CN100525395C - Pedestrian tracting method based on principal axis marriage under multiple vedio cameras - Google Patents

Abstract

The present invention relates to the technical field of computer vision, a pedestrian tracking method based on main axis matching under multi-camera, comprising the steps of: performing motion detection on sequence images; extracting the main axis features of people; performing tracking under single camera; establishing according to the main axis matching function Axis matching pair; fusion of multi-view information to optimize and update the tracking results. The invention proposes a new type of multi-camera matching technology, which overcomes the shortcomings of the traditional multi-camera matching method that requires calibration, adopts features that are easily affected by noise and viewing angle factors, and relies on accurate segmentation algorithms, and has a good application prospect.

Description

The pedestrian tracting method that mates based on main shaft under the multiple-camera

Technical field

The present invention relates to technical field of computer vision, particularly the multiple-camera matching process.

Background technology

Along with the development of society, it is important that safety problem becomes day by day.Vision monitoring becomes the requirement in a lot of large-scale places.These places comprise security department, the military affairs that country is important.Meanwhile in a lot of public places, it is most important that vision monitoring also becomes, and wherein typical application is a field of traffic.Because traditional visual monitor system all is a kind of monitoring afterwards, all accidents and abnormal conditions all are that the security personnel finds at that time or afterwards, and the behavior of various breach securitys is not had the prevention effect; Therefore traditional supervisory control system can not adapt to demands of social development.At this moment, intelligentized visual monitor system becomes the requirement in epoch and researcher's target.Intelligentized vision monitoring final purpose is to realize the behavior of Real time identification moving target, thereby abnormal behaviour is in time made judgement and warning.And depend on bottom layer treatment, i.e. the tracking of moving target with the solution of analyzing the contour level problem as the behavior identification of moving target.In monitoring scene, people's motion is the focus of our concern often.In addition, people's motion is a kind of non-movement of Rigid Body, and it is complicated more that other motion tracking (such as the motion tracking of vehicle) is compared in its tracking.Therefore people's tracking becomes vital part of Intellectualized monitoring.

Nowadays, people are more and more higher to the intellectuality and the security requirement of visual monitor system, use single camera that the people who moves is followed the tracks of far away and can not satisfy needs of society.In the last few years, utilized multiple-camera to come the people of pursuit movement to become the research focus.

Multiple-camera servant's tracking is the problem of a multiple-camera coupling in essence, promptly sets up the corresponding relation between the moving object under the different visual angles (video camera) at synchronization.The multiple-camera coupling is newer problem of computer vision.In this respect, abroad some universities and research institution (Maryland, Oxford MIT) have carried out some research work.Type according to selected feature can roughly be divided into related work two big classes: based on the method in zone with based on the method for characteristic point.

Matching process based on the zone is that the people is regarded as a moving region, utilizes the feature of moving region to set up different visual angles servant's corresponding relation.In these class methods, color is a most frequently used provincial characteristics.Orwell etc. and Krumm etc. utilize color histogram to estimate that people's field color distributes, and set up the coupling of multiple-camera by comparing color histogram; Mittal etc. have set up Gauss's color model for people's moving region, utilize these modelling couplings then; Another method is to utilize the distribution of cuclear density Function Estimation field color, sets up coupling on this basis.Though color characteristic is more directly perceived, however its robust not very aspect coupling.This be because, depend on the color of people's clothes based on the coupling of color, when two people's clothes color is identical, may produce wrong coupling; Secondly, owing to be subjected to the influence at illumination and visual angle, observed same individual's clothes color may change.

Method based on characteristic point is to regard the people as a series of characteristic point, and different visual angles servant's coupling just is converted into the coupling of characteristic point.The coupling of characteristic point is based on certain geometrical constraint.According to the difference of selected geometrical constraint, these class methods can be divided into two subclasses: three-dimensional method and two-dimension method.A.Utsumi chooses the barycenter of moving object as characteristic point, compares their tripleplane's point then; Q.Cai chooses some characteristic points of people's upper part of the body center line, and limit of utilization retrains and seeks coupling then.Three-dimensional method all needs video camera is demarcated.When employed video camera number is very big in the monitoring scene, camera calibration will be a huger task.In order to overcome the shortcoming of three-dimensional method, some researcher has proposed to utilize two-dimensional signal to set up coupling between the multiple-camera.Khan etc. utilize the constraint (homography constraint) of singly reflecting of ground level to come the characteristic point on the match people pin.But these characteristic points are subjected to The noise easily.Block or to detect effect bad when existence, when the people had only part to be observed, the characteristic point of extraction is robust not, and therefore the performance of coupling will variation.

Worldwide obtained paying close attention to widely and studying though it is emphasized that multiple-camera servant's tracking, still had many difficult points aspect the multiple-camera matching technique.Therefore how to choose more robust, feature is mated and is still a challenge accurately.

Summary of the invention

The objective of the invention is to avoid conventional method to need calibration, adopt feature to be subject to shortcomings such as noise and visual angle factor affecting, provide a kind of multiple-camera matching process simple, robust to be used for multiple-camera servant's tracking.

For achieving the above object, the pedestrian tracting method based on the main shaft coupling comprises step under the multiple-camera:

(1) moving Object Segmentation;

(2) extraction people's main shaft feature;

(3) tracking under the single camera;

(4) it is right to seek all optimum Match according to main shaft match likelihood function;

(5) merge various visual angles information updating tracking results.

Multiple-camera is meant that using two or more single cameras, the precondition of this method is to have a public plane in the viewing area of hypothesis different cameras, and generally speaking, this public plane is meant ground level.

The present invention extracts people's main shaft as feature, is a kind of novel multiple-camera matching process based on main shaft, has good application prospects.

Description of drawings

Fig. 1 is single people's main shaft detection example figure.

Fig. 2 is group's main shaft detection example figure.

Fig. 3 is the main shaft detection example figure of the people under the situation of blocking.

Fig. 4 is single camera servant's a tracking block diagram.

Fig. 5 is multiple-camera servant's main shaft projection relation figure.

Fig. 6 (a) is many people tracking test figure as a result under two video cameras of NLPR database.

Fig. 6 (b) is many people tracking test figure as a result under three video cameras of NLPR database.

Fig. 6 (c) is to the tracking test of single people under the situation of blocking figure as a result under two video cameras of PETS2001 database.

Fig. 6 (d) is to the tracking test of group under the situation of blocking figure as a result under two video cameras of PETS2001 database.

Fig. 7 is based on the general introduction figure of the pedestrian tracting method of main shaft coupling under the multiple-camera.

Embodiment

Main feature of the present invention is: 1) extract the feature of people's main shaft as coupling.Because people's main shaft is the symmetry axis of human region, according to symmetry, the point that is symmetrically distributed in the main shaft both sides their error of can cancelling each other, thus make main shaft robust more.And people's main shaft is subjected to motion detection and the result of cutting apart to influence less; 2) detection method of three kinds of different situations servants' main shaft has been proposed: the detection that promptly single people's main shaft detects, group's main shaft detects and blocks the main shaft of the people under the situation; 3) based on the geometrical relationship constraint of different visual angles lower main axis, defined main shaft match likelihood function and weighed the right similarity of different visual angles lower main axis; 4) merge various visual angles information according to matching result, tracking results is optimized renewal.

The general frame of scheme is seen accompanying drawing 7.At first the image sequence under the single camera is carried out motion detection, extract people's main shaft feature, carry out the tracking under the single camera; Then according to singly reflect relation constraint to the main shaft under the different visual angles to mating; Merge various visual angles information updating tracking results according to matching result at last.

Provide the explanation of each related in this invention technical scheme detailed problem below in detail.

(1) moving Object Segmentation

Moving Object Segmentation is the first step of motion tracking, and algorithm adopts is that the background method of wiping out of single Gauss model detects the moving region.In order to reduce the influence of illumination variation and shade, adopt normalized color rgs model, r=R/ (R+G+B) wherein, g=G/ (R+G+B), s=R+G+B.Earlier one section background image sequence is carried out filtering, obtain single Gauss's background model.Gauss's parameter of each point is U wherein _iAnd σ _i(i=r, g s) are the average and the variance of this background model respectively.Then present image and background image are carried out difference processing, can obtain binary image M in that difference image is carried out threshold process _Xy, the foregrounding pixel value is 1 in bianry image, background pixel value is 0.Its process is:

$M_{\mathrm{xy}}=\left\{\begin{array}{cc}0& \mathrm{otherwise}\\ 1& |I_i(x,y)-u_i(x,y)|>{\mathrm{\α}}_i{\mathrm{\σ}}_i(x,y),i\∈\{r,g,s\}\end{array}\right.---\left(1\right)$

I wherein _i(x, y), (i=r, g s) are pixel (x, current measured value y).α _iBe threshold parameter, can in experiment, determine by experience.After obtaining bianry image, it is being carried out corrosion and the next further filtering noise of expansive working in the morphological operator.Connected component analysis by binaryzation at last is used to extract simply connected foreground area, as the moving region after cutting apart.

(2) detection of people's main shaft

For simplicity, suppose that the people is upright walking, and people's main shaft exists.Suppose that based on this we discuss the detection method of three kinds of people's under the different situations main shaft, promptly single people's main shaft detects; Group's main shaft detects; Block the detection of situation lower main axis.These three kinds of situations can be judged according to corresponding relation in following the tracks of.

Single people's main shaft detects.For single people, adopt minimum intermediate value quadratic method (LeastMedianSequence) to detect its main shaft.Suppose i foreground pixel X _iTo the vertical range for the treatment of boning out l is D (X _i, l), according to minimum intermediate value quadratic method, then all foreground pixels are to square intermediate value minimum of the vertical range of main shaft.

$L=\arg \underset{i}{\min }{\mathrm{median}}_i\left\{D{(X_i,l)}^2\right\}---\left(2\right)$

Accompanying drawing 1 has provided single people's main shaft detection example.

Group's main shaft detects.Group's main shaft detection comprises that individuality is cut apart and main shaft detects the two large divisions.

For individuality is cut apart, introduce the upright projection histogram, the individuality among the group is corresponding to the histogrammic peak region of upright projection.Have only the peak region that satisfies certain condition just corresponding to single individuality, we call the obvious peak value zone to these peak regions.Two conditions must be satisfied in the obvious peak value zone:

A) maximum of peak region must be greater than a certain specific threshold value, peak threshold P _T

B) minimum value of peak region must be less than a certain specific threshold value, trough threshold value C _T

Suppose in a peak region P ₁, P ₂...., P _nIt is its local extremum.C _l, C _rBe respectively this regional left and right sides trough value, then above two conditions can be expressed as with mathematic(al) representation:

max(P ₁，P ₂，.....，P _n)>P _T (3)

C _l<C _T，C _r<C _T (4)

Trough threshold value C _TBe chosen for whole histogrammic average, crest threshold value P _TBe chosen for people in image coordinate height 80 percent.The method that in the second step main shaft detects, can adopt above-mentioned single people's main shaft to detect.

Fig. 2 has provided the example that a plurality of people's main shafts detect.(b) being detected foreground area, (c) is its upright projection histogram, (d) is the result after cutting apart.In upright projection histogram (c), three tangible peak regions are arranged, can be that three parts correspond respectively to three individualities thus with this Region Segmentation.At last, their main shaft provides in figure (e).

The detection of the main shaft of the people under the situation of blocking.Earlier the people is split, find out the pixel of its foreground area, detect people's main shaft in method the minimum intermediate value square error of foreground pixel utilization that splits.We adopt and realize cutting apart under the situation of blocking based on the color template method.This model comprises a color model and an additional probability mask.When occurring blocking, cutting apart of moving object can be expressed as classification problem, judges promptly which model which foreground pixel belongs to.This problem can solve with bayes rule.If satisfy following equation:

$k=\arg \underset{i}{\max }{P}_i\left(x\right)---\left(5\right)$

Then foreground pixel X belongs to k moving object (model).

The main shaft that blocks under the situation detects as shown in Figure 3.

(3) tracking under the single camera

Adopt Kalman filter to realize following the tracks of

Choose state vector X comprise the position of people in image (x, y) and movement velocity (v _x, v _y), observation vector Z be the people the position (x, y).The measured value of people's movement position (x y) is estimated as people pin point (" land-point in image "), the i.e. intersection point of the lower limb of people's main shaft and rectangle frame.Then:

Z _t＝[x _t，y _t] ^T X _t＝[x _t，v _x，t，y _t，v _y，t] ^T (6)

The speed of people's walking can be thought constant approx, and state-transition matrix and observing matrix are:

$\mathrm{\&Phi;}=\left[\begin{array}{cccc}1& \mathrm{\&Delta;t}& 0& 0\\ 0& 1& 0& 0\\ 0& 0& 1& \mathrm{\&Delta;t}\\ 0& 0& 0& 1\end{array}\right]$ $H=\left[\begin{array}{cccc}1& 0& 0& 0\\ 0& 0& 1& 0\end{array}\right]---\left(7\right)$

By the position of the position prediction present frame object of former frame object, the measured value (intersection point of people's main shaft and rectangle frame lower limb) with present frame compares then.If the distance between measured value and the predicted value is very little, show that then measured value is with a high credibility, measured value can be represented the position of present frame object; If both distances surpass certain threshold value (such as people's the Lower Half this situation that is blocked), then measured value is insincere, and the intersection point that we have defined main shaft and another line upgrades measured value, and this line is exactly the vertical line of future position to main shaft.Tracking framework under the whole single camera can be referring to accompanying drawing 4.

(4) it is right to seek all best main shaft couplings

At first need to calculate and singly reflect matrix between the different images plane.Singly reflected matrix description between the different images about the one-to-one relationship of the point on the same plane.Can find the solution the parameter of matrix by simultaneous equations by given corresponding points.The corresponding points of this method are obtained the employing manual mode: promptly in advance index point is set in scene or utilizes some special corresponding points in the scene.

Define main shaft match likelihood function then.

Main shaft match likelihood function is used for weighing the matching degree of different visual angles servant's main shaft.Before providing concrete definition, the geometrical relationship of different visual angles lower main axis can be referring to accompanying drawing 5.As shown in the figure, suppose to have two video camera i and j.Suppose that cameras view to the main shaft of people s is

,

Be

Projection on ground level.Observing people k for video camera j has accordingly With The video camera i plane of delineation to the matrix that singly reflects of the video camera j plane of delineation is

We will by singly reflecting matrix

Project in the plane of delineation coordinate system of video camera j and can obtain

With

Will intersect at a point

According to the character of singly reflecting matrix, if the people k that people s that video camera i observes and video camera j observe is corresponding to the same individual in the three dimensions, then

Corresponding to this people's " land-point ", the i.e. intersection point of people's main shaft and ground level.Therefore, the measured value of " land-point " and intersection point

Between distance can be used for weighing matching degree between the main shaft.Distance is mated between the bright main shaft of novel more more.

According to the geometrical relationship of different visual angles lower main axis, the match likelihood function between definition people s and the k main shaft is

$L(L_s^i,L_k^j)=p\left(X_s^i\right|Q_{\mathrm{ks}}^{\mathrm{ji}})p\left(X_k^j\right|Q_{\mathrm{sk}}^{\mathrm{ij}})---\left(8\right)$

Wherein

Be " land-point " that video camera i observes people s,

Be " land-point " that video camera j observes people k It is main shaft s is transformed into j visual angle and main shaft k from the i visual angle intersection point.

In order to be without loss of generality, suppose top two probability density functions clothes cloth, then:

$p\left(X_s^i\right|Q_{\mathrm{ks}}^{\mathrm{ji}})=2\mathrm{\&pi;}{\left(\left|{\mathrm{\&Sigma;}}_s^i\right|\right)}^{-1/2}\exp \{-\frac{1}{2}(X_s^i-Q_{\mathrm{ks}}^{\mathrm{ji}}){\left({\mathrm{\&Sigma;}}_s^i\right)}^{-1}{(X_s^i-Q_{\mathrm{ks}}^{\mathrm{ji}})}^T\}$

$p\left(X_k^j\right|Q_{\mathrm{sk}}^{\mathrm{ij}})=2\mathrm{\&pi;}{\left(\left|{\mathrm{\&Sigma;}}_k^j\right|\right)}^{-1/2}\exp \{-\frac{1}{2}(X_k^j-Q_{\mathrm{sk}}^{\mathrm{ij}}){\left({\mathrm{\&Sigma;}}_k^j\right)}^{-1}{(X_k^j-Q_{\mathrm{sk}}^{\mathrm{ij}})}^T\}---\left(9\right)$

At last find all optimum Match right according to matching algorithm

The coupling of multiple-camera in fact can modeling becomes the problem of maximum likelihood function, and the promptly corresponding mutually right main shaft match likelihood function of main shaft is maximum in the right main shaft match likelihood function of numerous main shafts.For the simplification problem, we have defined the main shaft matching distance, and the maximum likelihood function problem is converted into the minimal matching span problem. $\arg \underset{s,k}{\max }L(L_s^i,L_k^j)\&DoubleLeftRightArrow;\arg \underset{s,k}{\min }{D}_{\mathrm{sk}}^{\mathrm{ij}}$

Wherein

Be the main shaft matching distance, definition main shaft matching distance as shown in the formula:

$D_{\mathrm{sk}}^{\mathrm{ij}}=(X_s^i-Q_{\mathrm{ks}}^{\mathrm{ji}}){\left({\mathrm{\&Sigma;}}^i\right)}^{-1}{(X_s^i-Q_{\mathrm{ks}}^{\mathrm{ji}})}^T+(X_k^j-Q_{\mathrm{sk}}^{\mathrm{ij}}){\left({\mathrm{\&Sigma;}}^j\right)}^{-1}{(X_k^j-Q_{\mathrm{sk}}^{\mathrm{ij}})}^T---\left(10\right)$

The main shaft distance

More little, then main shaft mates each other more.

The main shaft matching algorithm is right in order to seek global best matches, makes theirs

The sum minimum.Main shaft matching algorithm between two video cameras is described below:

Suppose under video camera i, detect M main shaft:

Under video camera j, detect N main shaft: $L_1^j,L_2^j,.....,L_N^j.$

Step 1: detected main shaft under two visual angles is made up in twos, form the main shaft that might mate right.Be without loss of generality, suppose M≤N, then M main shaft selected M main shaft successively according to priority and matched formation altogether from N main shaft Plant combination.Each combination has following form:

${\mathrm{\&theta;}}_k=\{(L_1^i,L_{k1}^j),(L_2^i,L_{k2}^j).......,(L_M^i,L_{\mathrm{kM}}^j)\},k=1\&CenterDot;\&CenterDot;\&CenterDot;\&CenterDot;\&CenterDot;\&CenterDot;P_N^M;---\left(11\right)$

Step 2: for each the combination in each main shaft to { m, n} calculate its matching distance

And guarantee $D_{\mathrm{mn}}^{\mathrm{ij}}<D_T,$ D wherein _TBe the threshold value that draws by experience, be used for judging that main shaft is to { m, whether n} mates.If do not satisfy above-mentioned constraint, then main shaft to { m, n} is from θ _kIn delete.Step 3: choose the maximum combination Θ of coupling logarithm l _k, all Θ _kForm set Θ:

$\mathrm{\&Theta;}=\{{\mathrm{\&Theta;}}_k=(L_{k_1}^i,{\mathrm{<figure-callout\; id="1"\; label="L\; k"\; filenames="C200510108137E00171.png,C200510108137E00181.png"\; state="\{\{state\}\}">L</figure-callout>}}_{k_{}^{}}^{{}_1^{\&prime;}}),(L_{k_2}^i,{\mathrm{<figure-callout\; id="2"\; label="L\; k"\; filenames="C200510108137E00181.png,C200510108137E00191.png"\; state="\{\{state\}\}">L</figure-callout>}}_{k_{}^{}}^{{}_2^{\&prime;}}),...,(L_{k_l}^i,L_{k_l^{\&prime;}}^j)\}$ Wherein $k\&Element;P_N^M---\left(12\right)$

Step 4: in set Θ, seek global best matches combination λ, feasible wherein all matching distance Sum reaches minimum value, promptly satisfies following formula:

$\mathrm{\&lambda;}=\arg \underset{k}{\min }\left({\mathrm{\&Sigma;}}_{w=1}^l\left(D_{(k_w,k_w^{\&prime;})}^{(i,j)}\right)\right)---\left(13\right)$

Step 5: the Θ that finally obtains _λThen be the global best matches combination, wherein each is to then right for the main shaft of coupling.

Above-mentioned algorithm is easy to expand to the situation more than two video cameras.At first video camera makes up in twos, for a pair of video camera that public ground level zone is arranged, sets up matching relationship and can adopt the algorithm of introducing above; When if coupling produces contradiction between the video camera in twos, it is right that then matching relationship only considers to have the main shaft of minimal matching span.

(5) merge various visual angles information updating tracking results

When find all coupling main shafts to after, these match information just can be used for upgrading the tracking results under the single camera.At the situation of two video cameras, have only when the people who follows the tracks of is in public ground level zone, two visual angles, it is just effective to upgrade this step.

Suppose to find corresponding under same individual two visual angles by matching algorithm above-mentioned, the main shaft that is visual angle i and visual angle j respectively is right, main shaft under the j of visual angle is transformed in the plane of delineation of visual angle i by the relation of singly reflecting between two planes of delineation, main shaft under the then original visual angle i is exactly the position of final this person in plane of delineation i with the intersection point of the straight line that conversion is come, and is used for upgrading the tracking results under the original single-view i.For visual angle j in like manner as can be known.

As accompanying drawing 5, if the main shaft under two visual angles

With

Corresponding to same people, then will

Be transformed into visual angle j from visual angle i and obtain straight line

, then this straight line with

Meet at a little

This intersection point just corresponding " land-point " of this people under the j of visual angle, i.e. the intersection point of people's main shaft and ground level.

For the situation more than two video cameras, a people has two or more such intersection points, then chooses the final position " land-point " of the mean value of these intersection points as the people.

Situation about being blocked for people's Lower Half, based on the position of prediction and the main shaft of detection, the coupling that this algorithm still can robust estimates the position (" land-point ") of people in image accurately.

In order to implement concretism of the present invention, we have done a large amount of experiments on two databases, have realized that the pedestrian based on the main shaft coupling follows the tracks of under the multiple-camera.Experimental result has further been verified the validity and the robustness of this method.

Experimental result as shown in Figure 6, tracked human rectangle frame is represented, the sequence number that digital watch below the rectangle frame is leted others have a look at, the center line of rectangle frame is represented detected main shaft, the intersection point of the lower limb line of main shaft and rectangle frame is represented the position of people in image estimated.

Numeral 1,2,3,4 expressions in (a) are many people tracking test results under two video cameras of NLPR database.

(b) numeral 1,2,3,4 expressions in are many people tracking test results under three video cameras of NLPR database.

Numeral 1,2 expressions among Fig. 6 (c) are to the tracking test result of single people under the situation of blocking under two video cameras of PETS2001 database.

Numeral 1,2,3,4 expressions among Fig. 6 (d) are to the tracking test result of group under the situation of blocking under two video cameras of PETS2001 database.

Claims (8)

1. A pedestrian tracking method based on main axis matching under multi-camera, comprising steps: Multi-camera refers to the use of two or more single cameras. The premise of this method is to assume that there is a common plane in the visible areas of different cameras, and this common plane refers to the ground plane; (1) The image sequence is filtered to obtain a single Gaussian background model, and the moving target is segmented through the background pruning method to obtain the moving area; (2) Extract the main axis feature of the person in the motion area as the matching feature, and the main axis detection in the three cases is: the main axis detection of a single person, a group of people and people under occlusion; (3) Combined with the detected main axis, the predicted main axis position is obtained through the Kalman filter, and the tracking under a single camera is realized; (4) Calculate the homography matrix between different image planes, use the best matching pair algorithm, and find all the best matching pairs according to the predicted main axis position; (5) When all the best matching pairs are found, the matching information is fused to update the tracking results under the single view. 2. by the described method of claim 1, it is characterized in that, moving object segmentation, moving object segmentation is the first step of motion tracking, and what algorithm adopted is the background pruning method of single Gaussian model to detect moving area, adopts normalization The color rgs model, where r=R/(R+G+B), g=G/(R+G+B), s=R+G+B, first filter a sequence of background images to obtain a single Gaussian background model, and then perform differential processing on the current image and the background image, and perform threshold processing on the differential image to obtain a binary image M _xy . After obtaining the binary image, perform erosion and expansion operations on it in the morphological operator To further filter the noise, and finally use the binary connected component analysis to extract the single-connected foreground area as the segmented motion area. 3. by the described method of claim 1, it is characterized in that, the tracking under single camera adopts Kalman filter to realize tracking by the position of previous frame object prediction current frame object, then carries out with the observed value of current frame In comparison, if the distance between the observed value and the predicted value is small, it indicates that the observed value is highly reliable, and the observed value represents the position of the object in the current frame; if the distance between the two exceeds a certain threshold, the observed value is not credible, defined The intersection of the main axis and another line to update the observation value, this line is the perpendicular line from the predicted point to the main axis. 4. by the described method of claim 1, it is characterized in that, described searching for all optimal matches comprises the step: Calculate the homography matrix between different image planes; The homography matrix describes the one-to-one correspondence between different images on the same plane. The parameters of the matrix are solved by the given simultaneous equations of corresponding points. The corresponding points are obtained manually: that is, the marker points are set in the scene in advance. Or use some special corresponding points in the scene; Compute the matching likelihood function values for all pairs of principal axes; According to the geometric relationship of the main axes under different viewing angles, the matching likelihood function between the main axes s and k is defined as: $L({L^i}_{\mathrm{the\; s}},{L^j}_k)=p\left(x_{\mathrm{the\; s}}^i\right|Q_{\mathrm{ks}}^{\mathrm{the\; ji}})p\left(x_k^j\right|Q_{\mathrm{sk}}^{\mathrm{ij}})$ in

is the "foot point" of person s observed by camera i,

is the "foot point" of person k observed by camera j

is the intersection point of the main axis s from the angle of view i to the angle of view j and the main axis k; Find all the best matching pairs according to the matching algorithm; The matching modeling of multiple cameras becomes a problem of maximum likelihood function, that is, the principal axis matching likelihood function of the corresponding principal axis pair is the largest among the principal axis matching likelihood functions of many principal axis pairs. In order to simplify the problem, the maximum likelihood function problem Transformed into the minimum matching distance problem; The algorithm for finding all the best matching pairs is as follows: Under camera i, M spindles are detected:

Under camera j, N spindles are detected:

,...,

Step 1: Combine the spindles detected from the two perspectives in pairs to form all possible matching spindle pairs. Without loss of generality, assuming M≤N, the M spindles are sequentially selected from the N spindles. A main axis is matched with it to form a

combinations, each of which has the following form: ${\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; \{</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span><span\; class="notranslate">\; (</span>{{\mathrm{<span\; class="notranslate">\; L</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}}_{\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; kM</span>}}^{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \}</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; N</span>}}^{\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; ;</span>$ Step 2: For each principal axis pair {m, n} in each combination, calculate its matching distance

and guarantee ${\mathrm{D.}}_{\mathrm{mn}}^{\mathrm{ij}}<{\mathrm{D.}}_T,$ Where D _T is the threshold value, which is used to judge whether the main axis pair {m, n} matches. If the above constraints are not satisfied, the main axis pair {m, n} is deleted from θ _k ; Step 3: Select the combination Θ _k with the most matching pairs l, and all Θ _k form the set Θ: $\mathrm{\&Theta;}=\{{\mathrm{\&Theta;}}_k=(L_{k_1}^i,L_{k_1}^j),(L_{k_2}^i,L_{k_2}^j),...,(L_{k_1}^i,L_{k_1}^j)\}$ in $k\&Element;P_N^m$ Step 4: In the set Θ, find the global best matching combination λ, so that all the matching distances

The sum reaches the minimum value, which satisfies the following formula: $\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\underset{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; min</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}_{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; w</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; w</span>}}<span\; class="notranslate">\; )</span>}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>$ Step 5: The final Θ _{and λ} are the global best matching combination, and each pair of them is the matching principal axis pair. 5. according to the described method of claim 1, it is characterized in that, described fusion matching information updates the tracking result under single angle of view and comprises the steps: When all matching principal axis pairs are found, the matching information is used to update the tracking results under a single camera. For the case of two cameras, the update step is only available when the tracked person is in the common ground plane area of the two perspectives. efficient, Assuming that the matching algorithm mentioned above is used to find the main axis pair corresponding to the same person under the two perspectives, namely the perspective i and the perspective j, and convert the main axis under the perspective j to the perspective i through the homography relationship between the two image planes In the image plane of , the intersection point of the main axis under the original viewing angle i and the converted straight line is the final position of the person in the image plane i, which is used to update the tracking result under the original single viewing angle i. The same is true for the viewing angle j. For the case of more than two cameras, a person may have two or more such intersection points, and the average value of these intersection points is selected as the final position of the person. 6. The method according to claim 3, characterized in that the detection of the main axis of a single person, The main axis detection of a single person, for a single person, the least median square method is used to detect its main axis, assuming that the vertical distance from the ith foreground pixel X _i to the undetermined straight line l is D(X _i , l), according to the least median square method, the median of the squares of the vertical distances from all foreground pixels to the main axis is the smallest, $\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\underset{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; min</span>}}{\mathrm{<span\; class="notranslate">\; median</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; D.</span>}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; \}</span><span\; class="notranslate">\; .</span>$ 7. by the described method of claim 3, it is characterized in that, the main axis detection of a group of people comprises individual segmentation and main axis detection two major parts, in order to introduce individual segmentation into vertical projection histogram, the individual in a group of people corresponds to the vertical projection histogram The peak area of the graph, the peak area that satisfies the condition corresponds to a single individual is called the obvious peak area, and the obvious peak area must meet the two conditions of the peak area as follows: a) The maximum value of the peak area must be greater than a certain threshold, the peak threshold _PT ; b) The minimum value of the peak area must be less than a certain threshold, the valley threshold C _T ; Assuming that in a peak area, P ₁ , P ₂ ,..., P _n are its local extremums, C _l , C _r are the left and right valley values of the area respectively, then the above two conditions are expressed by mathematical expressions for: max(P ₁ ,P ₂ ,...,P _n )>P _T C _l < C _T , C _r < C _T The trough threshold C _T is selected as the mean value of the entire histogram, the peak threshold _PT is selected as 80% of the person's height in the image coordinates, and a single person's axis is detected in the axis detection. 8. by the described method of claim 3, it is characterized in that, in the detection of the people's main axis under occlusion situation, people is segmented out earlier, finds out the pixel of its foreground area, utilizes minimum in the foreground pixel that is segmented out The value squared error method is used to detect the main axis of the person, and the segmentation based on the color template method is used to realize the occlusion situation. The model includes a color model and an additional probability mask. When the occlusion occurs, the segmentation of the moving object is expressed as a classification The problem is to judge which foreground pixels belong to which model. The classification problem is solved by Bayesian rule, if the following equation is satisfied: $\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\underset{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; max</span>}}{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>$ Then the foreground pixel X belongs to the kth moving object.

Images