CN101271529A - Abnormity detection device and program - Google Patents
Abnormity detection device and program Download PDFInfo
- Publication number
- CN101271529A CN101271529A CNA2008100085095A CN200810008509A CN101271529A CN 101271529 A CN101271529 A CN 101271529A CN A2008100085095 A CNA2008100085095 A CN A2008100085095A CN 200810008509 A CN200810008509 A CN 200810008509A CN 101271529 A CN101271529 A CN 101271529A
- Authority
- CN
- China
- Prior art keywords
- unusual
- normal
- judging part
- monitored object
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Landscapes
- Image Analysis (AREA)
- Closed-Circuit Television Systems (AREA)
- Alarm Systems (AREA)
Abstract
The invention provides an anomaly detecting device which can reduce missing reports of wrongly detecting anomalous monitoring targets to natural monitoring targets via learning. The anomaly detecting device (20A) comprises: a photographing data obtaining department (21) obtaining photographing data of the monitoring targets; a normal learning data storing department (22A) storing normal learning data which processes to learn in advance for the normal monitoring targets to obtain; a normal judgment department (23A) which judges if the monitoring targets are normal according to the photographing data and the normal learning data; an anomaly learning data storing department (24A) storing anomaly learning data which processes to learn in advance for the anomalous monitoring targets to obtain; an anomaly judgment department (25A) which judges if the monitoring targets are anomalous according to the photographing data and the anomaly learning data; and a comprehensive judgment department (26A) which judges the monitoring targets are normal or anomalous according to the judgment result of the normal judgment department (23A) and the judgment result of the anomaly judgment department (25A).
Description
Technical field
The present invention relates to camera data that a kind of use taken by camera head abnormal detector that detects unusually and abnormality detection program to monitored object.
Background technology
In order to deal with the social wild effect of crime incidence increase etc., be that the platform number that is provided with of the video camera of purpose increases to monitor a suspect and suspect vehicle etc.When using numerous video cameras to monitor, need be used for the supervision support technique that effectively monitor area is monitored by less supervision personnel.
As above-mentioned supervision support technique; in patent documentation 1 (spy opens the 2006-79272 communique) a kind of detection technique is disclosed for example; its extraction is called as the image feature amount of the local autocorrelation characteristic of three-dimensional high order; use this characteristic quantity that the normal activity of the people in the image is learnt, and will be unusual action from the normal on movement detection of learning that departs from.
In addition, disclose a kind of detection technique in patent documentation 2 (spy opens flat 7-78239 communique), the image when it runs well shop equipment is learnt as normal condition, and the state-detection that will depart from from the normal condition of learning is an abnormality.In this detection technique, when detecting abnormality, by with prior learning to the pattern of multiple abnormality contrast, the kind of abnormality is discerned.
Patent documentation 2 spies open flat 7-78239 number openly
In patent documentation 1 disclosed detection technique, survey the wrong report of taking action for the flase drop of will normally taking action, by this image being appended study and the scope of normal action just being upgraded and can be tackled for unusual.But, in patent documentation 1 disclosed detection technique, there is following problem, that is, thereby will take action the flase drop survey unusually for normal failing to report of taking action for not detecting unusual action, even the image of normal action is appended study, also can't make and fail to report minimizing.
In addition, in patent documentation 2 disclosed detection techniques, carry out when detecting abnormality, the kind to abnormality to discern the detection of abnormality except adopting the method identical with patent documentation 1 disclosed detection technique.But, patent documentation 2 disclosed detection techniques are the same with patent documentation 1 disclosed detection technique, be will from prior learning to the technology that detects as abnormality of the state that departs from of normal condition, even, also can't make the problem of failing to report minimizing so exist image to append study to normal condition.
Summary of the invention
The present invention makes in order to address the above problem, and the object of the present invention is to provide a kind of abnormal detector and abnormality detection program, and it can survey failing to report for normal monitored object with unusual monitored object flase drop by learning to reduce.
To achieve these goals, abnormal detector of the present invention is characterised in that promptly, this abnormal detector has: the camera data acquisition unit of obtaining the camera data of monitored object; Normal learning data storage part, its storage are learnt in advance to normal described monitored object and the normal learning data that obtains; Normal judging part, it judges according to described camera data and described normal learning data whether the monitored object that is comprised in the described camera data is normal; Unusual study data store, its storage are learnt in advance to unusual described monitored object and the unusual learning data that obtains; Unusual judging part, it judges according to described camera data and described unusual learning data whether the monitored object that is comprised in the described camera data is unusual; With comprehensive judging part, it is according to described normal judgment result and described unusual judgment result, judges that the monitored object that is comprised in the described camera data is normally or unusual.
In addition, to achieve these goals, abnormality detection program of the present invention is characterised in that, that is, make robot calculator as following each performance function: the camera data acquisition unit of obtaining the camera data of monitored object; Normal judging part, it learns the normal learning data and the described camera data that obtain according to storage in advance in advance to normal described monitored object, judges whether the monitored object that is comprised in the described camera data is normal; Unusual judging part, its described monitored object to unusual according to storage are in advance learnt the unusual learning data and the described camera data that obtain in advance, judge whether the monitored object that is comprised in the described camera data is unusual; And comprehensive judging part, it is according to described normal judgment result and described unusual judgment result, judges that the monitored object that is comprised in the described camera data is normally or unusual.
Thereby, according to the present invention, can reduce unusual monitored object flase drop is surveyed failing to report for normal monitored object by study.
Description of drawings
Fig. 1 is the block diagram that expression has the abnormality detection system of the related abnormal detector of the 1st embodiment of the present invention.
Fig. 2 is the detailed diagram of the normal judging part of presentation graphs 1.
Fig. 3 is the detailed diagram of the unusual judging part of presentation graphs 1.
Fig. 4 (a) is the key diagram of camera data.
The key diagram of the calculating object of Fig. 4 (b) characteristic quantity.
The illustration figure of the mask pattern that Fig. 5 uses when being the three-dimensional high order of expression calculating part autocorrelation characteristic.
Fig. 6 is the block diagram that expression is used to generate the normal learning data generating apparatus of normal learning data.
Fig. 7 is the key diagram of the accumulation contribution rate of each principal component of obtaining in principal component analysis.
Fig. 8 is the key diagram of the computing of the positive normal manner of being undertaken by positive normal manner calculating part.
Fig. 9 is the block diagram that expression is used to generate the unusual study data generating device of unusual learning data.
Figure 10 is the key diagram of the computing of the abnormality degree that undertaken by the abnormality degree calculating part.
Figure 11 is the detailed diagram of the comprehensive judging part of presentation graphs 1.
Figure 12 (a) and (b) be the example key diagram of comprehensive judgement table.
Figure 13 is the process flow diagram that the action example to the abnormal detector of Fig. 1 describes.
Figure 14 is the process flow diagram that the action example to normal judging part describes.
Figure 15 is the process flow diagram that the action example to unusual judging part describes.
Figure 16 is the process flow diagram that the action example to feature value calculation unit describes.
Figure 17 is the process flow diagram that the action example to normal learning data generating apparatus describes.
Figure 18 is the process flow diagram that the action example to unusual study data generating device describes.
Figure 19 is the block diagram of the related abnormal detector of expression the 2nd embodiment of the present invention.
Figure 20 is the block diagram of the related abnormal detector major part of expression the 3rd embodiment of the present invention.
Figure 21 is the block diagram of the related abnormal detector major part of expression the 3rd embodiment of the present invention.
Figure 22 is the block diagram of the related abnormal detector major part of expression the 4th embodiment of the present invention.
Among the figure:
20A, 20B, 20C, 20D abnormal detector
21 camera data acquisition units
22A, 22B, the normal learning data storage part of 22C1,22C2
23A, 23B, the normal judging part of 23C
24A, 22B, 22C1,22C2 learn data store unusually
25A, 25B, the unusual judging part of 25C
The comprehensive judging part of 26A, 26D
Embodiment
The suitable accompanying drawing of following reference describes embodiments of the present invention.Give identical symbol to identical part, and omit repeat specification.
Abnormal detector of the present invention detected comprises from a plurality of frames (image) detected unusual action and these two notions of detected abnormality from single frame unusually.In the following embodiments, be that example describes with the situation that detects unusual action, but it does not constitute any restriction to technical scope of the present invention.In addition, include people, animal and vehicle etc. in the monitored object of the present invention.
＜the 1 embodiment 〉
" abnormality detection system "
Fig. 1 is the block diagram that expression has the abnormality detection system of the related abnormal detector of the 1st embodiment of the present invention.As shown in Figure 1, this abnormality detection system has camera head 10, unusual movement detection device 20A and external device (ED) 30.
External device (ED) 30 is to be used for the device (for example loudspeaker and display etc.) of testing result of output abnormality pick-up unit 20A.In addition, external device (ED) 30 can also constitute further possess security personnel from testing result to the security personnel that will notify with terminal, according to testing result to the mobile elevator control gear that limits of elevator and according to the device for controlling automatic door of testing result restriction automatically-controlled door action etc.That is to say, testing result is circulated a notice of to elevator control gear or device for controlling automatic door, thereby the action that just can limit elevator or automatically-controlled door prevents that the people as monitored object from being that suspicious personnel invade by abnormal detector 20A.
" detailed structure of abnormal detector 20A "
Below the detailed structure of abnormal detector 20A is described.As shown in Figure 1, abnormal detector 20A has camera data acquisition unit 21, normal learning data storage part 22A, normal judging part 23A, learns data store 24A, unusual judging part 25A, comprehensive judging part 26A and communicating department 27 as its function portion unusually.
Camera data acquisition unit 21 is obtained from the camera data of camera head 10 outputs.The camera data that is obtained is output to normal judging part 23A and unusual judging part 25A.
The prior learning result of the normal action of normal learning data storage part 22A storage monitored object is normal learning data.
Whether normal judging part 23A is normal action according to the action of the monitored object of judging in the camera data from the camera data of camera data acquisition unit 21 outputs and the normal learning data that reads from normal learning data storage part 22A to be comprised.At this, not (improper action) under the situation of normal action in the action of monitored object, normal judging part 23A is judged as unusual action with the action of monitored object.Judged result is output among the comprehensive judging part 26A.
The prior learning result of the unusual action of unusual study data store 24A storage monitored object is unusual learning data.
Whether unusual judging part 25A is unusual action according to the action of the monitored object of judging in the camera data from the camera data of camera data acquisition unit 21 outputs and the unusual learning data that reads from unusual study data store 24A to be comprised.At this, not (non-unusual action) under the situation of unusual action in the action of monitored object, unusual judging part 25A is judged as normal action with the action of monitored object.Judged result is output among the comprehensive judging part 26A.
Communicating department 27 will notify the notification data of this unusual action to output to external device (ED) 30 judging according to the judged result of comprehensive judging part 26A under the situation of action for action unusually of monitored object.
" detailed structure of normal judging part 23A "
The related normal judging part 23A of the 1st embodiment for example uses in the TOHKEMY 2006-79272 communique disclosed determination methods to judge whether the action of monitored object is normal action.At this moment, the normal learning data of being stored among the normal learning data storage part 22A is the inverse matrix of the matrix of a linear transformation, distribution center's (center of gravity) and variance-covariance matrix.Describe below about these.
Fig. 2 is the detailed diagram of the normal judging part of presentation graphs 1.As shown in Figure 2, normal judging part 23A has the 23a of extraction portion, feature value calculation unit 23b, characteristic quantity transformation component 23c, positive normal manner calculating part 23d, positive normal manner threshold value storage part 23e and positive normal manner judging part 23f.
The 23a of extraction portion extracts the part that motion is arranged from the camera data that is obtained by camera data acquisition unit 21.This carries out in order to delete with the camera data of judging irrelevant stationary part (for example background etc.).In order to extract the part that motion is arranged, the 23a of extraction portion can use disclosed portrait disposal route in the TOHKEMY 2005-92346 communique, for example, merely extract the method for 2 difference between the frame and handle the method for extracting the difference between 2 frames in the back etc. having implemented edge extracting.After the 23a of extraction portion has the part of motion in extraction, in order to remove the influence that waits the clutter that causes owing to illumination change, implementing binary conversion treatment for the part that motion is arranged, is 0 or 1 value to obtain pixel value.Camera data after having extracted the part that motion is arranged and having implemented binary conversion treatment is output among the feature value calculation unit 23b.
Feature value calculation unit 23b calculates the characteristic quantity of part that motion is arranged.Feature value calculation unit 23b can use in the TOHKEMY 2006-92346 communique the local autocorrelation characteristic of disclosed three-dimensional high order as characteristic quantity.The local autocorrelation characteristic of three-dimensional high order is that the geometric features of the voxel data (voxel data) that will be made up of the image of 3 successive frames is carried out Calculation Method as 251 proper vectors of tieing up.Computing method about this characteristic quantity will describe below.The characteristic quantity that calculates is that proper vector is output among the characteristic quantity transformation component 23c.
Characteristic quantity transformation component 23c use the transformation matrix stored in the normal learning data storage part 22 to the characteristic quantity that calculates by feature value calculation unit 23b, be that proper vector is carried out coordinate transform (linear transformation).This conversion is used for being extracted in the component of the normal action that proper vector comprises.At this, to be set at x by the proper vector that feature value calculation unit 23b calculates, the transformation matrix of being stored in the normal learning data storage part 22 is set at A, and in the time of will being set at x ' by the proper vector after the characteristic quantity transformation component 23c conversion, this conversion is represented by formula (1).
X '=Ax ... formula (1)
Transformation matrix A resolves the matrix that calculates by the multivariate of principal component analysis etc.Computing method about this transformation matrix A will describe below.When promptly the proper vector x of 251 dimensions use as the characteristic quantity of camera data with the local autocorrelation characteristic of high order, transformation matrix A be size for n * 251 (n=1,2 ..., 251) matrix.In addition, become the vector of n dimension by the proper vector x ' after the transformation matrix A linear transformation.Characteristic quantity after the conversion is that proper vector x ' is output among the positive normal manner calculating part 23d.
Positive normal manner calculating part 23d uses the proper vector x ' that has carried out conversion by characteristic quantity transformation component 23c to calculate to be used to the positive normal manner of the similar degree of representing the normal action of arriving with prior learning.At this, positive normal manner is a scalar, and this value is more little just represents that positive normal manner is high more; This value is big more just represents that positive normal manner is low more, represents that promptly it is unusual.Concrete computing method about positive normal manner will describe below.The positive normal manner that calculates is output among the positive normal manner judging part 23f.
Positive normal manner threshold value storage part 23e is used for storing positive normal manner threshold value.
Positive normal manner judging part 23f judges according to the positive normal manner that calculates by positive normal manner calculating part 23d whether the action of monitored object is normal action.The positive normal manner threshold value that positive normal manner judging part 23f will be stored among the positive normal manner threshold value storage part 23e is used as criterion.When positive normal manner when positive normal manner threshold value is following, positive normal manner judging part 23f judges that the action of monitored object be normally to take action.On the other hand, when positive normal manner had surpassed positive normal manner threshold value, positive normal manner judging part 23f judged that the action of monitored object is not that normal action (improper action) is promptly taken action unusually.Judged result is output in the comprehensive judging part 26.
" detailed structure of unusual judging part 25A "
The related unusual judging part 25A of the 1st embodiment for example uses in the TOHKEMY 2006-79272 communique disclosed determination methods to judge whether the action of monitored object is unusual action.At this moment, the unusual learning data of learning unusually to be stored among the data store 24A is the inverse matrix of the matrix of a linear transformation, distribution center's (center of gravity) and variance-covariance matrix (dispersion is divided into diffusing ranks).To describe below about these.
Fig. 3 is the detailed diagram of the unusual judging part of presentation graphs 1.As shown in Figure 3, unusual judging part 25A has the 25a of extraction portion, feature value calculation unit 25b, characteristic quantity transformation component 25c, abnormality degree calculating part 25d, abnormality degree threshold value storage part 25e and abnormality degree judging part 25f.
The 25a of extraction portion extracts the part that motion is arranged from the camera data that obtains by camera data acquisition unit 21.This carries out with the camera data of judging irrelevant stationary part in order to delete background etc.In order to extract the part that motion is arranged, the 25a of extraction portion can use disclosed portrait disposal route in the TOHKEMY 2005-92346 communique, for example, merely extract the method for 2 difference between the frame and handle the method for extracting the difference between 2 frames in the back etc. having implemented edge extracting.After the 25a of extraction portion has the part of motion in extraction, in order to remove the influence that waits the clutter that causes owing to illumination change, implementing binary conversion treatment for the part row that motion is arranged, is 0 or 1 value to obtain pixel value.Camera data after having extracted the part that motion is arranged and having implemented binary conversion treatment is output among the feature value calculation unit 25b.
Feature value calculation unit 25b calculates the characteristic quantity of part that motion is arranged.Feature value calculation unit 25b can use in the TOHKEMY 2006-92346 communique the local autocorrelation characteristic of disclosed three-dimensional high order as characteristic quantity.The local autocorrelation characteristic of three-dimensional high order is that the geometric features of the voxel data that will be made up of the image of 3 successive frames is carried out Calculation Method as 251 proper vectors of tieing up.Computing method about this characteristic quantity will describe below.The characteristic quantity that calculates is that proper vector is output among the characteristic quantity transformation component 25c.
It is that proper vector is carried out coordinate transform (linear transformation) to the characteristic quantity that calculates by feature value calculation unit 25b that characteristic quantity transformation component 25c uses the transformation matrix of being stored in the normal learning data storage part 22.This conversion is used for being extracted in the component of the unusual action that proper vector comprises.At this, when being set at y by the proper vector that feature value calculation unit 25b calculates, unusually the transformation matrix of learning to be stored in the data store 24 is set at B, and in the time of will being set at y ' by the proper vector that characteristic quantity transformation component 25c has carried out conversion, this conversion is represented by formula (2).
Y '=By ... formula (2)
Transformation matrix B resolves the matrix that calculates by the multivariate of principal component analysis etc.Computing method about this transformation matrix B will describe below.When promptly the proper vector y of 251 dimensions use as the characteristic quantity of camera data with the local autocorrelation characteristic of high order, transformation matrix B be size for n * 251 (n=1,2 ..., 251) matrix.In addition, the proper vector y ' that has carried out linear transformation by transformation matrix B becomes the vector of n dimension.Characteristic quantity after the conversion is that proper vector y ' is output among the abnormality degree calculating part 25d.
Abnormality degree calculating part 25d use by the proper vector y ' after the characteristic quantity transformation component 25c conversion calculate be used to represent with prior learning to the abnormality degree of similar degree of unusual action.At this, abnormality degree is a scalar, and the more little abnormality degree of just representing of this value is high more; The big more abnormality degree of just representing of this value is low more, represents that promptly it is normal.Concrete computing method about abnormality degree will describe below.The abnormality degree that calculates is output among the abnormality degree judging part 25f.
Abnormality degree threshold value storage part 25e is used for storing the abnormality degree threshold value.
Abnormality degree judging part 25f judges according to the abnormality degree that calculates by abnormality degree calculating part 25d whether the action of monitored object is unusual action.The abnormality degree threshold value that abnormality degree judging part 25f will be stored among the abnormality degree threshold value storage part 25e is used as criterion.When abnormality degree when the abnormality degree threshold value is following, abnormality degree judging part 25f judges that the action of monitored object be to take action unusually.On the other hand, when abnormality degree had surpassed the abnormality degree threshold value, abnormality degree judging part 25f judged that the action of monitored object is not the promptly normal action of unusual action (non-unusual action).Judged result is output in the comprehensive judging part 26.
" computing method of characteristic quantity "
Below to feature value calculation unit 23b, 25b carries out the computing method of characteristic quantity when calculating and is elaborated.And, because feature value calculation unit 23b is identical with the computing method of feature value calculation unit 25b, so following the computing method to feature value calculation unit 23b describe.Fig. 4 (a) is the key diagram of camera data, and Fig. 4 (b) is the key diagram of the calculating object of characteristic quantity.
Shown in Fig. 4 (a), the calculating object of characteristic quantity is image (animated image) i.e. continuous frame (image) group on sequential.In order to calculate the local autocorrelation characteristic of the related three-dimensional high order of present embodiment, need at least 3 frames.For example, in frame number be the frame F of m
mUnder the given situation, be the frame F of m-1 by the frame number that is positioned at this frame front and back position
M-1And frame number is the frame F of m+1
M+1And F
mThese 3 frames are formed 1 frame group G
F, and this frame group G
FCalculating object as characteristic quantity.
At this, when the resolution of frame is vertical h pixel and horizontal w pixel, frame group G
FConstitute 3 dimension frames of the voxel of h * w * 3.Feature value calculation unit 23b at whole key elements of voxel of 3 dimension frames, moves in turn and uses 3 * 3 * 3 mask pattern MP, extracts the local autocorrelation characteristic of three-dimensional high order thus.
And, in the present embodiment, with 3 continuous frame F
M-1, F
m, F
M+1The frame group G that forms
FAs process object, but also can be with the frame group formed by f frame arbitrarily as process object.At this moment, 3 dimension frames of the voxel of h * w * f become process object, and calculate the average characteristics amount of the image of f frame.
The illustration figure of the mask pattern that Fig. 5 uses when being the three-dimensional high order of expression calculating part autocorrelation characteristic.Mask pattern MP is used to calculate the local correlation feature of voxel, and in the present embodiment, mask pattern MP is made of 3 * 3 * 3 voxels.
The 1st mask pattern MP1 is used at sequential scanning during with respect to the voxel data of input picture, is the mask pattern that 1 o'clock number is counted to the voxel Bc1 at center.
The 2nd mask pattern MP2 is used for the mask pattern that the number when the voxel Bc2 on being positioned at voxel Bc1 also becomes 1 except that the voxel Bc1 at center is counted.
In the local autocorrelation characteristic of the three-dimensional high order relative, have 251 mask pattern MP (MP1 with 2 value images, MP2 ..., MP251), count by the number when satisfying each mask pattern MP, just the feature of the input picture proper vectors as 251 dimensions can be extracted.
That is to say that the number when satisfying i mask pattern MP becomes the i key element of proper vector.
" normal learning data generating apparatus 40 "
Then, the normal learning data generating apparatus that is used for generating the normal learning data that normal learning data storage part 22A stored is described.Fig. 6 is the block diagram that expression is used to generate the normal learning data generating apparatus of normal learning data.As shown in Figure 6, normal learning data generating apparatus 40 has study camera data acquisition unit 41, extraction portion 42, feature value calculation unit 43, principal component analysis portion 44 and segment space calculating part 45 as its function portion.
Study is obtained the normal action that camera head or image-reproducing apparatus are exported with camera data acquisition unit 41 and is learnt required study camera data.This study is with the normal action that includes monitored object in the camera data.And when hope is obtained a large amount of study when using camera data at short notice, study needs only from prior the record with camera data acquisition unit 41 and obtains study the recording medium (DVD etc.) of a large amount of study with camera data with camera data.The study of being obtained is output in the extraction portion 42 with camera data.
43 pairs of feature value calculation unit have the characteristic quantity of the part of motion to calculate.The action of feature value calculation unit 43 is identical with feature value calculation unit 23b.The characteristic quantity that calculates is a proper vector, is output in the principal component analysis portion 44.
Principal component analysis is carried out in the set of 44 pairs of proper vectors that calculate of principal component analysis portion.Principal component analysis is a kind of multivariate analytic method, and it generates compositional variable (being called as principal component) in incoherent mode each other from several variablees, the information that a plurality of variable had can be compiled thus.This principal component analysis method is widely used in the parsing of multivariate data, owing to for example in " parsing of The ぐ わ か Ru multivariate " East capital figure Books of the loyal husband's work in stone village (" multivariate that is easily understood is resolved " in October, 1992 is by the Tokyo bibliogony), detailed explanation is arranged, so be not described in detail at this.In the present embodiment, carry out principal component analysis by the set of 44 pairs 251 proper vectors of tieing up of principal component analysis portion, calculating 251 principal components thus is characteristic vector and eigenvalue.251 principal components that calculated are output in the segment space calculating part 45.
Segment space calculating part 45 calculates the segment space lower to the contribution rate of normal action according to the principal component that calculates.Then, segment space calculating part 45 calculates the matrix to this segment space with eigenvector projection.The matrix that is calculated is stored among the normal learning data storage part 22A.In addition, the inverse matrix of the variance-covariance matrix of the set of the center (center of gravity) of the set of the proper vector in the segment space calculating part 45 calculating section spaces and the proper vector in the segment space.The center of gravity that is calculated and the inverse matrix of variance-covariance matrix are stored among the normal learning data storage part 22A.
" computing of segment space "
Below the computing of the segment space that undertaken by segment space calculating part 45 is elaborated.Fig. 7 is the key diagram of the accumulation contribution rate of each principal component of obtaining in principal component analysis.The accumulation contribution rate obtains by in regular turn the contribution rate of each principal component being superposeed from big to small, and it is a kind of index data, represents that the quantity of information that principal component so far can have originally to the data of analytic target carries out the explanation of which kind of degree.
In the example of Fig. 7, the accumulation contribution rate till the third-largest principal component is 90%, and it means the total that the principal component of maximum is added up to the third-largest principal component, represent original data quantity of information 90%.On the other hand, the total till the principal component from the fourth-largest principal component to minimum only represented original data quantity of information 10%.
As mentioned above, can think big to the segment space that the third-largest principal component constitutes to the contribution rate of normal action by the principal component of maximum.In addition, can also think, little to the contribution rate of normal action by the fourth-largest principal component to the segment space that minimum principal component constitutes.So, be criterion with the accumulation contribution rate, can calculate the segment space little to the contribution rate of normal action.In addition, divide this item, then store in advance about accumulation contribution rate according to which number percent.
" computing of positive normal manner "
Below describe the computing of the positive normal manner of being undertaken in detail by positive normal manner calculating part 23d.Fig. 8 is the key diagram of the computing of the positive normal manner of being undertaken by positive normal manner calculating part.The calculating of positive normal manner is carried out in the less segment space of the contribution rate that aligns normal manner.This is because in this segment space, and the dispersion of the characteristic quantity of normal action is less, but not the normal i.e. bigger cause of dispersion of the characteristic quantity of unusual action of taking action.In the present embodiment, the segment space that uses the principal component of size below the big principal component of k+1 by principal component to constitute aligns normal manner and calculates.
This segment space is the segment space of 251-k dimension, but for convenience of explanation, selects big principal component and big these 2 axles of principal component of k+2 of k+1 to represent in Fig. 8.
The set C1 of characteristic quantity is the segment space that the characteristic quantity according to the normal action that is used to learn draws.In to the less segment space of the contribution rate of normal action, the center of gravity xc that characteristic quantity is distributed in set C1 is the near zone at center.Therefore, when the proper vector x ' relevant with the camera data of estimating was positioned near the center of gravity xc, the action of judging monitored object was normal action; As proper vector x ' during, the action of monitored object is judged as is not the promptly unusual action of normal action (improper action) away from center of gravity xc.
[0055]
At this, be positive normal manner with proper vector x ' relevant and the distance definition between the center of gravity xc with the camera data of estimating.This positive normal manner can be used as the less Euclidean distance that assesses the cost and calculates, but in the present embodiment, considers the deployment conditions of set C1, and calculates as Ma Shi (Mahalanobis) distance.
At this, the inverse matrix of the variance-covariance matrix of the set C1 of characteristic quantity is set at S
x -1The time, mahalanobis distance D1 is represented by following formula (3).
D1
2=(x '-xc)
tS
x -1(x '-xc) ... formula (3)
When the mahalanobis distance D1 that so calculates when normality threshold T1 is following, positive normal manner judging part 23f judges that the action of monitored object be normally to take action.On the other hand, as mahalanobis distance D1 during greater than positive normal manner threshold value T1, positive normal manner judging part 23f is judged as the action of monitored object and is not that normal action (improper action) promptly takes action unusually.At this, the positive normal manner threshold value T1 shown in Fig. 8 is the maximal value of selecting from each characteristic quantity of set C1 and the mahalanobis distance between the center of gravity xc.
" learning data generating device 50 unusually "
Below the unusual study data generating device that is used for generating the unusual learning data that unusual study data store 24A stored is described.Fig. 9 is the block diagram that expression is used to generate the unusual study data generating device of unusual learning data.As shown in Figure 9, learn data generating device 50 unusually and have study camera data acquisition unit 51, extraction portion 52, feature value calculation unit 53, principal component analysis portion 54 and segment space calculating part 55 as its function portion.
Study is obtained the unusual action that camera head or image-reproducing apparatus are exported with camera data acquisition unit 51 and is learnt required study camera data.This study is with the unusual action that includes monitored object in the camera data.And when hope is obtained a large amount of study when using camera data at short notice, study needs only from prior the record with camera data acquisition unit 51 and obtains study the recording medium (DVD etc.) of a large amount of study with camera data with camera data.The study of being obtained is output in the extraction portion 52 with camera data.
53 pairs of feature value calculation unit have the characteristic quantity of the part of motion to calculate.The action of feature value calculation unit 53 is identical with feature value calculation unit 25b.The characteristic quantity that calculates is a proper vector, is output in the principal component analysis portion 54.
Principal component analysis is carried out in the set of 54 pairs of proper vectors that calculate of principal component analysis portion.Principal component analysis is a kind of multivariate analytic method, and it generates compositional variable (being called as principal component) in incoherent mode each other from several variablees, the information that a plurality of variable had can be compiled thus.This principal component analysis method is widely used in the parsing of multivariate data, owing to for example in " parsing of The ぐ わ か Ru multivariate " East capital figure Books of the loyal husband's work in stone village (" multivariate that is easily understood is resolved " in October, 1992 is by the Tokyo bibliogony), detailed explanation is arranged, so be not described in detail at this.In the present embodiment, carry out principal component analysis by the set of 54 pairs 251 proper vectors of tieing up of principal component analysis portion, calculating 251 principal components thus is characteristic vector and eigenvalue.251 principal components that calculated are output in the segment space calculating part 55.
Segment space calculating part 55 calculates the segment space higher to the contribution rate of unusual action according to the principal component that calculates.Then, segment space calculating part 55 calculates the matrix to this segment space with eigenvector projection.The matrix that is calculated is stored among the unusual study data store 24A.In addition, the inverse matrix of the variance-covariance matrix of the set of the center (center of gravity) of the set of the proper vector in the segment space calculating part 55 calculating section spaces and the proper vector in the segment space.The center of gravity that is calculated and the inverse matrix of variance-covariance matrix are stored among the unusual study data store 24A.
" computing of segment space "
Next, segment space calculating part 55 can be that criterion calculates the segment space bigger to the contribution rate of unusual action with accumulation contribution rate shown in Figure 7.In addition, divide this item, then store in advance about accumulation contribution rate according to which number percent.
" computing of abnormality degree "
Below describe the computing of the abnormality degree that is undertaken by abnormality degree calculating part 25d in detail.Figure 10 is the key diagram of the computing of the abnormality degree that undertaken by the abnormality degree calculating part.The calculating of abnormality degree is carried out in to the bigger segment space of the contribution rate of abnormality degree.This is because in this segment space, and the dispersion of the characteristic quantity of unusual action is less, but not the unusual i.e. bigger cause of dispersion of the characteristic quantity of normal action of taking action.In the present embodiment, the segment space that uses the principal component of size more than the big principal component of k by principal component to constitute calculates abnormality degree.
This segment space is the segment space of k dimension, but for convenience of explanation, selects the largest principal component (the 1st principal component) and second largest these 2 axles of principal component (the 2nd principal component) to represent in Figure 10.
The set C2 of characteristic quantity is the segment space that the characteristic quantity according to the unusual action that is used to learn draws.In to the less segment space of the contribution rate of unusual action, the center of gravity yc that characteristic quantity is distributed in set C2 is the near zone at center.Therefore, when the proper vector y ' relevant with the camera data of estimating was positioned near the center of gravity yc, the action of judging monitored object was unusual action; As proper vector y ' during, the action of monitored object is judged as is not the promptly normal action of unusual action (non-unusual action) away from center of gravity yc.
At this, be abnormality degree with proper vector y ' relevant and the distance definition between the center of gravity yc with the camera data of estimating.This abnormality degree can be used as the less Euclidean distance that assesses the cost and calculates, but in the present embodiment, considers the dispersion of set C2 and calculates as mahalanobis distance D2.
At this, the inverse matrix of the variance-covariance matrix of the set C2 of characteristic quantity is set at S
y -1The time, mahalanobis distance D2 is represented by following formula (4).
D2
2=(y '-yc)
tS
y -1(y '-yc) ... formula (4)
When the mahalanobis distance D2 that so calculates when unusual threshold value T2 is following, abnormality degree judging part 25f judges that the action of monitored object be to take action unusually.On the other hand, as mahalanobis distance D2 during greater than abnormality degree threshold value T2, abnormality degree judging part 25f is judged as the action of monitored object and is not the promptly normal action of unusual action (non-unusual action).At this, the abnormality degree threshold value T2 shown in Figure 10 is the maximal value of selecting from each characteristic quantity of set C2 and the mahalanobis distance between the center of gravity yc.
" detailed structure of comprehensive judging part 26A "
Below the detailed structure of comprehensive judging part 26A is described.Figure 11 is the detailed diagram of the comprehensive judging part of presentation graphs 1.As shown in figure 11, comprehensive judging part 26A has comprehensive judgement table storage part 26a and normal/abnormal judging part 26b.
Comprehensive judgement table storage part 26a be used to store the judged result of the judged result of normal judging part 23A and unusual judging part 25A and comprehensively judged result set up comprehensive judgement table after related.
Normal/abnormal judging part 26b is with reference to the comprehensive judgement table of storing in comprehensive judgement table storage part 26a, according to the judged result of normal judging part 23A and the judged result of unusual judging part 25A, the action of judging monitored object is normal action or unusual action.This judged result is output in the communicating department 27.
" the comprehensive example of judging table "
Below the example of comprehensive judgement table is described.Figure 12 (a) and (b) be the example key diagram of comprehensive judgement table.
In the comprehensive judgement table 26a1 shown in Figure 12 (a), in the judged result of the judged result of normal judging part 23A and unusual judging part 25A all for just often, comprehensive judged result is judged as normally, when at least one judged result in the judged result of the judged result of normal judging part 23A and unusual judging part 25A is unusual, comprehensive judged result is judged as unusually.
In the comprehensive judgement table 26a2 shown in Figure 12 (b), at least one judged result is for just often in the judged result of the judged result of normal judging part 23A and unusual judging part 25A, comprehensive judged result is judged as normally,, comprehensive judged result is judged as unusually when all being unusual in the judged result of the judged result of normal judging part 23A and unusual judging part 25A.
The minimizing that focuses on of comprehensive judgement table 26a1 is failed to report, and focusing on of comprehensive judgement table 26a2 reduces in the wrong report.
" the action example of abnormal detector 20A "
Below the action example of the related abnormal detector 20A of the 1st embodiment of the present invention is described.Figure 13 is the process flow diagram that the action example to the abnormal detector of Fig. 1 describes.Below suitably describe moving example referring to figs. 1 through Figure 12.
At first, the camera data that will become process object by image acquiring unit 21 obtains (step S1) as numerical data.Next, the action of the monitored object of judging in the camera data to be comprised according to normal learning data by normal judging part 23A whether normally (step S2).In addition, the action of the unusual judging part 25A monitored object judging in the camera data to be comprised according to unusual learning data whether unusually (step S3).In addition, by the way, step S2 and step S3 can carry out by transpose, also can carry out side by side.
Next, according to the judged result of normal judging part 23A and the judged result of unusual judging part 25A, be normally or unusually to carry out comprehensive judgement (step S4) by comprehensive judging part 26A to the action of the monitored object that comprised in the camera data.The comprehensive result who judges is just often (in step S5 for No time), and communicating department 27 will exist unusual this situation of monitored object of taking action to circulate a notice of (step S6) to external device (ED) 30.Above-mentioned a series of processing is carried out repeatedly with the frequency of regulation, till user's end of input instruction of abnormal detector 20A (step S7).At this, the frequency of regulation for example is meant frequency with the frame frequency identical (1 second 30 times) of camera data etc.
" the action example of normal judging part 23A "
Below be that the action example of normal judging part 23A describes to the substep of step S2.Figure 14 is the process flow diagram that the action example to normal judging part describes.
At first, the 23a of extraction portion extracts the part that motion is arranged and carries out binary conversion treatment (step S21) from the camera data that obtains by camera data acquisition unit 21.Next, the characteristic quantity to the camera data after having extracted the part that motion is arranged and having carried out binary conversion treatment calculates (step S22) in feature value calculation unit 23b.Afterwards, being used the transformation matrix of normal learning data storage part 22A is that feature value vector is carried out coordinate transform by characteristic quantity transformation component 23c to the characteristic quantity that is calculated, and has carried out the new proper vector (step S23) of proper vector conversion with generation.After this, be used to represent to have carried out the proper vector of coordinate transform and the positive normal manner (step S24) of the similarity degree between the normal action by positive normal manner calculating part 23d calculating.Then, judge according to positive normal manner threshold value and positive normal manner whether the action of monitored object is normal action (step S25) by positive normal manner judging part 23f.
" the action example of unusual judging part 25A "
Below be that the action example of unusual judging part 25A describes to the substep of step S3.Figure 15 is the process flow diagram that the action example to unusual judging part describes.
At first, from the camera data that obtains by camera data acquisition unit 21, extract the part that motion is arranged and carry out binary conversion treatment (step S31) by the 25a of extraction portion.Next, the characteristic quantity to the camera data after having extracted the part that motion is arranged and having carried out binary conversion treatment calculates (step S32) in feature value calculation unit 25b.Afterwards, using the transformation matrix of unusual study data store 24A by characteristic quantity transformation component 25c is that feature value vector is carried out coordinate transform to the characteristic quantity that is calculated, and has carried out the new proper vector (step S33) of proper vector conversion with generation.After this, be used to represent to have carried out the proper vector of coordinate transform and the abnormality degree (step S34) of the similarity degree between the unusual action by abnormality degree calculating part 25d calculating.Then, judge according to abnormality degree threshold value and abnormality degree whether the action of monitored object is unusual action (step S35) by abnormality degree judging part 25f.
" feature value calculation unit 23a, the action example of 25a "
Below to step S22, the substep of S32 is feature value calculation unit 23a, the action example of 25a describes.Because the action example of feature value calculation unit 23a and feature value calculation unit 25a is identical, so, be that example describes at this with feature value calculation unit 23a.Figure 16 is the process flow diagram that the action example to feature value calculation unit describes.
At first, feature value calculation unit 23a to proper vector vn (n=1 ..., 251) and carry out initialization (step S101).Next, judge by feature value calculation unit 23a whether the voxel relative with i mask pattern MP all is 1 (step S102).When the voxel relative with i mask pattern MP all is 1 (in step S102 for Yes time), in the key element of the proper vector relative, add 1 (step S103) with i mask pattern MP by feature value calculation unit 23a.The processing of step S102 and S103 is carried out (circular treatment 1) to all mask pattern i, and all voxels of the frame group of process object are carried out (circular treatment 2).
By above-mentioned a series of processing, feature value calculation unit 23a can calculate the proper vector vn based on local autocorrelative 251 dimensions of three-dimensional high order.
" the action example of normal learning data generating apparatus 40 "
Below the action example of normal learning data generating apparatus 40 is described.Figure 17 is the process flow diagram that the action example to normal learning data generating apparatus describes.
At first, the part of motion is arranged and carry out binary conversion treatment (step S111) with extracting the camera data from the study of obtaining with camera data acquisition unit 41 by extraction portion 42 by study.Next, there is the characteristic quantity of the part of motion to calculate (step S112) by 43 pairs of feature value calculation unit.All study that the study normal row is employed are carried out the processing (circular treatment 3) of step S111 and S112 with camera data.Afterwards, carry out principal component analysis (step S113) with study with all relevant characteristic quantities of camera data by 44 pairs in principal component analysis portion.Then, by result calculating to the contribution rate of the normal action lower segment space of segment space calculating part 45, and calculate the matrix (step S114) to this segment space with eigenvector projection according to principal component analysis.
" learning the action example of data generating device 50 unusually "
Below the action example of unusual study data generating device 50 is described.Figure 18 is the process flow diagram that the action example to unusual study data generating device describes.
At first, the part of motion is arranged and carry out binary conversion treatment (step S121) by extraction portion 52 extraction from the camera data that obtains with camera data acquisition unit 51 by study.Next, there is the characteristic quantity of the part of motion to calculate (step S122) by 53 pairs of feature value calculation unit.All study to the usefulness of learning to take action are unusually carried out the processing (circular treatment 4) of step S121 and S122 with camera data.Afterwards, carry out principal component analysis (step S123) with study with all relevant characteristic quantities of camera data by 54 pairs in principal component analysis portion.Then, by result calculating to the contribution rate of the unusual action higher segment space of segment space calculating part 55, and calculate the matrix (step S124) to this segment space with eigenvector projection according to principal component analysis.
Because the related abnormal detector 20A of the 1st embodiment not only uses the judged result of making according to unusual learning data, but also use the judged result of making to carry out comprehensive judgement according to unusual learning data, so not only can reduce the wrong report of normal monitored object flase drop survey, unusual monitored object flase drop be surveyed failing to report for normal monitored object but also can reduce for unusual monitored object.
＜the 2 embodiment 〉
Below at the related abnormal detector of the 2nd embodiment, describe to attach most importance to the related abnormal detector 20A dissimilarity of the 1st embodiment.Figure 19 is the block diagram of the related abnormal detector of expression the 2nd embodiment of the present invention.
As shown in figure 19, the related abnormal detector 20B of the 2nd embodiment has normal learning data storage part 22B, normal judging part 23B, learns data store 24B and unusual judging part 25B unusually, replaces normal learning data storage part 22A, normal judging part 23A, learns data store 24A and unusual judging part 25A unusually.
Normal learning data storage part 22B has a plurality of normal learning data storage unit 22-1 and 22-2.Store diverse normal learning data respectively among a plurality of normal learning data storage unit 22-1 and the 22-2.
Normal judging part 23B has a plurality of normal judging unit 23-1 and 23-2.A plurality of normal judging unit 23-1 and 23-2 have the function that equates with normal judging part 23A respectively.Normal judging unit 23-1 judges that according to the normal learning data of being stored among the normal learning data storage unit 22-1 normal judging unit 23-2 judges according to the normal learning data of being stored among the normal learning data storage unit 22-2.
Unusual study data store 24B has a plurality of unusual study data storage cell 24-1 and 24-2.A plurality of unusual study data storage cell 24-1 and 24-2 store diverse unusual learning data respectively.
Unusual judging part 25B has a plurality of abnormal deciding means 25-1 and 25-2.A plurality of abnormal deciding means 25-1 and 25-2 have the function that equates with unusual judging part 25A respectively.Abnormal deciding means 25-1 judges according to the unusual learning data of being stored among the unusual study data storage cell 24-1, and abnormal deciding means 25-2 judges according to the unusual learning data of being stored among the unusual study data storage cell 24-2.
The normal/abnormal judging part 26b (with reference to Figure 11) of the comprehensive judging part 26A that the 2nd embodiment is related, at least one judged result in the judged result of a plurality of normal judging unit 23-1 and 23-2 is for just often, the judged result of normal judging part 23B is judged as normally, and in the judged result of a plurality of normal judging unit 23-1 and 23-2 when all being unusual, the judged result of normal judging part 23B is judged as unusually.So, because the action beyond a plurality of normal judging unit 23-1 and 23-2 will normally take action detects indirectly and is unusual action, so, when the judged result of a plurality of normal judging unit 23-1 and 23-2 when all being unusual, it is more appropriate that the normal/abnormal judgement judging part 26b of comprehensive judging part 26A is judged as anomaly ratio with the judged result of normal judging part 23B.
In addition, normal/abnormal judging part 26b, when at least one judged result in the judged result of a plurality of abnormal deciding means 25-1 and 25-2 is unusual, the judged result of unusual judging part 25B is judged as unusually, and all be just often in the judged result of a plurality of abnormal deciding means 25-1 and 25-2, the judged result of unusual judging part 25B is judged as normally.So, because a plurality of abnormal deciding means 25-1 and 25-2 directly detect unusual action, so, when at least one judged result in the judged result of a plurality of abnormal deciding means 25-1 and 25-2 when being unusual, it is more appropriate that the normal/abnormal judging part 26b of comprehensive judging part 26A is judged as anomaly ratio with the judged result of unusual judging part 25B.
Because the related abnormal detector 20B of the 2nd embodiment has the storage unit with the result of independent study stores according to the kind classification (clustering) of normal action and unusual action, so can therefore can improve the judgement precision effectively according to different kind component part spaces.
＜the 3 embodiment 〉
Below at the related abnormal detector of the 3rd embodiment, describe to attach most importance to the related abnormal detector 20A dissimilarity of the 1st embodiment.Figure 20 and Figure 21 are the block diagrams of the related abnormal detector major part of expression the 3rd embodiment of the present invention.
As Figure 20 and shown in Figure 21, the related abnormal detector 20C of the 3rd embodiment has a plurality of normal learning data storage part 22C1 and 22C2, normal judging part 23C, a plurality of unusual study data store 24C1 and 24C2 and unusual judging part 25C, replaces normal learning data storage part 22A, normal judging part 23A, learns data store 24A and unusual judging part 25A unusually.In addition, abnormal detector 20C also has the clock portion 28 of informing the moment.
A plurality of normal learning data storage part 22C1 store different normal learning datas respectively with 22C2.
Normal judging part 23C also has normal learning data selection portion 23g.Normal learning data selection portion 23g is chosen in the normal learning data that uses among feature value calculation unit 23b and the characteristic quantity transformation component 23c according to the moment of reading from a plurality of normal learning data storage part 22C1 and 22C2 from clock portion 28.
A plurality of unusual study data store 24C1 store different unusual learning datas respectively with 24C2.
Unusual judging part 25C also has unusual learning data selection portion 25g.Unusual learning data selection portion 25g is chosen in the unusual learning data that uses among feature value calculation unit 25b and the characteristic quantity transformation component 25c according to the time of reading from a plurality of unusual study data store 24C1 and 24C2 from clock portion 28.
Because the learning data that the related abnormal detector 20C of the 3rd embodiment comes switching judging to use according to the moment, so, for example, also can access good comprehensive judged result in relevant with the time to a great extent occasion of the use form of monitored objects such as office block.
＜the 4 embodiment 〉
Below at the related abnormal detector of the 4th embodiment, describe to attach most importance to the related abnormal detector 20A dissimilarity of the 1st embodiment.Figure 22 is the block diagram of the major part of the related abnormal detector of expression the 4th embodiment of the present invention.
As shown in figure 22, the related abnormal detector 20D of the 4th embodiment has comprehensive judging part 26D, replaces comprehensive judging part 26A.In addition, abnormal detector 20D also has in order to inform clock portion 28 constantly.
Comprehensive judging part 26D has comprehensive judgement table selection portion 26c.In addition, a plurality of comprehensive judgement table 26a1 and 26a2 have been stored among the related comprehensive judgement table storage part 26a of present embodiment.Comprehensive judgement table selection portion 26c is chosen in employed comprehensive judgement table among the normal/abnormal judging part 26b according to the moment of reading from a plurality of comprehensive judgement table 26a1 and 26a2 from clock portion 28.
Because the related abnormal detector 20D of the 4th embodiment switches comprehensive judgement table according to the moment, so, for example, also can access good comprehensive judged result in relevant with the time to a great extent occasion of the use form of monitored objects such as office block.
More than embodiments of the present invention are illustrated, but the present invention is not limited in above-mentioned embodiment, and can suitably carry out design alteration without departing from the spirit and scope of the present invention.For example, abnormal detector 20A that can be related, 20B, the combination that 20C, 20D suit to each embodiment.In addition, also can and learn data generating device 50 unusually with normal learning data generating apparatus 40 is combined into one with abnormal detector 20A.And each selection portion 23g, 25g, 26c can be configured to not select according to the moment, and according to for example selecting by the selection instruction of user's input.In addition, the present invention can also specific implementation make the abnormality detection program of robot calculator as described abnormal detector performance function.
Claims (13)
1. abnormal detector has:
The camera data acquisition unit, it obtains the camera data of monitored object;
Normal learning data storage part, its storage are learnt in advance to normal described monitored object and the normal learning data that obtains;
Normal judging part, it judges according to described camera data and described normal learning data whether the monitored object that is comprised in the described camera data is normal;
Unusual study data store, its storage are learnt in advance to unusual described monitored object and the unusual learning data that obtains;
Unusual judging part, it judges according to described camera data and described unusual learning data whether the monitored object that is comprised in the described camera data is unusual; With
Comprehensive judging part, it is according to described normal judgment result and described unusual judgment result, judges that the monitored object that is comprised in the described camera data is normally or unusual.
2. abnormal detector as claimed in claim 1, it is characterized in that, in described normal judgment result is that normal and described unusual judgment result is non-when unusual, and the monitored object that is comprised in the described camera data of described comprehensive judgement section judges is for normal; In described normal judgment result is improper or described unusual judgment result when being unusual, and the monitored object that is comprised in the described camera data of described comprehensive judgement section judges is for unusual.
3. abnormal detector as claimed in claim 1, it is characterized in that, in described normal judgment result is that normal or described unusual judgment result is non-when unusual, and the monitored object that is comprised in the described camera data of described comprehensive judgement section judges is for normal; In described normal judgment result is improper and described unusual judgment result when being unusual, and the monitored object that is comprised in the described camera data of described comprehensive judgement section judges is for unusual.
4. abnormal detector as claimed in claim 1 is characterized in that,
Described normal learning data storage part has a plurality of normal learning data storage unit that normal learning data is according to type stored,
Described normal judging part has corresponding with described a plurality of normal learning data storage unit; and kind according to the normal learning data of being stored in this normal learning data storage unit; the action of judging the monitored object that is comprised in the described camera data is normally or unusual a plurality of normal judging unit
When described comprehensive judging part all is improper in the judged result of described a plurality of normal judging units, judge that described normal judgment result is improper.
5. abnormal detector as claimed in claim 1 is characterized in that,
Described unusual study data store has a plurality of unusual study data storage cell that unusual learning data is according to type stored,
Described unusual judging part has corresponding with described a plurality of unusual study data storage cells; and kind according to the unusual learning data of being stored in this unusual study data storage cell; judge that the monitored object that is comprised in the described camera data is unusually or normal a plurality of abnormal deciding means
When described comprehensive judging part at least 1 in the judged result of described a plurality of abnormal deciding means is unusual, judge that described unusual judgment result is for unusual.
6. abnormal detector as claimed in claim 1 is characterized in that,
Have a plurality of described normal learning data storage parts,
Described normal judging part selects 1 to be used for judging from a plurality of described normal learning data storage parts.
7. abnormal detector as claimed in claim 6 is characterized in that, described normal judging part is based on selecting 1 constantly from a plurality of described normal learning data storage parts.
8. abnormal detector as claimed in claim 1 is characterized in that,
Have a plurality of described unusual study data store,
Described unusual judging part selects 1 to be used for judging from a plurality of described unusual study data store.
9. abnormal detector as claimed in claim 8 is characterized in that, described unusual judging part is based on selecting 1 constantly from a plurality of described unusual study data store.
10. abnormal detector as claimed in claim 1 is characterized in that,
Described comprehensive judging part has:
Comprehensive judge the table storage part, its storage is set up comprehensive judgement table after related with described the 1st judgment result and described the 2nd judgment result and this comprehensive judgment result; With
Normal/abnormal judging part, it is according to described the 1st judgment result, described the 2nd judgment result and described comprehensive judgement table, judges that the monitored object that is comprised in the described camera data is normally or unusual.
11. abnormal detector as claimed in claim 10 is characterized in that,
Described comprehensive judgement table storage part has a plurality of described comprehensive judgement tables,
Also have comprehensive judgement table selection portion, it selects 1 from described a plurality of comprehensive judgement tables,
Described normal/abnormal judging part uses the comprehensive judgement table of being selected by described comprehensive judgement table selection portion to judge.
12. abnormal detector as claimed in claim 11 is characterized in that, described comprehensive judgement table selection portion is based on selecting 1 constantly from described a plurality of comprehensive judgement tables.
13. an abnormality detection program makes computing machine as following each performance function:
The camera data acquisition unit, it obtains the camera data of monitored object;
Normal judging part, it learns the normal learning data and the described camera data that obtain according to storage in advance in advance to normal described monitored object, judges whether the monitored object that is comprised in the described camera data is normal;
Unusual judging part, its described monitored object to unusual according to storage are in advance learnt the unusual learning data and the described camera data that obtain in advance, judge whether the monitored object that is comprised in the described camera data is unusual; And
Comprehensive judging part, it is according to described normal judgment result and described unusual judgment result, judges that the monitored object that is comprised in the described camera data is normally or unusual.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2007-076472 | 2007-03-23 | ||
| JP2007076472A JP4369961B2 (en) | 2007-03-23 | 2007-03-23 | Abnormality detection device and abnormality detection program |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN101271529A true CN101271529A (en) | 2008-09-24 |
| CN101271529B CN101271529B (en) | 2010-09-29 |
Family
ID=39907217
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2008100085095A Active CN101271529B (en) | 2007-03-23 | 2008-01-23 | Abnormity detection device and program |
Country Status (3)
| Country | Link |
|---|---|
| JP (1) | JP4369961B2 (en) |
| CN (1) | CN101271529B (en) |
| HK (1) | HK1122633A1 (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110494861A (en) * | 2017-03-31 | 2019-11-22 | 横河电机株式会社 | Method for detecting abnormality and system based on image |
| CN110998654A (en) * | 2017-07-27 | 2020-04-10 | 三菱日立电力系统株式会社 | Model learning device, learning completion model generation method, program, learning completion model, monitoring device, and monitoring method |
| CN111164575A (en) * | 2017-10-11 | 2020-05-15 | 三菱电机株式会社 | Sample data generation device, sample data generation method, and sample data generation program |
| CN111902712A (en) * | 2018-03-29 | 2020-11-06 | 三菱电机株式会社 | Abnormality detection device and abnormality detection method |
| CN112771462A (en) * | 2018-10-01 | 2021-05-07 | 三菱电机株式会社 | Degradation detection system |
| CN113396369A (en) * | 2019-02-07 | 2021-09-14 | 株式会社电装 | Abnormality detection device |
| CN113542671A (en) * | 2020-04-21 | 2021-10-22 | 株式会社日立制作所 | Event analysis system and event analysis method |
Families Citing this family (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP4842197B2 (en) * | 2007-04-17 | 2011-12-21 | 財団法人ソフトピアジャパン | Abnormal operation detection device using multiple divided images, abnormal operation detection method, and abnormal operation detection program |
| JP5197343B2 (en) * | 2008-12-18 | 2013-05-15 | 綜合警備保障株式会社 | Registration apparatus and registration method |
| JP5214533B2 (en) * | 2009-05-21 | 2013-06-19 | 富士フイルム株式会社 | Person tracking method, person tracking apparatus, and person tracking program |
| JP5496566B2 (en) * | 2009-07-30 | 2014-05-21 | 将文 萩原 | Suspicious behavior detection method and suspicious behavior detection device |
| JP5358851B2 (en) * | 2009-11-12 | 2013-12-04 | 将文 萩原 | Suspicious behavior detection method and suspicious behavior detection device |
| JP5455787B2 (en) * | 2010-05-27 | 2014-03-26 | パナソニック株式会社 | Motion analysis apparatus and motion analysis method |
| JP5738028B2 (en) * | 2011-03-25 | 2015-06-17 | セコム株式会社 | Video processing device |
| JP5764592B2 (en) * | 2013-02-22 | 2015-08-19 | 東京エレクトロン株式会社 | Substrate processing apparatus, substrate processing apparatus monitoring apparatus, and substrate processing apparatus monitoring method |
| JP6401648B2 (en) * | 2015-03-31 | 2018-10-10 | 株式会社Ｓｃｒｅｅｎホールディングス | Defect classification apparatus and defect classification method |
| JP6708385B2 (en) | 2015-09-25 | 2020-06-10 | キヤノン株式会社 | Discriminator creating device, discriminator creating method, and program |
| JP6910786B2 (en) * | 2016-11-30 | 2021-07-28 | キヤノン株式会社 | Information processing equipment, information processing methods and programs |
| JP6823501B2 (en) * | 2017-03-03 | 2021-02-03 | 日本電信電話株式会社 | Anomaly detection device, anomaly detection method and program |
| JP7035395B2 (en) * | 2017-09-13 | 2022-03-15 | 沖電気工業株式会社 | Anomaly detection system, information processing device, and anomaly detection method |
| JP7059846B2 (en) * | 2018-07-24 | 2022-04-26 | 日本製鉄株式会社 | Tuyere monitoring device, tuyere monitoring program, and tuyere monitoring method |
| JP6876219B2 (en) * | 2018-11-29 | 2021-05-26 | ホームネットカーズ株式会社 | Refueling system, refueling method, program, and recording medium |
| US20220390272A1 (en) * | 2019-07-05 | 2022-12-08 | Nec Corporation | Optical fiber sensing system, optical fiber sensing equipment, and underground action monitoring method |
| JPWO2021015256A1 (en) * | 2019-07-25 | 2021-01-28 | ||
| JP7375934B2 (en) * | 2020-06-24 | 2023-11-08 | 日本電気株式会社 | Learning device, estimation device, learning method and program |
| JP6979660B1 (en) * | 2020-09-30 | 2021-12-15 | 株式会社エクサウィザーズ | Information processing method, learning model generation method, information processing device and computer program |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7624174B2 (en) * | 2003-05-22 | 2009-11-24 | Microsoft Corporation | Self-learning method and system for detecting abnormalities |
| JP4718253B2 (en) * | 2005-06-17 | 2011-07-06 | 株式会社日立ビルシステム | Image abnormality detection device for surveillance camera |
-
2007
- 2007-03-23 JP JP2007076472A patent/JP4369961B2/en active Active
-
2008
- 2008-01-23 CN CN2008100085095A patent/CN101271529B/en active Active
-
2009
- 2009-01-21 HK HK09100646.9A patent/HK1122633A1/en unknown
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110494861A (en) * | 2017-03-31 | 2019-11-22 | 横河电机株式会社 | Method for detecting abnormality and system based on image |
| CN110494861B (en) * | 2017-03-31 | 2023-09-22 | 横河电机株式会社 | Image-based anomaly detection method and system |
| CN110998654A (en) * | 2017-07-27 | 2020-04-10 | 三菱日立电力系统株式会社 | Model learning device, learning completion model generation method, program, learning completion model, monitoring device, and monitoring method |
| CN110998654B (en) * | 2017-07-27 | 2024-03-26 | 三菱重工业株式会社 | Model learning device, learning completion model generation method, program, learning completion model, monitoring device, and monitoring method |
| CN111164575B (en) * | 2017-10-11 | 2023-08-22 | 三菱电机株式会社 | Sample data generating device, sample data generating method, and computer-readable storage medium |
| CN111164575A (en) * | 2017-10-11 | 2020-05-15 | 三菱电机株式会社 | Sample data generation device, sample data generation method, and sample data generation program |
| CN111902712A (en) * | 2018-03-29 | 2020-11-06 | 三菱电机株式会社 | Abnormality detection device and abnormality detection method |
| CN111902712B (en) * | 2018-03-29 | 2023-07-25 | 三菱电机株式会社 | Abnormality inspection device and abnormality inspection method |
| CN112771462B (en) * | 2018-10-01 | 2023-06-09 | 三菱电机株式会社 | Degradation detection system |
| CN112771462A (en) * | 2018-10-01 | 2021-05-07 | 三菱电机株式会社 | Degradation detection system |
| CN113396369A (en) * | 2019-02-07 | 2021-09-14 | 株式会社电装 | Abnormality detection device |
| CN113396369B (en) * | 2019-02-07 | 2024-06-14 | 株式会社电装 | Abnormality detection device |
| CN113542671A (en) * | 2020-04-21 | 2021-10-22 | 株式会社日立制作所 | Event analysis system and event analysis method |
Also Published As
| Publication number | Publication date |
|---|---|
| JP4369961B2 (en) | 2009-11-25 |
| JP2008234551A (en) | 2008-10-02 |
| HK1122633A1 (en) | 2009-05-22 |
| CN101271529B (en) | 2010-09-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN101271529B (en) | Abnormity detection device and program | |
| CN101571914B (en) | Abnormal behavior detection device | |
| CN206601173U (en) | Use view data calculated weight | |
| CN101025825B (en) | Abnormal action detecting device | |
| EP3038028B1 (en) | Sequencing products recognized in a shelf image | |
| US20170295953A1 (en) | Monitoring device, monitoring system, and monitoring method | |
| US20170300938A1 (en) | Commodity monitoring device, commodity monitoring system, and commodity monitoring method | |
| US20160300181A1 (en) | Product monitoring device, product monitoring system, and product monitoring method | |
| JP6549558B2 (en) | Sales registration device, program and sales registration method | |
| US20160086149A1 (en) | Merchandise item registration apparatus | |
| JP6464612B2 (en) | Data analysis device, data analysis system, sales prediction device, sales prediction system, data analysis method, sales prediction method, program and recording medium | |
| WO2016199405A1 (en) | Shelf allocation assistance device, shelf allocation assistance system, shelf allocation assistance method, and recording medium | |
| JP6147676B2 (en) | Information processing apparatus, store system, and program | |
| US20100272350A1 (en) | Methods and apparatus to perform image classification based on pseudorandom features | |
| WO2011148439A1 (en) | Operation analysis device and operation analysis method | |
| US20090322875A1 (en) | Surveillance system, surveillance method and computer readable medium | |
| US20140340425A1 (en) | Display control apparatus, control method of display control apparatus, and storage medium | |
| JP7213294B2 (en) | Image/video analysis method using activity signature and its analysis system | |
| JP2018185660A (en) | Video inspection and recognition device | |
| EP4174788A1 (en) | Learning program, learning method, and information processing apparatus | |
| JP6869742B2 (en) | Information processing equipment and programs | |
| KR102597692B1 (en) | Method, apparatus, and computer program for measuring volume of objects by using image | |
| CN111339957B (en) | Method, system and medium for detecting state of money bundle in vault based on image recognition | |
| JP2007179138A (en) | Image retrieval method and image retrieval device | |
| KR20220067363A (en) | Image analysis server, object counting method using the same and object counting system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| REG | Reference to a national code |
Ref country code: HK Ref legal event code: DE Ref document number: 1122633 Country of ref document: HK |
|
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| REG | Reference to a national code |
Ref country code: HK Ref legal event code: GR Ref document number: 1122633 Country of ref document: HK |