JPH0844874A - Image change detector - Google Patents

Abstract

PURPOSE:To distinguishedly detect only a moving object by calculating the probability of an abnormal change from the occurance probability distribution of a feature vector and the feature vector calculated from a still state and a normal change state from a timewise image sequence. CONSTITUTION:An image sequence providing means 11 provides an image (image sequence) to a feature extracting means 1. The feature extracting means 1 calculates the feature vector for each picture element or previously decided small area in the applied image sequence. A probability distribution calculating means 2 calculates the occurance probability distribution of the feature vector in the time-space area of that picture element or small area. A change probability calculating means 3 calculates the probability of change presence/absence from the feature vector of a change detecting object image and the probability distribution stored in the probability distribution calculating means 2. A change presence/absence discriminating means 4 discriminates the change presence/ absence from the calculated probability value. A result display means 12 displays the result of the change detection by the change discriminating means 4 for a user. Thus, the abnormal change and normal change of the mobile object can be distinguishedly detected.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION The present invention, in any image,
The present invention relates to an image change detecting device for detecting only a moving object.

[0002]

2. Description of the Related Art As a method for detecting only a moving object in a video image captured by a TV camera or the like, as a conventional method, a method for obtaining a difference between a plurality of continuously captured images, There is a method of obtaining the difference between the background image created in advance and the current image.

These methods are, for example, for fixed TVs.
It is used in a device or the like for detecting an intruder or an intruder in a monitoring target area, such as an image monitoring device using an imaging device such as a camera.

However, in these methods, it is a major premise that the imaging device does not move and that there is no environmental change in the monitored area other than the intruding object. Therefore, there is a drawback that if this premise is broken, false detection is made for anything other than an intruder.

[0005] For example, when an environment such as the outdoors that changes greatly is monitored, intruders cannot be detected due to changes in illumination intensity due to changes in weather, fluctuations of trees due to wind, and irregular reflection of sunlight on the water surface. It was impossible to distinguish the object from the intruding object and it was very difficult to improve the detection accuracy of the moving object.

As a method for avoiding erroneous detection due to such environmental changes, reference document 1 (Maeda, Tanaka, Shio, Ishii: Object extraction method using distribution in multidimensional luminance space,
IEICE Technical Research Report Pattern Recognition / Understanding
PRU91-68 pp.67-74) and reference 2 (JP-A-6-893).
No. 39 Koike, Adachi: Object detection method) is known.

According to this method, various statistical feature amounts in a small area in an image (referred to as a luminance distribution width in Reference Document 2)
Is obtained and compared with the background image or that at the previous time point to realize object detection that is not affected by environmental changes. Reference 1 suggests a stable detection method using a time-series image by considering an image sequence as a multidimensional space, but unfortunately, no specific method is described.

In addition to the method of detecting by detecting the difference in spatial feature amount as described above, Reference 3 (Sato, Mase, Suenaga: x) for detecting change from a plurality of background images.
-t Robust object extraction method from spatio-temporal images, Proceedings of the 1991 IEICE Fall National Conference D-198 (6-200)) and Reference 4 (KANETA, KANOH, KANEMARU, NAGAI: Image)
Processing Method for Intruder Detection around Po
wer Line Towers, proc. of 1992 IAPR Workshop on Ma
The method of chine Vision Applications (MVA'92) pp.353-356) is known.

In Reference Document 3, a luminance distribution in the time direction is obtained for each pixel, and by assuming that this distribution is close to a normal distribution, pixels outside the center of the distribution are detected as moving object areas. is there. Also, reference document 4
Is to obtain the distribution of the difference value between the photographed image and the background image on the entire screen, and to imitate this distribution as a single-peaked distribution so as to detect pixels that are off the center of the distribution as changes.

The above-mentioned conventional methods mainly use the case where the change occurs relatively slowly in a wide area of the screen such as the change of the illumination intensity, or the brightness changes around a certain true brightness value. Then, it is a method of reducing the detection error of a moving object such as a pedestrian in response to an expected environmental change.

However, in addition to changes in illumination intensity, changes in the environment include those that cause steady changes in brightness such as trees confusing moving objects and fluctuations in the water surface (hereinafter, this change is called steady changes). Yes, in the method mentioned above, the change of the moving object (hereinafter, this change is called non-stationary change)
It was not possible to distinguish between and steady changes.

[0012]

As described above, in the conventional change detecting method for detecting a moving object such as an intruder, when there is fluctuation in brightness or the like other than the detection target in a captured image, However, there is a problem that it is not possible to distinguish and detect only a moving object to be detected.

The present invention has been made in consideration of such circumstances, and an object thereof is a change of a moving object even when there is a brightness change confusing with a detection target due to a change of a photographing environment. An object of the present invention is to provide an image change detecting device capable of distinguishing and detecting a steady change and a steady change such as a fluctuation of a tree or a water surface.

[0014]

An image change detecting apparatus of the present invention calculates a reference feature vector for each area of a temporal image sequence composed of a stationary state and a steady changing state, and at the same time, calculates an image of an object to be measured. Feature extraction means for calculating the measurement target feature vector for each area, probability distribution calculation means for obtaining the occurrence probability distribution of the reference feature vector for each area calculated by this feature extraction means, and measurement calculated by the feature extraction means A change probability calculating means for calculating a probability value of the presence or absence of a non-steady change from the target feature vector and the occurrence probability distribution obtained by the probability distribution calculating means, and a non-steady state from the probability value calculated by the change probability calculating means. And a change presence / absence determining means for determining the presence / absence of a dynamic change.

[0015]

[Operation] The image change detection device will be described.

The feature extraction means calculates a reference feature vector for each region of a temporal image sequence composed of a stationary state and a steady change state, and also calculates a measurement target feature vector for each region of the image to be measured. .

The probability distribution calculating means obtains the occurrence probability distribution of the reference feature vector for each area calculated by the feature extracting means.

The change probability calculating means calculates a probability value of the presence or absence of a non-steady change from the measurement target feature vector calculated by the feature extracting means and the occurrence probability distribution obtained by the probability distribution calculating means.

The change presence / absence determining means determines the presence / absence of an unsteady change from the probability value calculated by the change probability calculating means.

As a result, even if there is a steady change such as a brightness change in the image, it is possible to distinguish and detect the non-steady change of only the moving object. That is, the occurrence probability distribution of the feature vector composed of the stationary state and the steady change state is calculated from the captured temporal image sequence, and the non-steady change is calculated from the occurrence probability distribution and the observed image feature vector. By calculating the probability of, only moving objects can be distinguished and detected.

[0021]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

First Embodiment FIG. 1 is a schematic configuration diagram of a moving object detecting apparatus for detecting a moving object to which the image change detecting apparatus of the present invention is applied.

The moving object detecting device comprises a feature extracting means 1, a probability distribution calculating means 2, a change probability calculating means 3, a change determining means 4, an image sequence providing means 11 and a result displaying means 12.

The feature extracting means 1 calculates a feature vector for each pixel in a given image sequence or for each predetermined small area.

The probability distribution calculating means 2 obtains the occurrence probability distribution of the feature vector in the spatiotemporal area of the pixel or the small area.

The change probability calculating means 3 calculates the probability of presence or absence of change from the feature vector of the change detection target image and the probability distribution stored in the probability distribution calculating means 2.

The change presence / absence determining means 4 determines presence / absence of change from the calculated probability value.

The image sequence providing means 11 provides a video (image sequence) to the feature extracting means 1.

The result display means 12 displays the result of the change detection by the change determination means 4 to the user.

The change detection method is roughly divided into a two-step process of a probability calculation process and a change determination process.

Probability Calculation Process Before the change due to a moving object is detected, first, the feature extraction means 1 and the probability distribution calculation means 2 determine the probability distribution of the feature vector. FIG. 2 shows the flow of processing, and FIG. 3 shows a flowchart of processing in the following embodiment.

Change Discrimination Processing A change probability is obtained by the change probability calculating means 3 by obtaining a feature vector in the image to be subjected to the detection processing and comparing it with the above probability distribution. To determine. FIG. 4 shows the flow of processing, and FIG. 5 shows a flowchart of processing in the following embodiment.

The reference feature vector and the feature vector to be measured extracted by the feature extracting means 1 are scalar quantities (one-dimensional vector) having only the brightness value I of each pixel as an element.
The case will be described.

(1) The probability calculation process will be described.

It is assumed that there is an image sequence S (the total number of frames N, N = 1, 2, 3, ...) Of a video image photographed in advance. The image sequence handled here is represented by a rectangular parallelepiped as shown in FIG. The horizontal and vertical directions of the rectangular parallelepiped are spatial positions (x, y) on the image plane.
[Pixel], and the depth direction represents time (t) [frame number]. The pixel at position (i, j) in frame n is (i, j; n)
Shall be represented by. Further, as shown in FIG. 7, in the image of each frame, when the region Rxy [k × k pixel rectangular region (k = 1, 2, 3, ...), k = 1 at the position (x, y), 1 pixel] is defined. The image sequence S is assumed to be composed of a stationary scene and a scene of a steady change such as a fluctuation of a tree or a water surface, and does not include a scene of a non-steady change of a moving object.

When the image sequence provided by the image sequence providing means 11 is only a gray scale luminance image, the feature extracting means 1 does nothing concretely, and the probability after the previous image sequence providing means 11 is simply changed. Intensity I of an arbitrary pixel
Is simply passed as the reference feature vector.

However, when the image sequence is composed of images having color information such as RGB, the brightness I is calculated from these elements and the brightness value is passed to the probability distribution calculating means 2 as the reference feature vector (see FIG. 3). Steps 101 and 102).

Next, the probability distribution calculating means 2 obtains the probability distribution p (I) of the brightness I in the image sequence S.

Here, a method for obtaining p (I) as a statistical probability will be described.

The region Rxy of frame n (n = 1, 2, ... N) is defined as Rxy (n), and the union (image sequence of Rxy) Uxy of n of Rxy (n) is

[Equation 1] (See FIG. 7). All pixels included in Uxy

[Equation 2] About, the cumulative histogram fxy of the brightness I in Rxy
(I) is obtained (steps 103 and 104 in FIG. 3).

[0041]

(Equation 3) That is, by accumulating all the pixels in Uxy having the same brightness, fxy is obtained as a histogram as shown in FIG.
(I) is obtained. This fxy (I) divided by the total cumulative frequency Nxy in Uxy is the probability distribution p of I in the region Rxy.
xy (I) (step 105 in FIG. 3).

[0042]

[Equation 4] The probability distribution calculating means 2 calculates this probability distribution pxy (A) as
All regions Rxy are calculated for x and y, and this is stored.

(2) The change determination process will be described.

In addition to the image sequence S provided in advance, an image sequence S'to be measured for change detection (total frame numbers N ', N'
= 1, 2, 3, ...). S and S'may be completely different image sequences or may partially overlap each other. It is assumed that the image sequence S'includes a scene in a stationary state, a scene of a steady change such as a fluctuation of a tree or a water surface, and a scene of a non-steady change of a moving object to be detected.

Hereinafter, a method of discriminating a change in the region Rxy of S'will be described.

Method of calculating probability of presence / absence of change in change probability calculating means 3 A method of calculating probability of presence / absence of change in the change probability calculating means 3 will be described. Here, description will be given along with a method of calculating the probability of change presence or absence as a posterior probability.

The brightness I (i, j; n ') of the pixel (i, j; n') of the frame n'of S'is obtained by the feature extraction means I as the feature vector to be measured. When this pixel is in the region Rxy, the occurrence probability of the luminance I (i, j; n ') is pxy from the probability distribution at S calculated and stored by the probability distribution calculation unit 2 described above.
It shall be given by (I (i, j; n ')).

At each pixel of S ', the "no change" event θ0 in which the state of the pixel is a static state or a steady change
And, there are two events, “changed” event θ1 that there is an unsteady change. In other words, even if there is some brightness change in a pixel, if it is constant, the state of that pixel is a stationary state or a constant change and there is no change θ 0, otherwise, That is, when there is a non-stationary change, the event θ1 is that there is a change.

Let w (i, j; n ') be the prior probability that the pixel is θ1 before observing the brightness I of the pixel (i, j; n). The prior probability w can be set to any real number from 0 to 1. For example, assume that w (i, j; n ') = 1/2 if θ0 and θ1 are expected to occur with similar probabilities.

This prior probability w (i, j; n ') and the occurrence probability pxy of the observed brightness I (i, j; n') (step 201 in FIG. 5).
From (I (i, j; n ')), the posterior probability w' (i, j; n ') of being θ1 after the observation is obtained (step 202 in FIG. 5).

If the probability distribution at the event θ 0 is given by pxy (I), while the probability distribution at I under the event θ 1 is given by qxy (I), the posterior probability w ' (i,
j; n ') is obtained by the following formula from Bayes' theorem (step 203 in FIG. 5). Since it is complicated, w (i, j; n ') is simply referred to as w and I (i, j; n') is simply referred to as I in the following equation.

[0052]

(Equation 5) Here, qxy (I) assumes that all distributions are the same when the distribution of I is unknown under θ1. Therefore, when the range of the distribution of I is V,

(Equation 6) It can be melted.

For example, when I is given as an integer value of 256 levels from 0 to 255, V = 256, that is, qxy (I) = 1/256. The posterior probability w ′ (i, j; n ′) of the event θ 1 with change calculated in this way is
It is used as the probability that a change has occurred at pixel (i, j; n ').

When the change detection is performed in consecutive frames in the image sequence S ', the event .theta.1 obtained by the change probability calculating means 3 is detected.
The posterior probability w ′ (i, j; n ′) of can be used as the prior probability w (i, j; n ′ + 1) at the time of processing in the next frame. That is,

(Equation 7) , The previous posterior probability is used as the current prior probability in frame order. On the other hand, the value of w initially determined may be used as it is.

Discrimination Method of Presence or Absence of Change by Means for Discriminating Change 4 Next, a method of discriminating whether or not there is a change in the discrimination means 4 will be described. Here, description will be given along the method of assuming the loss of discrimination error.

Here, the loss L in the case where the discrimination is erroneous is assumed. Let L0 be the loss when it is discriminated that there is a change in a certain pixel which is actually θ0, but L1 is the loss when it is discriminated as θ0 although it is the opposite θ1. At this time, the criterion for discrimination that minimizes the expected value of loss after discrimination is

(Equation 8) Given in.

That is, by using the discrimination criterion of the above equation, it is possible to minimize the influence of failure after discrimination. Here, the respective losses L0 and L1 may be set to arbitrary values in accordance with the capability of the change detection method, and if it is assumed that both losses are equal,

[Equation 9] Is the criterion for discrimination. When a pixel (i, j; n ') satisfies the above discrimination criteria, it is determined that the pixel has an abnormal change,
It will be detected as a change area (step 204 in FIG. 5).

As described above, it is checked whether all the pixels (i, j; n ') or all the small regions Rxy satisfy the above-mentioned discrimination condition, and the pixels and the small regions satisfying this condition are selected. Assuming that there is an abnormal change, the change region is detected as shown in FIG.

In the original image shown in FIG. 9A, a constantly changing object such as a tree or a water surface is shown in FIG.
As shown in (b), it was inevitable to detect the change as in the case of a running person, but by using this device, only the non-steady change in which the steady change was removed as shown in FIG. 9 (c). Only people running can be detected.

Second Embodiment In the first embodiment, the feature extraction means 1 has been described as calculating the brightness I of a pixel, but various feature vectors may be used instead of the brightness I, and further, There is a method of combining two or more of these to construct a feature vector of two or more dimensions and using the feature vector.

For example, when RGB color element values can be obtained for each pixel of an image sequence, these elements may be used instead of the brightness I, or RGB can be viewed as a three-dimensional vector to obtain a feature vector. A as A [R,
A method of obtaining a three-dimensional feature vector G, B] for each pixel can be considered. Alternatively, using only two of the elements, for example, as a two-dimensional vector A [R, G], or using a luminance I together, a four-dimensional vector A [R, G, B, I] is also used. Conceivable.

Furthermore, not only the color element values are used as they are, but also element values such as hue H and saturation S in the Munsell color system which can be calculated from them can be used as elements of the feature vector A.

In addition to the RGB color system and Munsell color system described above, generally known XYZ, UCS, CMY, Y
The color systems such as IQ, Ostwald, L * u * v *, L * a * b * are specified. The element values of each color system are combined in the same way as described above, and the feature vector of any dimension can be obtained. It can be used. For the conversion method between the respective color systems, refer to Reference 5 (Takaki, Shimoda, supervised "Image Analysis Handbook", University of Tokyo Press ISBN 4-13-061107-0 C3050).
There is a method described in.

Third Embodiment For each element of the feature vector including the luminance,
It is also conceivable to use the result of the differential operation or the integral operation spatially or temporally as an element of the feature vector.

Specifically, as a spatial differential operation, generally known ones are sobel, Roberts, Robinso.
There are differential operators such as n, Prewitt, Kirsch, and canny, and ▽ ^ 2G (Laplacian Gaussian) and moment operators. The image edge strength as a result of applying these operators is used as an element of the feature vector.

It is also conceivable to use the result of noise removal processing such as an integral averaging filter or median filter for the feature vector.

Regarding the above operators and filters,
See Reference 5 for more details.

Fourth Embodiment Not only such spatial inter-pixel processing, but also calculation between corresponding pixels in the time axis direction of an image sequence, for example, difference (differential) processing in a continuous frame or fixed time It is also possible to use the result of the integration processing or the like, in which the weighted addition is performed, as an element of the feature vector.

Fifth Embodiment In the above, the feature vector uniquely determined for each pixel is described. However, the statistic that can be obtained in a predetermined minute area in the image is also an element of the feature vector. Can be used as

Examples of the statistic include average value, median value, mode value, range, variance, standard deviation, average deviation, and the like. These statistic values may be calculated in the small area Rxy described above, or some smaller areas r may be determined in Rxy and then calculated.

Sixth Embodiment Of the values that can be used as the elements of the feature vector described above, only one value may be used instead of the brightness I, or a plurality of arbitrary elements may be combined to make a large number. It can also be used as the dimensional feature vector A.

Here, the change detection method of the present invention will be described by taking the case where the feature vector is a two-dimensional vector composed of two elements of the spatial differential value D and the hue H as an example.

It is assumed that the image sequence providing means 11 provides images of three colors of RGB for one frame. In the feature extraction means 1, the HSI conversion converts the RGB value into H.
Calculate the SI value. Sobel for this I (luminance value) image
The operator is operated to obtain the spatial differential value (edge strength) D and the hue H for each pixel. HSI conversion and sobel
Since the operator is generally well known, detailed description thereof will be omitted here, and only a reference document in which a concrete method is described is given (for the HSI conversion, reference document 5 pp. .486-489, s
Also for the obel operator, refer to Reference 5 pp.553).

As in the case of only the brightness I described above, it is assumed that the small region Rxy and the region row Uxy are determined in the image row S, and all pixels included in Uxy

[Equation 10] For the above, the probability distribution calculating means 2 obtains the cumulative histogram fxy (A) of the feature vector A (i, j; n) in the feature vector space.

[0075]

[Equation 11] That is, the number of all A (i, j; n) having the same D and H values is accumulated to obtain a two-dimensional cumulative histogram fxy (A) in the DH space as shown in FIG. As in the first embodiment, the probability distribution pxy (A) is obtained by dividing the histogram fxy (A) by the cumulative frequency Nxy.

[0076]

(Equation 12) Further, in the same manner as in the first embodiment, the change probability calculating means 3 obtains the posterior probability w'of the event θ1 with change in each pixel. In this example, the probability distribution qxy (A) of A under θ1 is A according to the method of the first embodiment.
Is distributed in the two-dimensional space of DH, V is DH
If the range of D is VD and the range of H is VH, qxy (A) becomes

(Equation 13) Becomes For example, in 256 steps of D from 0 to 255,
When H is given in 360 steps from 0 to 359, qxy (A) = 1 / (256 × 360) = 1/921
It becomes 60. The formula for calculating the posterior probability w'is the same as in the first embodiment, and the prior probability of θ1 is w,

[Equation 14] Given in. As described above, even when the feature vector A is given as a multidimensional vector, the change presence / absence determining means 4 determines the determination criterion in the same manner as in the first embodiment based on the posterior probability w'of θ1 thus obtained, and determines the change. Can be done.

Seventh Embodiment In the embodiments described above, it is assumed that the probability distribution qxy (A) of A under θ 1 is uniform in the distribution space of A, but it is desired to detect it. If the distribution of the feature vector A on the target object can be investigated in advance, it is not necessary to assume that it is uniform.

For example, when the area of the target object in the image or the image sequence different from S'can be taken out by a method different from the present invention or a method in which the operator designates pixel by pixel. , The probability distribution qxy (A) of A in this target object area can be obtained. In other words, in the extracted region or region sequence, a cumulative histogram is obtained by the same method as that when the cumulative histogram fxy (A) is obtained from the previous S, and this histogram is divided by the cumulative frequency to obtain the statistics. Qxy as the target probability
(A) can be obtained.

Further, among all the spaces in which A is distributed, pxy is selected according to the subspace desired by the operator to detect.
(A) and qxy (A) can be set to completely arbitrary values.

Eighth Embodiment In the embodiments described above, the change probability calculating means 3 obtains the posterior probability of θ1, but the posterior probability calculation part may be omitted. Assuming that the probability of no change θ 0 under A and pxy (A) are equal, 1-p
Use xy (A) as the probability of θ1.

At this time, the discrimination method by the change presence / absence discriminating means 4 is as follows:

(Equation 15) When the condition is satisfied, it is detected that the pixel has changed. The reference value C in this case is arbitrary [0. ． 1] can be set to a real value.

The determination criterion by the change presence / absence determining means 4 in the ninth embodiment does not need to be uniform in all pixels or all small areas of the image to be processed or the image sequence S ', and the detection performance desired by the operator can be obtained. It can be changed together.

That is, the discrimination reference value is small in the image area where it is desired to correctly detect even a small change, whereas the discrimination reference value is large in the area where it is desired to detect a small change. The judgment reference value can be set to various values.

[0084]

According to the image change detecting device of the present invention, when detecting a moving object in a video such as an image monitoring device for detecting an intruder using an image pickup device such as a TV camera, Even if there are various changes in the environment in which the image was captured, and not only the changes in lighting conditions but also the constantly changing object that is confusing with the moving object to be detected, the purpose is to detect these environmental changes. It is possible to eliminate erroneous detection such as erroneously detecting as a moving object.

[Brief description of drawings]

FIG. 1 is a schematic configuration diagram showing a first embodiment of the present invention.

FIG. 2 is an explanatory diagram showing a flow of processing in probability calculation processing.

FIG. 3 is a flowchart of a probability calculation process.

FIG. 4 is an explanatory diagram showing a flow of processing in change determination processing.

FIG. 5 is a flowchart of change determination processing.

6 is an explanatory diagram showing an image sequence S. FIG.

7 is an explanatory diagram showing regions Rxy and Uxy in an image sequence S. FIG.

FIG. 8 is an explanatory diagram showing a probability distribution pxy when the feature vector is a scalar quantity.

FIG. 9 is an image diagram showing a result of change detection. (A) It is an original image figure. (B) It is an image figure processed by the conventional method. (C) is an image diagram processed by this apparatus.

FIG. 10 is an explanatory diagram showing an example of a probability distribution pxy when the feature vector is a two-dimensional vector.

[Explanation of symbols]

 1 Feature Extracting Means 2 Probability Distribution Calculating Means 3 Change Probability Calculating Means 4 Change Existence Discriminating Means 11 Image Sequence Providing Means 12 Result Displaying Means

Continuation of the front page (51) Int.Cl. ⁶ Identification code Office reference number FI technical display location H04N 7/18 K 9061-5H G06F 15/70 460 B

Claims (10)

[Claims] 1. Feature extraction means for calculating a reference feature vector for each region of a temporal image sequence composed of a stationary state and a steady change state, and for calculating a measurement target feature vector for each region of a measurement target image. , A probability distribution calculating means for obtaining the occurrence probability distribution of the reference feature vector for each area calculated by the feature extracting means, a measurement target feature vector calculated by the feature extracting means, and a probability distribution calculating means Change probability calculating means for calculating a probability value of presence / absence of non-steady change of the measurement target image from the occurrence probability distribution, and presence / absence of non-steady change of the measurement target image from the probability value calculated by the change probability calculating means An image change detecting device, comprising: a change presence / absence determining unit for determining the change. 2. The image according to claim 1, wherein the reference feature vector and the measurement target feature vector calculated by the feature extracting means are one-dimensional vectors composed of elements having only a luminance value at each pixel. Change detection device. 3. Elements of a reference feature vector and a measurement target feature vector calculated by the feature extraction means are one-dimensional or more vectors having element values of an arbitrary color system of RGB and HSI in each pixel. The image change detection apparatus according to claim 1, wherein the image change detection apparatus is a one-dimensional or more-dimensional vector composed of a combination of at least some of the values calculated from 4. The feature extracting means calculates the elements of the reference feature vector and the measurement target feature vector, and further calculates the spatial differentiation of the elements of the reference feature vector and the measurement target feature vector as a result of spatial calculation. The at least one of the result of the integral operation, the result of the temporal differential operation, or the result of the temporal integral operation is newly used as an element of the reference feature vector and the measurement target feature vector. Alternatively, the image change detection device according to claim 3. 5. The feature extraction means sets a region in an image in advance and uses statistics of elements of the reference feature vector and the measurement target feature vector in the region. Alternatively, the image change detection device according to claim 3. 6. The probability distribution calculating means obtains the occurrence probability distribution of the reference feature vector as a statistical probability from the frequency distribution of the reference feature vector in the spatiotemporal region of a given image sequence. The image change detection device according to claim 1. 7. The change probability calculating means determines a specific pixel or image from the occurrence probability distribution of the reference feature vector obtained by the probability distribution calculating means and the measurement target feature vector calculated by the feature extracting means. 2. The image change detection device according to claim 1, wherein a posterior probability of an event of having a change is calculated in an inner region, and the posterior probability is used as a probability value of whether or not there is a non-stationary change in the measurement target image. 8. The change probability calculating means extracts the probability distribution of the measurement target feature vector under the event of “change” in advance, and the frequency distribution of the feature vector in the spatiotemporal region where there is a change. The image change detecting apparatus according to claim 7, wherein the posterior probability is calculated from the statistical probability, and the posterior probability is calculated from the statistical probability. 9. The image change detecting apparatus according to claim 1, wherein the change presence / absence determining unit determines that there is a change when the probability of the event having a change is larger than a determination reference value. 10. The change presence / absence determining means determines a discrimination reference value on the assumption of a loss due to a discrimination error, and the discrimination reference value is an arbitrary real value between 0 and 1. 9. The image change detection device according to item 9.