divided into the zone of a fixed number firmly, and each pixel belongs to a certain specific cluster according to its gray feature; Earlier image is carried out over-segmentation, extract by class then and carry out piecemeal and mark, obtain markd zonule, adopt many tolerance fuzzy criterions to merge to mark zonule again, obtained segmentation effect preferably.

Need to continue development and satisfy actual application.

Three, summary of the invention

The present invention seeks to propose a kind of be applied to motion target detection and recognition methods in the video monitoring system, can adapt to the moving object detection and the identification in video monitoring place under the different complex environments.

Motion target detection and recognition methods in a kind of video monitoring system, the moving object detection that combines with background subtraction based on the inter-frame difference of piecemeal and based on the pretreated support vector machine moving target identification of rough set; Employing is carried out in two steps based on the moving target detecting method that the inter-frame difference (two frames each sub-piece in front and back subtracts each other) of piecemeal combines with background subtraction, at first video image is carried out piecemeal, with the inter-frame difference method each sub-piece is carried out calculus of differences then; Realize moving image is carried out coarse segmentation according to difference result, isolate background area and foreground area; The background area is used for constructing and update background module; After the background image structure is finished, utilize background image that difference is carried out in the moving region that coarse segmentation in every two field picture goes out, the pixel that gray-scale value satisfies certain threshold value (threshold value is generally between 20～40) in the difference image is considered to the pixel of moving target, then all moving target pixels are carried out cluster, and utilize in the HSV space feature of shade to carry out Shadows Processing, finally obtain accurate movement target in the video image.

In the inventive method, moving object detection is carried out in two steps, at first with inter-frame difference moving image is carried out coarse segmentation, isolates background area and foreground area, and the background area is used for constructing and update background module.In case after the background structure is finished, just can adopt the background subtraction method to segment and cut, detect the accurate movement target to the moving region that coarse segmentation goes out.Background area structure and the renewal background of using inter-frame difference to obtain can incorporate background to the moving target that stops effectively, also can remove the target origin-location left " ghost " of unexpected motion in the scene simultaneously.Owing to be to carry out background subtraction in the moving region that coarse segmentation goes out, can effectively overcome ground unrest, segment the zone of handling when cutting simultaneously and dwindle greatly, so the integral operation amount of this algorithm is less.Cut apart during target segments cut at background subtraction, we separate shade according to the feature in the HSV space of shade.

Experimental result shows, this method overcomes inter-frame difference and extracts the not meticulous shortcoming of target, solved the difficulty that the background structure upgrades in the background subtraction method, can be to moving target in the monitoring scene be accurately detected, to light change, background interference is insensitive, has robustness and real-time preferably.

In the Target Recognition stage,, comprise that the focus point of the contour images of length breadth ratio, simple shape degree and a plurality of directions arrives the distance of image boundary according to the various characteristic informations of the profile extraction target that detects target.Utilize rough set theory then, the target signature of extracting is carried out yojan, extract the favourable feature of classifying.Adopt the support vector machine of multiclass at last, construct a multiple goal sorter, carry out quick identification detecting moving target.The advantage of this method is, in different application scenarios, under the situation at different video camera visual angles, can extract automatically the favourable target signature of classifying, and constructs effective support vector machine, and various particular surroundingss are had adaptability preferably.

Experimental result shows that this method can detect and Classification and Identification vehicle, pedestrian, crowd and the bicycle etc. that move rapidly and accurately in monitor video, illegal target invasion in the monitoring scene is reported to the police automatically.

Four, description of drawings

Fig. 1 is a moving object detection algorithm flow chart of the present invention

Fig. 2 is that the present invention carries out image coarse segmentation synoptic diagram

Fig. 3 is the Classification and Identification process flow diagram that the present invention is based on the moving target of support vector machine and rough set

Fig. 4 is that the contour shape of different target among the present invention is given an example

Fig. 5 is a characteristic component extracting method among the present invention

Fig. 6 is a multi-class support vector machine structure of the present invention, and wherein 1 represents pedestrian, 2 vehicles, 3 crowds, 4 bicycles

Fig. 7 is video image of the present invention and makes up the background experimental result

Fig. 8 is the result that coarse segmentation of the present invention and segmentation are cut

Fig. 9 is that the present invention removes target segmentation result behind the shade

Five, specific implementation

1 moving object detection that combines with background subtraction based on the inter-frame difference of piecemeal

1.1 method general introduction

Although frame difference method can effectively be removed static background, often the target of Ti Quing is more slightly made, and is bigger than the moving target profile of reality, and cavitation can occur in the target.And in the background subtraction method background modeling and context update are had relatively high expectations.Need set up background image fast, and background is upgraded in time, to guarantee that to illumination interference waits the environmental change adaptive faculty.We utilize the effectively advantage of separating background of frame difference method, utilize the frame difference method that background is carried out modeling.With the background subtraction method target being carried out essence again extracts.Concrete grammar is at first with inter-frame difference moving image to be carried out coarse segmentation, isolates background area and foreground area, and the background area is used for constructing and update background module.In case after the background structure is finished, just can adopt the background subtraction method to segment and cut, detect the accurate movement target to the moving region that coarse segmentation goes out.

Cut apart during target segments cut at background subtraction, we adopt the dual threshold mode that foreground area and background are reduced, and obtain the actual target area and the mixed zone of target shadow, utilize the position feature of shade again, and shade is separated.Detailed process can be referring to accompanying drawing 1.

The general coloured image that obtains represents that with rgb space RGB has critical role in computer realm, is widely used among computer graphical and the imaging.Yet when handling the image of real world, RGB is not very effective, because it all uses R, the G of isometric pixel, B three looks to be synthesized to all colors.This just makes each pixel have identical pixel depth and display resolution on R, G, three compositions of B.In addition, the vision system with the people is relative exactly, and human eyes have higher susceptibility to low frequency signal comparison high-frequency signal, and the many of sensitivity are wanted in the change that human eyes are also compared color to the change of legibility.For the needs of back Shadows Processing, we are converted into the HSV space to rgb space, and H represents chrominance space, and S represents the color saturation space, and V represents brightness space.

In the processing of back, if do not specify, all processing are all carried out in brightness V space.1.2 the foundation of background model and renewal

Set { fk (x, y) }, k=1,2 ... be an image sequence, (x, y) expression k two field picture is in that (x, gray-scale value y) when not producing ambiguity, also are used for representing the k two field picture to fk.In order to obtain the background image of scene, we at first adopt the consecutive frame difference to come the background image modeling.

Be every frame sequence image division m * n sub-piece at first,, divide each fritter of back and be expressed as S for example to the k two field picture _Ij ^k(x, y), i=1,2 .., m j=1,2 ..., n, promptly

$f_k(x,y)=\underset{i=1}{\overset{m}{U}}\underset{j=1}{\overset{n}{U}}{S}_{\mathrm{ij}}^k(x,y)---\left(1\right)$

When carrying out the inter-frame difference coarse segmentation, to two adjacent two field picture f _K-1(x, y), f _k(x, y) pairing sub-piece carries out difference, and obtains may be judged to be in the sub-piece pixel of prospect or background according to a threshold value T, and the number of statistics foreground pixel point.Promptly statistics satisfies formula

$|S_{\mathrm{ij}}^k(x,y)-S_{\mathrm{ij}}^{k-1}(x,y)|>T---\left(2\right)$

The number of pixel, be designated as MVCount.

$\mathrm{MVCount}=\mathrm{count}\left(\right\{(x,y)\∈S_{\mathrm{ij}}^k^|S_{\mathrm{ij}}^k(x,y)-S_{\mathrm{ij}}^{k-1}(x,y)|>T\left\}\right)---\left(3\right)$

If in sub-piece, the number M VCount that is judged to be the foreground point reaches certain threshold value δ, for example reaches 5% of all pixel numbers in the sub-piece, then we to be judged to be this sub-piece be the sport foreground piece, otherwise be the background piece.

Can see that such cutting apart slightly made, in fact under most of situation, the moving region amplified.Finally, we have obtained of image cuts apart roughly, sees accompanying drawing 2, and wherein (1) is the background area, and (2) are the moving region.

If k frame background image constantly is expressed as B _k(x, y), similar in appearance to the processing of prior figures picture frame, background image is also corresponding is divided into the sub-piece of m * n, and each sub-piece is expressed as B _Ij ^k(x, y), and when initial, setting all gray values of pixel points of background image is-1, i.e. B _o(x, y)=-1, the expression background is not initialised.According to the result of frame difference coarse segmentation, we use the sub-piece S that is judged as background _Ij ^k(x y) constructs or background image updating.Concrete rule is that if the corresponding sub-piece of background was not initialised, promptly its pixel value is-1, and is then directly alternative with the gray-scale value of the sub-piece that is judged to be background.If be initialised, then suitably upgrade with the sub-piece of current background.

$B_{\mathrm{ij}}^k(x,y)=\left\{\begin{array}{ccc}{S}_{\mathrm{ij}}^k(x,y)& \mathrm{if}& B_{\mathrm{ij}}^k(x,y)=-1\\ \mathrm{\α}{B}_{\mathrm{ij}}^k(x,y)+(1-\mathrm{\α})S_{\mathrm{ij}}^k(x,y)& \mathrm{if}& B_{\mathrm{ij}}^k(x,y)\≠-1\end{array}\right.---\left(4\right)$

Wherein 0≤α≤1 is a renewal rate, the renewal speed of expression background model, and α is big more, and the expression context update is slow more.Background model is not upgraded when α=1, when α=0, directly replaces respective regions in the background model with the sub-piece area grayscale of the background of present frame value.

When illumination variation is strong, should upgrade the variation of background as early as possible, we adopt the variation update strategy relevant with image averaging brightness.Rule of thumb we know, when ambient lighting changes, uniform when the brightness of background area changes in the general scene, so our variation by the mean value of background area brightness in the frame of the front and back variation that comes the perceived light photograph, the renewal rate α of background also regulated simultaneously according to this information.

If Lk is the sub-piece S that has powerful connections in the k two field picture _Ij ^k(x, y) mean value of middle grey scale pixel value is represented the brightness of k two field picture background.

$L_k=\frac{1}{N}\underset{(x,y)\∈S_{\mathrm{ij}}^k}{\mathrm{\Σ}}{S}_{\mathrm{ij}}^k(x,y)\mathrm{if}{S}_{\mathrm{ij}}^k\∋\mathrm{background}---\left(5\right)$

N for the number of pixel in the sub-piece of having powerful connections.

And the mean flow rate of the sub-piece of background of former frame correspondence is designated as L _K-1

$\mathrm{\α}=\mathrm{\η}(1-\frac{|L_k-L_{k-1}|}{\max (L_k,L_{k-1})})---\left(6\right)$

0≤η≤1st, a constant of setting can be arranged in 0.05～0.4 scope.When illumination variation was strong, α diminished as can be seen, and context update is also fast.

Those are judged to be the zone of prospect, and the corresponding sub-piece of background remains unchanged.

Along with moving of target, target background B _k(x, y) the sub-piece of in all finally all is initialised, and also is successfully constructed with regard to the presentation video reference background.

As can be seen, by inter-frame difference, our substantial separation background area and moving region, the noise of the background of non-moving region has been carried out effective compacting.Dwindled the scope of target detection, good help has been cut in the target segments of back.

Because inter-frame difference can't detect static object, static target can be taken as background, therefore this method of subtracting each other in conjunction with front and back two frames is constructed and background image updating, can effectively be included in the static then moving target of setting in motion in the background image model, also can solve smoothly since target by the static phenomenon that stays " ghost " in the motion place background image that changes into.

1.3 moving target extracts

In case the background structure is finished, and just can carry out the segmentation of background subtraction and cut.Because the antithetical phrase piece carries out difference when obtaining the sub-piece of prospect, may be taken as the background piece to the point at moving target edge and lose, therefore, we do some expansions a little to the moving region, the next-door neighbour is judged as the sub-piece of foreground blocks,, also is labeled as foreground blocks as long as comprise with respect to threshold value δ foreground point still less, front, at last, we carry out difference to the sub-piece of these prospects and its corresponding background.Construct the binaryzation template M of the moving region of the sub-piece of each prospect of k two field picture _Ij ^k(x, y).

$M_{\mathrm{ij}}^k(x,y)=\left\{\begin{array}{cc}1& |S_{\mathrm{ij}}^k(x,y)-B_{\mathrm{ij}}^k(x,y)|>T\\ 0& \mathrm{otherwise}\end{array}---\left(7\right)\right.$

T is a segmentation threshold, can rule of thumb obtain, and generally between 20～40, also can obtain by the OSTU method.The binaryzation template M of these all moving regions _Ij ^k(x y) carries out sub-piece and merges, and is last, obtained the moving target binaryzation template M of whole k two field picture _k(x y), carries out and operation with this template and original image, just can extract moving target.

1.4 moving target Shadows Processing

Shade mainly causes owing to light source (as sunlight, light etc.) is blocked by object.In scene, when light source was strong, the shade of moving target will occur, and moves along with the motion of target.The process of the target detection by the front as can be seen because shade has changed pixel intensity in the background image, so shade meeting and the target of moving are taken as the sport foreground zone together and detect.

In order to obtain accurate target, also must be shadow removal.Visual signature according to shade, the shadow region can be counted as translucent zone, when background dot is covered by shade, its brightness value diminishes, and the chromatic value size remains unchanged substantially, and during background dot passive movement target coverage, its brightness value may become and greatly also may diminish, but chromatic value generally alters a great deal.According to these features, this paper further detects processing to detected target.

Those are judged as the point of moving target, are judging, remove shadow spots according to following formula ^[8]

$M_{\mathrm{ij}}(x,y)=\left\{\begin{array}{cc}0& \mathrm{if}({\mathrm{\&beta;}}_1\&le;\frac{S_{\mathrm{ij}}^Y(x,y)}{B_{\mathrm{ij}}^Y(x,y)}\&le;{\mathrm{\&beta;}}_2)^|S_{\mathrm{ij}}^S(x,y)-B_{\mathrm{ij}}^S(x,y)|\&le;{\mathrm{\&Gamma;}}^S^|S_{\mathrm{ij}}^H(x,y)-B_{\mathrm{ij}}^H(x,y)|\&le;{\mathrm{\&Gamma;}}^{\mathrm{<figure-callout\; id="1"\; label="H"\; filenames="A20071002006700121.png,A20071002006700141.png"\; state="\{\{state\}\}">H</figure-callout>}}\\ 1& \mathrm{otherwise}\end{array}---\left(8\right)\right.$

M _Ij(x is the sub-piece of corresponding two-value template of finally removing the moving target of shade y), and subscript V, S, H distinguish the value of presentation video in the HSV space, β in the following formula ₁, β ₂, Γ ^S, Γ ^HBe respectively threshold value, β ₁＜β ₂＜1, their value need be in the test decision.

After the segmentation of process background difference is cut, still comprise possible noise spot in the binaryzation template of the moving target that obtains.Simultaneously because the gray-scale value difference of background and some position of moving target may differ very little, may exist cavitation in the moving target binaryzation template that obtains at last, therefore also must carry out subsequent treatment such as morphological operation, just can obtain real accurate movement To Template.

Because foreground area may be to comprise a plurality of moving targets, adopt clustering method, extract each movement destination image respectively.

2 Target Recognition based on rough set and support vector machine

2.1 method general introduction

In the Target Recognition stage, the present invention utilizes rough set theory then according to the various characteristic informations of the profile extraction target that detects target, and the target signature of extracting is carried out yojan, extracts the favourable feature of classifying.Adopt support vector machine at last, carry out quick identification detecting moving target.Detailed process as shown in Figure 3.

2.2 feature extraction

Target signature is the foundation of identification, and the selection of target signature is the committed step of carrying out target classification identification.The selection of feature is very big to the Target Recognition influence, and good feature can be easy to target is made a distinction, and the little feature of discrimination tends to cause the False Rate of target to increase, even can't discern.For example in the traffic monitoring scene, the employing color characteristic distinguishes vehicle and single pedestrian is inappropriate, and employing area features and length breadth ratio are relatively good.But area features and length breadth ratio can not fine differentiation vehicle and crowds, just must seek better target signature.Target signature selection at present generally is rule of thumb to carry out.

According to the appearance profile of moving target moving target being carried out the Classification and Identification conventional method is, at first obtain the two-value template of moving target by method for testing motion, also just get access to simultaneously the outline model of target, extract the feature of moving target outward appearance according to the outline model, the area of target for example, length breadth ratio etc.According to these characteristic informations, differentiate then in conjunction with experimental knowledge.

Some target signatures commonly used are described.

(1) length breadth ratio feature

The length breadth ratio of target is meant the length breadth ratio of the boundary rectangle that comprises target, long is meant the shared length of target on the y direction of principal axis, widely is meant the shared length of target on the x direction of principal axis, with formula (9) expression.

$f_{h/w}=\frac{|y_{\mathrm{botom}}-y_{\mathrm{top}}|}{|x_{\mathrm{right}}-x_{\mathrm{left}}|}---\left(9\right)$

(2) Qu Yu simple shape degree

The simple degree of shape can represent that wherein, P represents the girth in zone with formula (10), and A represents the area in zone.

$f_c=\frac{4\mathrm{\&pi;A}}{P^2}---\left(10\right)$

More than several are target geometric properties relatively commonly used.Under some occasion, these features often can't be finished identification of targets work.For example, target has a strong impact on the use of area features from the pick-up lens far and near distance, when a pedestrian when camera lens is nearer than automobile, often the area of pedestrian's target is also bigger than automobile.And crowd's the length breadth ratio that comprises many people is just similar with vehicle, sees Fig. 4.Therefore need to seek better feature and go to distinguish different targets.

By analysis as can be seen, the area features that the front is introduced, length breadth ratio features etc. do not reflect the contour shape characteristics of different target well.Therefore for the better contour shape characteristics that embody target, we must adopt the characteristic parameter of the shape of the fine reflection target of energy.We adopt a kind of heart-shaped configuration feature as shown in the figure among the present invention for this reason, and concrete grammar is successively every certain angle $\mathrm{\&alpha;}=\frac{2\mathrm{\&pi;}}{d},$ Get distance as a characteristic component r from the focus point of objective contour image to image boundary _i(i=1,2 ..., d).These components radially distribute at interval, targeted outline line.If certain vector passes the border of objective contour more than once, then chosen distance farthest, the length that each is vectorial is expressed as the proper vector of target shape in order.See accompanying drawing 5.

For fear of the feature difference that target causes from the camera lens distance, we carry out normalization to above-mentioned feature, promptly

$r_i^{*}=\frac{r_i}{\underset{j=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,d}{\max }\left\{r_j\right\}}---\left(11\right)$

For convenience, we are r _i ^*Still note is r.Do normalization and handle, make proper vector have flexible unchangeability.As can be seen, proper vector is only relevant with the profile of target, might can not impact the clarification of objective vector in the target internal cavity in the target leaching process from the extraction process of proper vector.

We are the length breadth ratio of introducing previously, complex boundary degree feature and expression objective contour shape facility r _i(i=1,2 ..., d) together as the target classification feature, note is done

F＝(f _h/w，f _c，r ₁，r ₂，...，r _d) (12)

This proper vector dimension is d+2.

2.3 feature reduction

In the target signature that is adopted, may there be redundancy in different occasions, this structure and the identification in later stage influence to sorter is very big.Therefore we at first utilize the rough set theory method of introducing previously that target signature is carried out yojan.This yojan has two benefits, and the one, simplified intrinsic dimensionality, the secondth, deleted unnecessary useless sample.

We are expressed as 1,2 to four class targets of needs classification, 3,4 respectively.The 1 single pedestrian of expression, 2 expression vehicles, 3 expression crowds, 4 expression bicycles.From comprise n sample set, extract feature then, construct an information table

Utilize then and slightly make diversity method this information table is carried out feature reduction, finally obtain an information table after the yojan, comprised less characteristic component and sample data, i.e. m≤n, and k≤d+2 in this table.

2.4 the structure of support vector machine

With the structure of the sample data after rough set method yojan support vector machine.Dividing pedestrian, vehicle, crowd, bicycle 4 classes with target is example, and we adopt DDAG-SVMs method construct sorter among the present invention ^[7]Specific practice is to construct 6 sorters successively, constitutes multicategory classification device as shown in Figure 6.Sorter adopts radial basis function, σ ²=0.5.

When training, these support vector machine only need the relative sample of part, for example concerning support vector machine SVMs2-3, only need vehicle and crowd's sample data to construct.

In case after all vector machine structures are finished, enter the Classification and Identification stage, any target must just can obtain correct result through 3 subseries.

3 experimental result explanations

We adopt one section resolution is that 320 * 240 highway video HighwayI is AMD3000 at storer, in save as on the PC of 512MB and do experiment.

In the experiment, every frame video image is divided into 32 * 24,100 pixels of promptly every block size.The background of image has been set up and has been finished, and the background of foundation is very clean, sees Fig. 7.

Fig. 8 is that the result is cut in result and the segmentation that the coarse segmentation that adopts this paper method to be carried out roughly extracts the moving region of moving target, and Fig. 9 is moving target two-value template and a moving target of removing shade.

Following table be adopt the inventive method effect that people, vehicle, crowd and bicycle are discerned.

The target classification	Detect number of times	Correct number of times	Accuracy	The erroneous judgement number of times	False Rate
						1-pedestrian	63	60	95.24％	3	4.76％
The 2-vehicle	82	79	96.34％	3	3.66％
						3-crowd	47	43	91.49％	4	8.51％
The 4-bicycle	40	38	95.00％	2	5.00％
						Add up to	232	220	94.83％	12	5.17％

Involved in the present invention is during video image handles the moving object detection and the recognition technology of basic most critical play a part very crucial in the video monitoring technology.The present invention can be widely used in fields such as intelligent transportation, security protection, military affairs.Also can be applied to fields such as object-based video coding, video frequency searching, video conference, machine vision.

Claims (4)

1. motion target detection and recognition methods in the video monitoring system, it is characterized in that the moving object detection that the inter-frame difference based on piecemeal combines with background subtraction: adopt inter-frame difference based on piecemeal, each sub-piece of two frames subtracts each other promptly, the moving target detecting method that combines with background subtraction, carry out in two steps, at first video image is carried out piecemeal, with the inter-frame difference method each sub-piece is carried out calculus of differences then; Realize moving image is carried out coarse segmentation according to difference result, isolate background area and foreground area; The background area is used for constructing and update background module; After the background image structure is finished, utilize background image that difference is carried out in the moving region that coarse segmentation in every two field picture goes out, gray-scale value satisfies certain threshold value in the difference image, threshold value is between 20～40, the pixel that satisfies threshold value is considered to the pixel of moving target, then all moving target pixels are carried out cluster, and utilize in the HSV space feature of shade to carry out Shadows Processing, finally obtain accurate movement target in the video image.

2. motion target detection and recognition methods in the video monitoring system according to claim 1, it is characterized in that recognition methods is when cognitive phase based on the pretreated support vector machine moving target of rough set, according to the various characteristic informations of the profile extraction target that detects target, comprise that the focus point of the contour images of length breadth ratio, simple shape degree and a plurality of directions arrives the distance of image boundary.These characteristic use rough set theories that obtain carry out yojan, extract the favourable feature of classifying; Adopt the support vector machine of multiclass at last, construct a multiple goal sorter, carry out quick identification detecting moving target.

3. motion target detection and recognition methods in the video monitoring system according to claim 1 is characterized in that in the Target Recognition stage, adopted the heart-shaped configuration characteristic recognition method, successively every certain angle $\mathrm{\&alpha;}=\frac{2\mathrm{\&pi;}}{d},$ Get from objective contour

The focus point of image arrives the distance of image boundary as a characteristic component r _i(i=1,2 ..., d); These components radially distribute at interval, targeted outline line; If certain vector passes the border of objective contour more than once, then chosen distance farthest, the length that each is vectorial is expressed as the proper vector of target shape in order; And above-mentioned feature carried out normalization, promptly

$r_i^{*}=\frac{r_i}{\underset{j=\mathrm{1,2},...,d}{\max }\{r_j\}}$

R _i ^*Still note is r.Do normalization and handle, make proper vector have flexible unchangeability; As can be seen, proper vector is only relevant with the profile of target, might can not impact the clarification of objective vector in the target internal cavity in the target leaching process from the extraction process of proper vector; The length breadth ratio of introducing previously, complex boundary degree feature and expression objective contour shape facility r _i(i=1,2 ..., d) together as the target classification feature, note is done

F＝(f _h/w,f _c,r ₁,r ₂,...，r _d)

This proper vector dimension is d+2.Utilize the rough set theory method of introducing previously that target signature is carried out yojan, respectively four class targets of needs classification are expressed as 1,2,3,4:1 represents single pedestrian, 2 expression vehicles, 3 expression crowds, 4 expression bicycles; From comprise n sample set, extract feature then, construct an information table, utilize then and slightly make diversity method this information table is carried out feature reduction, information table after yojan of final acquisition, a less k characteristic component and m sample data have been comprised in this table, i.e. m≤n, and k≤d+2.

4, motion target detection and recognition methods in the video monitoring system according to claim 1 is characterized in that with sample data structure support vector machine after the rough set method yojan; Adopt DDAG-SVMs method construct sorter; Sorter adopts radial basis function, σ ²=0.5; When training, support vector machine only needs the relative sample of part, in case after all vector machine structures are finished, enter the Classification and Identification stage, arbitrary target must obtain correct result through 3 subseries.

Images