CN105405150B

CN105405150B - Anomaly detection method and device based on fusion feature

Info

Publication number: CN105405150B
Application number: CN201510689357.XA
Authority: CN
Inventors: 许健; 郑慧; 万定锐
Original assignee: Netposa Technologies Ltd
Current assignee: Netposa Technologies Ltd
Priority date: 2015-10-21
Filing date: 2015-10-21
Publication date: 2019-04-30
Anticipated expiration: 2035-10-21
Also published as: CN105405150A

Abstract

The present invention provides a kind of anomaly detection method and device based on fusion feature determines the behavior type of moving target including the detecting and tracking processing result according to moving target in video to be tested；Extract the various dimensions feature in moving target: the variation degree of pixel, the arrangement tightness degree of pixel, global shape, the similarity degree of frame image, motion feature, position feature and morphological feature；Various dimensions feature is analyzed and processed according to behavior type corresponding Fusion Features model, and judges moving target with the presence or absence of abnormal behaviour according to processing result；The inventive features of above-mentioned multiple abnormal behaviours, it is capable of the robustness and stability of effective boosting algorithm, the Fusion Features model obtained according to the learning training of a large amount of abnormal behaviours, which is analyzed and processed above-mentioned various dimensions feature, can effectively avoid algorithm over-fitting or poor fitting problem, it is suitable for Various Complex application scenarios, a large amount of time cost and human cost are saved, has very high promotional value.

Description

Anomaly detection method and device based on fusion feature

Technical field

The present invention relates to field of video monitoring, in particular to based on fusion feature anomaly detection method and Device.

Background technique

Currently, many public places such as bank, market, station and traffic intersection are provided with video monitoring apparatus, it is used for The video image in default monitoring area is acquired in real time, and above-mentioned video image is sent to monitoring center, forms monitoring system. But there is still a need for more to be accomplished manually for corresponding many actual monitor tasks in above-mentioned monitoring system, and existing view Frequency monitoring system is generally only the video image of storage acquisition, and any processing will not be carried out to the video image, therefore These are typically only capable to be used as post-mordem forensics without the video image of explanation (handling), are unable to give full play the real-time of monitoring And initiative.

As in recent years, the pressure of fight against terrorism and violence incrementally increases, for specific public place (such as subway station, railway station, quotient Etc.) in the abnormal behaviour that occurs be measured in real time and the technology alarmed seems ever more important, and above-mentioned abnormal behaviour is general Finger runs, suddenly rapidly, fight, behaviors or the event such as crowd is restless.

In order to give full play to monitoring system for the real-time of above-mentioned detection abnormal behaviour and the advantage of initiative, related skill Art provides a kind of anomaly detection method, comprising: selects moving target in all targets, then extracts moving target Certain a kind of single features (such as displacement of movement velocity, moving target size and movement), then divides the single features Analysis processing, finally judges whether the behavior of the moving target belongs to abnormal behaviour scope according to analysis and processing result；By above-mentioned Method can be analyzed in real time, track and differentiate monitored object (video image acquired), and the superior when anomalous event occurs Prompted and reported when department or government department, be government department, associated safety administrative department to anomalous event carry out and When decision and take correct action provide support.

Inventor has found that often there are significant limitations in practical application for above-mentioned single features under study for action, is easy Lead to failing to report or reporting by mistake for abnormal behaviour.

Summary of the invention

The purpose of the present invention is to provide anomaly detection methods and device based on fusion feature, can be more accurate And it is comprehensive abnormal behaviour is detected, testing result reliability is higher, and effectively improve algorithm robustness and Stability is suitable for Various Complex application scenarios, has very high promotional value.

In a first aspect, the embodiment of the invention provides a kind of anomaly detection methods based on fusion feature, comprising:

According to the detecting and tracking processing result of moving target in video to be tested, the behavior class of the moving target is determined Type；

Extract the various dimensions feature in the moving target；The various dimensions feature includes the spy of moving target as described below It is multiple in sign: the variation degree of pixel, the arrangement tightness degree of pixel, global shape, the similarity degree of frame image, fortune Dynamic feature, position feature and morphological feature；

The various dimensions feature is analyzed and processed according to the behavior type corresponding Fusion Features model, and according to Analysis and processing result judges the moving target with the presence or absence of abnormal behaviour.

With reference to first aspect, the embodiment of the invention provides the first possible embodiments of first aspect, wherein mentions The various dimensions feature in the moving target is taken to include:

Calculate the variation degree DisturbRate=FC/FA of the pixel of the moving target in predetermined period；Wherein, FA indicates the quantity of all pixels point in the corresponding foreground area of the predetermined period；FC indicates certain time week in foreground area The quantity of the pixel in the foreground area was not in front of phase；；

Calculate the arrangement tightness degree CompactRate=CS/ of the pixel of the moving target in the predetermined period CN；Wherein, CN indicates that all four neighborhoods in the foreground area are in the quantity of the foreground pixel point of the foreground area；CS In expression CN there is the quantity of the foreground pixel point of luminance difference in all four neighborhoods；

Calculate circularity Ω=P/ (2*sqrt (π A)) of the global shape of the moving target in predetermined period；Wherein, A Indicate the area of the foreground area；P indicates the perimeter of the foreground area；

Calculate the similarity degree AreaSimilarRate=AND/ of frame image in the moving target in the predetermined period OR；Wherein, AND indicates that consecutive frame in the foreground area is in the number of the pixel of foreground area；The OR indicates institute State the number for being not in the pixel of the foreground area in foreground area in consecutive frame.

With reference to first aspect, the embodiment of the invention provides second of possible embodiments of first aspect, wherein mentions The various dimensions feature in the moving target is taken to include:

Extract the motion amplitude of the moving target in predetermined period in the foreground area；The motion amplitude packet Include: length travel, lateral displacement, the height of current foreground area and current foreground area width；

And/or

Calculate the direction of motion OffRate=Offset/Route of the moving target in the predetermined period；Wherein, Offset indicates the actual displacement of the moving target, which includes lateral displacement and length travel；Route indicates described pre- If the accumulative stroke of all actual displacements in the foreground area in the period.

With reference to first aspect, the embodiment of the invention provides the third possible embodiments of first aspect, wherein mentions The various dimensions feature in the moving target is taken to include:

The change in displacement value in predetermined period in the moving target between every two field pictures is extracted, and is become in the displacement When change value is greater than preset threshold, the variation frequency of the position feature of the moving target is added 1, until the predetermined period knot Beam, using the final variation frequency as the position feature of the moving target.

With reference to first aspect, the embodiment of the invention provides the 4th kind of possible embodiments of first aspect, wherein mentions The various dimensions feature in the moving target is taken to include:

Calculate the ratio of width to height of the video area where the video to be tested；

According to the ratio of width to height of the video area, the ratio of width to height of actual foreground area is calculated；

The ratio of width to height of the actual foreground area and default the ratio of width to height threshold value are compared, and sentenced according to comparing result The corresponding actual form of moving target in the actual foreground area of breaking；The actual form includes: single form and more Humanoid state.

With reference to first aspect, the embodiment of the invention provides the 5th kind of possible embodiments of first aspect, wherein root The various dimensions feature is analyzed and processed according to the behavior type corresponding Fusion Features model, and according to analysis processing knot Fruit judges that the moving target includes: with the presence or absence of abnormal behaviour

All feature vectors in the various dimensions feature are extracted, all feature vectors indicate are as follows: F=[F1, F2, F3, F4, F5, F6, F7]；Wherein, F indicates the set of all feature vectors；Feature vector F1, F2, F3, F4, F5, F6 and F7 are respectively The variation degree, the arrangement tightness degree, the global shape, the similarity degree of the frame image, the motion feature, Above-mentioned position feature and the morphological feature；

The corresponding weight coefficient of all feature vectors in the Fusion Features model is obtained, above-mentioned weight coefficient indicates are as follows: M =[w1, w2, w3, w4, w5, w6, w7]；Wherein, M indicates the set of all weight coefficients；Weight coefficient w1, w2, w3, w4, w5, W6 and w7 be respectively the variation degree, it is described arrangement tightness degree, the global shape, the frame image similarity degree, The motion feature, above-mentioned position feature and the corresponding weight coefficient of the morphological feature；

According to formulaTo all feature vectors and the corresponding weight coefficient It is calculated, and when calculated result meets abnormal behaviour threshold value, determines the moving target for abnormal behaviour and alarm；Its In, label indicates calculated result；What T was represented is matrix transposition；Indicate each feature vector pair in the Fusion Features model The weight coefficient answered.

With reference to first aspect, the possible embodiment of the first of first aspect is appointed into possible embodiment in the 5th It anticipates a kind of possible embodiment, the embodiment of the invention provides the 6th kind of possible embodiments of first aspect, wherein institute Stating Fusion Features model is generated previously according to following methods:

The video resource for containing abnormal behaviour is chosen as positive sample and contains the video resource of normal behaviour as negative Sample；

Calculating is trained to the positive sample and the negative sample by support vector machines, establishes default Fusion Features mould Type；

More wheel repetitive exercises are carried out to the default Fusion Features model according to the negative sample, and in the default feature When Fusion Model reaches preset threshold, the Fusion Features model by corresponding default Fusion Features model for practical application is determined.

With reference to first aspect, the embodiment of the invention provides the 7th kind of possible embodiments of first aspect, wherein institute Before the behavior type for stating the determining moving target, further includes:

Classify to the moving target of detecting and tracking, determines the characteristic type of the moving target, and detecting When characteristic type to the moving target is the behavior of people, the step of executing the behavior type for determining the moving target.

Second aspect, the embodiment of the invention also provides a kind of unusual checking device based on fusion feature, comprising:

Determination unit determines the movement for the detecting and tracking processing result according to moving target in video to be tested The behavior type of target；

Extraction unit, for extracting the various dimensions feature in the moving target；The various dimensions feature includes following institute It states multiple in the feature of moving target: the variation degree of pixel, the arrangement tightness degree of pixel, global shape, frame figure Similarity degree, motion feature, position feature and the morphological feature of picture；

Analysis and processing unit, the corresponding Fusion Features model of the behavior type for being determined according to the determination unit The various dimensions feature extracted to the extraction unit is analyzed and processed, and judges the movement according to analysis and processing result Target whether there is abnormal behaviour.

In conjunction with second aspect, the embodiment of the invention provides the first possible embodiments of second aspect, wherein institute Stating extraction unit includes:

First computation subunit, for calculating the variation degree of the pixel of the moving target in predetermined period DisturbRate=FC/FA；Wherein, FA indicates the quantity of all pixels point in the corresponding foreground area of the predetermined period；FC Indicate the quantity for the pixel being not in front of a certain period of time in the foreground area in foreground area；；

Second computation subunit, the close journey of arrangement for calculating the pixel of the moving target in the predetermined period Spend CompactRate=CS/CN；Wherein, CN indicates that all four neighborhoods are in the foreground area in the foreground area Foreground pixel point；In CS expression CN there is the foreground pixel point of luminance difference in all four neighborhoods；

Third computation subunit, for calculating circularity Ω=P/ of the global shape of the moving target in predetermined period (2*sqrt(πA))；Wherein, A indicates the area of the foreground area；P indicates the perimeter of the foreground area；

4th computation subunit, for calculating the similarity degree of frame image in the moving target in the predetermined period AreaSimilarRate=AND/OR；Wherein, AND indicates that consecutive frame in the foreground area is in the pixel of foreground area The number of point；The OR indicates the number for being not in the pixel of the foreground area in the foreground area in consecutive frame.

In conjunction with second aspect, the embodiment of the invention provides second of possible embodiments of second aspect, wherein institute State extraction unit further include:

First extracts subelement, for extracting the movement width of the moving target in predetermined period in the foreground area Degree；The motion amplitude include: length travel, lateral displacement, the height of current foreground area and current foreground area width；

And/or

5th computation subunit, for calculating the direction of motion OffRate=of the moving target in the predetermined period Offset/Route；Wherein, Offset indicates the actual displacement of the moving target, which includes lateral displacement and longitudinal position It moves；Route indicates the accumulative stroke of all actual displacements in the foreground area in the predetermined period.

In conjunction with second aspect, the embodiment of the invention provides the third possible embodiments of second aspect, wherein institute State extraction unit further include:

Second extracts subelement, becomes for extracting the displacement in predetermined period in the moving target between every two field pictures Change value；Operation subelement will when the change in displacement value for extracting subelement extraction described second is greater than preset threshold The variation frequency of the position feature of the moving target adds 1；Subelement is set, it, will be described for terminating in the predetermined period Position feature of the final variation frequency that operation subelement is calculated as the moving target.

In conjunction with second aspect, the embodiment of the invention provides the 4th kind of possible embodiments of second aspect, wherein institute State extraction unit further include:

6th computation subunit, for calculating the ratio of width to height of the video area where the video to be tested；

7th computation subunit, the ratio of width to height of the video area for being calculated according to the 6th computation subunit, Calculate the ratio of width to height of actual foreground area；

Contrast subunit, the ratio of width to height of the actual foreground area for calculating the 7th computation subunit with Default the ratio of width to height threshold value compares；

Judgment sub-unit, for being judged in the actual foreground area according to the comparing result of the contrast subunit The corresponding actual form of moving target；The actual form includes: single region and more people regions.

In conjunction with second aspect, the embodiment of the invention provides the 5th kind of possible embodiments of second aspect, wherein institute Stating analysis and processing unit includes:

Third extracts subelement, for extracting all feature vectors in the various dimensions feature, all feature vector tables It is shown as: F=[F1, F2, F3, F4, F5, F6, F7]；Wherein, F indicates the set of all feature vectors；Feature vector F1, F2, F3, F4, F5, F6 and F7 be respectively the variation degree, it is described arrangement tightness degree, the global shape, the frame image it is similar Degree, the motion feature, above-mentioned position feature and the morphological feature；

Subelement is obtained, extracts all features that subelement extracts for obtaining third described in the Fusion Features model The corresponding weight coefficient of vector sum, above-mentioned weight coefficient indicate are as follows: M=[w1, w2, w3, w4, w5, w6, w7]；Wherein, M is indicated The set of all weight coefficients；Weight coefficient w1, w2, w3, w4, w5, w6 and w7 are respectively that the variation degree, the arrangement are tight Close degree, the global shape, the similarity degree of the frame image, the motion feature, above-mentioned position feature and the form The corresponding weight coefficient of feature；

8th computation subunit, for extracting son to all feature vectors and the corresponding third according to formula The weight coefficient that unit obtains is calculated, and when calculated result meets first threshold, determines that the moving target is Abnormal behaviour；Wherein, label indicates calculated result；What T was represented is matrix transposition；Indicate each in the Fusion Features model The corresponding weight coefficient of a feature vector.

Appoint in conjunction with the first possible embodiment of second aspect, second aspect into the 5th kind of possible embodiment It anticipates a kind of possible embodiment, the embodiment of the invention provides the 6th kind of possible embodiments of second aspect, wherein institute The unusual checking device based on fusion feature stated, further includes:

Selection unit, for choosing the video resource for containing abnormal behaviour as positive sample and containing the view of normal behaviour Frequency resource is as negative sample；

Training computing unit, positive sample and the negative sample for being chosen by support vector machines to the selection unit It is trained calculating, establishes default Fusion Features model；

Repetitive exercise computing unit, the negative sample for being chosen according to the selection unit calculate the training single The default Fusion Features model that member is established carries out more wheel repetitive exercises, and reaches default in the default Fusion Features model When threshold value, the Fusion Features model by corresponding default Fusion Features model for practical application is determined.

In conjunction with second aspect, the embodiment of the invention provides the 7th kind of possible embodiments of second aspect, wherein institute The unusual checking device based on fusion feature stated further include:

Taxon classifies for the moving target to detecting and tracking, determines the feature of the moving target Type；

The determination unit is specifically used for, and when the characteristic type for detecting the moving target is the behavior of people, determines The behavior type of the moving target.

A kind of anomaly detection method and device based on fusion feature provided in an embodiment of the present invention, comprising: first According to the detecting and tracking processing result of moving target in video to be tested, the type of moving target is determined；Then above-mentioned fortune is extracted Various dimensions feature in moving-target；Wherein, above-mentioned various dimensions feature includes multiple in following characteristics: the variation journey of moving target Degree, arrangement tightness degree, global shape, the similarity degree of frame image, motion feature, position feature and morphological feature；Last root Various dimensions feature is analyzed and processed according to the above-mentioned type corresponding Fusion Features model, and judges to transport according to analysis and processing result Moving-target whether there is abnormal behaviour, often there is very big office in practical application using single features in the prior art It is sex-limited, it is easy to cause failing to report or report by mistake and comparing for abnormal behaviour, proposes multiple inventive features for abnormal behaviour, energy The robustness and stability of enough effective boosting algorithms, and the best features obtained according to the learning training to a large amount of abnormal behaviours Fusion Model is analyzed and processed above-mentioned various dimensions feature, can effectively avoid the over-fitting of algorithm in analytic process or owe quasi- Conjunction problem, it is suitable for Various Complex application scenarios, save a large amount of time cost and human cost, have very high promotion price Value.

To enable the above objects, features and advantages of the present invention to be clearer and more comprehensible, preferred embodiment is cited below particularly, and cooperate Appended attached drawing, is described in detail below.

Detailed description of the invention

In order to illustrate the technical solution of the embodiments of the present invention more clearly, below will be to needed in the embodiment attached Figure is briefly described, it should be understood that the following drawings illustrates only certain embodiments of the present invention, therefore is not construed as pair The restriction of range for those of ordinary skill in the art without creative efforts, can also be according to this A little attached drawings obtain other relevant attached drawings.

Fig. 1 shows a kind of process of the anomaly detection method based on fusion feature provided by the embodiment of the present invention Figure；

The stream of anomaly detection method Fig. 2 shows another kind provided by the embodiment of the present invention based on fusion feature Cheng Tu；

Fig. 3 shows the stream of anomaly detection method of the another kind provided by the embodiment of the present invention based on fusion feature Cheng Tu；

Fig. 4 shows a kind of structure of the unusual checking device based on fusion feature provided by the embodiment of the present invention Schematic diagram；

Fig. 5 shows in a kind of unusual checking device based on fusion feature provided by the embodiment of the present invention and extracts The structural schematic diagram of unit；

Fig. 6 shows in a kind of unusual checking device based on fusion feature provided by the embodiment of the present invention and analyzes The structural schematic diagram of processing unit；

Fig. 7 shows the knot of unusual checking device of the another kind provided by the embodiment of the present invention based on fusion feature Structure schematic diagram.

Main element symbol description:

11, determination unit；22, extraction unit；33, analysis and processing unit；44, selection unit；55, training computing unit； 66, repetitive exercise computing unit；221, the 6th computation subunit；222, the 7th computation subunit；223, contrast subunit；224, Judgment sub-unit；331, third extracts subelement；332, subelement is obtained；333, the 8th computation subunit.

Specific embodiment

Below in conjunction with attached drawing in the embodiment of the present invention, technical solution in the embodiment of the present invention carries out clear, complete Ground description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.Usually exist The component of the embodiment of the present invention described and illustrated in attached drawing can be arranged and be designed with a variety of different configurations herein.Cause This, is not intended to limit claimed invention to the detailed description of the embodiment of the present invention provided in the accompanying drawings below Range, but it is merely representative of selected embodiment of the invention.Based on the embodiment of the present invention, those skilled in the art are not doing Every other embodiment obtained under the premise of creative work out, shall fall within the protection scope of the present invention.

In order to give full play to monitoring system for the real-time of above-mentioned detection abnormal behaviour and the advantage of initiative, existing skill Art provides a kind of anomaly detection method based on monitoring system, but above-mentioned detection method is according to moving target Certain a kind of single feature (such as the displacement of movement velocity and movement) analysis and processing result judged, and use above-mentioned Single features are easy to cause failing to report or reporting by mistake for abnormal behaviour, and single often in practical application there are significant limitation The algorithm robustness that one feature normally results in analysis processing is poor, it is also relatively high to be applicable in scene requirement to algorithm, to also make It is higher to promote difficulty.

It also, is using average weighted method to the process that single features are analyzed and processed in above-mentioned detection method And linear classification is calculated, the former is that the average value of weighted value is set as to each single features (such as movement velocity With the displacement of movement) corresponding weighted value, the latter be priority level initializing is carried out to single features by artificial experience, so that it is determined that The corresponding weighted value of each single features；And it is above-mentioned using average weighted method and by artificial experience be arranged feature weight ginseng Several processes, it is not only higher to the experience and knowledge level requirement of staff, and easily lead to the over-fitting of algorithm or owe quasi- Conjunction problem.

The above problem of anomaly detection method based on the prior art based on monitoring system, the embodiment of the present invention provide A kind of anomaly detection method based on fusion feature, with it is in the prior art using single features often in practical application On there are significant limitation, be easy to cause failing to report or report by mistake and comparing for abnormal behaviour, propose multiple for abnormal behaviour Inventive features, be capable of effectively boosting algorithm robustness and stability, and according to a large amount of abnormal behaviours study instruct The best features Fusion Model got is analyzed and processed above-mentioned various dimensions feature, can effectively avoid calculating in analytic process The over-fitting or poor fitting problem of method, it is suitable for Various Complex application scenarios, save a large amount of time cost and human cost, Has very high promotional value.

For the ease of the understanding to the embodiment of the present invention, letter is carried out to monitoring system involved in the embodiment of the present invention first Illustrate, above-mentioned monitoring system includes: monitor camera device, and it is many to be pre-installed in bank, market, station and traffic intersection etc. Public place, for acquiring the video image of above-mentioned public place in real time；Monitoring center, including server and monitor terminal, if It sets in corresponding government department or relevant safety management department, passes through the above-mentioned monitoring camera dress of server real-time reception The video image of transmission is set, and the video image is analyzed and processed, to judge whether there is abnormal row in above-mentioned video image For, and warning device alarm is controlled when detecting abnormal behaviour；Its by the above-mentioned video image of monitor terminal real-time exhibition and The analysis of above-mentioned video image is as a result, with for user's real time inspection and corresponding operation processing is carried out.

With reference to a kind of flow chart of the anomaly detection method based on fusion feature shown in FIG. 1, the method includes Following steps:

S102, according to the detecting and tracking processing result of moving target in video to be tested, determine the row of the moving target For type.

In above-mentioned monitoring system, monitoring center determines moving target first from video to be tested, can be in order It treats moving target all in test video and successively carries out detection and tracking processing, it can also be first according to the movement of moving target State and motion amplitude etc. are extraordinarily thick slightly to detect the moving target of suspicion abnormal behaviour, and successively examines to these moving targets It surveys and tracking is handled.

Then the behavior type of moving target is determined according to the result that above-mentioned detection and tracking is handled, above-mentioned behavior type can With include: fight, run, the behavior in natural calamity (such as earthquake, fire) and theft；And in practice, the embodiment of the present invention In above-mentioned behavior type be not limited to above-mentioned concrete behavior type.

In addition, the moving object detection and tracking technique in the present embodiment be realize unusual checking technology basis and Premise, wherein moving object detection can be divided into the movement under static background according to the relationship between target and video camera Detection and the motion detection under dynamic background；

1, for the motion detection under static background: only target is moving during entire monitoring, mainly include with Lower several method.

(1) background subtraction；During entire monitoring, need ceaselessly to safeguard one " pure background ", for any one frame For monitored picture, it is subjected to difference with pure background, to obtain appearing in the moving target in current picture.(2) interframe Calculus of finite differences；It is calculated by difference between consecutive frame, come the method for the information such as the position, the shape that obtain moving target, this method It is very strong to the adaptability of illumination；The difference that specifically can use between adjacent three frame calculates, the detection of Lai Jinhang moving target.

(3) optical flow method；In space, movement can be described with sports ground；And on a plane of delineation, the movement of object It is embodied often by the difference that image grayscale in image sequence is distributed, so that the sports ground in space be made to be transferred on image It is indicated as optical flow field.Optical flow field reflects the variation tendency of every bit gray scale on image, can regard the picture with gray scale as Vegetarian refreshments moves on the image plane and the instantaneous velocity field that generates, and the approximate evaluation of a kind of pair of real motion field；Relatively managing In the case where thinking, it is able to detect any information of the object of self-movement without scene is known in advance, it can be very accurate Ground calculates the speed of moving object, and the case where can be used for dynamic scene.

2, for the motion detection under dynamic background: during monitoring, all movement or variation are occurring for target and background；? In the application environment of moving object detection, dynamic background is in comparison more complicated.According to the forms of motion of camera, can be divided into Following two:

(1) camera support is fixed；But camera can be rotated with the movement of holder, the movement such as inclination.In addition, phase Machine can also control lens focusing according to remote computer instruction, to generate distant view and close shot scaling movement.

(2) camera is placed on mobile device (for example, in-vehicle camera)

For any one of both the above camera motion form, before carrying out moving object detection, require Overall motion estimation and compensation are carried out according to certain method.In general, can use the progress such as block matching method, Feature Points Matching method The estimation of amount of exercise.It is of course also possible to establish light stream field model using optical flow method, image slices vegetarian refreshments is solved using optical flow equation Movement velocity.

And motion target tracking is exactly to find in each frame monitored picture interested in a continuous videos sequence Moving target (for example, vehicle, pedestrian, object (such as animal)).Tracking can be roughly divided into following steps:

1, effective description of target；The tracking process of target needs effectively to describe it as target detection, That is, it needs to clarification of objective be extracted, so as to express the target；In general, the edge, profile, shape of image can be passed through The modes such as shape, texture, region, histogram, moment characteristics, transformation coefficient carry out clarification of objective description；

2, similarity measurement calculates；Common method has: Euclidean distance, mahalanobis distance, chessboard distance, Weighted distance, phase Like coefficient, related coefficient etc.；

3, target area search matching；If all carrying out feature extraction, similitude meter to all targets occurred in scene It calculates, then the spent calculation amount of system operation is very big.Therefore moving target may be gone out by generalling use certain mode at present Existing region is estimated, to reduce redundancy, accelerates the speed of target following；Common prediction algorithm has: Kalman filter, Particle filter, average drifting etc..

Various dimensions feature in S104, the extraction moving target；The various dimensions feature includes movement mesh as described below It is multiple in target feature: the variation degree of pixel, the arrangement tightness degree of pixel, global shape, frame image it is similar Degree, motion feature, position feature and morphological feature.

Specifically, the above-mentioned various dimensions feature extracted in the present embodiment can be multiple and above-mentioned in features described above It can be in any combination between feature；Wherein, above-mentioned motion feature includes motion amplitude and the movement of the direction of motion of moving target Direction, above-mentioned position feature include the variation frequency of displacement；Above-mentioned two feature and other 5 features are in single features It is targetedly designed and is transformed on the basis of the feature and abnormal behaviour feature of (displacement of such as movement velocity and movement) Completely new characteristic model, the embodiment of the present invention effectively prevents single features by comprehensively utilizing a variety of above-mentioned behavioural characteristics Disadvantage improves the robustness and stability of algorithm.

S106, the various dimensions feature is analyzed and processed according to the behavior type corresponding Fusion Features model, And judge the moving target with the presence or absence of abnormal behaviour according to analysis and processing result.

Specifically, needing when carrying out anomaly analysis to moving target according to the upper various dimensions feature of extraction to movement mesh Each above-mentioned feature of target is individually analyzed, in the case where judging that all various dimensions features are abnormal behaviour, really Determining moving target is abnormal behaviour.

A kind of anomaly detection method based on fusion feature provided in an embodiment of the present invention, is adopted in the prior art With single features often in practical application there are significant limitation, it is easy to cause failing to report or report by mistake and comparing for abnormal behaviour, It proposes multiple inventive features for abnormal behaviour, is capable of the robustness and stability of effective boosting algorithm, and root The best features Fusion Model obtained according to the learning training to a large amount of abnormal behaviours is analyzed and processed above-mentioned various dimensions feature, The over-fitting of algorithm or poor fitting problem in analytic process can be effectively avoided, it is suitable for Various Complex application scenarios, save A large amount of time cost and human cost, have very high promotional value.

In view of above-mentioned various dimensions are characterized on the basis of the feature of single features (displacement of such as movement velocity and movement) The upper completely new characteristic model for targetedly being designed and being transformed, thus in the embodiment of the present invention respectively to above-mentioned 7 multidimensional The extraction process of degree feature is described in detail:

The first, for the variation degree of the pixel of moving target, above-mentioned variation degree can also be expressed as restless degree, tool Body extraction process is as follows:

The change of the pixel of the moving target in predetermined period is calculated using following formula DisturbRate=FC/FA Change degree；Wherein, the FA in above-mentioned formula indicates the quantity of all pixels point in the corresponding foreground area of the predetermined period；FC Indicate the quantity for the pixel being not in front of a certain period of time in the foreground area in foreground area.

Specifically, foreground area refers to the zone of action of moving target, the region will not generally occur under normal behaviour Variation, and the region can change under abnormal behaviour, therefore corresponding foreground area can not in the period in different times Together；In view of the above problem, in the embodiment of the present invention, with the same predetermined period, (the predetermined period number can be according to reality first Border needs any setting, using predetermined period as a cycle in the present invention, hereinafter collectively referred to as period 1, and 5 frame images are one Period is illustrated) for be illustrated, FA is then the quantity of all pixels point in the foreground area in the period 1, and FC indicates the picture being not in the foreground area before a certain period of time (such as a cycle) in all pixels point in FA The quantity of vegetarian refreshments, the ratio of calculating FC and FA are the variation degree or restless degree of the pixel of moving target.

The second, for the arrangement tightness degree of the pixel of moving target, above-mentioned arrangement tightness degree can also be expressed as Compactness, specific extraction process are as follows:

Using following formula CompactRate=CS/CN, the pixel of the moving target in the period 1 is calculated Arrangement tightness degree CompactRate=CS/CN；Wherein, CN indicates that all four neighborhoods are in institute in the foreground area State the quantity of the foreground pixel point of foreground area；In CS expression CN there is the number of the foreground pixel point of luminance difference in all four neighborhoods Amount.

Specifically, containing two characteristic points in above-mentioned arrangement tightness degree, first is the quantity of pixel, and second is picture There is luminance difference between vegetarian refreshments；It is to distinguish illumination and tree in video image by the purpose that above-mentioned two characteristic point carries out operation The erroneous judgement of the behaviors such as leaf shaking, wherein when the closeness of the behavior of people is larger, it is big that corresponding pixel appears like one Piece, and the closeness of the behavior of the closeness and people of the pixel of illumination is more similar, it appears that and it is a sheet of, but illumination Pixel between brightness be it is the same, this is different from the behavior of people；And the behavior of closeness and people that corresponding leaf shakes Closeness it is also more similar, it appears that and it is a sheet of, but leaf shake in pixel between difference it is smaller and be not All the points can be avoided the erroneous judgement feelings of above-mentioned illumination and leaf shaking by above-mentioned two characteristic point all in foreground area Condition.

Wherein, four neighborhoods in above-mentioned foreground area refer in foreground area four around centered on some pixel The circle of a point composition or the section of ball.

Third, the global shape for moving target, above-mentioned global shape can be expressed as circularity, specifically extract Journey is as follows:

The circle of the global shape of the moving target in predetermined period is calculated using following formula Ω=P/2.sqrt (π A) Shape degree；Wherein, A indicates the area of the foreground area；P indicates the perimeter of the foreground area.

Specifically, the global shape of the corresponding video image of the behavior of people is round or ellipse, and the video figure of vehicle The global shape of picture is rectangle, therefore the purpose of circularity by calculating is to further determine that the behavior extracted is people Behavior, so that the abnormal behaviour to people is analyzed.

4th, for the similarity degree of frame image in moving target, above-mentioned similarity degree can also be expressed as similitude, Specific extraction process is as follows:

According to formula AreaSimilarRate=AND/OR, frame figure in the moving target is calculated in the period 1 The similarity degree of picture；Wherein, AND indicates that consecutive frame in the foreground area is in the number of the pixel of foreground area；Institute Stating OR indicates the number for being not in the pixel of the foreground area in the foreground area in consecutive frame.

Specifically, such as fighting in the abnormal behaviour of people, the similarity of the multiple image in one period is very high , this feature can judge the abnormal behaviour of people in conjunction with other features, further such that the abnormal behaviour of analysis people Algorithm accuracy and robustness are preferable.

Wherein, above-mentioned foreground area remains as period 1 corresponding foreground area, and AND then indicates phase in the foreground area Adjacent frame is in the number of the pixel of foreground area；And OR then indicates to be not in the prospect in the foreground area in consecutive frame The number of the pixel in region.

5th, for the motion feature of moving target, above-mentioned motion feature includes motion amplitude and the direction of motion, and above-mentioned The direction of motion can be expressed as one-way；Specifically, the specific extraction process of above-mentioned motion feature is as follows:

And/or according to formula OffRate=Offset/Route, the moving target in the predetermined period is calculated The direction of motion；Wherein, Offset indicates the actual displacement of the moving target, which includes lateral displacement and length travel； Route indicates the accumulative stroke of all actual displacements in the foreground area in the predetermined period.

Specifically, above-mentioned motion feature may include two features of motion amplitude and the direction of motion, foreground zone is extracted first The length travel of moving target, lateral displacement, width of the height of current foreground area and current foreground area etc. move in domain Amplitude characteristic；Furthermore it is also possible to according to above-mentioned motion amplitude feature calculation direction of motion feature (i.e. one-way)；

Wherein, the period 1, molecule Offset was to subtract first frame from the current displacement point of the 5th frame by taking 5 frame images as an example Current displacement point, the displacement actually occurred for moving target；Denominator Route is the absolute value of the intrinsic displacement of every frame within 5 frames It is cumulative；Then when obtained OffRate is 1, then it is determined as one-way, therefore running facing one direction is corresponding unidirectional OffRate is then close to 1 for property.

Wherein, lateral one-way is laterally calculated with formula ColOffRate=ColOffset/ColRoute quantization: its In, ColOffset is the lateral actual displacement of moving target, and ColRoute is the lateral frame that foreground area is in the period 1 Between add up stroke；And transverse direction one-way ColOffRate is close to 1 expression transverse direction one-way movement.

Similarly, longitudinal with formula RowOffRate=RowOffset/RowRoute quantization, calculate longitudinal one-way: its In, RowOffset is longitudinal actual displacement of moving target, and RowRoute is longitudinal frame that foreground area is in the period 1 Between add up stroke；And longitudinal direction one-way RowOffRate is close to 1 expression longitudinal direction one-way movement.

6th, for the position feature of moving target, the specific extraction process of above-mentioned motion feature is as follows:

Specifically, as long as the displacement of moving target changes and all calculates shift in position in practice, but in the present embodiment, in advance It is first provided with movement threshold, when shift in position is more than the movement threshold, is just determined as shift in position, movement threshold is set Purpose in order to distinguish the object as time display screen etc., position remains that text constant but thereon is to become always Change, if the time is ceaselessly dynamic.

Wherein, the specific calculation of the above-mentioned variation frequency is as follows:, will be in the period 1 still by taking the period 1 as an example Subtraction is done in the position between every two adjacent moving target in foreground area, if difference is greater than movement threshold 5, then sentences Disconnected shift in position changes the frequency and adds 1, until the final variation frequency is as the moving target after period 1 Position feature.

It should be noted that above-mentioned preset threshold can according to need any setting, the present invention does not do its specific value Concrete restriction.

7th, for the morphological feature of moving target, the specific extraction process following steps of above-mentioned morphological feature, with reference to figure 2:

The ratio of width to height of video area where S202, the calculating video to be tested.

S204, according to the ratio of width to height of the video area, calculate the ratio of width to height of actual foreground area.

S206, the ratio of width to height of the actual foreground area and default the ratio of width to height threshold value are compared, and according to comparison As a result judge the corresponding actual form of moving target in the actual foreground area；The actual form includes: single shape State and how humanoid state.

Specifically, single form and how humanoid state form the width in foreground target (moving target i.e. in foreground area) region For height than different, this is related with the ratio of width to height of video itself；It is such as compressed in video height itself, one width will be made high Than the ratio of width to height close to more people；Such as normal video, it is 1:2 after being compressed that single the ratio of width to height, which is 1:4,；And normally more people are wide Height than also be 1:2, so when also need to judge single form and how humanoid state according to the ratio of width to height of video itself.

Specific calculating process is i.e.: calculating the ratio of width to height of video area first, is then calculating actual foreground area The ratio of width to height compares the ratio of width to height in actual foreground region and default form threshold value, is such as to determine being less than default form threshold value It is to be determined as how humanoid state being greater than default form threshold value for single form.

It should be noted that above-mentioned default form threshold value can according to need any setting, the present invention is to its specific value It is not particularly limited.

Restless degree, compactness, circularity, similitude, motion feature, position are indicated with F1~F7 respectively in the embodiment of the present invention Set the corresponding characteristic value of above-mentioned 7 multidimensional characteristics such as feature and shape feature.

After being extracted above-mentioned 7 multidimensional characteristics, the corresponding Fusion Features model of housing choice behavior type, and according to selection Corresponding features described above Fusion Model is analyzed and processed above-mentioned various dimensions feature, and described in being judged according to analysis and processing result Moving target whether there is abnormal behaviour, wherein above-mentioned analysis handles and judges that the specific implementation of abnormal behaviour is as follows:

All feature vectors in the various dimensions feature are extracted first, and above-mentioned all feature vectors can indicate are as follows: F= [F1, F2, F3, F4, F5, F6, F7]；Wherein, F indicates the set of all feature vectors；Feature vector F1, F2, F3, F4, F5, F6 It is respectively the variation degree, the arrangement tightness degree, the global shape, the similarity degree of the frame image, institute with F7 State motion feature, above-mentioned position feature and the morphological feature；

According to formulaTo all feature vectors and the corresponding weight coefficient It is calculated, and when calculated result meets first threshold, determines the moving target for abnormal behaviour；Wherein, label is indicated Calculated result；What T was represented is matrix transposition；Wi indicates the corresponding weight system of each feature vector in the Fusion Features model Number.

Specifically, the corresponding Fusion Features model of behavior type of selection can indicate are as follows: M=[w1, w2, w3, w4, w5, W6, w7], above-mentioned 7 feature vector F1~F7 and corresponding above-mentioned 7 weight coefficient w1~w7 are updated to formulaIn, it is corresponding as a result, and when label=-1 (i.e. abnormal behaviour threshold that label can be calculated Value) when, then judge that moving target is normal condition；When label=1 (i.e. normal behaviour threshold value), then detect in moving target There is abnormal behaviour, and alarm, to remind staff to safeguard.

In view of using average weighted method in the prior art and passing through the mistake of artificial experience setting feature weight parameter Journey easily leads to the over-fitting or poor fitting problem of algorithm, and the embodiment of the present invention uses support vector machines, on a large amount of training samples Training is carried out, theoretical best features Fusion Model is obtained, this method is weighed it is possible to prevente effectively from feature is arranged by experience Algorithm over-fitting or poor fitting problem caused by weight parameter.With reference to Fig. 3, the Fusion Features model in the embodiment of the present invention is preparatory It is generated, is specifically comprised the following steps: according to following methods

S302, selection contain the video resource of abnormal behaviour as positive sample and the video resource containing normal behaviour is made For negative sample.

Specifically, choosing the video clip containing abnormal behaviour, the upper of the moving target during abnormal behaviour occurs is extracted State 7 class multidimensional characteristics, and the positive sample as training sample set；It chooses and contains other normal video clips, it is random to extract arbitrarily Above-mentioned 7 category feature of the moving target at moment, and the negative sample as training sample set.

S304, calculating is trained to the positive sample and the negative sample by support vector machines, establishes default feature Fusion Model.

Specifically, using support vector machines to be trained for classifier to above-mentioned positive sample and negative sample, and by taking turns more Repetitive exercise obtains Fusion Features model；Wherein, the Fusion Features model trained according to different behavior types is different, such as beats Frame corresponds to a kind of Fusion Features model, and corresponding another Fusion Features model of running；And the feature of different behavior types Fusion Model comprising feature vector corresponding to weighted value (or claim weight coefficient) it is different.

S306, more wheel repetitive exercises are carried out to the default Fusion Features model according to the negative sample, and described pre- If Fusion Features model reaches preset threshold, the Fusion Features by corresponding default Fusion Features model for practical application are determined Model.

In view of the integrity problem of features described above Fusion Model that training obtains, according to negative sample on established It states Fusion Features model and is carrying out more wheel repetitive exercises, negative sample is such as constantly substituted into the characteristic model that above-mentioned training obtains In, if judging result matches with practical abnormal behaviour in the first preset times (such as 100 times), then it is assumed that features described above mould Type is reliable, and as the Fusion Features model of practical application；If judging result be in the preset times (such as 20 times) with Time anomaly behavior mismatches (i.e. output is normal behaviour), then it is assumed that features described above model is unreliable, at this point, whenever output Judging result and actual result mismatch when, support vector machines is to the weight for stating each feature vector in features described above model Value repeatedly adjustment, until obtain matching with practical abnormal behaviour in the first preset times (such as 100 times) as a result, then As the Fusion Features model of practical application.

In addition, the abnormal behaviour of anomaly detection method in the embodiment of the present invention primarily directed to people, in order to more Moving target is determined as people, in step 102 " before the behavior type for determining the moving target ", further includes: to detection with The moving target of track is classified, and determines the characteristic type of the moving target, then, is detecting the moving target Characteristic type be people behavior when, in the step of being determined the behavior type of the moving target.

It, first can be according to the movement velocity of moving target, moving target size and movement purpose in the embodiment of the present invention The characteristic points such as displacement, classify to the moving target of detecting and tracking, judge that it belongs in people's behavior or vehicle on earth In behavior or object (such as illumination and leaf) behavior.And when detecting above-mentioned moving target is the behavior of people, carrying out really The step of behavior type of the fixed moving target.

With reference to Fig. 4, the embodiment of the present invention also provides a kind of unusual checking device based on fusion feature, comprising:

Determination unit 11 determines the fortune for the detecting and tracking processing result according to moving target in video to be tested The behavior type of moving-target；

Extraction unit 22, for extracting the various dimensions feature in the moving target；The various dimensions feature includes following It is multiple in the feature of the moving target: variation degree, the arrangement tightness degree of pixel, the global shape, frame of pixel Similarity degree, motion feature, position feature and the morphological feature of image；

Analysis and processing unit 33, the corresponding Fusion Features mould of the behavior type for being determined according to the determination unit Type is analyzed and processed the various dimensions feature that the extraction unit extracts, and judges the fortune according to analysis and processing result Moving-target whether there is abnormal behaviour.

A kind of unusual checking device based on fusion feature provided in an embodiment of the present invention, is adopted in the prior art With single features often in practical application there are significant limitation, it is easy to cause failing to report or report by mistake and comparing for abnormal behaviour, It proposes multiple inventive features for abnormal behaviour, is capable of the robustness and stability of effective boosting algorithm, and root The best features Fusion Model obtained according to the learning training to a large amount of abnormal behaviours is analyzed and processed above-mentioned various dimensions feature, The over-fitting of algorithm or poor fitting problem in analytic process can be effectively avoided, it is suitable for Various Complex application scenarios, save A large amount of time cost and human cost, have very high promotional value.

In view of above-mentioned various dimensions are characterized on the basis of the feature of single features (displacement of such as movement velocity and movement) The upper completely new characteristic model for targetedly being designed and being transformed, therefore also need respectively in the embodiment of the present invention to above-mentioned 7 Various dimensions feature extracts, and corresponding extraction unit 22 includes:

The first, for the variation degree of the pixel of moving target, above-mentioned variation degree can also indicate restless degree；

The second, for the arrangement tightness degree of the pixel of moving target, above-mentioned arrangement tightness degree can also indicate tight Cause property；

Third, the global shape for moving target, above-mentioned global shape can indicate to calculate the circularity of moving target；

4th, for the similarity degree of frame image in moving target, above-mentioned similarity degree can also be expressed as similitude；

Specifically, above-mentioned restless degree, compactness, the extraction process of circularity and similitude are as follows: the extraction unit 22 Include:

First computation subunit, for calculating the variation degree of the pixel of the moving target in predetermined period DisturbRate=FC/FA；Wherein, FA indicates the quantity of all pixels point in the corresponding foreground area of the predetermined period；FC Indicate the quantity for the pixel being not in front of a certain period of time in the foreground area in foreground area；

5th, for the motion feature of moving target, above-mentioned motion feature includes: motion amplitude and the direction of motion, and on One-way can be expressed as by stating the direction of motion；Specifically, the specific extraction process of above-mentioned motion feature is as follows: the extraction unit 22 further include:

And/or

6th, for the position feature of moving target, the specific extraction process of above-mentioned motion feature is as follows: the extraction is single Member 22 further include:

7th, for the morphological feature of moving target, the specific extraction process of above-mentioned morphological feature is as follows: referring to Fig. 5, institute State extraction unit 22 further include:

6th computation subunit 221, for calculating the ratio of width to height of the video area where the video to be tested；

7th computation subunit 222, the width of the video area for being calculated according to the 6th computation subunit 221 High ratio, calculates the ratio of width to height of actual foreground area；

Contrast subunit 223, the width of the actual foreground area for calculating the 7th computation subunit 222 Height with default the ratio of width to height threshold value than comparing；

Judgment sub-unit 224, for judging the actual foreground zone according to the comparing result of the contrast subunit 223 The corresponding actual form of moving target in domain；The actual form includes: single region and more people regions.

Wherein, above-mentioned restless degree, compactness, circularity, similitude, motion feature, position feature and the shape of said extracted The corresponding characteristic value of above-mentioned 7 multidimensional characteristics such as shape feature is indicated with F1~F7 respectively.

After being extracted above-mentioned 7 multidimensional characteristics, the corresponding Fusion Features model of housing choice behavior type, and according to selection Corresponding features described above Fusion Model is analyzed and processed above-mentioned various dimensions feature, and described in being judged according to analysis and processing result Moving target whether there is abnormal behaviour, wherein the specific implementation of corresponding analysis and processing unit is as follows:

Further, with reference to Fig. 6, in the above-mentioned unusual checking device based on fusion feature, the analysis processing is single First 33 include:

Third extracts subelement 331, for extracting all feature vectors in the various dimensions feature, all feature vectors It indicates are as follows: F=[F1, F2, F3, F4, F5, F6, F7]；Wherein, F indicates the set of all feature vectors；Feature vector F1, F2, F3, F4, F5, F6 and F7 are respectively the variation degree, the arrangement tightness degree, the global shape, the frame image Similarity degree, the motion feature, above-mentioned position feature and the morphological feature；

Subelement 332 is obtained, extracts all of subelement extraction for obtaining third described in the Fusion Features model Feature vector and corresponding weight coefficient, above-mentioned weight coefficient indicate are as follows: M=[w1, w2, w3, w4, w5, w6, w7]；Wherein, M Indicate the set of all weight coefficients；Weight coefficient w1, w2, w3, w4, w5, w6 and w7 are respectively the variation degree, the row Column tightness degree, the global shape, the similarity degree of the frame image, the motion feature, above-mentioned position feature and described The corresponding weight coefficient of morphological feature；

8th computation subunit 333, for according to formulaTo all features to Amount and the corresponding third are extracted the weight coefficient that subelement obtains and are calculated, and meet abnormal row in calculated result When for threshold value, the moving target is determined for abnormal behaviour and is alarmed；Wherein, label indicates calculated result；What T was represented is square Battle array transposition；w_iIndicate the corresponding weight coefficient of each feature vector in the Fusion Features model.

In view of using average weighted method in the prior art and passing through the mistake of artificial experience setting feature weight parameter Journey easily leads to the over-fitting or poor fitting problem of algorithm, and the embodiment of the present invention pre-establishes Fusion Features mould by following device Type.Specifically, with reference to Fig. 7, the above-mentioned unusual checking device based on fusion feature further include:

Selection unit 44, for choosing the video resource for containing abnormal behaviour as positive sample and containing normal behaviour Video resource is as negative sample；

Training computing unit 55, positive sample and the negative sample for being chosen by support vector machines to the selection unit Originally it is trained calculating, establishes default Fusion Features model；

Repetitive exercise computing unit 66, the negative sample for being chosen according to the selection unit calculate the training The default Fusion Features model that unit is established carries out more wheel repetitive exercises, and reaches pre- in the default Fusion Features model If when threshold value, determining the Fusion Features model by corresponding default Fusion Features model for practical application.

In addition, the abnormal behaviour of anomaly detection method in the embodiment of the present invention primarily directed to people, in order to more Moving target is determined as people, in the above-mentioned unusual checking device based on fusion feature, further includes: taxon is used for Classify to the moving target of detecting and tracking, determines the characteristic type of the moving target；

The determination unit 11 is specifically used for, when the characteristic type for detecting the moving target is the behavior of people, really The behavior type of the fixed moving target.

The computer program that the anomaly detection method based on fusion feature is carried out provided by the embodiment of the present invention produces Product, the computer readable storage medium including storing program code, before the instruction that said program code includes can be used for execution Method described in the embodiment of the method for face, specific implementation can be found in embodiment of the method, and details are not described herein.

It is apparent to those skilled in the art that for convenience and simplicity of description, the system of foregoing description, The specific work process of device and unit, can refer to corresponding processes in the foregoing method embodiment, and details are not described herein.

In several embodiments provided herein, it should be understood that disclosed systems, devices and methods, it can be with It realizes by another way.The apparatus embodiments described above are merely exemplary, for example, the division of the unit, Only a kind of logical function partition, there may be another division manner in actual implementation, in another example, multiple units or components can To combine or be desirably integrated into another system, or some features can be ignored or not executed.Another point, it is shown or beg for The mutual coupling, direct-coupling or communication connection of opinion can be through some communication interfaces, device or unit it is indirect Coupling or communication connection can be electrical property, mechanical or other forms.

The unit as illustrated by the separation member may or may not be physically separated, aobvious as unit The component shown may or may not be physical unit, it can and it is in one place, or may be distributed over multiple In network unit.It can select some or all of unit therein according to the actual needs to realize the mesh of this embodiment scheme 's.

It, can also be in addition, the functional units in various embodiments of the present invention may be integrated into one processing unit It is that each unit physically exists alone, can also be integrated in one unit with two or more units.

It, can be with if the function is realized in the form of SFU software functional unit and when sold or used as an independent product It is stored in a computer readable storage medium.Based on this understanding, technical solution of the present invention is substantially in other words The part of the part that contributes to existing technology or the technical solution can be embodied in the form of software products, the meter Calculation machine software product is stored in a storage medium, including some instructions are used so that a computer equipment (can be a People's computer, server or network equipment etc.) it performs all or part of the steps of the method described in the various embodiments of the present invention. And storage medium above-mentioned includes: that USB flash disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), arbitrary access are deposited The various media that can store program code such as reservoir (RAM, Random Access Memory), magnetic or disk.

The above description is merely a specific embodiment, but scope of protection of the present invention is not limited thereto, any Those familiar with the art in the technical scope disclosed by the present invention, can easily think of the change or the replacement, and should all contain Lid is within protection scope of the present invention.Therefore, protection scope of the present invention should be based on the protection scope of the described claims.

Claims

1. a kind of anomaly detection method based on fusion feature characterized by comprising

According to the detecting and tracking processing result of moving target in video to be tested, the behavior type of the moving target is determined；

Extract the various dimensions feature in the moving target；The various dimensions feature includes in the feature of moving target as described below It is multiple: the variation degree of pixel, the arrangement tightness degree of pixel, global shape, the similarity degree of frame image, movement are special Sign, position feature and morphological feature；

The various dimensions feature is analyzed and processed according to the behavior type corresponding Fusion Features model, and according to analysis Processing result judges the moving target with the presence or absence of abnormal behaviour；

The various dimensions feature extracted in the moving target includes:

Calculate the variation degree DisturbRate=FC/FA of the pixel of the moving target in predetermined period；Wherein, FA table Show the quantity of all pixels point in the corresponding foreground area of the predetermined period；FC is indicated in foreground area before a certain period of time It is not in the quantity of the pixel in the foreground area；

Calculate the arrangement tightness degree CompactRate=CS/CN of the pixel of the moving target in the predetermined period；Its In, CN indicates that all four neighborhoods in the foreground area are in the quantity of the foreground pixel point of the foreground area；CS is indicated In CN there is the quantity of the foreground pixel point of luminance difference in all four neighborhoods；

Calculate circularity Ω=P/ (2*sqrt (π A)) of the global shape of the moving target in predetermined period；Wherein, A is indicated The area of the foreground area；P indicates the perimeter of the foreground area；

Calculate the similarity degree AreaSimilarRate=AND/OR of frame image in the moving target in the predetermined period； Wherein, AND indicates that consecutive frame in the foreground area is in the number of the pixel of foreground area；OR indicates the foreground zone The number of the pixel of the foreground area is not in domain in consecutive frame；

The change in displacement value in predetermined period in the moving target between every two field pictures is extracted, and in the change in displacement value When greater than preset threshold, the variation frequency of the position feature of the moving target is added 1, until the predetermined period terminates, it will Position feature of the final variation frequency as the moving target；

Calculate the ratio of width to height of the video area where the video to be tested；According to the ratio of width to height of the video area, calculate real The ratio of width to height of the foreground area on border；The ratio of width to height of the actual foreground area and default the ratio of width to height threshold value are compared, and The corresponding actual form of moving target in the actual foreground area is judged according to comparing result；The actual form packet It includes: single form and how humanoid state；

Extract the motion feature in the moving target, comprising:

Extract the motion amplitude of the moving target in predetermined period in the foreground area；The motion amplitude includes: vertical To displacement, lateral displacement, the height of current foreground area and current foreground area width；

And/or

Calculate the direction of motion OffRate=Offset/Route of the moving target in the predetermined period；Wherein, Offset Indicate the actual displacement of the moving target, including lateral displacement and length travel；Before Route is indicated in the predetermined period The accumulative stroke of all actual displacements in scene area.

2. the anomaly detection method according to claim 1 based on fusion feature, which is characterized in that according to the row The various dimensions feature is analyzed and processed for type corresponding Fusion Features model, and institute is judged according to analysis and processing result Moving target, which is stated, with the presence or absence of abnormal behaviour includes:

All feature vectors in the various dimensions feature are extracted, all feature vectors indicate are as follows: F=[F1, F2, F3, F4, F5, F6, F7]；Wherein, F indicates the set of all feature vectors；Feature vector F1, F2, F3, F4, F5, F6 and F7 are respectively the picture The variation degree of vegetarian refreshments, the arrangement tightness degree of the pixel, the global shape, the similarity degree of the frame image, institute State motion feature, above-mentioned position feature and the morphological feature；

The corresponding weight coefficient of all feature vectors in the Fusion Features model is obtained, above-mentioned weight coefficient indicates are as follows: M= [w1, w2, w3, w4, w5, w6, w7]；Wherein, M indicates the set of all weight coefficients；Weight coefficient w1, w2, w3, w4, w5, w6 It is respectively the variation degree of the pixel, the arrangement tightness degree of the pixel, the global shape, the frame figure with w7 The similarity degree of picture, the motion feature, above-mentioned position feature and the corresponding weight coefficient of the morphological feature；

According to formulaAll feature vectors and the corresponding weight coefficient are carried out It calculates, and when calculated result meets abnormal behaviour threshold value, determines the moving target for abnormal behaviour and alarm；Wherein, Label indicates calculated result；What T was represented is matrix transposition；F_iIndicate any feature vector in all feature vector F；I=1, 2,3,4,5,6,7；w_iIndicate the corresponding weight coefficient of each feature vector in the Fusion Features model.

3. the anomaly detection method according to claim 1 or 2 based on fusion feature, which is characterized in that the spy Levying Fusion Model is generated previously according to following methods:

Calculating is trained to the positive sample and the negative sample by support vector machines, establishes default Fusion Features model；

More wheel repetitive exercises are carried out to the default Fusion Features model according to the negative sample, and in the default Fusion Features When model reaches preset threshold, the Fusion Features model by corresponding default Fusion Features model for practical application is determined.

4. the anomaly detection method according to claim 1 based on fusion feature, which is characterized in that the determining institute Before the behavior type for stating moving target, further includes:

Classify to the moving target of detecting and tracking, determines the characteristic type of the moving target, and detecting When the characteristic type for stating moving target is the behavior of people, the step of executing the behavior type for determining the moving target.

5. a kind of unusual checking device based on fusion feature characterized by comprising

Determination unit determines the moving target for the detecting and tracking processing result according to moving target in video to be tested Behavior type；

Extraction unit, for extracting the various dimensions feature in the moving target；The various dimensions feature includes fortune as described below It is multiple in the feature of moving-target: the variation degree of pixel, the arrangement tightness degree of pixel, global shape, frame image Similarity degree, motion feature, position feature and morphological feature；

Analysis and processing unit, the corresponding Fusion Features model of the behavior type for being determined according to the determination unit is to institute The various dimensions feature for stating extraction unit extraction is analyzed and processed, and judges the moving target according to analysis and processing result With the presence or absence of abnormal behaviour；

The extraction unit includes:

Second computation subunit, for calculating the arrangement tightness degree of the pixel of the moving target in the predetermined period CompactRate=CS/CN；Wherein, before CN indicates that all four neighborhoods are in the foreground area in the foreground area The quantity of scene vegetarian refreshments；In CS expression CN there is the quantity of the foreground pixel point of luminance difference in all four neighborhoods；

Third computation subunit, for calculating circularity Ω=P/ (2* of the global shape of the moving target in predetermined period sqrt(πA))；Wherein, A indicates the area of the foreground area；P indicates the perimeter of the foreground area；

4th computation subunit, for calculating the similarity degree of frame image in the moving target in the predetermined period AreaSimilarRate=AND/OR；Wherein, AND indicates that consecutive frame in the foreground area is in the pixel of foreground area The number of point；OR indicates the number for being not in the pixel of the foreground area in the foreground area in consecutive frame；

Second extracts subelement, for extracting the change in displacement in predetermined period in the moving target between every two field pictures Value；Operation subelement, when the change in displacement value for extracting subelement extraction described second is greater than preset threshold, by institute The variation frequency for stating the position feature of moving target adds 1；Subelement is set, for terminating in the predetermined period, by the fortune Position feature of the final variation frequency that operator unit is calculated as the moving target；

6th computation subunit, for calculating the ratio of width to height of the video area where the video to be tested；7th calculates son list Member, the ratio of width to height of the video area for being calculated according to the 6th computation subunit, calculates actual foreground area The ratio of width to height；Contrast subunit, the ratio of width to height of the actual foreground area for calculating the 7th computation subunit with Default the ratio of width to height threshold value compares；Judgment sub-unit, for judging the reality according to the comparing result of the contrast subunit The corresponding actual form of moving target in the foreground area on border；The actual form includes: single region and more people regions；

The extraction unit, further includes:

First extracts subelement, for extracting the motion amplitude of the moving target in predetermined period in the foreground area； The motion amplitude include: length travel, lateral displacement, the height of current foreground area and current foreground area width；

With,

5th computation subunit, for calculating the direction of motion OffRate=of the moving target in the predetermined period Offset/Route；Wherein, Offset indicates the actual displacement of the moving target, including lateral displacement and length travel； Route indicates the accumulative stroke of all actual displacements in the foreground area in the predetermined period.

6. the unusual checking device according to claim 5 based on fusion feature, which is characterized in that at the analysis Managing unit includes:

Third extracts subelement, and for extracting all feature vectors in the various dimensions feature, all feature vectors are indicated are as follows: F=[F1, F2, F3, F4, F5, F6, F7]；Wherein, F indicates the set of all feature vectors；Feature vector F1, F2, F3, F4, F5, F6 and F7 are respectively the variation degree of the pixel, the arrangement tightness degree of the pixel, the global shape, institute State similarity degree, the motion feature, above-mentioned position feature and the morphological feature of frame image；

Subelement is obtained, extracts all feature vectors that subelement extracts for obtaining third described in the Fusion Features model Corresponding weight coefficient, above-mentioned weight coefficient indicate are as follows: M=[w1, w2, w3, w4, w5, w6, w7]；Wherein, M indicates ownership The set of weight coefficient；Weight coefficient w1, w2, w3, w4, w5, w6 and w7 are respectively the variation degree of the pixel, the pixel The arrangement tightness degree of point, the global shape, the similarity degree of the frame image, the motion feature, above-mentioned position feature Weight coefficient corresponding with the morphological feature；

8th computation subunit, for according to formulaTo all feature vectors and correspondence The weight coefficient calculated, and when calculated result meets first threshold, determine the moving target for abnormal behaviour； Wherein, label indicates calculated result；What T was represented is matrix transposition；F_iIndicate any feature vector in all feature vector F； I=1,2,3,4,5,6,7；w_iIndicate the corresponding weight coefficient of each feature vector in the Fusion Features model.

7. the unusual checking device according to claim 5 or 6 based on fusion feature, which is characterized in that further include:

Selection unit, for choose the video resource for containing abnormal behaviour as positive sample and the video containing normal behaviour money Source is as negative sample；

Training computing unit, positive sample and negative sample progress for being chosen by support vector machines to the selection unit Training calculates, and establishes default Fusion Features model；

Repetitive exercise computing unit, the negative sample for being chosen according to the selection unit build the trained computing unit The vertical default Fusion Features model carries out more wheel repetitive exercises, and reaches preset threshold in the default Fusion Features model When, determine the Fusion Features model by corresponding default Fusion Features model for practical application.

8. the unusual checking device according to claim 5 based on fusion feature, which is characterized in that further include:

Taxon classifies for the moving target to detecting and tracking, determines the characteristic type of the moving target；

The determination unit is specifically used for, when the characteristic type for detecting the moving target is the behavior of people, described in determination The behavior type of moving target.