JP4573857B2 - Sequential update type non-stationary detection device, sequential update type non-stationary detection method, sequential update type non-stationary detection program, and recording medium recording the program - Google Patents

Description

  The present invention records a sequential update type non-stationary detection device, a sequential update type non-steady state detection method, a sequential update type non-steady state detection program, and a program thereof that perform statistical analysis from a large number of input data to detect a non-steady state. The recording medium.

In the conventional technology for discriminating non-stationary patterns from a large number of patterns,
(1) Matching with a dictionary composed of learned samples, and if there are similar samples in the dictionary (or not), determine that they are non-stationary (2) Discriminant function generated by learning many samples There are methods such as discriminating by the output of

  When handling online input data, in order to ensure high speed, there is a method of performing high-speed matching by the method (1) or using a fixed discriminant function by the method (2). Has been taken.

  However, (a) the cost of labeling a large number of samples as stationary and non-stationary cannot be applied, (b) it is difficult to obtain a sufficient number of samples for each camera in advance, (c ) There are many cases where the situation changes after setting, and it is not easy to improve the accuracy of the identification dictionary.

  In particular, when the input is an image, and the situation where unsteady pattern discrimination is applied to an image of a remote monitoring system composed of a large number of cameras, both problems can occur.

  For this reason, it is difficult to discriminate a stable unsteady pattern unless the environment is relatively stable indoors. For these reasons, currently unsteady and practical systems are limited to applications in which the environment and the situation of people coming and going are relatively stable and the feature extraction function and the identification function do not need to be updated sequentially.

  Non-Patent Document 1 is an example of non-stationary detection in the monitoring in an elevator by the method (2) described above. However, “Non-Patent Document 1 has a favorable condition that the movement range of people is narrow and the number of appearances is small. It is stated in the literature that stable results can be obtained.

  As a solution to the problem (a), it is necessary to make the learning method an unsupervised method. Further, as a means for solving the above (b) and (c), it is necessary to update the identification axis every time a sample is added by using a method of performing learning sequentially.

  Considering such a background, sequential learning that can improve identification accuracy without manpower by online operation even if only a small number of samples can be obtained in the initial state for unsteady detection for application to monitoring system etc. It is required to be a technique that has functions and is premised on unsupervised learning.

  The conventional method for unsteady detection using unsupervised learning includes a method using a probabilistic model, such as a method for estimating a probabilistic model of spatio-temporal feature quantities in a change region of an image and detecting non-stationary motion such as falling Has been proposed.

  The probabilistic model can be updated sequentially, but if all the inputs are added to the model update as they are, if the newly input sample is an outlier, the model will be updated under the influence of the outlier There is a fear. Therefore, it is considered necessary to determine whether or not to add to the update data after determining whether or not it is an outlier, but it is difficult to determine whether or not the sample is an outlier while there are a small number of samples. The model must be generated with the number of stationary data.

  Regarding sequential learning, since it is important for a moving monitoring robot to update the steady state of a visual field image, several methods have been proposed, and some use the occurrence probability of a pixel value. In addition, in the intrusion detection system, there is a method of identifying samples input online and adding the samples identified as stationary to weight and update the background of the mask area. However, in order to deal with more complicated feature amounts, a method for updating the identification plane in the feature space must be applied.

  It is effective to use spatio-temporal features to perform non-stationary determination from video with high accuracy. This is because a non-stationary state is assumed due to the position of the person or the object, and a non-stationary state is assumed depending on the direction of the movement or the speed of the movement.

  Non-patent documents 2 and 3 use spatio-temporal features for this reason. However, there is no sequential update function of the identification space. This is because spatio-temporal features have a large number of dimensions, but there is no means for performing sequential update using an identification method suitable for identifying feature quantities having a large number of dimensions.

An incremental PCA (principal component analysis) algorithm related to the present invention is described in Non-Patent Document 4, and an incremental SVM (Support Vector Machine) algorithm is described in Non-Patent Document 5.
Kazuhiko Sumi, Masato Seki, Hideki Shiozaki, “Technology for detecting abnormalities in elevators using images”, Information Processing, Vol. -48, No. ˜1, pp. -17-22, 2006. Kazufuji, Wakabayashi, Arakawa, Anno, “Detection of unsteady sequences from long-term surveillance video”, Information Research Report CVIM-151, pp. 77-82, 2005. Kazufuji, Osawa, Wakabayashi, Anno, “Estimation of non-stationarity from surveillance video with a change area in video space-time”, IEICE Tech. 106, no. 75, pp. 49-54, 2006. Juyang Weng, Yilu Zhang and Wey-Shiuan Hwang, “Candid Covariance-free Incremental Component Analysis,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. -25, no. -8, pp. ˜1-16, ˜2003. G. Cauwenbergs, T .; Poggio, “Incremental and Decremental Support Vector Machine Learning,” Proc. Neural Information Processing, no. ˜13, pp. 409-415, 2000. URL: http: // siteseeer. ist. psu. edu / 454996. html Last access date: May 15, 2007

  In view of the above technical background, there is a need for a non-stationary detection method that can handle a feature quantity having a large number of dimensions as an input and that can sequentially update the identification system by additionally inputting a new sample. is there.

  The present invention solves the above-described problems, and the object of the present invention is to treat a feature quantity having a large number of dimensions as an input, and to sequentially update the identification system by newly inputting a new sample. An object of the present invention is to provide a sequential update type non-stationary detection device, a sequential update type non-stationary detection method, a sequential update type non-stationary detection program, and a recording medium on which the program is recorded.

As the main method of transient detection,
(1) A method of matching an input pattern with a steady pattern or an unsteady pattern itself (2) A method of probabilistically modeling a steady pattern or an unsteady pattern and obtaining a probability that the input pattern matches the model (3 ) There is a method that uses SVM (Support Vector Machine), which is a non-linear identification method, to identify whether an input pattern is on the stationary or non-stationary side of the identification space created by learning.

  Considering the identification accuracy, it is suitable for handling input patterns having a smaller number of dimensions in the order of (1), (2), and (3).

  For example, when using spatio-temporal features in non-stationary detection from video, the number of dimensions of the input pattern is (number of pixels in one frame) × (number of frames). In such a case, (2) and (3) are more suitable than (1) above (1).

  However, considering the calculation efficiency, if the number of dimensions is large, the cost of optimization calculation time and memory capacity is required even if the above (3) is used, so it is necessary to handle it efficiently to ensure high speed. is there.

  In such cases, it is generally considered appropriate to apply dimension compression by principal component analysis (PCA).

  Therefore, an identification system is used that performs learning by SVM using a dimension compression feature by principal component analysis as an input and updates the identification axis by sequentially reflecting input samples.

However, construction of such an identification system is not easy due to the following two problems.
Problem 1
In order to update the PCA subspace, it is usually necessary to store all the input data so far. The same applies to the update of the identification axis of the SVM. For this reason, a large amount of memory is required.
Problem 2
When sequential learning is performed on a PCA subspace, vectors (PCA features) projected onto different subspaces before and after sequential learning are generated. That is, a PCA feature generated by projection onto a partial space before sequential learning and a PCA feature generated by projection onto a partial space after sequential learning cannot be compared on the same scale.

Therefore, in the present invention,
As means for solving the problem 1,
The partial space of the PCA is updated by the incremental PCA, and the identification space of the SVM is updated by the incremental SVM. In this case, it is only necessary to store a support vector expressed with a small number of dimensions by PCA, and it is not necessary to store all high-dimensional input features.

As a means for solving the problem 2,
Under the condition that a sufficient contribution rate can be obtained with a small number of upper eigenvalues, the fact that the distance measure between vectors in the optimization calculation of the evaluation function of the SVM does not change is used.

That is, the sequential update type non-stationary detection apparatus according to claim 1 of the present invention is an apparatus that detects an unsteady state by statistical analysis from data sequentially input for each sample or for a small number of samples. , A feature vector generating means for generating a feature vector by statistical processing from input data using an algorithm of incremental PCA (principal component analysis), and an algorithm of incremental SVM (support vector machine) using the generated feature vector The feature vector generation means has a feature space model storage section, the identification means has an identification dictionary model storage section, and the feature vector generation means has input data. Each time is input, the predetermined frame constituting it is stored in the feature space model memory. In addition to the feature space model stored in the memory, the feature space model is updated, and the feature vector of the input data is projected onto the updated feature space model to obtain a dimension vector, and the identification is performed. Updating the support vector group in the form of projecting the input vector onto the feature space model before update stored in the dictionary model storage unit to the support vector group projecting onto the updated feature space model; Each time the dimension-compressed feature vector is input, the support vector group stored in the identification dictionary model storage unit is updated, and the unsteady state is detected based on the function of the incremental SVM for unsteady determination based on the update. It is characterized in that.

According to a second aspect of the present invention , the sequential update type non-stationary detection device according to the first aspect is characterized in that the input data is spatiotemporal data based on extraction of a change area of a video.

In addition, the sequential update type non-stationary detection method according to claim 3 of the present invention includes a feature vector generation unit and an identification unit, and includes incremental PCA (mainly from data sequentially input for each sample or a small number of samples. The non-stationary state is detected by an incremental SVM (support vector machine) algorithm by statistical analysis using a component analysis algorithm, and the feature vector generation means is configured to input the input data every time input data is input. A first update step of updating the feature space model by newly adding a predetermined frame constituting it to the feature space model stored in the feature space model storage unit, and the feature vector generation means includes the input A feature vector obtained by projecting the feature vector of the data onto the updated feature space model and dimensionally compressing it is used. And the feature vector generation means stores a support vector group in a form in which the input vector is projected onto the feature space model before being updated, which is stored in the identification dictionary model storage unit. A second updating step for updating to the support vector group projected on the screen, and the identification unit updates the support vector group stored in the identification dictionary model storage unit each time the dimension-compressed feature vector is input. A third update step and the identification means include a step of detecting a non-steady state based on a function of an incremental SVM for non-steady state determination by updating the third update step .

According to a fourth aspect of the present invention , the sequential update type non-stationary detection method according to the third aspect is characterized in that the input data is spatiotemporal data based on extraction of a change area of a video.

The sequential update type non-stationary detection program according to claim 5 of the present invention is a sequential update type non-stationary detection program that causes a computer to function as the feature vector generation unit and the identification unit according to claim 1 or 2 . It is characterized by that.

A recording medium according to claim 6 of the present invention is a computer-readable recording medium on which the sequential update type non-stationary detection program according to claim 5 is recorded.

(1) According to the inventions described in claims 1 to 6 , a feature quantity having a large number of dimensions can be treated as an input, and the identification system can be sequentially updated by additionally inputting a new sample. Possible non-stationary detection can be performed.

For example, when detecting a non-stationary scene from a video such as a surveillance camera image, the learning of the steady and non-stationary scenes is performed in advance by sequentially learning the accumulated scene and using it for calculation of the degree of normality. It is possible to detect a non-stationary scene without performing it.
(2) In addition, it is possible to construct an identification system capable of updating the identification axis by sequentially reflecting the input samples and performing learning by SVM using the dimension compression feature by PCA as input. Since the PCA subspace can be updated sequentially by the incremental PCA and the SVM identification space by the incremental SVM, the accuracy increases as the number of input samples increases.

  Conventionally, when performing sequential learning on a PCA subspace, vectors (PCA features) projected onto different subspaces before and after the sequential learning are generated and generated by projection onto the subspace before sequential learning. PCA features generated by projection to a partial space after sequential learning cannot be compared on the same scale, so PCA and SVM could not be used in a system that performs incremental input. .

  According to the present invention, the fact that the distance measure between vectors in the optimization calculation of the evaluation function of the SVM does not change under the condition that a sufficient contribution rate can be obtained with a small number of higher eigenvalues is used. It is possible to reduce the amount of calculation by expressing with a small number of dimensions by PCA before applying, and to save the memory capacity by storing SVM support vectors in a dimension-compressed form.

  For this reason, unsteady detection by sequential learning can be performed even when the input is a high-dimensional feature quantity, such as spatiotemporal data obtained from video.

  In non-stationary detection from video surveillance, non-stationary data is often not obtained enough for learning, but by using the present invention, even if only a small number of samples can be obtained in the initial state, online data cannot be obtained. The identification accuracy can be improved without manpower by unsupervised sample input.

  In addition, it is often difficult to define steady and non-stationary beforehand, or the cost of labeling steady and non-stationary is not high, but in the algorithm of the present invention, SVM is unsupervised (1 Class SVM), it is possible to improve the accuracy of the identification system simply by continuing to take data without performing steady and non-steady labeling for learning.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings, but the present invention is not limited to the following embodiments.

  FIG. 1 is a block diagram showing the overall configuration of the present embodiment, and 11 is a feature vector generation unit as feature vector generation means of the present invention. Reference numeral 110 denotes a feature space model storage unit included in the feature vector generation unit.

  Reference numeral 12 denotes an identification unit as identification means of the present invention, and 120 denotes an identification dictionary model storage unit included in the identification means.

  The feature vector generation unit 11, the feature space model storage unit 110, the identification unit 12, and the identification dictionary model storage unit 120 are configured by, for example, a computer, and each process described later is executed by the computer.

  In FIG. 1, when data is input online, a feature vector is generated in the feature vector generation unit 11. The feature vector is subjected to identification by the identification unit 12 and a non-stationary determination result is output.

  The feature vector generation unit 11 sequentially updates the feature space model using the feature space model stored in the feature space model storage unit 110, the identification dictionary stored in the identification dictionary model storage unit 120, and new input data, New input data is converted into a feature vector using the updated feature space model.

  The identification unit 12 sequentially updates and updates the identification dictionary model using the feature space model stored in the feature space model storage unit 110, the identification dictionary stored in the identification dictionary model storage unit 120, and new input data. The identification dictionary stored in the identification dictionary model storage unit 120 is used for identification to detect an unsteady state.

  Next, an embodiment of a specific method for realizing feature vector generation and identification will be described with reference to FIG. When samples are input online, it is considered that not only the SVM identification dictionary but also the subspace of principal component analysis for compressing input feature values changes as the number of samples increases. Therefore, the partial space can be obtained by incremental PCA (IPCA), and the result can be used as an input of incremental SVM (ISVM).

  First, a video (a time series of images of size w × h) is input to the feature vector generation unit 11. Next, after performing moving region extraction and binarization, each t frame is considered as one sample, and is set as a spatiotemporal feature of the order of w × h × t.

  If this is the case, the number of dimensions is large and it takes a high cost for identification. Therefore, principal component analysis is performed. Each time one sample of the time series of the change region is input, the t-frame constituting it is newly added to update the subspace (update (1) in FIG. 2). For updating the subspace, the IPCA algorithm disclosed in Non-Patent Document 4 is used.

  By projecting onto the partial space generated in this way, the spatiotemporal feature of the order of w × h × t is compressed p × t. Hereinafter, this p × t-order compressed feature value is referred to as a principal component feature.

  Next, unsteady detection is performed using this principal component feature as an input to the discriminator (identification unit 12).

  Each time one sample of principal component features is input, this sample is newly added to update the identification axis (update (3) in FIG. 2). One class SVM which is one of the outlier detection methods is used for the discriminator, and the ISVM algorithm disclosed in Non-Patent Document 5 is used for updating the discriminating axis.

  However, when the input data is obtained with supervision, the SVM discriminant can be a discriminant of two-class identification.

  When the condition that a sufficient contribution rate can be obtained with a small number of higher eigenvalues is satisfied, the distance between two original feature vectors can be expressed as a distance between principal component features with a small amount of data.

  Therefore, the calculation at the time of learning of SVM can be made efficient without changing the distance measure between samples used for the optimization calculation of the evaluation function.

  When performing non-stationary determination from monitoring data, the condition that the majority of data is stationary and a sufficient contribution rate can be obtained with a small number of higher eigenvalues is often satisfied. it can.

  At this time, the support vector used for the identification calculation is stored in the model of the identification dictionary in a form in which the original input vector is projected onto the partial space at the previous increment. This needs to be reprojected into the updated subspace (update (2) in FIG. 2).

A specific calculation formula is as shown in FIG. In FIG. 3, x is an input vector, U is a projection matrix into a subspace, X _SV is a support vector set of ISVM, α and ρ are parameters of 1 class SVM, and f (x) is an identification function of 1 class SVM. . The number (n) on the right shoulder indicates the number of increments. _xn is the nth input vector.

FIG. 4 shows a process flow for incrementing the subspace and SVM discriminant functions. The calculation formulas (update formulas) in each process correspond to the calculation formulas (1) to (4) in FIG. _{When n} is input, the subspace U ^n-1 is updated to U ⁿ (step S1; update (1)). The support vector group necessary for the SVM discriminant function is stored in a dimensionally compressed state by projection onto the subspace. This support vector group U ^n-1 X _SV ^n-1 is converted to U ⁿ X _SV ⁿ Update to ^-1 (step S2; update (2)).

Further, by projecting the feature vector x _n in the subspace U ^n, obtains a feature vector x ^_'n = U _n X _n that dimensional compression (Step S3).

Next, ISVM processing is performed. First, the support vector group X _SV ^n-1 and the dimension-compressed support vector group U ⁿ X _SV ^n-1 are updated to store U ⁿ X _SV ⁿ , and then the support vector parameters α ^n-1 and ρ ^n-1 is updated to α ⁿ and ρ ⁿ (step S4; update (3)).

Using the ISVM function f ⁿ (x) updated with these parameters, a vector x ′ _n obtained by compressing the input feature vector x _n is input to the function (step S5), and the unsteady determination result f ⁿ ( x ′ _n ) is output (calculation formula (4)).

In the above process, it is also possible to reverse the order of the update of the identification system and the output for the latest input data. In this case, non-stationary using f ^n-1 (x) before the update After the determination result f ^n-1 (x ′ _n ) is output, f ^n-1 (x) is updated to f ⁿ (x) in the same manner as described above (step S5 in FIG. 4 and output of the unsteady determination result). Reverse processing order).

The flow of the above processing will be described with reference to FIG. 3. First, a subspace U ^{(n-1) as} a result of IPCA at the ^{(n-1) th} step is generated, and when a new input vector x _n is added, U ^{( n−1)} is incremented using x _n to generate U ^{(n) (} formula (1) in FIG. 3).

As a result of the n-1 step, a support vector set X _SV ^{(n-1) of} ISVM is obtained. It is projected this because it is a vector consisting of a subset of U ^(n-1) in the projection of the _{x n-1 x 'n-} 1 = U (n-1) x n-1 to the current subspace The vector is corrected (Equation (2) in FIG. 3), and the new input x ′ _n = U ⁽ⁿ⁻¹⁾ x _n is added to increment ISVM to update α and ρ (Equation in FIG. 3). (3)) Using these updated parameters, the discrimination function f (x ′ _n ) of the one-class SVM is obtained (formula (4) in FIG. 3).

  The output of the discriminator (identification unit 12) is the output (scalar) of the discriminant function of 1 class SVM. The value of the scalar a is a> 0 when it is regarded as a stationary sample, and a <0 when it is regarded as a non-stationary sample.

  When a <0, it can be considered that the larger the value of | a |, the greater the deviation from the steady state. This is called the output non-stationary degree in the present invention.

  Next, the relationship between the number of input samples and the output of the discriminator (identification unit 12) will be specifically described with reference to FIGS.

  FIG. 5 is a schematic diagram showing what kind of change occurs depending on the number of increments when a non-stationary degree is output for a time series of input samples when a discriminant function is incremented by sequentially inputting a certain time series feature. FIG.

  FIG. 6 is a graph showing the specific number of samples and output values in FIG. 5, and the number of samples corresponds to A, B, and C in FIG.

  FIG. 7 is a monitoring image taken in the same set as the actual ATM.

  5 to 7, the start frame is shifted and a certain number of frames are input as one sample. A sample in which a frame of a non-stationary scene is included in one sample is non-stationary, and a sample that is not included is stationary. Ideally, the output values for a plurality of consecutive samples are close unless a sudden video change occurs, and the sign of the output value does not change abruptly between samples with adjacent start times.

  As shown in FIGS. 5 and 6, when the number of learning samples is small (illustration (A)), a phenomenon in which the output value fluctuates finely and positively is observed, and all the samples corresponding to those negative values are not unsteady. . However, by increasing the number of learning samples in the order of B and C in the figure, it is expected that false detection that becomes unsteady will decrease and the output graph will become smooth as shown in FIG.

  As a monitoring image, a large number of people were photographed as steady scenes (FIGS. 7 (a) and 7 (b)) in which a person came in front of the machine in order, operated, and left. In addition, a scene (FIG. 7 (c)) that was taken with a portable camera from behind the operator was photographed as non-stationary, and these were edited into one video.

  Non-stationary labels are around 200 samples, all others are stationary. Around 20 samples are scenes where a person first appears in the background, around 200 samples are voyeur scenes, and the middle is a machine operation scene.

  5 and 6, it can be seen that as the number of learning samples A, B, and C increases, unsteady erroneous detection of the output around 20 samples decreases. In addition, it can be seen that by increasing the number of B, C and learning samples, unsteady false detections around 40 samples are reduced.

  A scene in which a person appears is erroneously detected as non-stationary because there is no other similar scene, but it is considered that if a similar scene is added by increasing the number of input samples, it is learned as stationary.

  In addition, the present invention can be realized by configuring some or all of the functions of each means in the sequential update type non-stationary detection device of the present embodiment with a computer program and executing the program using the computer. It goes without saying that the procedure in the sequential update type non-stationary detection method of the present embodiment can be configured by a computer program, and the program can be executed by the computer. Readable recording media such as FD (Floppy (registered trademark) Disk), MO (Magneto-Optical disk), ROM (Read Only Memory), memory card, CD (Compact Disk) -ROM, DVD (Digital) (Versatile Disk) -ROM, CD-R, CD-RW, HDD, removable disk, etc. can be recorded and stored or distributed. It is also possible to provide the above program through a network such as the Internet or e-mail.

The block diagram which shows the structure of one embodiment of this invention. The block diagram which shows the structure of the specific Example of this invention. Explanatory drawing of the update type | formula and calculation type | formula processed by the feature vector production | generation means and identification means of this invention. The flowchart which shows the flow of the process which each means of this invention performs. The characteristic view which shows the relationship between the number of input samples and the output of a discriminator. The graph which shows the relationship between the number of input samples and the output of a discriminator. Explanatory drawing which shows the monitoring image | video used as input data.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 11 ... Feature vector production | generation part, 12 ... Identification part, 110 ... Feature space model memory | storage part, 120 ... Identification dictionary model memory | storage part.

Claims (6)

An apparatus for detecting a non-steady state by statistical analysis from data sequentially input for each sample or a small number of samples,
Feature vector generation means for generating a feature vector by statistical processing from input data using an algorithm of incremental PCA (principal component analysis), and an incremental SVM (support vector machine) algorithm using the generated feature vector Comprising an identification means for detecting the unsteady state,
The feature vector generation means has a feature space model storage unit,
The identification means has an identification dictionary model storage unit,
Each time input data is input , the feature vector generation means adds a predetermined frame constituting the input data to the feature space model stored in the feature space model storage unit and updates the feature space model. The feature vector of the input data is projected onto the updated feature space model to obtain a dimension vector, and the feature vector is obtained by dimensional compression. The input vector stored in the identification dictionary model storage unit is used as the feature space model before the update. Update the projected support vector group to the support vector group projected onto the updated feature space model,
The identification unit updates the support vector group stored in the identification dictionary model storage unit each time the dimension-compressed feature vector is input, and the non-steady determination based on the function of the incremental SVM for non-stationary determination. A sequential update type unsteady state detecting device characterized by detecting a steady state . The sequential update type non-stationary detection device according to claim 1, wherein the input data is spatiotemporal data based on video change area extraction .   Incremental SVM (support vector machine) by means of statistical analysis using incremental PCA (principal component analysis) algorithm from data sequentially input for each sample or a small number of samples. ) To detect the unsteady state by the algorithm of
  The feature vector generation means updates the feature space model each time input data is input by newly adding a predetermined frame constituting the input data to the feature space model stored in the feature space model storage unit. 1 update step;
  The feature vector generation means obtains a feature vector obtained by dimension-compressing the feature vector of the input data by projecting it onto the updated feature space model;
  The support vector group in which the feature vector generation unit projects the support vector group in the form of projecting the input vector onto the feature space model before the update, which is stored in the identification dictionary model storage unit, onto the updated feature space model A second update step for updating to
  A third updating step in which the identification unit updates a support vector group stored in the identification dictionary model storage unit each time the dimension-compressed feature vector is input;
  A step of detecting an unsteady state based on a function of an incremental SVM for unsteady determination by updating the third updating step;
  A sequential update type non-stationary detection method comprising:   4. The sequential update type non-stationary detection method according to claim 3, wherein the input data is spatio-temporal data based on video change area extraction.   A sequential update type non-stationary detection program for causing a computer to function as the feature vector generation unit and the identification unit according to claim 1.   A computer-readable recording medium on which the sequential update type non-stationary detection program according to claim 5 is recorded.