CN101739690B - Method for detecting motion targets by cooperating multi-camera - Google Patents

Abstract

The invention discloses a method for detecting motion targets by cooperating multi-camera and solves the technical problems that the method for detecting motion targets by cooperating multi-camera of the prior art has high error rate. The invention utilizes multi-plane limited multi-camera cooperation to detect motion targets and solves the shelter problem by using multi-visual redundancy. Without increasing the number of cameras, the redundancy of multi-camera is efficiently removed by limiting multi-plane in order to remove false-alarm and decrease the error rate for detection. In the scene of three cameras and six persons, the average error rate of the prior art is decreased to 21.8% from 69.8%.

Description

Method for detecting motion targets by cooperating multi-camera

Technical field

The present invention relates to a kind of object detection method, particularly method for detecting motion targets by cooperating multi-camera.

Background technology

Document " January 2006 for A multiview approach to tracking people in crowded scenes using a planar homographyconstraint; Lecture Notes in Computer Science; Computer Vision-ECCV 2006; 3954LNCS (7), p133-146 " discloses a kind of polyphaser moving object detection tracking based on the Planar Mapping constraint.The at first selected some camera perspectives of this method are the visual angle as a reference; Background subtraction through based on mixed Gauss model carries out foreground extraction to the view data that each visual angle obtains, and each prospect probabilistic image that will obtain then projects under the coordinate system of reference viewing angle according to surface constraints.Each visual angle probability graph after the conversion connected to take advantage of obtain comprehensive prospect probability graph; Adopt a given threshold value that comprehensive prospect probability graph is cut apart the position at the pin place that can obtain final goal, can solve occlusion issue preferably thereby be less than in the target number under the situation of camera number.Under the crowd scene, target numbers is the unknown and dynamic change, when target number during greater than the camera number, a large amount of false-alarms can occur in this method testing result, and in 6 people's of 3 cameras scene, error rate is up to 69.8%.

Summary of the invention

In order to overcome the high deficiency of prior art polyphaser moving object detection tracking error rate, the present invention provides a kind of method for detecting motion targets by cooperating multi-camera, on the basis that makes full use of various visual angles information; Do not rely on calibration information, solve the target number more than moving object detection orientation problem under the camera said conditions, promptly in certain scope; Replace increasing the camera number with increasing the constraint plane number; Realize the removal of false-alarm, can reduce the faults rate, improve accuracy of detection.

The technical solution adopted for the present invention to solve the technical problems: a kind of method for detecting motion targets by cooperating multi-camera is characterized in comprising the steps:

(a) select visual angle as a reference, visual angle 1, at first calculate the mapping matrix H between reference viewing angle and the top plan view, select equally distributed K surface level π from ground to the crown _i(i=1,2 ..., K) as the projection constraint plane, and calculate at π respectively _i(i=1,2 ..., K) following j the visual angle of constraint is to the mapping parameters H of reference viewing angle _Ij, i.e. i surface level π _iRetrain the mapping matrix of following j visual angle to reference viewing angle;

(b) adopt background subtraction algorithm based on mixed Gauss model to each visual angle input picture IS in viewnum the camera perspective _jAccording to formula

Carry out foreground extraction, obtain foreground picture I _j, j=1,2 ..., viewnum;

In the formula (1),<w _i, Model _i>Be mixed Gauss model Model={<w _i, Model _i>, i=1,2 ..., i Gauss model among the num}, weight is w _i, θ is the foreground extraction threshold value, num representes the number of single Gauss model that mixed Gauss model comprises;

(c) the mapping parameters H that utilizes step (a) to calculate _Ij, under K different level projection constraint, the foreground image at each visual angle is mapped to reference viewing angle successively

$\left(\begin{array}{c}u\\ v\\ 1\end{array}\right)=H_{\mathrm{ij}}\&CenterDot;\left(\begin{array}{c}x\\ y\\ 1\end{array}\right)---\left(2\right)$

I _ij(x，y)＝I _j(u，v) (3)

Foreground picture I after obtaining shining upon _Ij, and will be according to the foreground image I after the same plane restriction conversion _Ij, j=2,3 ..., viewnum connects and takes advantage of, thereby obtains the comprehensive foreground picture Foregroundmap of K surface level under retraining _i, i=1,2 ..., K;

${\mathrm{Foregroundmap}}_i=I_1\&CenterDot;\underset{j=2}{\overset{\mathrm{viewnum}}{\mathrm{\&Pi;}}}{I}_{\mathrm{ij}}---\left(4\right)$

(d) utilize H to project to top plan view said K comprehensive foreground picture

$\left(\begin{array}{c}x\\ y\\ 1\end{array}\right)=H\&CenterDot;\left(\begin{array}{c}m\\ n\\ 1\end{array}\right)---\left(5\right)$

Overlookingview _i(m，n)＝Foregroundmap _i(x，y) (6)

Obtain K and overlook foreground picture Overlookingview _i, i=1,2 ..., K;

(e) m multiplied each other before calculating was overlooked in the foreground picture respectively

${\mathrm{MultipleForegroundmap}}_m=\underset{i=1}{\overset{m}{\mathrm{\&Pi;}}}{\mathrm{overlookingview}}_i,m=\mathrm{1,2},...,K---\left(7\right)$

M=1 in the formula (7), 2 ..., K, multiplied result MultipleForegroundmap _mDo weighted sum again

$\mathrm{FinalForegroundmap}=\underset{m=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_m\&times;{\mathrm{MultipleForegroundmap}}_m---\left(8\right)$

Obtain final foreground picture FinalForegroundmap, weighting coefficient satisfies: α _m＞α _M-1

(f) carry out threshold process to overlooking prospect probability graph FinalForegroundmap, obtain overlooking foreground picture FinalForegroundmap ':

${\mathrm{FinalForegroundmap}}^{\&prime;}(m,n)=\left\{\begin{array}{cc}1& \mathrm{FinalForegroundmap}(m,n)>\mathrm{thresh}\\ 0& \mathrm{else}\end{array}\right.---\left(9\right)$

FinalForegroundmap ' is carried out connected component labeling, and area, is accomplished and is detected as detected moving target greater than the connected region of 100 pixels; Thresh is a binary-state threshold.

The invention has the beneficial effects as follows:, solve occlusion issue with the various visual angles redundant information owing to adopted the polyphaser cooperative motion target detection of multilayer planar constraint; Under the prerequisite that does not increase number of cameras; Effectively remove the redundant information of polyphaser with the constraint of multilayer planar, reach the purpose of removing false-alarm, thereby reduce the faults rate; Under 6 people's of 3 cameras scene, the average error rate is reduced to 21.8% from 69.8% of prior art.

Below in conjunction with embodiment the present invention is elaborated.

Embodiment

Concrete grammar step of the present invention is following:

(1) initialization and CALCULATION OF PARAMETERS.

Suppose viewnum camera arranged; Present embodiment is chosen viewnum=3, and selected visual angle 1 is the visual angle as a reference, at first calculates the mapping matrix H between visual angle 1 and the top plan view; Selected then from ground to the crown an equally distributed K surface level be designated as π as the projection constraint plane _i, i=1,2 ..., K (present embodiment is got K=5), and calculate at π respectively _i(i=1,2 ..., K) following j the visual angle of constraint is to the mapping parameters H of reference viewing angle _Ij, i represents i plane, and j represents j visual angle, adopts Perspective transformation model to carry out H here _IjCalculating (H _IjRepresent i surface level π _iRetrain the mapping matrix of following j visual angle) to reference viewing angle.

(2) foreground extraction.

Employing is carried out foreground extraction based on the background subtraction algorithm of mixed Gauss model to each visual angle, and the foreground picture that obtains is designated as I _j, j=1,2 ..., viewnum.<w, Model>Represent a single Gauss model that weight is w, suppose coordinate in the image for (x, some place mixed Gauss model y) is Model={<w _i, Model _i>, i=1,2 ..., num} (num representes the number of single Gauss model that mixed Gauss model comprises), the foreground extraction formula is following so:

In the formula (1), I0 _jBe j the current input original image in visual angle, θ is the foreground extraction threshold value, can be fixed threshold, and present embodiment is taken as θ adaptive, gets the sub-minimum in all weights.

(3) collaborative in foreground picture mapping and the plane.

To i=1 arbitrarily, 2 ..., K, j=2,3 ..., viewnum utilizes H _IjAccording to surface level π _iConstraint is with I _jProject to the foreground picture I after reference viewing angle obtains conversion _IjLocate for , have:

$\left(\begin{array}{c}u\\ v\\ 1\end{array}\right)=H_{\mathrm{ij}}\&CenterDot;\left(\begin{array}{c}x\\ y\\ 1\end{array}\right)---\left(2\right)$

I _ij(x，y)＝I _j(u，v) (3)

For each visual angle j, will be according to the foreground image I after the same plane restriction conversion _Ij, j=2,3 ..., viewnum connects and takes advantage of, thereby obtains the comprehensive foreground picture Foregroundmap of K surface level under retraining _i, i=1,2 ..., K;

${\mathrm{Foregroundmap}}_i=I_1\&CenterDot;\underset{j=2}{\overset{\mathrm{viewnum}}{\mathrm{\&Pi;}}}{I}_{\mathrm{ij}}---\left(4\right)$

(4) the reference viewing angle foreground picture is to the mapping of overlooking a foreground picture.

Utilize mapping matrix H with K comprehensive foreground picture Foregroundmap _iProject to top plan view, obtain overlooking foreground picture Overlookingview _i, i=1,2 ..., K.Locate for

, have:

$\left(\begin{array}{c}u\\ v\\ 1\end{array}\right)=H\&CenterDot;\left(\begin{array}{c}x\\ y\\ 1\end{array}\right)---\left(5\right)$

Overlookingview _i(x，y)＝Foregroundmap _i(u，v) (6)

(5) working in coordination with between the foreground picture of Different Plane constraint condition nutation visual field.

M links to each other and takes advantage of before overlooking K in the foreground picture respectively, m=1 wherein, and 2 ..., K, multiplied result is designated as MultipleForegroundmap respectively _m

${\mathrm{MultipleForegroundmap}}_m=\underset{i=1}{\overset{m}{\mathrm{\&Pi;}}}{\mathrm{overlookingview}}_i,m=\mathrm{1,2},...,K---\left(7\right)$

To MultipleForegroundmap _m, m=1,2 ..., K does weighted sum again and obtains final prospect probability graph, and weight coefficient is respectively 0＜α _m＜1, m=1,2 ..., K, weighted sum:

$\mathrm{FinalForegroundmap}=\underset{m=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_m\&times;{\mathrm{MultipleForegroundmap}}_m---\left(8\right)$

Because constraint is many more, the foreground point of judgement is credible more, so weighting coefficient satisfies: α _m＞α _M-1, get A wherein _mBe normalized factor, promptly

(6) threshold process and testing result show output.

Carry out threshold process to overlooking prospect probability graph FinalForegroundmap, obtain overlooking foreground picture FinalForegroundmap ', thresh is a binary-state threshold, and present embodiment gets 0.7:

${\mathrm{FinalForegroundmap}}^{\&prime;}(x,y)=\left\{\begin{array}{cc}1& \mathrm{FinalForegroundmap}(x,y)>\mathrm{thresh}\\ 0& \mathrm{else}\end{array}\right.---\left(9\right)$

FinalForegroundmap ' is carried out UNICOM's zone marker, and area as detected moving target, shows the output testing result greater than the UNICOM zone of 100 pixels.

Present embodiment reaches 21.8% average error rate on the basis that the actual video data of 3000 frames is tested, than the average error rate of prior art 69.8%, detecting performance has great raising.

Claims (1)

1. a method for detecting motion targets by cooperating multi-camera is characterized in that comprising the steps:

(a) select visual angle as a reference, visual angle 1, at first calculate the mapping matrix H between reference viewing angle and the top plan view, select equally distributed K surface level π from ground to the crown _i, i=1,2 ..., K is as the projection constraint plane, and calculates at π respectively _i, i=1,2 ..., following j the visual angle of K constraint is to the mapping parameters H of reference viewing angle _Ij, i.e. i surface level π _iRetrain the mapping matrix of following j visual angle to reference viewing angle;

(b) adopt background subtraction algorithm based on mixed Gauss model to each visual angle input picture IS in viewnum the camera perspective _jAccording to formula

Carry out foreground extraction, obtain foreground picture I _j, j=1,2 ..., viewnum;

In the formula (1),<w _k, Model _k>Be mixed Gauss model Model={<w _k, Model _k>, k=1,2 ..., k Gauss model among the num}, weight is w _k, θ is the foreground extraction threshold value, num representes the number of single Gauss model that mixed Gauss model comprises;

(c) the mapping parameters H that utilizes step (a) to calculate _Ij, under K different level projection constraint, the foreground image at each visual angle is mapped to reference viewing angle successively

I _ij(x，y)＝I _j(u，v) (3)

Foreground picture I after obtaining shining upon _Ij, and will be according to the foreground image I after the same plane restriction conversion _Ij, j=2,3 ..., viewnum connects and takes advantage of, thereby obtains the comprehensive foreground picture Foregroundmap of K surface level under retraining _i, i=1,2 ..., K;

(d) utilize H to project to top plan view said K comprehensive foreground picture

Overlookingview _i(m，n)＝Foregroundmap _i(x，y) (6)

Obtain K and overlook foreground picture Overlookingview _i, i=1,2 ..., K;

(e) m multiplied each other before calculating was overlooked in the foreground picture respectively

Formula (7) multiplied result MultipleForegroundmap _mDo weighted sum again

Obtain overlooking prospect probability graph FinalForegroundmap, weighting coefficient satisfies: α _m＞α _M-1

(f) carry out threshold process to overlooking prospect probability graph FinalForegroundmap, obtain overlooking foreground picture FinalForegroundmap ':

Finalforegroundmap ' is carried out connected component labeling, and area, is accomplished and is detected as detected moving target greater than the connected region of 100 pixels; Thresh is a binary-state threshold.