CN101635835A - Intelligent video monitoring method and system thereof - Google Patents

Abstract

The invention relates to an intelligent video monitoring method and a system thereof. The system comprises a camera, a first alarm device, a coding processor, a data analyzing processor and a display device. The method comprises the steps of detecting and tracking move targets. The intelligent video monitoring method and the system are executed on the base of common network television monitoring, not only have the advantages of common intelligent video monitoring, but also can bring better income for users. The invention mainly has the following advantages of (1) being capable of reliably monitoring all day long with 24*7; (2) being capable of improving alarm precision due to a strong intelligent characteristic; (3) being capable of improving response speed due to the strong intelligent characteristic; and (4) being capable of effectively enlarging the use of video resources.

Description

Intelligent video monitoring method and system

Technical field

The present invention relates to method for supervising and system, more particularly, relate to a kind of intelligent video monitoring method and system.

Background technology

The present video monitoring equipment that is used widely in various occasions such as teaching, security protection, industrial production, major part is carried out the analog video signal transmission based on closed circuit cable TV (CCTV) mode.Such video camera has certain limitation, shows as: when 1, adopting the analog video signal transmission, because signal is subjected to the decay of various interference and signal inevitably in transmission course, make have certain difficulty aspect the long Distance Transmission.Usually the mode that adopts relaying to amplify solves this problem, but can increase the cost of system.2, because analog signal transmission adopts the base band mode to carry out usually, make the structure of video camera and monitoring terminal can only adopt man-to-man form.This makes the topological structure of whole system become complicated, has increased the cost of comprehensive wiring.3, when video camera is connected to controllable The Cloud Terrace, camera lens or other input/output signals, need add control circuit usually, make that the structure of system is complicated more.Particularly in the security protection occasion, because image transmitted is clear inadequately between monitoring scene and the Control Room, the real-time, interactive performance reliability is not high, cause on duty in the sentry attacked, gun are robbed and indivedual sentry sleeps hilllock, take the problem that rifle leaves the post etc. to be difficult in time find and take place by the bad person.In addition, existing supervisory control system accuracy is not high, and Changes in weather, the false alarm that the shade shadow is produced take place easily because toys such as mosquito, moth, kitten, doggie disturb and the branch swing.

The modern management of current social all trades and professions needs the scientific and technical means of uses advanced, and electronic technology and computer control are integrated in the rounded system.In army, safety is the problem that primarily needs guarantee, uses existing monitoring security device, can effectively strengthen the management to personnel, intuitively reflects the field condition of primary location timely, strengthens safety precautions, is the strong instrument of army's modern management.

Army is a special place, in order to protect the national property safety and the personal safety of army personnel. one cover height intelligent monitoring system is very important for army building; So-called intelligent video analysis monitoring technology, sometimes also claim " behavior monitoring technology ", just be meant and adopt intelligentized video analysis algorithm, utilize computer that the specific behavior of target within the vision is analyzed and extracted, meet certain regular behavior (as directed movement, cross the border, go around, leave over etc.) when taking place when find existing, send cue from the trend supervisory control system, take certain countermeasure (reporting to the police) or notify the monitor staff to carry out manual intervention etc. as audible-visual annunciator.

Along with the continuous development of computer technology, the communication technology, image processing techniques, more and more universal based on the detection and the tracking system of video.Native system has merged the knowledge of many association areas such as Computer Image Processing, pattern recognition, artificial intelligence and automatic control, adopt advanced algorithm both at home and abroad, utilize the method for computer vision and video analysis to analyze the image sequence of video camera recording automatically, realization is to location, identification and the tracking of target in the scene, and analyzes and judge the behavior of target on this basis.

At present, mostly the security protection video monitoring system is that with video monitoring and alarm linkage be the system for crime prevention and control of representative, and a kind of instrument of evidence obtaining video recording afterwards that provides almost is provided.In this safety defense monitoring system, generally adopt traditional system configuration pattern: promptly build Surveillance center centralizedly, video data is uploaded to the video server of centralized setting, and carry out storage in this centralized node; On video acquisition point, generally adopt video collector, control multi-channel video camera.This video collector adopts broadband line, or adopts wireless network card, under the situation of bandwidth deficiency, may adopt the mode of multiple wireless line bundle to connect public the Internet, carries out transfer of data with the same video monitoring center that is connected on the internet.But this supervisory control system all has some intrinsic limitation:

1, owing to mankind itself's weakness, easily cause failing to report:

Generally speaking, the mankind are not to be an observer that can trust fully, they are when observing real-time video flowing or observing playing back videos, because the difference of monitor staff's individual condition and self physiological weakness, often security threat can't be perceived, thereby the generation of (False Negatives) phenomenon may be caused failing to report.

2, each control point can not all be in monitoring all the time:

Except the less TV monitoring and controlling of some scales was used, it can be rig camera configuration monitoring device according to 1: 1 ratio that TV monitor system is seldom arranged.Therefore, each control point is not all to be in the middle of the monitoring all the time.

3, easily cause the wrong report and fail to report:

Reporting (False-Positive) by mistake and failing to report is modal two large problems in the video monitoring system.Fail to report and refer to when control point generation security threat, this threat does not have monitored system or Security Officer to find.It is security threat that wrong report refers to that the security activity that is positioned at the control point is mistaken as, thereby produces wrong warning.

4, because the lack of wisdom factor makes the data analysis difficulty:

The back report to the police to take place the video recording data are analyzed one of work that Security Officer normally must do, wrong report and fail to report phenomenon and then further aggravated demand to data analysis.The Security Officer often is required to find out the Video Document relevant with alert event, finds the troublemaker, determines accident responsibility or assesses the security threat of this incident.

Because traditional video surveillance system lack of wisdom factor, the video recording data can't be by effectively classification storage, can only stamp time tag at most, therefore data analysis work becomes extremely consuming time, and be difficult to obtain whole relevant informations, and recurrent wrong report phenomenon further increases hash, thereby brings bigger difficulty to data analysis work.

5, owing to the lack of wisdom factor, the response time is long:

The overall performance that is related to a safety system for the response speed of security threat.Traditional TV monitor system is all responded to security threat by the trouble free service personnel usually and handles, and this is general for handling, real-time response requires the lower security threat enough.But under a lot of situations, when threatening generation, need a plurality of funtion parts of safety system, even the relevant department's cooperation in the shortest time of a plurality of safety, crisis handled jointly.At this time, the response speed of supervisory control system will be directly connected to the damaed cordition of user's the person or property.

Summary of the invention

The technical problem to be solved in the present invention is, and is poor at the not enough mood of above-mentioned image, the real-time, interactive of prior art, be easy to generate defectives such as false alarm, and a kind of intelligent video monitoring method and system are provided.

The technical solution adopted for the present invention to solve the technical problems is: construct a kind of intelligent video monitoring method, may further comprise the steps:

S1: detect moving target:

S11: gather and preliminary treatment M frame image sequence, wherein, M is a natural number;

S12: according to described M frame image sequence, initialization background model;

S13: gather the M+1 frame image sequence,, described M+1 frame image sequence is carried out inter-frame difference and background subtraction divisional processing, detect moving target based on described background model;

S2: pursuit movement target:

S21:, set up object module to detected moving target;

S22: according to described object module, detected moving target is carried out continuous K frame handle, set up the target following information table, wherein K is a natural number;

S23: set up the object matching matrix between adjacent two frames, matrix element is the matching degree of adjacent two interframe object modules, obtains the match condition of moving target;

S24: according to object module and object matching matrix, analyze present frame moving target state, and upgrade object module.

In intelligent video monitoring method of the present invention, in described step S11, described preliminary treatment comprises to be handled and the filtering and noise reduction processing the gray processing of M frame image sequence.

In intelligent video monitoring method of the present invention, in described step S12, also comprise each pixel in the background model is carried out Gauss's modeling.

In intelligent video monitoring method of the present invention, in described step S13, the M+1 frame image sequence is carried out inter-frame difference when handling, with the area update that do not change in background model; The zone and the background model that change are carried out match, appear district and background model to distinguish.

In intelligent video monitoring method of the present invention, in step S13, also comprise detected moving target is gone Shadows Processing.

In intelligent video monitoring method of the present invention, described object module comprises: area similarity function, shape similarity function, direction of motion consistency function, displacement reliability function and fuzzy similarity function.

According to another aspect of the present invention, provide a kind of intelligent video monitoring system, comprise camera, first warning device, encode processor, data analysis processor, display; Wherein, camera is connected with encode processor respectively with alarm, and encode processor is connected with the data analysis processor respectively with display;

It is vision signal that encode processor is used for the video flowing that camera sends is carried out compressed encoding;

The data analysis processor is used for that the vision signal that encode processor sends is sent to display and shows, and carries out analyzing and processing, when abnormal conditions occurring, by encode processor alarm signal is sent to first warning device to report to the police.

In intelligent video monitoring system of the present invention, described intelligent video monitoring system also comprises: pick-up and audio amplifier, and wherein pick-up is connected with encode processor, and audio amplifier is connected with the data analysis processor;

It is audio signal that encode processor also is used for the audio stream that pick-up sends is carried out compressed encoding;

The data analysis processor is used for that also the audio signal that pick-up sends is sent to audio amplifier and plays.

In intelligent video monitoring system of the present invention, described intelligent video monitoring system also comprises: the foot-operated alarm and second warning device; Wherein, foot-operated alarm is connected with encode processor, and second warning device is connected with the data analysis processor;

The data analysis processor also is used to receive the alarm signal that foot-operated alarm sends by encode processor, and controls second warning device and report to the police.

In intelligent video monitoring system of the present invention, described intelligent video monitoring system also comprises: microphone and loudspeaker; Wherein, microphone is connected with the data analysis processor; Loudspeaker are connected with encode processor; Microphone sends to loudspeaker by data analysis processor and encode processor with the audio stream that collects and plays.

In intelligent video monitoring system of the present invention, first warning device and second warning device are buzzer, multitone alarm or audible-visual annunciator.

In intelligent video monitoring system of the present invention, encode processor carries out MPEG4, H.263, H.264 or the M-JPEG compressed encoding to picture signal.

In intelligent video monitoring system of the present invention, camera comprises: at least one camera lens and imageing sensor.

In intelligent video monitoring system of the present invention, imageing sensor is ccd image sensor or cmos image sensor.

Implement intelligent video monitoring system of the present invention, have following beneficial effect: engineering construction is easy, and system expands convenient; Realize trans-regional remote monitoring, make picture control not be subjected to distance limit, and clear picture, reliable and stable; And monitoring is on-the-spot and Control Room can carry out information exchange in real time.And supervisory control system accuracy height can be avoided because toys such as mosquito, moth, kitten, doggie disturb and the branch swing Changes in weather, the false alarm that the shade shadow is produced.Intelligent television monitoring is based on common Web TV monitoring, and except the advantage that possesses the Web TV monitoring that is widely known by the people, the intelligent television monitoring system can also bring bigger income for the user.Its main advantage is:

1, can carry out 24 * 7 round-the-clock reliably monitorings

The intelligent television monitoring system, thoroughly changed the pattern that by the trouble free service personnel monitored picture was monitored and analyzed fully in the past, it is by being embedded in the intelligent video module in the headend equipments such as web camera or video server, the picture of being monitored is uninterruptedly analyzed, and adopt intelligent algorithm and user-defined security model to compare, in case finding has security threat, at once to Surveillance center's early warning or warning.

2, because powerful intelligent characteristic is arranged, can improve the warning accuracy

The intelligent television monitoring system can improve the warning accuracy effectively, greatly the generation that reduces wrong report and fail to report phenomenon.Because headend equipments such as the web camera of intelligent television monitoring system or video servers, integrated powerful image processing, Identification And Traceability, and operation high-grade intelligent algorithm make the user can more accurately define the feature of security threat.Thereby can reduce wrong report effectively and fail to report phenomenon, to reduce the hash amount.Can define virtual warning lines one as the user, and regulation has only to cross over and just produces warning when this warning line enters or walks out, from the warning line next door through then not producing warning.Also we can say, only pass through movable just generation of door as user definition and report to the police, and do not produce warning etc. through the activity of door.

3, because powerful intelligent characteristic is arranged, can improve response speed

The intelligent television monitoring system, have the intelligent characteristic more powerful than general network TV monitor system, it can discern suspicious activity, left over suspicious object (explosive) in public places if any the people, overlong time that the people stops in the sensitizing range etc. is perhaps arranged, therefore just can point out the Security Officer to pay close attention to relevant monitored picture before security threat takes place, making security department that time enough be arranged is the potential ready work of threat.In addition; can also make the user more definite be defined in specific security threat and occur the time; the action that should take, and guarantee that by supervisory control system itself the crisis treatment step can accurately carry out according to definite plan, prevent the delay that in confusion, causes effectively owing to human factor.

4, the effective purposes of extending video resource

No matter be traditional TV monitor system or network TV monitoring system, the video pictures that it monitored all can only be applied in the security monitoring field, and in the intelligent television monitoring system, these video resources can also have more purposes.As video resource being applied to non-security fields: the surveillance video of hall of bank can be used for strengthening the service to the client, and the intelligent television monitoring system can discern VIP user's feature automatically, and notifies the contact staff in time to carry out services; When having the people to fall accidentally among the discovery crowd, in time near the staff the notice offers help; The intelligent television monitoring system, client's quantity that can also help to add up the same day and patronize is in order to analyze turnover etc.

Description of drawings

The invention will be further described below in conjunction with drawings and Examples, in the accompanying drawing:

Fig. 1 is the structured flowchart of intelligent video monitoring system of the present invention;

Fig. 2 is the flow chart of intelligent video monitoring method of the present invention;

Fig. 3 is the original image that collects of intelligent video monitoring method of the present invention;

Fig. 4 is Fig. 3 background modeling figure;

Fig. 5 is moving object detection figure shown in Figure 3;

Fig. 6 is that figure is cut apart in moving object shown in Figure 3;

Fig. 7 is that detection figure behind the shade is removed in moving object shown in Figure 3;

Fig. 8 is that moving object shown in Figure 3 goes to cut apart figure behind the shade;

Fig. 9 is a moving object trajectory diagram shown in Figure 3.

Embodiment

As shown in Figure 1, in intelligent video monitoring system of the present invention, comprise camera 31, first warning device 33, encode processor 2, data analysis processor 1, display 42; Wherein, camera 31 is connected with encode processor 2 respectively with first warning device 33, and encode processor 2 is connected with data analysis processor 1 respectively with display 42; It is vision signal that encode processor 2 is used for the video flowing that camera 31 sends is carried out compressed encoding, and especially, 2 pairs of picture signals of encode processor are carried out MPEG4, H.263, H.264 or the M-JPEG compressed encoding; Data analysis processor 1 is used for that the vision signal that encode processor 2 sends is sent to display 42 and shows, and carries out analyzing and processing, when abnormal conditions occurring, by encode processor 2 alarm signal is sent to first warning device 33 to report to the police.In addition can be according to actual needs and the customer requirements flexible configuration for the quantity of camera 31, first warning device 33 and display 42.

When specific design, native system adopts the operational mode of C/S model and the combination of B/S pattern.Development environment is: operating system: Microsoft Windows XP.Database software: Microsoft SQL Server2005, Microsoft Access 2003.Developing instrument: Microsoft Visual C++2005, Microsoft Asp.Net 2005.Software requirement: DirectX 9.0, DonetFramwork 2.0.Hardware environment is that CPU:Pentium IV is more than 3.0, more than the RAM 1G.Other equipment: video acquisition video camera, digital hard disc video recorder, intelligent-tracking clipping the ball, controlling alarm unit and alarm detector, warning lamp warning signal, on-the-spot display device.

In addition, camera 31 is embodied as anti-riot PTZ semicircle video camera, has the automatic following function of PTZ, omnibearing secure site arranges and overcomes the monitoring blind area that when target was invaded, the PTZ automatic tracking module can aim at the mark automatically, target is remained in the monitoring range, when target was hided into barrier behind, video camera can be aimed at barrier always, and the related personnel can rush towards the scene rapidly and handle.The staff can manual operation PTZ control lever, perhaps carries out automatically-monitored tracking by the automation mechanized operation module of software.Left the monitoring range of certain video camera when target, other video camera or transducer can carry out relay to it on every side, as long as specific objective does not leave the whole zone of deploying troops on garrison duty, target can be in monitoring range always.The monitoring mode that this is different from the past when crime dramas takes place, does not need manual the operating of staff, and target is followed the tracks of.Native system can be created alert event automatically, and automatically target is followed the tracks of, fast the concrete position of display alarm scene.In addition, encode processor 2 can be embodied as video capture card, analog video signal is converted to the special equipment of digital video signal.It can receive the analog video signal from video inputs, digital signal is gathered, is quantized into to this signal, compressed encoding becomes digital video then, it is the interface of ccd video camera and computer, what this image capturing system was selected is the DHCG300 of Daheng video frequency collection card, it has acted on the characteristics of PCI image card, be that the IMAQ transmission does not take the CPU time substantially, in the course of the work, video image is delivered to data buffer through multi-channel switcher, decoder, A/D converter with digitized view data.After reduction, ratio compression and Data Format Transform, cover and transfer of data by inner DSP control figure, the transfer of data target location is determined by software, can directly be sent to calculator memory or video memory.It is suitable for image processing, Industry Control, multimedia monitoring, fields such as office automation.

Aspect the serial communication between data analysis processor and ancillary equipment, when data when serial port sends, byte data is converted to the position of serial; When receiving data, the position of serial is converted into byte data.Under the Windows environment, serial ports is the part of system resource.Application program must propose resource bid to operating system and require to open serial ports if will use serial ports to communicate before using, must discharge resource after communication is finished and close serial ports.Native system utilizes the WindowsAPI function to realize communication function.API is an extremely important part that attaches in Windows inside.The API of Windows mainly is a series of very complicated functions and massage set.It can be regarded as the Windows system and strengthens interface for the open general utility functions that provides at its various development systems that move down.

Data analysis processor 1 is to utilize computer vision technique, to the process that video pictures is analyzed, handles, used, generally comprises following four levels: moving target extracts, tracking, target identification and the behavioural analysis of moving target; Wherein, the purpose that moving target extracts is to get rid of external interference effectively, finds and extract the object that moves in the picture, and in other words, it is the process of an evidence obtaining, obtains the required evidence of our video analysis.Exactly because so, his stability and robustness have directly determined tracking, the identification of back, and the performance of behavioural analysis, we can say that it is the basic data analysis of data analysis analysis processor 1.From the angle that technology realizes, it can be divided into three levels: the mutation analysis of video pictures, filtered noise and extracted region.The mutation analysis of video pictures is that (compression or non-compression) carries out simple video analysis to original video stream, obtains some along with the time, the zone of variation relatively took place.Usually the algorithm that adopts comprise consecutive frame do difference or set up background model do poor, and optical flow method or the like.The purpose of filtered noise is to get rid of the disturbance that light variation and nature and non-natural environment change, and therefore how eliminating these interference of noise is effectively to extract a vital task of moving target.Substantially, the reason of noise appearance can be divided into three kinds.One, camera self noise, signal disturb, and DE Camera Shake is tiny and don't be that very continuous bright spot belongs to this class substantially as in the foreground picture some.Its two, light change comprise indoor, the variation of UV light.Outdoor light change comprise Changes in weather (by the cloudy day sky that clears up, clear to overcast, position of sun moves), change round the clock, the moving of shade (cloud, building etc.); Indoor light changes the light and shade variation that comprises light, the position of light source and the variation of direction.And the noise that the light variation is caused is often apparent in view, can appear as the wrong report of large stretch of area in foreground picture.Its three, natural environment disturbs.It comprises the ripple that shakes the water surface, wave of leaf, unsteady cloud, rain, snow; Also have the interference of some non-natural environment to comprise waving of flag, vertically hung scroll, curtain, and the reflection of glass of building wall or the like.Therefore through the foreground picture of denoising, compare and will have greatly improved with the source foreground picture, particularly the general shape of pedestrian and Che has been tending towards obviously, and whole noise is also little a lot.In the extracted region step, handling resulting foreground image by last two links is unit often with the pixel, the global concept of neither one " object ".On the other hand, there are many spaces probably in the foreground area inside of handling like this, makes troubles for the shape of describing object.In this link, the main purpose of extracted region utilizes the Processing Algorithm of some basic bianry images (B﹠W) that the foreground picture that obtains is processed with regard to seeming, plug the gap, and the zone that will connect is distinguished, do as a whole at last, its content can comprise area size, position, shape, color, pattern or the like key feature descriptor, analyzes targetedly for next step.To be added through most of space that this step object the inside comprises, and the global shape of object becomes more level and smooth.

Then, to the tracking of target is to realize the needed prerequisite of any one intelligent video analysis function (cross the border, invade, leave over, steal, pace up and down, traffic statistics or the like), because we must know is for which object, when, any place occurred, and how long had occurred, and travel direction how, or the like information, and these all can only obtain by tracking.The relevant static state of a series of and presentation that has obtained moving target by extracted region is described, as shape, color or the like.Yet the movable information of wanting tracking target and understanding them must utilize these descriptions to set up motion model, promptly carries out object representation, and the method for setting up motion model has a lot, decide according to different needs.The simplest can be the central point or the center of mass point of target, and its benefit is can very clear and definite ground to observe the periodicity of target travel.In addition can also be with the external figure (rectangle of object edge, oval or the like), be used for simply describing shape, size and the position of target object resolving into many rectangles that join like this, thereby can describe the motion conditions of limbs well, be used for analyzing individual action behavior.Specifically, the extraction of moving target is the process of two mutual reciprocity and mutual benefit with following the tracks of in fact.On the one hand, if extract do very accurate, it is very simple that tracking will become, as long as just can in the center of select target; On the other hand, if follow the tracks of do very desirable, we just can extract emphatically in the place that next time point may occur at moving target, and the result who obtains like this can be more accurate.Yet, all there is very big uncertainty just because of this respect, we need weigh both sides and obtain best performance.Certainly, a stable track algorithm is the prerequisite that is preferably showed.The algorithm of following the tracks of has very and arrives, have based on the object color position, and with good grounds movement direction of object has that other objects of cascade are auxiliary to be followed the tracks of, adopt in addition template or the like.But speech and in a word, purpose has only, that is exactly to infer the next position that it is possible according to the motion state (comprising speed, acceleration, direction etc.) before the mobile object.Correct compensation by the moving area information of extracting previously again, the motion state of confirming the final position then and upgrading object is handled for next time point.

More than be the simple scenario of some tracking, often only relate to tracking one or several pinpoint targets.Yet it is complicated a lot of that reality is wanted.This polymerization that comprises blocking, disappear, reappearing of single target and a plurality of targets with separate or the like.We not only need to realize individual tenacious tracking, and need make judgement to these complex situations, thereby take appropriate measures to guarantee can not occur obscuring, careless mistake, wrong phenomenon such as to repeat.The major premise of the video monitoring that the front is involved is single static camera, in addition, the video analysis technology is applied to the direction that a plurality of or Pan/Tilt/Zoom camera also is an awfully hot door.Wherein, autonomous type PTZ follows the tracks of the autonomous focusing that can realize interesting target, mobile and stretching, and does not need the auxiliary of other video camera.Algorithm of using and front we introduced closely similar, just need regulate the PTZ parameter extraly and consider that the PTZ motor moves required time-delay or the like.In addition, also have the relay-type tracking and the master-slave mode video camera of a plurality of video cameras to follow the tracks of or the like, give unnecessary details no longer one by one here.

Identification to moving target is important process, the stability that it not only can enhanced system, reduces rate of false alarm, raises the efficiency, and lay the first stone for next step behavioural analysis.Identification comprises two processes, and one is the process of machine learning, and another is based on result after the study to the identification process of emerging target.Machine learning comprises training and testing.Training is to utilize the information of having known to come guidance machine, makes it have the ability of differentiating object.And test is to utilize known result to test the machine of succeeding in school, and estimates its performance and also relearns after adjusting where necessary again.For example car and people's identification (classification), at first we need car and people's sample set, do training and testing respectively telling training set test set from sample.The method of machine learning has a lot, comprises neural net, Support Vector Machine, data qualification (linear with nonlinear), probability (Bayes, Bayesian network, Markov model, CRF, graphical model or the like).The classification basis can be shape, size, color, pattern, the symmetry of target object, also can be the direction of motion, speed, the acceleration of target object, the rigidity of motion, periodically.Can construct corresponding model, template, distribution or subspace through the machine of learning uses for identification.

In identification process, for a given new object, system compares it and the model of having set up, and selects the label (people, car etc.) of immediate coupling as it.Among perhaps can being mapped to the space of learning well to it or distributing, select the maximum or nearest classification of probability to make label.The purpose of behavioural analysis is to utilize the result of identification, for different targets (people, car etc.), carries out behavior targetedly and judges.It is time of occurrence, direction, position, speed, size, target distance according to one or more target from relative direction etc., realize different functions by different rules.Its basic function that can realize comprises crosses the border, and hides, and hypervelocity is lost, and leaves over, and is detained or the like; Premium Features comprise traffic statistics, and people's individual behavior for example speed is fallen down, and bends over, and sits down; And some and other people or object mutual, for example join article, traffic accident, get on or off the bus etc.The implementation pattern that the behavioural analysis neither one is fixing.Simply can be a rule, as the restriction of speed limit, direction complicated can be a model, as people's limbs model, many people interaction models.

Above system configuration has realized that Control Room carries out analyzing and processing to the video of monitoring on-site transfer, and one side will be judged by the relevant personnel after can will monitoring on-the-spot video image demonstration by display 42; Undertaken judging behind the intellectual analysis by data analysis processor 1 on the other hand; Then if the relevant personnel or data analysis processor 1 judge when abnormal conditions occurring, can transmit control signal to the first on-the-spot warning device 33 of monitoring, report to the police to start.

In addition, in order further to strengthen the function of native system, also can be according to actual needs or customer requirements carry out the expansion of ancillary equipment, for example, intelligent video monitoring system also comprises: pick-up 32 and audio amplifier 43, wherein pick-up 32 is connected with encode processor 2, and audio amplifier 43 is connected with data analysis processor 1; It is audio signal that encode processor 2 also is used for the audio stream that pick-up 32 sends is carried out compressed encoding; Data analysis processor 1 is used for that also the audio signal that pick-up 32 sends is sent to audio amplifier 43 and plays.Under this configuring condition, not only can carry out collection analysis to the video at scene, can also carry out collection analysis to audio frequency, thereby avoid camera not capture video image and the actual situation that has an accident, thus further perfect native system.

In another embodiment, intelligent video monitoring system also comprises: the foot-operated alarm 35 and second warning device 44; Wherein, foot-operated alarm 35 is connected with encode processor 2, and second warning device 44 is connected with data analysis processor 1; Data analysis processor 1 also is used to receive the alarm signal that foot-operated alarm 35 sends by encode processor 2, and controls second warning device 44 and report to the police.Under this configuring condition, it is on-the-spot when abnormal conditions occurring further to have strengthened monitoring, initiatively report to the police to Control Room, thus the intelligent behaviour of further enhanced system.

In a further embodiment, this intelligent video monitoring system also comprises: microphone 41 and loudspeaker 34; Wherein, microphone 41 is connected with data analysis processor 1; Loudspeaker 34 are connected with encode processor 2; Microphone 41 sends to loudspeaker 34 by data analysis processor 1 and encode processor 2 with the audio stream that collects and plays.In this embodiment,, can remind the on-the-spot personnel of monitoring,, thereby avoid accident to take place with the measure of taking to be correlated with by microphone if when the relevant personnel of Control Room find the situation of particularly urgent.

In addition, for the setting of various accessories in the system, first warning device 33 and second warning device 44 are buzzer, multitone alarm or audible-visual annunciator.Camera 31 comprises: at least one camera lens and imageing sensor, and imageing sensor is ccd image sensor or cmos image sensor.2 pairs of picture signals of encode processor are carried out MPEG4, H.263, H.264 or the M-JPEG compressed encoding.

In intelligent video monitoring method of the present invention, mainly comprise two steps: i.e. the tracking of motion target detection and moving target, wherein detect moving target and comprise: S11: gather and preliminary treatment M frame image sequence, wherein, M is a natural number; S12: according to described M frame image sequence, initialization background model; S13: gather the M+1 frame image sequence,, described M+1 frame image sequence is carried out inter-frame difference and background subtraction divisional processing, detect moving target based on described background model; In addition, the pursuit movement target comprises: S21: to detected moving target, set up object module; S22: according to described object module, detected moving target is carried out continuous K frame handle, set up the target following information table, wherein K is a natural number; S23: set up the object matching matrix between adjacent two frames, matrix element is the matching degree of adjacent two interframe object modules, obtains the match condition of moving target; S24: according to object module and object matching matrix, analyze present frame moving target state, and upgrade object module.

As shown in Figure 2, in the implementation process of whole method for supervising, can be divided into three parts, i.e. collection, detection and tracking; Wherein, by high resolution CCD video camera photographic images sequence, adopt video frequency collection card to convert video signal to digital image sequence and input data analysis processor 1.Then, the preceding M frame by video flowing extracts background, then after the difference binaryzation operation through present frame and background, with shape filtering the segmentation result of target is carried out reprocessing, the influence that elimination noise and background disturbance bring.Carry out connected region at last and detect, demarcate moving target accurately.Then, extract the center and the area size information of moving region, set up the target chained list, by Kalman filtering the tracked target regional movement trend in the target chained list is carried out position prediction then, in the predicted position scope, carry out the target area match search, setting up the incidence relation of target, and Kalman filtering is carried out real-time update with the moving region of optimum Match.

Shown in Fig. 4～8, when the initialization background model, relatively the consecutive frame image can find that the background pixel point is slowly to change in time, and difference is little in the regular hour.And the pixel of object of which movement region of variation correspondence alters a great deal.Therefore can utilize the Gauss model modeling to each pixel of the background frames chosen.Here we have become the hsv color pattern to the RGB mode-conversion, use V=(R+G+B)/3 as average, and variance are 0 in a frame, later on according to the variation of light, calculates different values.

In the image preprocessing process, when carrying out the gray processing processing, the GRB of image divides measures equal value, and image have maximum value process, mean value method, weighted average method etc. by the common gray processing method of the inevitable lost part information of one-dimensional characteristic that original three-dimensional feature drops to behind the gray processing.No matter take the sort of method as can be seen, its original color characteristic often is changed or loses, and makes with a kind of binarization method because the different gray processing processing procedure of a width of cloth coloured picture usually obtains different results.

Consider the reasonability of image, select for use following formula to carry out gradation conversion:

Gray＝0.30×R+0.59×G+0.11×B

R＝G＝B＝Gray

The gray value of pixel in the Gray presentation video wherein, R represents the red component of this pixel, G represents green component, the weights of the most reasonable gray level image that B represents that blue component 0.30,0.59,0.11 is respectively that experiment and theoretical derivation proof draw.

When carrying out binarization processing of images, image simply is divided into background and target object, the most frequently used method is chosen a threshold xi exactly, with ξ image is divided into the two large divisions, greater than ξ zone (being generally target object) with less than the zone (being generally background) of ξ, if input picture is f (x, y), output image be g (x, y), then

$g(x,y)=\left\{\begin{array}{c}1,f(x,y)\&GreaterEqual;\mathrm{\&theta;}\\ 0,f(x,y)<\mathrm{\&theta;}\end{array}\right.$

When carrying out the filtering and noise reduction processing of image, the image of importing is carried out filtering remove noise, strengthen image, sharpening.After video image generally passes through preliminary treatment, make the interesting areas effect of visualization improve, help the further work of image.Mainly contain mean filter, medium filtering, morphologic filtering etc.

In the testing process of region of variation, inter-frame difference can detect adjacent two interframe the zone that changes has taken place.In fact this territory comprises that the motion thing promptly appears the district in the zone that former frame covered, and the present zone that covers of moving object promptly is exactly moving object itself in present frame.

This two two field picture is carried out difference processing, and the amount of making difference can be gray scale, brightness, chromatic value or other parameters, and we adopt gray value to carry out difference.At first set a threshold value, after the gray value of former frame subtracts a frame gray value less than this threshold value, then be background, otherwise, then be prospect.This threshold value comprises two parts, and a part is a gray threshold, and another part is represented the variation of light, and it is that all pixel grey scales are worth mean value, with the mean value of this gray scale, makes this difference threshold value can adapt to the variation of light.This object space does not have change in adjacent two frames down if moving object is static in scene, so be detected as the point in the background in difference processing.Can not enter among the subsequent treatment, can be not moving object by flase drop just also.Because region of variation need be further processed with background frames and be partitioned into moving object, thereby here needn't be accurate to choosing of threshold value, accommodation is very wide.

Image motion means image change.In the moving object detection algorithm one basic according to being the variation of image intensity, can represent the relative variation of intensity with the difference of a pair of image of adjacent time in the image sequence, and the image difference operation definition is

f _d(p，t ₁，t ₂)＝f(p，t ₂)-f(p，t ₁)

F in the formula _dBe difference image, and p=(x, y).The following formula computing relates to the additive operation of corresponding pixel intensity, therefore this algorithm is quite simple, and be suitable for Parallel Implementation, image difference has reflected the higher level character of scenery to a certain extent or has lain in the variation of the sensor movement on the plane of delineation.If have several relatively independent moving objects and a mobile transducer in the scenery, difference image is the combination of these motions.Analysis to difference image can draw as drawing a conclusion:

(1) image difference can be used as and image function is carried out a kind of of time differentiate approaches.Simple 2 finite differences are to time interval t ₂-t ₁The intermediate point place

A kind of approaching.

(2) difference image has the character of edge image, and this is because the difference algorithm and the image gradient function operator of image have similar character.

(3) in real image, difference image is the same with the stagnating margin image, is not to be made up of the contour area of ideal sealing, and expresses incomplete change information often.For example, when an object when similarly moving in the background with its plane of delineation intensity (or texture), just can not obtain useful difference image information.The entrained information of difference image is not that absolute image intensity changes just, and it relates to the type of variation.

Travel direction can also be simply estimated in the variation of two two field picture intensity before and after difference image had reflected.But also there is following limitation in difference image; The first, the difference image of front and back two frames can only reflect that the relative position of moving object in this two two field picture changes; The second, it has also ignored slow moving target and moving wisp.

In the motion target detection process, pixel in the motion change zone that splits and Gauss model are separately gone match.The characteristic that we measure this point with the gray scale of pixel, less than a specific threshold value, then being judged as is to appear the district as if gray scale, otherwise is moving object.

Suppose that the video sequence of being studied is { f _k(x, y) } _K-1 ⁿ(K is the frame preface, and N is the totalframes of video sequence), note

f _k(x，y)＝b _k1k+1(x，y)+a(x，y)+u _k(x，y)+n _k(x，y)

f _k+1(x，y)＝b _k1k+1(x，y)+a(x+Δx，y+Δy)+v _k+1(x，y)

B wherein _K1k+1(x, y) expression k, the common background district between the k+1 frame, (x, y), a (x+ Δ x, y+ Δ y) is corresponding K respectively, the motion target area in the K+1 frame, the displacement vector of (Δ x, Δ y) expression moving target from the K frame to the K+1 frame, u for a _k(x, y), v _K+1(x, y) represent respectively by moving target cause at K, the background area that is capped and appears in the K+1 frame, n _k(x, y), n _K+1(x y) represents K respectively, the noise in the K+1 frame.

Difference image between the consecutive frame

D(x，y)＝f _k+1(x，y)-f _k(x，y)

＝[a(x+Δx，y+Δy)-a(x，y)]+v _k+1(x，y)-v _k(x，y)

+n _k+1(x，y)-n _k(x，y)

In the following formula, and a (x+ Δ x, y+ Δ y)-a (x, y), v _K+1(x, y) and v _k(x, y) all belong to the motion change zone make MR (x, y)=[a (x+ Δ x, y+ Δ y)-a (x, y)]+v _K+1(x, y)-v _k(x, y) expression motion change zone, n (x, y)=+ n _K+1(x, y)-n _k(x, y) relative noise between adjacent two frames of expression then can obtain

D(x，y)＝MR(x，y)+n(x，y)

By following formula as can be known, comprise motion change zone and noise two parts that moving target causes in the difference image, wherein the motion change zone comprises real motion target area again, background area three parts that are capped and appear, in order to detect moving target exactly, especially very fast when the movement velocity of target, when the moving displacement between consecutive frame causes the background area that is capped and appears in the motion change zone big greatly, can consider the boundary information of moving target.The border of moving target can be united the result of present frame rim detection and motion change zone and be obtained.At last, detect moving target according to filling the motion target area that obtains by moving object boundary.

In the process of context update, upgrade with different renewal frequencies with background area appearing the district.Certainly, appear the renewal in district frequently greater than background area.Because appear the district, revealed again at present frame, so will obtain faster upgrading for previous frame is the zone that moving object covers.When modeling, there is the object of motion to exist even this processing policy makes, also can obtains clean background frames model rapidly along with moving of moving object.

In order to reflect of the influence of factors such as illumination variation, noise to real background, can periodically carry out the context update operation, context update can adopt this algorithm, establishes I _CbBe the background image of current preservation in the system, I _C1Be the image of the current collection of system, calculate the difference image I of the two _Di, I _Di=| I _Ci-I _Cb|, self adaptation is calculated I _DiThreshold value T, it is carried out binaryzation obtains I _Mark:

$I_{\mathrm{mark}}(x,y)=\left\{\begin{array}{c}1,I_{\mathrm{di}}(x,y)\&GreaterEqual;T,\\ 0,I_{\mathrm{di}}(x,y)<T.\end{array}\right.$

With I _MarkBe switch function, construct instant background I _b:

$I_b(x,y)=\left\{\begin{array}{c}{I}_{\mathrm{cb}}(x,y),I_{\mathrm{mask}}(x,y)=1,\\ I_{\mathrm{cn}}(x,y),I_{\mathrm{mask}}(x,y)=0.\end{array}\right.$

Then with the following formula background image updating:

I _cb(x，y)＝a*I _b(x，y)+(I-a)*I _cb(x，y).

Wherein, a is a weight coefficient.

In the noise remove process, because the form of noise is varied: isolated noise point has the noise than small size, the noise of inside, moving region etc.Noise for isolated noise point and small size can adopt morphological erosion and the removal of expanding, but for the noise of componental movement object of which movement intra-zone and to have a noise piece of certain area undesirable as morphological erosion and morphology expansion effect, the processing of this partial noise mainly adopts the limit to lead to the zone marker method, count the number of all connected components on the binary map by the logical zone marker in limit, and count the area of each connected component, features such as length and width.According to these statistical property filtering noise pieces, and fill some cavities that may occur in the vehicle.

The two-value form expands and corrosion is the most basic two-value morphological transformation operation, can constitute other complicated morphological transformation thus, as the form opening operation, and form closed operation, skeletal extraction, computings such as refinement.In preliminary treatment, use the erosion of form mansion, expansion, open and close computing, therefore only be introduced at these several basic morphological operations.

A set A translation distance X can be expressed as A+X, and it is defined as:

A+x＝{a+x；a∈A}

From how much shown in figure (7-2), A+X represent A along vector X translation one segment distance.

Definition B is a bianry image, and S is given structural element.Obtain basic two-value morphological transformation thus.The corrosion of two-value form:

BθS＝∩{B-s，s∈S}

Corrosion can be by with input picture translation-S, and calculates the common factor of all translations and obtain, and wherein structural elements is defined as flat structure.

The two-value form expands:

$B\&CirclePlus;S=\&cup;\{B+s;s\&Element;S\}---(7-12)$

Expansion can have a translation input picture by the institute of relative structural element, calculates its union then to obtain, and wherein structural elements is defined as disc structure.

The two-value form is opened:

$\mathrm{BoS}=\left(\mathrm{B\&Theta;S}\right)\&CirclePlus;S---(7-13)$

The two-value form is closed:

$\mathrm{BoS}=(B\&CirclePlus;S)\mathrm{\&Theta;S}---(7-14)$

Erosion operation in the two-value morphological transformation is a contracted transformation, and it makes that target obtains shrinking, and make hole obtain expansion, and dilation operation is an expansion conversion, and it makes target obtain expansion, and hole obtains shrinking.Therefore will expand and corrosion combines, one side can but be removed noise spot, can fill hole on the other hand.Be the effect and the effect of form open and close computing, concrete effect is shown in figure (7-5).Earlier input picture is carried out erosion operation, again the image after the corrosion is carried out dilation operation, then interference such as noise spot that can some are little, protruding point but remove.While is the interior hole of constriction zone well, and this also is the result of form opening operation.After above-mentioned processing, oneself removes the noise in the background, and the moving target of Ti Quing is more reasonable like this.

Under the stronger situation of illumination, detected moving object meeting includes its shade, and the demanding occasion of the moving object accuracy of separation is also needed shade is removed.

In obtaining the process of moving region, native system adopts the inter-class variance method of genetic optimization threshold value, and step is as follows:

1, parameter coding: because genetic algorithm can not directly be handled the data of separating of understanding the space, therefore must they be expressed as the genotype string structure data in hereditary space by coding, parameter coding adopts binary coding and two kinds of forms of real coding usually.Because the gradation of image value is between 0-255, so available 8 binary code 00000000-11111111 represent the branch threshold value, candidate's threshold value is promptly between 0-255.

2, initialization of population: between 0-255, produce 20 individualities at random, it is encoded as the initial population of genetic algorithm by binary form.

3, population fitness function: definition population fitness function is:

$f={\mathrm{\&alpha;}}_1{(u_1-u)}^2+{\&PartialD;}_2{(u_2-u)}^2$

Make the fitness function value obtain the optimal threshold T that maximum T is difference image ^*, the present frame difference image is divided into prospect and background two parts with him.

4, selection operation: selection operation is how to be used for determining that from parent colony the operation purpose is for fear of gene delection, improves global convergence and computational efficiency by choosing a kind of genetic algorithm of which individual inheritance in the colony of future generation someway.Ratio is selected to have adopted steamer gambling mode to realize.

5, interlace operation: the individuality at first the group being estimated takes out a pair of individuality that will match at random, according to bit string length, to the individual picked at random crossover location that will match, implement interlace operation according to crossover probability, subsequently the chromosome of two mutual fetters is pressed single-point intersection its portion gene of exchange mutually at crossover location, thus two new individualities of form.

6, mutation operation: variation be with less probability change randomly on the chromosome string some the position.The variation probability here gets 0.01, there are two kinds of situations in binary string genetic mutation: 01,10.

7, termination condition: genetic algebra reaches at 60 o'clock, finish the present frame difference image based on the determining of the dynamic threshold of genetic algorithm, with the threshold value obtained as optimal threshold to the difference image binaryzation.

The connected region that keeps in the bianry image is analyzed, judged the attribute of target object, accurately the setting movement object space.Calculate the center of gravity of moving object in the plane of delineation, obtain the coordinate of center of gravity in the plane of delineation.After obtaining the coordinate of moving object in the plane of delineation, it is mapped in the true environment, just can obtains the position of pedestrian at real world.Utilize bianry image in the horizontal direction with on the vertical direction location on border up and down to be carried out in the moving region, this method can well be oriented the scope of moving region.

In the target's feature-extraction process, in the recognition and tracking process,, determine that exactly target signature is the key of motion tracking and coupling for effectively identification and tracking, the target signature quality of extraction directly has influence on the precision and the speed of target identification.The principal character of target comprises: the border of target, and it comprises four borders up and down, it is to be determined in the projection of X-axis and Y-axis by target; The area of target, it is the pixel number that object boundary surrounds; The gravity center characteristics of target, it is the position of moving target in present frame; The principal axis of inertia of target, its direction are the bearing of trends that is used for describing target.

Setting up in the process of object module is according to the target signature of extracting from bianry image, for detected moving target is set up corresponding feature templates.Same moving target is in two adjacent two field pictures, and changing features such as movement position, shape, area are little.Utilize a little characteristics, can set up feature templates for detected target.

As shown in Figure 9, carry out in the Kalman filtering position prediction numerous advantages such as the Kalman filter has simply, real-time is good, widespread usage in engineering for moving target.In Target Tracking System, particularly under the complex background situation on a surface target the tracking, the correlation tracking algorithm is a kind of algorithm of using always.But problem is the method for the color global search of traditional related algorithm, makes amount of calculation quite big, is difficult for real-time implementation, and when the target partial occlusion took place, target was lost easily.For addressing this problem, this article adopts a kind of target correlation tracking method based on the Kalman filter, the forecast function that makes full use of the Kalman filter is predicted the zone that the next frame target may occur, in less estimation range, carry out the relevant matches computing then, find best relevant matches point, the target correlation tracking has more initiative.

Wherein, the filter principle is carried out the algorithm of linear minimum variance estimation error for the Kalman filter is one to the status switch of dynamical system, comes descriptive system by dynamic state agenda and observation agenda.It can begin observation as starting point more arbitrarily, adopts the method for recursive filtering to calculate.

If the state agenda of linear system and observation agenda are respectively:

State equation: x _k=AX _K-1+ W _K-1

Observational equation: X _K=AX _K-1+ W _K-1

Here, XK is that constantly to maintain system state vector: ZK be K timetable m * 1 dimension observation vector to K in n * 1; A is that n * n maintains the system state-transition matrix; HK is that m * n maintains overall view survey matrix; WKJ is that K ties up the systematic observation noise vector that random disturbances noise vector: VK is K moment m * 1 dimension in the n of process * 1 constantly.WK herein, VK is assumed to be mutually independently zero-mean white Gaussian noise vector usually, and we make QK and RK be respectively their covariance matrixes:

Q _k＝E{W _KW _K ^r}

R _K＝E{V _KV _K ^r}

Because system is definite, then A and HK are known, and WK-1 and the satisfied certain hypothesis of VK, and also known, establishing PK is the covariance matrix of XK, P _k ^rBe XK and

The error covariance matrix.

The Kalman filter equation: the error covariance that the Kalman filter is calculated the estimation of the posteriority of the system mode of each moment point K reduces to minimum, and its divides prediction and correction two parts to finish, and the Kalman filter equation is as follows:

The state update equation:

Kalman gain coefficient equation: $k_k=p_k^r{H}_K^T{(H_K{P}_K^r{H}_k^t+R)}^{-1}$

The state update equation: ${\hat{x}}_k=\widehat{x_k^r}+k_k(Z_K-H_K{\stackrel{{^}^{\&prime;}}{X}}_k)$

The covariance update equation: $P_K=(I-K_k{H}_k)p_k^1$

Above-mentioned recursion equation can be described intuitively by Fig. 8-2, and it is very beneficial for the computer programming realization as can be seen.

Kalman filtering utilizes feedback control system to estimate motion state, can estimate state sometime, and obtain the predicted value of this state.Kalman filtering formula is told two parts: prediction and correction.Wherein, predicted portions is responsible for handy current state and error covariance is estimated next state constantly, obtains prior estimate; Retouch is responsible for feedback, and new actual observed value is considered with priori estimates, estimates thereby obtain posteriority, after finishing prediction at every turn and revising, predict next prior estimate constantly by posterior estimate, repeat above step, the recurrence operation principle of Here it is Kalman filter.

For the application of Kalman filter in trajectory predictions, the motion state parameters of hypothetical target is the position and the speed of a certain moment target.In tracing process, because the time interval of adjacent two two field pictures is shorter, target changes smaller in so short time interval internal state, can hypothetical target be uniform motion in the unit interval.

Definition Kalman filter system state x _kBe one 4 dimensional vector (XS _k, YS _k, XV _K, YV _K) ^rXS _k, YS _k, XV _K, YV _KBe respectively position and the speed of target on X-axis and Y direction.By images match, so can only obtain dwelling, the position definition two-dimensional observation vector Z of target _K=(XW _K, YW _K) ^rThe coordinate that the expression coupling obtains.

Because target is uniform motion in the unit interval, definition status transfer matrix A is:

$A=\left(\begin{array}{ccc}1& 0& \mathrm{\&Delta;t}\\ 0& 1& 0\\ 0& 0& 1\end{array}\right)$

Wherein Δ t represents the time interval between the two continuous frames image.

By the relation of system mode and observer state as can be known, observing matrix H _KFor:

$H_K=\left(\begin{array}{cc}1& 0\\ 0& 1\end{array}\right)$

Above supposed w _k, V _KBe generally mutually independently zero-mean white Gaussian noise vector, the covariance matrix of therefore establishing them is respectively:

$Q_k=\left[\begin{array}{cccc}1& 0& 0& 0\\ 0& 1& 0& 0\\ 0& 0& 1& 0\\ 0& 0& 0& 1\end{array}\right],$ $R_k=\left[\begin{array}{cc}1& 0\\ 0& 1\end{array}\right];$

In tracing process, use the motion of Kalman filter estimating target to be divided into four-stage, be respectively initialization, status predication, coupling and the state correction of filter.The specific implementation step is as follows :-

The first step: initialization.Will carry out initialization to filter when using the Kalman filter for the first time, be the first position and the speed of target with the X0 initialize.Under the speed condition of unknown, can be made as 0, and the record present image constantly, establishes initial error covariance P0=0 simultaneously.

Second step: prediction.Before in every two field picture of new input, carrying out match search, the time interval Δ t of record and previous frame image, the motion state of prediction current goal The error of prediction is designated as Δ p _k=w _k-S _K, be used for the calculating of next frame region of search.Further predict new error covariance.

The 3rd step: coupling.Setting with

In (xsk, ysk) for the zone at center be the region of search, in this zone, seek best match position, find optimal moving target, target area image is duplicated to TK+1, and first pixel coordinate of the upper left corner, target area is two-dimensional observation vector (XWK, YWK), substitution state update equation obtains that (XSK+1 YSK+1), calculates measuring speed VK+1=(the SK+1-SK)/Δ T of target simultaneously.

The 4th step: revise.Obtain Kalman filter gain coefficient.(XWK YWK), obtains by the revised state vector of current actual observation, simultaneously the round-off error variance matrix according to ZK=.

For the coupling of object module, be feature to the processing of recognition and tracking, template matches is finished tracking, and the similarity criterion between the mainly leading different images feature is mated.Provide following several similarity to pass judgment on: the area similarity function; The shape similarity function; Direction of motion consistency function; The displacement reliability function; The fuzzy similarity function; Set a multifactorial evaluation function for this reason, do not belonged to the target signature function.Two sizes that target is the possibility of same object of its expression two continuous frames.Its value depends on top 5 decision functions.

For target following, the result according to the To Template coupling associates the same moving target in the video image, sets up corresponding relation between the target in image sequence, sets up object key, obtains the complete movement locus of each target.Tracking comprises the steps: 1, the setting of tracing area, and the motion of object is ever-changing, and we only need to follow the tracks of our interesting areas and get final product.Tracing area is set will be in conjunction with the shooting situation of CCD and the position of area-of-interest; Distance between the end lines of main consideration start line wants enough, guarantees to carry out the tracking of continuous multiple frames to target.Its shape is provided with as required, can be rectangle, also trapezoidal or polygon.Help like this getting rid of and disturb and the raising processing speed.2, set up the trace information table, target following is exactly to extract the result according to target, the target that enters tracing area is carried out successive frame to be handled, according to the detection target location of extracting in each two field picture, realize behavioral analysis of moving targets, therefore need set up the target following information table, preserve the characteristic information that trace information writes down tracked target in each two field picture, reconstruction of objects is at the motion track of tracing area, and object constantly changes in the tracing area, so the record of tables of data also must dynamically update with the variation of object.

Feature in the every row of Track Table determines as required, can be the information such as position, movement locus of characteristic information such as jump target sequence number, border, length and width, area, center of gravity and the principal axis of inertia and target.Follow the tracks of sequence number and represent that object enters tracking area arrangement sequence number successively, same target keeps same tracking sequence number in tracing process.

Characteristic information is the important information of following the tracks of in order to ensure target reliability.Wherein initiation feature is represented the characteristic information that extracts when target enters tracking area first; Because the continuity of target travel, target knot on the characteristic of every frame performance also has changes, and therefore need extract (or renewal) target signature in every two field picture, guarantees the stability of following the tracks of.Trace information is the basis of behavioural analysis, and the prediction of no matter taking exercises, analysis all need the positional information of target in each two field picture.The predicted position of target is to the predicted value of position in the next frame according to its historical position information; The movement locus of target is generated by historical position information, is the main foundation of behavioural analysis; Dbjective state is meant the state of target in historical frames.Utilize the feature of target in this two field picture can obtain the state of target in the previous frame image, the state of target mainly comprises normal condition and transitional states.After every two field picture is finished dealing with, need to upgrade the trace information table.To newly enter tracing area moving target information and add the trace information table, the target of rolling tracing area away from will be deleted from information table.After tracked moving target is finished the tracking of this frame, target is added information table at the characteristic information of this frame; Follow the tracks of step number and add 1 automatically; Step number and the state of target in historical frames according to current tracking determine whether to calculate the center of gravity predicted value of moving target in next frame; Determine the state of this target at previous frame.

For tracking strategy, the method that we have adopted characteristic matching to add trajectory predictions is followed the tracks of.The tracking hunting zone of characteristic matching is big, and in order to reduce the target search scope, we adopt the tracking strategy of trajectory predictions, have significantly reduced the target search time, has improved the stability and the reliability of following the tracks of.Trajectory predictions is predicted target location in the next frame image by the historical position information of moving target.After next image frame grabber is finished, at first near this position more among a small circle in ferret out.If find this target, the record object positional information; If do not find, then enlarge the hunting zone.

When moving target has target overlapping, need do merging or separating treatment in two continuous frames.The situation that needs can be done merging and separating treatment is divided into two types: the first kind is merging or the separation between the old and new's target, occurs between the target and present frame fresh target in the previous frame; Second class is merging or the separation between old target, occurs between the old target in the previous frame.

When the previous frame target is followed the tracks of disappearance in present frame, and should not leave tracing area according to trajectory predictions, at this moment we can have target overlapping by preliminary judgement, record disappearance target identification, and do following processing:

Searching position target in the certain limit of estimation range, and predict the outcome according to crafty plot and preliminary to judge that having in this zone denys that a plurality of known target disappear; If have, then give this unknown object a plurality of disappearance target identifications simultaneously, wait until target and separate the back affirmation; If do not have, then except that giving this unknown object new logo, also to give the sign of disappearance target.Here it should be noted that: determining of region of search is mainly definite according to the maximum of former step predicated errors; Here can only do preliminary judgement, final result also will be treated could determine after target is separated, can be referred to as doubtful merging phenomenon.Down several frames are followed the tracks of this doubtful merging targets, and emphasis is judged whether rapid expanding of target area or length, with the attribute of the doubtful merging target of this further judgement.Violent if expand, just can conclude to have to merge and take place; Otherwise still be doubtful merging target.

If the merging target, after the tracking of number frame, merging target must separate.At this moment need the objective attribute target attribute definite, determine the character (new, old target) of each target by characteristic matching according to sign; Because therefore characteristic the unknown of fresh target need be handled respectively these two types.For fresh target, extract this clarification of objective parameter, and in the trace information table, increase fresh target.For old target, need the characteristic information in the lastest imformation table.Carrying out the tracking of fresh target again handles.No matter new, old target do initial tracking processing again.When old object matching, find apart from merging this preceding target of the nearest pairing merging of that frame, fuzzy similarity functional value between the target after calculating it and separating, principle by maximum fuzzy characteristic similarity functional value, find the corresponding relation that separates between back target and the preceding target of separation, and a fresh target characteristic information that separates is joined in the information row that separates preceding target place by corresponding relation.

The present invention describes by several specific embodiments, it will be appreciated by those skilled in the art that, without departing from the present invention, can also carry out various conversion and be equal to alternative the present invention.In addition, at particular condition or concrete condition, can make various modifications to the present invention, and not depart from the scope of the present invention.Therefore, the present invention is not limited to disclosed specific embodiment, and should comprise the whole execution modes that fall in the claim scope of the present invention.

Claims (10)

1, a kind of intelligent video monitoring method is characterized in that, may further comprise the steps:

S1: detect moving target:

S11: gather and preliminary treatment M frame image sequence, wherein, M is a natural number;

S12: according to described M frame image sequence, initialization background model;

S13: gather the M+1 frame image sequence,, described M+1 frame image sequence is carried out inter-frame difference and background subtraction divisional processing, detect moving target based on described background model;

S2: pursuit movement target:

S21:, set up object module to detected moving target;

S22: according to described object module, detected moving target is carried out continuous K frame handle, set up the target following information table, wherein K is a natural number;

S23: set up the object matching matrix between adjacent two frames, matrix element is the matching degree of adjacent two interframe object modules, obtains the match condition of moving target;

S24: according to object module and object matching matrix, analyze present frame moving target state, and upgrade object module.

2, intelligent video monitoring method according to claim 1 is characterized in that, in described step S11, described preliminary treatment comprises to be handled and the filtering and noise reduction processing the gray processing of M frame image sequence.

3, intelligent video monitoring method according to claim 1 and 2 is characterized in that, in described step S12, also comprises each pixel in the background model is carried out Gauss's modeling.

4, intelligent video monitoring method according to claim 3 is characterized in that, in described step S13, the M+1 frame image sequence is carried out inter-frame difference when handling, with the area update that do not change in background model; The zone and the background model that change are carried out match, appear district and background model to distinguish.

5, intelligent video monitoring method according to claim 4 is characterized in that, in step S13, also comprises detected moving target is gone Shadows Processing.

6, intelligent video monitoring method according to claim 5 is characterized in that, described object module comprises: area similarity function, shape similarity function, direction of motion consistency function, displacement reliability function and fuzzy similarity function.

7, a kind of intelligent video monitoring system is characterized in that, comprises camera, first warning device, encode processor, data analysis processor, display; Wherein, camera is connected with encode processor respectively with first warning device, and encode processor is connected with the data analysis processor respectively with display;

It is vision signal that encode processor is used for the video flowing that camera sends is carried out compressed encoding;

The data analysis processor is used for that the vision signal that encode processor sends is sent to display and shows, and carries out analyzing and processing, when abnormal conditions occurring, by encode processor alarm signal is sent to first warning device to report to the police.

8, intelligent video monitoring system according to claim 7 is characterized in that, described intelligent video monitoring system also comprises: pick-up and audio amplifier, and wherein pick-up is connected with encode processor, and audio amplifier is connected with the data analysis processor;

It is audio signal that encode processor also is used for the audio stream that pick-up sends is carried out compressed encoding;

The data analysis processor is used for that also the audio signal that pick-up sends is sent to audio amplifier and plays.

9, according to claim 7 or 8 described intelligent video monitoring systems, it is characterized in that described intelligent video monitoring system also comprises: the foot-operated alarm and second warning device; Wherein, foot-operated alarm is connected with encode processor, and second warning device is connected with the data analysis processor;

The data analysis processor also is used to receive the alarm signal that foot-operated alarm sends by encode processor, and controls second warning device and report to the police.

10, intelligent video monitoring system according to claim 9 is characterized in that, described intelligent video monitoring system also comprises: microphone and loudspeaker; Wherein, microphone is connected with the data analysis processor; Loudspeaker are connected with encode processor; Microphone sends to loudspeaker by data analysis processor and encode processor with the audio stream that collects and plays; First warning device and second warning device are buzzer, multitone alarm or audible-visual annunciator; Encode processor carries out MPEG4, H.263, H.264 or the M-JPEG compressed encoding to picture signal; Camera comprises: at least one camera lens and imageing sensor; Imageing sensor is ccd image sensor or cmos image sensor.

Images