KR101192163B1 - Method and apparatus for detecting objects in motion through background image analysis by objects - Google Patents

Abstract

The object detecting method of the present invention comprises the steps of extracting objects from the input images, determining whether there is a similar object among the objects constituting the background model with respect to the extracted objects, and if there is a similar object, Updating the object with the extracted similar object, otherwise adding the extracted object to the background model, and among the objects updated or added to the background model, the object is not updated for a certain time by the object extracted from the new input image. And removing from the model.

Description

TECHNICAL AND APPARATUS FOR DETECTING OBJECTS IN MOTION THROUGH BACKGROUND IMAGE ANALYSIS BY OBJECTS

The present invention relates to image analysis, and more particularly, to a technique for detecting a moving object in an image.

Conventionally, detection techniques of moving objects have been divided into units of pixels or blocks (sets of pixels), and then an independent variation of each pixel or block is observed and analyzed. For example, in the case of using the learned background image, the variation width of the pixel values is learned in the pixel set of the input images, and then whether the object is moved is detected based on the difference between the training data and the input image. In this pixel-based object detection method, since each pixel-based process is learned independently, malfunction may occur when an unlearned portion appears, and sometimes only a part of an actual object area is detected. In addition, in the case of a dynamic background such as a tree swaying in the wind or the wind, the conventional technique is significantly less capable of detecting a moving object.

Even if the background is not dynamic, conventional techniques are difficult to apply when the camera moves, i.e., in an environment using a PTZ (pan, tilt, zoom) camera.

SUMMARY OF THE INVENTION An object of the present invention is to provide an object detection method and apparatus capable of solving a malfunction occurring in a pixel detection method by analyzing an image in object units and modeling a background based on the divided object unit information. It's there.

Object detection method according to an aspect of the present invention,

Extracting objects from input images;

Determining whether there is a similar object among objects constituting a background model with respect to the extracted objects;

In a learning mode, if there is a similar object as a result of the determination, updating the original object in the background model with the extracted similar object; otherwise, adding the extracted object to the background model; And

In a learning mode, the method may include removing from the background model an object that is not updated for a predetermined time by an object extracted from a new input image among objects updated or added to the background model.

According to one embodiment, the object detection method,

In the detection mode, the method may further include determining that the foreground object is the foreground object if there is no similar object to the background model.

According to one embodiment, the step of extracting the object,

Subdividing successive input images into regions having similar characteristics;

Obtaining regions corresponding to each other among the subdivided regions of the input images;

Clustering regions of high similarity based on the similarity calculated for the subdivided regions; And

Optimizing labeling of the clustered regions and defining an object with regions of the same label.

According to an embodiment, the subdividing the continuous input images into regions having similar characteristics may include:

The method may include subdividing regions having similar color values of pixels of the input images using a mean-shift method.

According to an embodiment, the step of obtaining regions corresponding to each other among the subdivided regions of the input images may include:

Calculating an optical flow voting between the plurality of pixels in a specific region of the previous image and the plurality of pixels in the region of the current image, and comparing the operation result to a threshold value;

Calculating a battacharrya distance with respect to appearance matching descriptors of a specific region of the previous image and a specific region of the current image, and comparing the calculation result to a threshold value;

Calculating distortion occurring in the representative movement direction of the peripheral regions, and determining that the tracking is successful with respect to the specific region in the continuous image when the calculation is less than or equal to the predetermined value; And

If the region of the current image does not have a correspondence between the regions of the previous image among the regions that are not determined to be the hidden or disappeared region, the method may include classifying the region of the current image as a newly appeared region. .

According to one embodiment, the step of clustering the areas of high similarity based on the similarity calculated for the subdivided areas,

Calculating similarity with respect to transfer characteristics, color characteristics, scale spatial homogeneity of the label, and temporal homogeneity of the label; And

It may include the step of armalizing the areas of high similarity.

Object detection method according to another aspect of the present invention,

Subdividing successive input images into regions having similar characteristics;

Obtaining regions corresponding to each other among the subdivided regions of the input images;

Clustering regions of high similarity based on the similarity calculated for the subdivided regions; And

Optimizing labeling of the clustered regions and defining an object with regions of the same label.

Object detecting apparatus according to another aspect of the present invention,

An object extracting unit which extracts objects from input images; And

It includes a background model learning unit for learning a background model,

The background model learning unit,

It is determined whether there is a similar object among the objects constituting the background model with respect to the extracted objects,

If it is determined that there is a similar object, the original object in the background model is updated with the extracted similar object; otherwise, the extracted object is added to the background model,

Among objects updated or added to the background model, an object that is not updated for a predetermined time by an object extracted from a new input image may be removed from the background model.

According to one embodiment, the object detection device,

The apparatus may further include a foreground determiner configured to determine whether there is a similar object among the objects constituting the background model with respect to the extracted objects, and determine that the object is a foreground object when there is no similar object in the background model with respect to the extracted object. Can be.

According to the object detection method and apparatus of the present invention, it is possible to perform background image modeling in units of objects in a dynamic background environment or a PTZ camera environment, and to detect a moving object from such a background image model.

1 is a flow chart illustrating an object detection method according to an embodiment of the present invention.
2 is a flowchart illustrating an object extraction step in detail in the object detection method according to an embodiment of the present invention.
3 is a diagram illustrating an image of a result of performing an area segmentation process of an object extraction step in an object detection method according to an embodiment of the present invention.
4 is a flowchart illustrating an area tracking process of an object extraction step in the object detection method according to an embodiment of the present invention.
5 is a diagram illustrating a result of the area tracking process of the object extraction step of the object detection method according to an embodiment of the present invention.
FIG. 6 is a diagram illustrating an area clustering process of an object extraction step in an object detection method according to an embodiment of the present invention.
7 is a diagram illustrating code words based on object information analyzed in the background modeling step of the object detection method according to an embodiment of the present invention.
8 is a diagram illustrating a pseudo code for explaining a background modeling step of an object detecting method according to an embodiment of the present invention.
9 is a diagram illustrating a result of a background modeling step of an object unit in an object detecting method according to an exemplary embodiment of the present invention.
10 is a block diagram illustrating an object detecting apparatus according to an embodiment of the present invention.

For the embodiments of the invention disclosed herein, specific structural and functional descriptions are set forth for the purpose of describing an embodiment of the invention only, and it is to be understood that the embodiments of the invention may be practiced in various forms, The present invention should not be construed as limited to the embodiments described in Figs.

Hereinafter, with reference to the accompanying drawings, it will be described in detail a preferred embodiment of the present invention. The same reference numerals are used for the same constituent elements in the drawings and redundant explanations for the same constituent elements are omitted.

Background Art [0002] Conventional object detection techniques mainly perform background modeling based on pixels, and compare the input image with the background modeled once.

On the other hand, the object detection method of the present invention first classifies areas of high correlation among the input images as objects, and models the background based on objects that do not move among the classified objects. Subsequently, if there is an object that does not belong to the objects constituting the background among the input images, it may be determined as an object belonging to the foreground rather than the background.

Since the object detection method of the present invention models the background based on the objects, even if there is a change in the input image due to panning, tilting and zooming of the camera, homogeneity of the objects constituting the background can be maintained. Since the background is robustly modeled, an object that does not belong to the background among objects analyzed in the input image may be easily detected as a moving object in the foreground.

1 is a flow chart illustrating an object detection method according to an embodiment of the present invention.

Referring to FIG. 1, the object detecting method of the present invention may be divided into a learning mode for largely modeling a background model and a detection mode for detecting a moving object by comparing an input image with the modeled background.

In operation S11, the method may start from extracting objects from input images.

If certain areas in the image have the following characteristics, they can be assumed to be areas belonging to an object.

1.Movement trajectory characteristics: have similar movement characteristics in time

2. Color characteristics: have similar color values

3. Label similarity in scale space: labels remain even when scale changes

4. Temporal Label Similarity: The same label is maintained over time

On the basis of these four characteristics, if certain regions of the input image show the same in the criteria of the above characteristics, such regions constitute one object. Therefore, an object may be extracted by analyzing regions of the input image. A process of extracting an object by dividing and clustering regions from the input image will be described with reference to FIGS. 2 to 5.

In step S12, it is determined whether there is a similar object among the objects constituting the background model with respect to the objects extracted from the input image.

For the object extracted in step S11, if it is determined in step S12 that there is a similar object in the background model and is also in the learning mode, in step S13, by updating the corresponding object in the original background model with the extracted object. Update the background model. If an object extracted from a new frame updates an existing background model, it means that the object is actually a background.

For the object extracted in step S11, if it is determined that the object was not in the background model in step S12 and is also in the learning mode, in step S14, the extracted object is added to the background model.

In the learning mode, since the background model is initially initialized, all the extracted objects will be added to the background model, after which only new objects will be added to the background model.

Subsequently to step S13 or step S14, in step S15, among the objects in the previously updated or added background model, if not updated for a predetermined time or more by the object extracted in the new frame, the background model is excluded.

The process of building and updating the learning mode, that is, the background model, will be described with reference to FIGS. 6 to 8.

In the learning mode, the background model is constructed and updated by repeating steps S11 to S15.

Step S16 is a step of comparing the object and the background model using the background model constructed in the learning mode as the detection mode to determine whether the object is a foreground object. In step S16, the objects extracted in step S11 are determined to be objects that are not in the background model in step 12, and in the case of detection mode, are determined to be foreground objects.

2 is a flowchart illustrating an object extraction step in detail in the object detection method according to an embodiment of the present invention.

Referring to FIG. 2, the object extracting step S11 may have the following detailed steps based on the characteristics of the object assumed above.

First, in step S21, successive input images are subdivided into regions having similar characteristics. Subdivided areas are the smallest units that make up an object. For example, an area may be subdivided into a set of pixels having the same to similar colors. Even if an object can have several colors, it is unlikely that an area of the same color belongs to two different objects.

According to an exemplary embodiment, segmentation of the input image may be performed by dividing into sets of pixels having similar color values using a mean-shift technique. The mean shift technique is a statistical technique that detects the center mode through a non-parametric method in probability space. When the average shift technique is applied to image analysis, the pixel value of pixels is regarded as a probability density function, and the average position and average color value of surrounding data having a color distribution similar to the specific data among the surrounding data existing near the specific data are calculated. After calculating and repeating the calculation of the average position and average color value of the surrounding data at the average position again, the average position converges to a certain coordinate. By repeating the assignment of the position of for all the pixels in the input image, it is possible to group pixels having similar color values.

Referring to FIG. 3 for a while, FIG. 3 is a diagram illustrating an image of a result of performing an area segmentation process of an object extraction step in an object detection method according to an exemplary embodiment of the present invention. It can be seen that pixels of similar color in the input image are divided into finely divided regions.

Referring back to FIG. 2, in step S22, regions corresponding to each other are found between the subdivided regions of each temporally consecutive input image.

This is to analyze what areas are temporally related to each other for the subsequent similarity calculation, based on the assumption that areas belonging to an object will have the same label and the same label in time.

4 is a flowchart illustrating an area tracking process of an object extraction step in the object detection method according to an embodiment of the present invention.

Referring to FIG. 4 for a moment, the area tracking process of the object extraction step includes, in step S221, several pixels in a certain region R _i (t-1) of a previous image and a pixel in a certain region R _j (t) of a current image. Compute optical flow voting between them and compare the result to the threshold.

In step S222, a batattacharrya distance of an appearance matching descriptor of a region R _i (t-1) of a previous image and a region R _j (t) of the current image is calculated. , Compare the result of the operation to the threshold.

Visual flow votes must _satisfy both the forward direction (from the area of the past image to the area of the current image, V _Fopt ) as well as the reverse (from the area of the current image to the area of the past image, V _Bopt ). The batatcharya distance value (BhattachDist) of the fields AppearDesc _i (t-1) and AppearDesc _j (t) should also be greater than or equal to a predetermined value.

These conditions are shown in equation (1).

According to an embodiment, the area tracking may be completed with only one of the two conditions of Equation 1. In other words, both step S221 and step S222 may be performed, or only one step may be performed.

Referring to FIG. 5 for a while, FIG. 5 is a diagram illustrating a result of an area tracking process of an object extraction step in an object detection method according to an embodiment of the present invention.

(A) of FIG. 5 illustrates visual flow votes between regions of the previous video and regions of the current image, and (b) illustrates the correspondence between the regions determined by such visual flow votes. .

By the external matching operation procedure of step S222, the representative movement direction W _j (t) of the region R _j (t) may be determined as in Equation 2.

Where u (x, y) is the motion vector of the region and V _opt is the visual flow votes.

On the other hand, in a case where an object is covered by another object, or when an area in the previous image disappears from the screen in the current image, the correspondence of the above-described steps S221 and S222 may appear incorrectly. To deal with this false correspondence, the tracking results of the surrounding areas of a certain area can be used. In view of the similarity of the movement characteristics among the properties assumed to appear in the case of one object, the areas considered to be the same object will have similar representative directions of movement. In contrast, the areas covering an object or the areas remaining after the object disappears tend to be different from the representative direction of movement of the other areas belonging to the object.

Therefore, in step S223, as shown in Equation 3, if the distortion generated in the representative movement direction of the peripheral areas is calculated, and if the calculated distortion is less than or equal to a predetermined value, the tracking succeeds with respect to the specific area in the continuous image. Determined to be.

Here, f is a vector defined by the magnitude (m) and direction (θ) of the optical flow.

Further, in step S224, among the areas of the current image not previously determined to be the area that is hidden or disappeared in step S223, an area of a current image corresponds to an area of the previous image in step S221. If no relationship is found, the area of the current image is classified as a newly appearing area of the current image.

The newly appeared area is not related to the area of the past image, but can find a temporal correspondence with the areas of the image afterwards.

Through this process, the movement characteristics of the regions of the current image can be known.

2, in step S23, regions with high similarity are clustered based on similarity calculated with respect to movement characteristic, color characteristic, scale spatial homogeneity of the label, and temporal homogeneity of the label.

Conventional object detection techniques usually adopt a method of subdividing an image and detecting an object based on a pre-learned object model. These methods compare an image captured in a real environment with a limited object model even though it works well in a laboratory environment. There were many errors in the operation.

In contrast, in the exemplary embodiments of the present invention, the structured data is clustered in the image obtained each time regardless of the object model, so that the object detection speed is fast and strong in the real environment.

First, a graph is generated based on the previously subdivided regions, movement characteristics, color characteristics, and temporal correspondences, and each region is a node and the similarity of the regions is an edge.

In this case, the similarity between the regions may be based on an appropriately defined or selected similarity matrix. In embodiments of the present invention, the similarity matrix A _ij may be used as shown in Equation 4 based on the first four assumptions.

Here, the four similarity matrices are similarity matrices regarding the movement trajectory, the color characteristic, the scale space homogeneity of the label, and the temporal homogeneity of the label, respectively.

Here, f is a vector defined by the magnitude (m) and direction (θ) of the optical flow. BD is a Battacharrya distance comparing the difference in image distribution as is well known, SL means label value on scale space, and OL means past object label value. [sigma] is a bandwidth for defining the shape of the Gaussian distribution of each similarity term shown in Equation 4, and can be determined empirically.

Calculate the Laplacian matrix L of the graph thus generated, obtain the eigen vectors of the Laplacian matrix L, and then use the k-means algorithm to convert the eigenvectors of the Laplacian matrix L into k clusters. Can be clustered

For reference, FIG. 6 is a diagram illustrating an area clustering process of an object extraction step in an object detection method according to an exemplary embodiment of the present invention.

Referring to FIG. 6, regions of the current image are transformed into graphs having nodes and edges, and are clustered by cutting the edges of the graph to minimize the energy of the graph based on the similarity matrix.

In step S24, the labeling of the previously clustered areas is optimized, and the object is defined as the same labeled areas.

Once, the area that succeeded in tracking in the preceding image is labeled by the object unit, while the areas that fail in tracking are not labeled.

In the object modeling step of step S24, the object-specific label is optimized for the entire image based on the labels of the previously labeled clustered areas, and at this time, the unlabeled areas are also labeled due to the tracking failure. .

Specifically, the clustered regions represented by the graphs above are optimized for object labeling using a graph cut theory.

In one embodiment of the present invention, object labeling may be optimized by finding an assignment of an object label for a particular region that minimizes the energy function comprised of the energy of the specific region and the energy between the surrounding region.

According to Equation 5, the energy function E is composed of the continuous energy E _smooth according to the relationship between the energy E _data of the specific region and the surrounding regions.

In Equation 6, the energy E _data of a specific region is defined as the consistency of the moving characteristics of the region R _i in a continuous image. The movement characteristic refers to the movement of partial regions given the same label by region clustering, and is defined as an affine transform in successive images. That is, in order to calculate the consistency of the moving characteristics, the current partial region is changed by affine transformation of the region cluster, and the image distribution of the partial region at the moved position is compared with the image distribution of the past partial region. The comparison of the image distribution is obtained by the Batchaya distance.

The similarity H (Affine (R _i )) of the moving characteristics is obtained by obtaining a two-dimensional affine transformation such as RANSAC (Random SAmple Consensus) technique of clusters obtained through the clustering procedure and inverting from the current image to the previous image. Means similarity.

The continuous energy E _smooth may be expressed as in Equation 7 below.

In Equation 7, the correlation between neighboring regions is used. If the histogram distribution with the neighboring regions is similar, the energy becomes small, and if the assigned labels are the same, the energy becomes small.

In this way, object labeling is performed to cut the graph so that energy is minimized with respect to the graph generated based on the regions and the similarity, and regions assigned the same label are defined as one object model.

Now, the background can be modeled based on the models of the extracted object.

7 is a diagram illustrating code words based on object information analyzed in the background modeling step of the object detection method according to an embodiment of the present invention.

The present invention extracts objects from successive input images and aligns the extracted objects to model a background in order to correspond to the PTZ camera environment for panning, tilting and zooming. Thus, the modeled background is not limited to the field of view facing the current camera.

If the label of each object extracted earlier is like one code word, the background is a large codebook composed of such code words.

Shapes in successive images are extracted into objects through segmentation, association tracking, clustering, and labeling, respectively, and FIG. 7 illustrates how the extracted objects are managed as code words.

8 is a diagram illustrating a pseudo code for explaining a background modeling step of an object detecting method according to an embodiment of the present invention.

Referring to FIG. 8, first, the background model BG is initialized, and the number L of objects constituting the background model is also initialized to zero.

Thereafter, the following learning operation may be performed on all objects obtained by analyzing the image through step S11.

Since the background is composed of objects that are almost unchanged, if the extracted object is a background, it will be continuously extracted from the image and recognized as a background. Furthermore, if the extracted object is slightly different from the objects belonging to the background, the background will be slightly different. It can be determined that there has been a change, and if there is little change since the new extraction, a new background can be determined.

On the contrary, if an extracted object is completely different from the objects in the background or changes greatly, it should be excluded from the background. In order to exclude an object from the background, the number of times determined that the object extracted from the continuous images and the object belonging to the background is different.

If there is an object Obj _i extracted from the image, if the object BG _j and the similarity (Dist (Obj _i , BG _j )) included in the background model in the previous step are less than or equal to the predetermined value Th _obj , the object Obj _i As a result, the object BG _j constituting the background model is updated. That is, if the object is a background, the object constituting the background model is continuously updated.

For this object Obj _i , some background BG _j is initialized to zero to keep the background BG _j not updated to remove it from the background model.

If the extracted object Obj _i is the foreground or no longer the background, the distance to any of the object BG _j constituting the current background model will be determined to be greater than or equal to a predetermined value. This object Obj _i is added as a new background component instead of updating the existing background component BG _j (Obj _i → BG _L , L = L + 1), and the receipt increases by 1 for these objects. In the learning mode, a completely new extracted object is temporarily added to the background model once.

If the receipt is above a certain value (Th _age ), this means that the object Obj _i is essentially unlikely to belong to the background or is almost unlikely to be a slightly changed object. do.

9 is a diagram illustrating a result of a background modeling step of an object unit in an object detecting method according to an exemplary embodiment of the present invention.

Referring to FIG. 9, object models of (b) were generated by segmenting, corresponding relationship tracking, clustering, and labeling the input image of (a).

As the object models of (b), the background model of (c) may be constructed through the process of FIG. 8. It can be seen that the rightmost human model of the object models of (b) was excluded from the background model of (c).

10 is a block diagram illustrating an object detecting apparatus according to an embodiment of the present invention.

Referring to FIG. 10, the object detecting apparatus 10 may include a camera 11, an object extractor 12, a background model learner 13, a foreground determiner 14, and a memory 15.

The image photographed by the camera 11 may be provided to the object extractor 12.

The object extractor 12 divides an image into regions subdivided based on color characteristics, tracks correspondences between regions from successive images, moves characteristics, color characteristics, scale spatial homogeneity and temporality of labels. The regions are clustered based on homogeneity and then objects are repeatedly extracted for each image as a set of identically labeled regions by optimizing the labeling for the clustered regions.

The background model learner 13 compares the objects extracted by the object extractor 12 with the background model, and updates an object in the background model with the extracted object if an object similar to the extracted object exists in the background model. Otherwise, the extracted object is added to the background model, and if the object in the background model is not updated for more than a predetermined time as the extracted object, the object will not be called the background as the object does not remain in a constant shape, so the object is added to the background model. Can be excluded.

The learned background model is stored in the memory 15 and managed by update, add and remove operations by the background model learner 13.

The foreground determination unit 14 may compare the object extracted by the object extraction unit 12 with the background model, and determine the object as the foreground if it does not belong to the background model. The foreground determination unit 14 may store the foreground determined object in the memory 15.

The object detecting apparatus according to the embodiments of the present invention may be used for detecting and tracking an object in a PTZ surveillance camera system.

As described above, although the present invention has been described by way of limited embodiments and drawings, the present invention is not limited to the above-described embodiments, which can be variously modified and modified by those skilled in the art to which the present invention pertains. Modifications are possible. Accordingly, the spirit of the invention should be understood only by the claims set forth below, and all equivalent or equivalent modifications will fall within the scope of the invention.

In addition, the apparatus according to the present invention can be embodied as computer readable codes on a computer readable recording medium. A computer-readable recording medium includes all kinds of recording apparatuses in which data that can be read by a computer system is stored. Examples of the recording medium include ROM, RAM, optical disk, magnetic tape, floppy disk, hard disk, nonvolatile memory, and the like, and also include a carrier wave (for example, transmission through the Internet). The computer readable recording medium can also be distributed over network coupled computer systems so that the computer readable code is stored and executed in a distributed fashion.

10 Object detection device 11 Camera
12 Object Extraction Unit 13 Background Model Learning Unit
14 foreground judgment unit 15 memory

Claims (9)

Extracting objects from each of successive input images;
Judging whether any of the objects constituting the background model are similar to each of the extracted objects because a similarity is greater than a predetermined value;
In a learning mode, if there is an object determined to be similar as the result of the determination, updating the original object in the background model with the extracted object; otherwise, adding the extracted object to the background model; And
And removing, from the background model, an object that is not updated for a predetermined time by an object extracted from a new input image among objects updated or added to the background model in a learning mode. The object detection method according to claim 1, further comprising: in the detection mode, determining that the object is a foreground object if none of the objects constituting the background model with respect to the extracted object are similar. The method of claim 1, wherein the extracting of the objects comprises:
Subdividing each of the consecutive input images into a plurality of regions based on a characteristic of a pixel;
Obtaining regions corresponding to each other among the subdivided regions of each of the input images;
Clustering regions having a similarity higher than a predetermined value based on the similarity calculated for the subdivided regions; And
Optimizing labeling of the clustered areas and defining an object with areas of the same label. The method of claim 3, wherein the subdividing the continuous input images into a plurality of regions based on a characteristic of a pixel comprises:
And subdividing the color values of the pixels of the input images using a mean-shift method. The method of claim 3, wherein the obtaining of regions corresponding to each other among the subdivided regions of the input images comprises:
Calculating an optical flow voting between the plurality of pixels in a specific region of the previous image and the plurality of pixels in the region of the current image, and comparing the operation result to a threshold value;
Calculating a battacharrya distance with respect to appearance matching descriptors of a specific region of the previous image and a specific region of the current image, and comparing the calculation result to a threshold value;
Calculating distortion occurring in the representative movement direction of the peripheral regions, and determining that the tracking is successful with respect to the specific region in the continuous image when the calculation is less than or equal to the predetermined value; And
If the area of the current image does not have a correspondence between the areas of the previous image among the areas determined to be unsuccessful in tracking because the distortion calculated with respect to the representative movement direction is greater than a predetermined value, the area of the current image And classifying the into newly appearing areas. The method of claim 3, wherein the clustering of regions having similarity higher than a predetermined value based on the similarity calculated for the subdivided regions includes:
Calculating similarity with respect to each or a combination of moving characteristics, color characteristics, scale spatial homogeneity of the label, and temporal homogeneity of the label; And
And clustering regions having a similarity higher than a predetermined value. Subdividing each of the successive input images into a plurality of regions based on a characteristic of a pixel;
Obtaining regions corresponding to each other among the subdivided regions of the input images;
Clustering regions having a similarity higher than a predetermined value based on the similarity calculated for the subdivided regions; And
Optimizing labeling of the clustered areas and defining an object with areas of the same label. An object extracting unit which extracts objects from each of the input images; And
It includes a background model learning unit for learning a background model,
The background model learning unit,
It is determined whether there is an object determined to be similar because the similarity is greater than a predetermined value among the objects constituting the background model with respect to the extracted objects,
If there is an object determined to be similar as the result of the determination, the original object in the background model is updated with the extracted object; otherwise, the extracted object is added to the background model,
And removing an object from the background model that is not updated for a predetermined time by an object extracted from a new input image among objects updated or added to the background model. The method of claim 8, wherein it is determined whether there is an object determined to be similar among the objects constituting the background model with respect to the extracted objects, and the determination result determines that the object is similar among the objects in the background model with respect to the extracted object. And a foreground determiner which determines that the object is a foreground object if there is no object.

Images