Summary of the invention
Problem to be solved
For existing monitoring system exist monitoring dead angle, complex situations are needed special messenger do not stop to check monitor, to invasion object without
The problems such as doscrimination, the present invention provides a kind of intrusion detection warning system based on multiple-camera data fusion and detection method thereof,
By control system and common camera, monopod video camera, zoom wide-angle multi-target tracking system, multi-level alarm processing module phase
Connect, it is possible to realize monitoring region is covered without dead angle, the object invaded is carried out Intelligent Measurement, takes not according to testing result
Same monitor mode, makes monitoring system more flexible and intelligent.
2. technical scheme
In order to solve the problems referred to above, the technical solution adopted in the present invention is as follows:
A kind of intrusion detection warning system based on multiple-camera data fusion, including fore-end, hop, control part,
Display part, multi-level alarm processing module and supply module, described fore-end include common camera, monopod video camera and
Zoom wide-angle multi-target tracking system, is used for obtaining shooting information and being transferred to control part;Described control part includes PC
Formula DVR, control system, cradle head controllor and invader identification and processing module, PC formula DVR obtains front end
Carry out the camera picture after processing and processing after the shooting information that fractional transmission comes and be transferred to control system;Described control system
Camera picture information after PC formula DVR is processed by tracking invader identification and processing module judges, completes many
The response of level warning processing module and the display of display picture;Described hop be used for transmitting fore-end, control part,
Data between display part and multi-level alarm processing module;Described supply module is that whole system is powered.
Preferably, described common camera be arranged on same position corresponding with monopod video camera, and its position is numbered.
Preferably, described hop uses Optical Fiber Transmission, has the advantages that loss is low, transmission quality is high.
Preferably, described zoom wide-angle multi-target tracking system includes zoom wide-angle multi-target tracking video camera, ultrahigh speed intelligence
Ball and embedded video server.
Preferably, described invader identification and processing module include that the contrast of sound state recognizer module, stationary body judges journey
Sequence module, stationary body contour detecting program module, recognition of face contrast program module, target finder module.
The detection method of above-mentioned a kind of based on multiple-camera data fusion intrusion detection warning system, detection process is: detection
The image information of shooting is transferred to PC formula DVR by the fore-end in warning system, and PC formula DVR is to shooting
Picture shows in display part after processing;When there being invader to occur, control system is called in invader identification and processing module
Target finder its positional information of module Real-time Feedback, call sound state recognizer module simultaneously and common camera clapped
Take the photograph image information and carry out sound state identification;When invader is motionless object, control system is called stationary body contrast and is judged journey
Invader is compared with preset threshold value and judges by sequence module, is then directly displayed and is deposited by display part if less than threshold value
Storage information captured by common camera, and respond multi-level alarm processing module and send 3 grades of alarm signals;If greater than threshold value,
Control system is then called stationary body contour detecting program module and is judged the specific profile of invader, then sends command adapted thereto and controls
Its profile is patrolled and examined shooting by the monopod video camera of reference numeral position, returns photographing information in real time, and response multi-level alarm processes simultaneously
Module sends 2 grades of alarm signals;When the object that invader is movement, control system calls recognition of face contrast program module pair
It carries out recognition of face contrast, if comparing result is not staff or can't detect face information, then by the many mesh of zoom wide-angle
Mark tracing system is tracked shooting to invader, and responds multi-level alarm processing module and send 1 grade of alarm signal.
Preferably, described target finder module uses the Camera Calibration Algorithm of radial arrangement restraint (RAC) to obtain
The inside and outside parameter of monocular-camera, then recovers target object three-dimensional coordinate in world coordinate system, concretely comprises the following steps:
(1) camera model that the present invention uses is the pinhole camera model with lens radially single order distortion, if
(xw,yw,zw) it is the coordinate of object P in three-dimensional world coordinate system, (x, y z) are P three-dimensional coordinate in camera coordinate system.
XOiY is that center is at OiPoint is parallel to x, the image coordinate system of y-axis.(xu,yu) P point under preferable pinhole camera model
Image coordinate, (xd,yd) it is the deviation (x caused by lens distortionu,yu) real image coordinate.Image seat in a computer
Mark unit is pixel count, so need to be by the three-dimensional coordinate of object point is transformed into plane of delineation coordinate.
(2) RAC calibrating camera parameters.First choose N number of coplanar characteristic point, determine image coordinate and the generation of this N number of point
Boundary's coordinate.Owing to needing the external parameter demarcated and inner parameter to be respectively 6, and proportionality coefficient and picture centre point coordinates etc.
Parameter is the most fixed, and RAC two-stage calibration method includes obtaining most calibrating parameters by analytic solutions and processing radial distortion generation with iterative method estimation
Effective focal length, three parameters such as degree of depth constituent element of translation vector.RAC is utilized to may determine that the spin matrix R in rigid body translation
T with translation matrix Tx, TyComponent.RAC is represented by direction
IfBy rigid body staticsWith in RACCan obtain:
[xwYd ywYd zwYd Yd -xwYd -xwYd -zwYd]·[r1/Ty r2/Ty r3/Ty Tx/Ty r4/Ty r5/Ty r6/Ty]T=Xd
Row vector is it is known that column vector is parameter to be asked.Use conplane spatial point coordinate fixed, choose world coordinate system, make
zw=0, then above formula is represented by:
Then can solve r1,r2,r4,r5Four independent variables, orthogonal matrix adds a ratio (1/Ty) also there are four independent variables, therefore
Can determine that spin matrix R and translational component Tx,Ty。
(3) effective focal length f, T are calculatedzComponent and distortion coefficient k.To each characteristic point PiCan have:
If wi=r7xwi+r8ywi, it is contemplated that zi=wi+Tz, Yu=(Yf-Yc)/NY, above formula is represented by:
[yi -(Yf-Yc)/NY]·[f Tz]T=(Yf-Yc)wi/NY
Make k=0, solve and can try to achieve effective focal length f, the T of translation matrix T respectivelyzComponent.
(4) according to target position (X in the picturef,Yf) and step (2) (3) in try to achieve camera parameters and can solve target
Position (x in world coordinate systemw,yw,zw)。
Preferably, described sound state recognizer module uses background differential technique, concretely comprises the following steps:
(1) initialization background: the image choosing one section of video is averaged, then with the gray value of the average image and each frame figure
The gray value of picture subtracts each other gained absolute value, and the subregion of least absolute value correspondence image forms current background image;
Seek the mean value of 3 two field picture gray valuesThe absolute value that the gray value of each two field picture and mean value subtract each other,
Formula is:
Background image is considered as the image section that minimum of a value is corresponding, wherein i≤3, and formula is:
(2) obtain when former frame continuous print images, try to achieve current background image F according to step (1)i+1(x,y)
F (x, y)=λ Fi(x,y)+(1-λ)Fi+1(x,y)
In formula: (x y) is up-to-date background image, F to Fi(x y) is background image last time, Fi+1(x, y) for currently newly to obtain Background
Picture, Fj(x, y) and Fj+1(x, y) is a few two field picture of continuous print, and λ is weights.The impact changed in view of noise and background, when
|Fj+1(x,y)-Fj(x, y) | < T, current frame image is taken as background image.
(3) image currently gathered and background image are made calculus of differences, when result exceedes certain threshold value, can determine whether in image
Corresponding pixel belongs to motion target area, produces corresponding binary image, and uses frame differential method further, will
Current frame image and previous frame image subtraction, if difference is less than certain threshold value, then count this pixel.Work as counting
When value Count exceedes certain threshold value, illustrate that this pixel is own through incorporating background, it is judged that for static state;Otherwise then judge that target is fortune
Move.
Preferably, the step that is embodied as of described stationary body contrast determining program module is:
(1) first two frames adjacent in video sequence are carried out difference, obtain difference image, be described as follows with formula:
Dk(x, y)=| Ik+1(x,y)-Ik(x,y)|
Dk(x,y)、Ik+1(x,y)、Ik(x, y) the most corresponding difference image, kth+1 frame and kth frame original image;
(2) difference image is carried out threshold division and obtain the foreground image of binaryzation, when foreground image is less than segmentation threshold,
It is labeled as 0, is otherwise 1;
(3) the ratio M of binaryzation foreground image contour area and background image entire area is calculated.
Preferably, described stationary body contour detecting program module uses Canny edge detection algorithm, and it is embodied as step
For:
(1) eliminate noise: generally, use Gaussian filter convolution to image noise reduction.
(2) gradient magnitude and direction are calculated:
UseCalculate gradient magnitude and direction.
(3) utilize non-maxima suppression to get rid of non-edge pixels, only retain candidate edge.
(4) hysteresis threshold: if a certain location of pixels amplitude exceedes high threshold, this pixel is left edge pixel;If it is little
In Low threshold, then get rid of;If between two threshold values, this pixel is only when being connected to the pixel that is higher than high threshold
Retained.
Preferably, described monopod video camera and zoom wide-angle multi-target tracking system all receive the instruction that control system sends, cloud
Concrete physical features is shot by platform video camera more than the static invader of preset threshold value;Zoom wide-angle multi-target tracking system is same
Time the multiple mobile invader of tracking lock, and can further tracking invader to obtain feature, display picture is in multiple targets
Between switch.
Preferably, described monopod video camera has self adaptation zoom function, according to size and the distance of camera lens of invader, automatically
Adjusting focal length, obtains high-resolution image information;Monopod video camera has two differences respectively both horizontally and vertically mobile
Motor-driven, when needs call, under control system, issue a command to cradle head controllor, and then control monopod video camera and carry out
Shooting.
In the present invention, PC formula DVR can utilize the mobile device such as mobile phone, flat board to remotely access after connecing switch;
Display part can be the mobile device such as display or mobile phone plane plate;Region to be monitored is carried out without dead angle all standing by fore-end
Lay, fore-end shooting picture through PC formula DVR reception, process (include invader dynamic Static Detection,
Stationary body contrast judges, mobile object recognition of face and target location) and transmission etc. operate after show over the display, and
Real-time storage is convenient checks afterwards and manages, and multi-level alarm processing module is connected with PC formula DVR.Control system is to general
Abnormal object in logical video camera carries out sound state identification, when invader is motionless object, by its concrete physical features such as
The threshold value preset with system such as size, length compares, and then directly displays and store common camera institute if less than threshold value
The information of shooting, and respond multi-level alarm processing module and send 3 grades of alarm signals;If greater than threshold value, control system controls
Its specific profile is patrolled and examined shooting by the monopod video camera of reference numeral position, returns photographing information in real time and responds multistage announcement simultaneously
Alert processing module sends 2 grades of alarm signals.When the object that invader is movement, control system carries out recognition of face contrast to it,
If not staff or can't detect facial image, then control zoom wide-angle multi-target tracking system and it is shot, real
Time return photographing information and respond multi-level alarm processing module simultaneously and send 1 grade of alarm signal.
Monopod video camera is to be realized by two operating motors both horizontally and vertically rotated, and motor accepts from controlling system
The signal of system is accurately positioned and rotates.The motorized zoom lens of monopod video camera can realize the pH effect such as aperture, focusing, will
Photographed scene furthers, push away remote with this to obtain the image information of fine definition.
Zoom wide-angle multi-target tracking system is by third party's video camera, a ultrahigh speed intelligent sphere and embedded video clothes
Business device composition.When there being multiple target to occur, embedded video server controls ultrahigh speed intelligence by running embedded algorithm routine
Feature video and the photo of each target is followed the trail of in energy ball timesharing, and system can automatically control the movement of ultrahigh speed intelligent sphere and select
Good focal length realizes picture and optimizes display, and simultaneously monitoring picture freely can switch between multiple targets.The many mesh of zoom wide-angle
Whether mark tracing system starts and is determined the feature recognition result of invader by control system.
Multi-level alarm processing module is connected in PC formula DVR, and control system determines according to picture captured by common camera
The corresponding warning level of response.
3. beneficial effect
Compared to prior art, the invention have the benefit that
(1) the present invention is directed to current monitor system needs special messenger ceaselessly to check monitor, easily lose the situation of important information,
By control system, picture captured by common camera is carried out sound state identification, stationary body contrasts judgement and recognition of face etc.
Reason, takes different monitoring strategies for different situations, makes intelligence degree be improved;Monopod video camera is possible not only to
Horizontal and vertical directions adjusts angle shot, and camera lens has automatic focusing function, improves resolution ratio so that afterwards
Analyze.
(2) present invention is compared to existing technology, and zoom wide-angle multi-target tracking system can be detected simultaneously by multiple target, and right
Target carries out Intelligent Recognition, and display picture freely can also switch between multiple targets;Multi-level alarm processing module is divided into 3
Individual grade, each grade represents different monitored object respectively, has separated the order of importance and emergency, simple and clear.
(3) present invention uses a kind of method recovering target object three-dimensional coordinate in world coordinate system to come invader location,
Its coordinate information of Real-time Feedback, facilitates doubtful situations quickly to respond, and ensures the efficiency of emergency processing.
Embodiment 1
As depicted in figs. 1 and 2, intrusion detection warning system based on multiple-camera data fusion includes: control system 1, aobvious
Show device 2, PC formula DVR 3, cradle head controllor 4, common camera A1-An, monopod video camera B1-Bn, zoom
Wide-angle multi-target tracking system (includes zoom wide-angle multi-target tracking video camera C, ultrahigh speed intelligent sphere and embedded video service
Device), multi-level alarm processing module (including 3 grades of alarm device S, 2 grades of alarm device T, 1 grade of alarm device U).Common camera A1-An
Is deployed to ensure effective monitoring and control of illegal activities without dead angle in monitoring region, monopod video camera B1-BnWith common camera A1-AnCorrespondence is arranged on same position right
Its position is numbered, zoom wide-angle multi-target tracking video camera C can overall view monitoring Zone Full, the vision signal of three is the most straight
Access the high-performance video and audio compress card of PC formula DVR 3.Video/audio signal after video and audio card Compress softwares by soft
Part carries out processing showing on a display 2, and can realize showing invader 1,4,9,16 picture and to mobile target
Tracking, amplification process.3 grades of alarm device S, 2 grades of alarm device T, 1 grade of alarm device U receive external various alarm control unit
Input, control bus by RS485 and pass to the controlling alarm inquiry mouth realize alarm linkage of PC formula DVR.
The image information that PC formula DVR 3 is transmitted by control system 1 judges, and determines monopod video camera B1-Bn, zoom
Whether wide-angle multi-target tracking video camera C works, and controls PC formula DVR 3 simultaneously and responds the multi-level alarm process of correspondence
Module.Video flowing through overcompression is used the modes such as multicast to pass to each sub-control of network by network interface card by PC formula DVR 3
In terminal, the mobile device such as mobile phone and flat board 8 is facilitated to check whenever and wherever possible.When there being invader to occur, control system 1
Call target finder its positional information of module Real-time Feedback in invader identification and processing module.
As in figure 2 it is shown, sound state recognizer module uses background differential technique, it is as follows that it is embodied as step:
(1) initialization background.The image choosing one section of video is averaged, then with the gray value of the average image and each frame figure
The gray value of picture subtracts each other gained absolute value, and the image utilizing the subregion composition of the correspondence image of absolute value minimum is exactly current
Background image.
(2) obtain when former frame continuous print images, try to achieve current background image F according to step (1)i+1(x,y)
F (x, y)=λ Fi(x,y)+(1-λ)Fi+1(x,y)
In formula: (x y) is up-to-date background image, F to Fi(x y) is background image last time, Fi+1(x, y) for currently newly to obtain Background
Picture, Fj(x, y) and Fj+1(x, y) is a few two field picture of continuous print, and λ is weights.The impact changed in view of noise and background, when
|Fj+1(x,y)-Fj(x, y) | < T, current frame image is taken as background image.
(3) image currently gathered and background image are made calculus of differences, when result exceedes certain threshold value, can determine whether in image
Corresponding pixel belongs to motion target area, produces corresponding binary image, and determines whether that this pixel is at one section
The most also changing in time, if continuing to change, then judging that target is dynamic, be otherwise static.
When invader is mobile object, control system 1 calls recognition of face contrast program module, and this module uses eigenface side
Method, it is as follows that it is embodied as step:
(1) obtain the set S of staff's facial image and calculate its average image Ψ, being described as follows with formula:
S={ Г1,Г2,......ГM}
(2) calculate difference Φ of every image and the average image, i.e. each element in S set deducts the mean value in Ψ
Φi=Гi-Ψ
(3) with M orthogonal unit vector unΦ distribution, u are describednMiddle kth (k=1,2,3...M) individual vector ukPass through following formula
Can be calculated,
Wherein ukFor unit orthogonal vectors, calculate ukI.e. calculate the characteristic vector of the covariance matrix of C, wherein
A={ Φ1,......Φn}
(4) recognition of face.When there being people to occur, from a frame of video flowing, first detect face and separate, then carrying
Take out feature wk
Wherein k=1,2...M, for kth eigenface uk, above formula can calculate the weight of its correspondence, and M weight may be constructed
Vector ΩT。
ΩT=[w1,w2,......wM]
Wherein Ω is face to be differentiated, Ω k represents certain face in training set, both represents by the weight of eigenface.
Both are sought Euclidean distance, illustrates that when distance is less than threshold value the kth face in face to be differentiated and training set is same person.
If comparison result is not staff or can't detect facial image, control C pair, zoom wide-angle multi-target tracking video camera
Object is tracked shooting, returns photographing information in real time and responds 1 grade of alarm signal simultaneously.When multiple mobile target occurs, can
Picture and feature with the multiple target of manual switching.
When invader is static state, control system 1 obtains corresponding common camera A1-AnPositional information, and call stationary body
Contrast determining program module.It is embodied as step:
(1) first two frames adjacent in video sequence are carried out difference, obtain difference image, be described as follows with formula:
Dk(x, y)=| Ik+1(x,y)-Ik(x,y)|
Dk(x,y)、Ik+1(x,y)、Ik(x, y) the most corresponding difference image, kth+1 frame and kth frame original image;
(2) difference image is carried out threshold division and obtain the foreground image of binaryzation, when foreground image is less than segmentation threshold,
It is labeled as 0, is otherwise 1;
(3) the ratio M of binaryzation foreground image contour area and background image is calculated.
If M is less than threshold value, directly displays and store common camera A1-AnCaptured information, and respond 3 grades of reports
Alert signal;If M is more than threshold value, control system 1 first determines whether the specific profile of invader, then controls reference numeral position
Monopod video camera B1-BnIts specific profile is patrolled and examined shooting, returns photographing information in real time and respond 2 grades of alarm signals simultaneously.
Common camera A in system1-AnNeed shooting free of discontinuities, zoom wide-angle multi-target tracking video camera C and monopod video camera
B1-BnStart working after receiving instruction.Each shot by camera information can be by real time record to PC formula DVR 3
In, conveniently check afterwards and manage.PC formula DVR 3 is by common camera A1-An, monopod video camera B1-Bn、
Zoom wide-angle multi-target tracking video camera C shooting picture information shows over the display in real time.
Comprehensive refering to shown in Fig. 1 and Fig. 2, the multistage intelligent monitoring and alarming system with multiple-camera is arranged on one 1000 flat
The factory of side's rice, after startup, common camera A1-AnStart working.At a time, common camera A2Detect certainly
There is a unknown object in dynamic production equipment side, control system 1 feed back its more specific location information and by this article size be
System preset threshold value compares.If M is less than (1/30), display shows common camera A2Image information, and respond 3
Level alarm signal;If M is more than (1/30), control system first determines whether the specific profile of this object, then controls correspondence and compiles
The monopod video camera B of number position2Its profile is patrolled and examined shooting.Automatic zoom camera lens can be with far and near, automatically according to target and lens location
Suitable focal length is selected to obtain picture clearly.Common camera A is shown on display2Image information, and respond 2 grades of reports
Alert signal.If now there being people's process, control system 1 intercepts one section of video flowing, detects face from which and carry out in a frame
Separate, then the feature extracted is compared with face characteristic in existing face database.Comparison result shows it is not work people
Member, control system 1 sends instruction control zoom wide-angle multi-target tracking video camera C and target is tracked shooting, returns in real time
Photographing information responds 1 grade of alarm signal simultaneously, simultaneously Real-time Feedback target position information and can carrying out it over the display
Feature amplification processes.When there being many people to occur, utilize existing many pictures monitoring software can realize picture and the feature of multiple target
Freely switch.