JP2962548B2 - Moving object region division method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a moving object region dividing method and, more particularly, to a moving object region dividing method which divides a moving object region and a background region in an image adaptable to environmental changes such as background and lighting. About the method.

[0002]

2. Description of the Related Art The inventor of the present application is studying a non-contact hand gesture recognition apparatus using image processing as an example of a man-machine interface. The purpose of this device is to acquire information on the posture and shape of a target object (hand) from an input image based on a prepared model. One of the most straightforward methods for this is to assume no variable conditions such as lighting camera parameters, and from raw images input from the camera, based on an elaborate model of the target object prepared in advance, The aim is to find a suitable posture and shape.

[0003]

However, this method has the following problems. First, in order to enable model fitting in various environments, it is necessary to create not only information on shapes and movements but also precise modeling including the reflectance and texture of the target object. However, not only is such a model difficult to create, but versatility is reduced by specializing to a specific object. Furthermore, as the model information used becomes more sophisticated, the amount of calculation required to apply the model increases.

For this reason, it is common practice to actually assume a certain degree of environmental conditions in advance and to simplify the processing based on the prior knowledge. For example, preprocessing such as extraction of a silhouette image using prior knowledge about a luminance value of a target object and a background image is an example thereof. Such processing based on prior knowledge is effective as long as the assumed conditions are satisfied, but is naturally vulnerable to unknown environments or changes in environmental conditions. To overcome this, a method that can be dynamically applied to changes in the environment is needed.
In practice, it is rare to be able to fix environmental conditions, and such processing vulnerabilities have often been a problem.

[0005] Therefore, a main object of the present invention is to provide a moving object region dividing method which can be adapted to environmental changes such as the background and lighting, and which can divide a moving object region and a background region in an image. That is.

[0006]

The invention according to claim 1 is
A moving object region dividing method for dividing a moving object region and a background region in an image photographed by a camera, wherein the moving object region is extracted from a difference between an image of an original frame photographed by the camera and images of preceding and succeeding frames. By obtaining the pixel value distribution for each of the target object and the background based on the step and the rough knowledge about the extracted moving object region and the target, and expressing the obtained pixel value distribution by a beta distribution,
Generating a probability image indicating the probability that the target object exists for each pixel on the newly obtained image.

In the invention according to claim 2, the rough knowledge of claim 1 is the movement and the size of the target object.

[0008]

FIG. 1 is a block diagram of an apparatus for implementing an embodiment of the present invention. In FIG. 1, a camera 1 captures a moving image. In this embodiment, a hand movement is photographed as a moving image. The photographing output of the camera 1 is provided to an image processing device 2, which extracts an area of a hand movement that is a moving object.

FIG. 2 is a flowchart for explaining the operation of one embodiment of the present invention, and FIG. 3 is a diagram for explaining a method of cutting out a part of a moving object in an image by an inter-frame difference. .

Next, the operation of one embodiment of the present invention will be described with reference to FIGS. The image processing device 2 extracts a moving area as shown in FIG. The extraction of the moving area is performed using the difference between consecutive frames as shown in FIG. That is, the original frame t and the previous and next frames t−
A part of the moving area is extracted by binarizing the difference image between 1... Tm, t + 1. Using the image as a mask, information (average, variance, texture) on the moving area is obtained.

Considering an image of N × M pixels, a set of pixel positions is defined as Ω (Ω = ｛(i, j) | 1 <i <N, 1 <d
<M｝). An input image at time t and Y _t (Y _t
= {Y _{l, t} | _l } Ω, y _{l, t} {1,2, ..., L},
Here, L is the number of gradations.

The difference image between Y _i and Y _j is represented by S _{i, j} = ｛S ^l
_{i, j} | l {Ω}, each S ^l _{i, j} is

[0013]

(Equation 1) I can write The mask image M _t = ｛m _{l, t} obtained at this time
| L {Ω} is given by the following equation (2).

[0014]

(Equation 2) Next, region division is performed based on luminance values using rough knowledge of the target object. That is, a pixel value distribution for only the target object at each position on the image and a pixel value distribution for the whole including the background are obtained, and the probability that the target object exists at each position in the input image is calculated based on the distribution. By binarizing the obtained probability image with an appropriate threshold, a silhouette image of the target object can be obtained.

The input image at time t is represented by Y _t = ｛y _{l, t}
| L {Ω}. y _{l, t} indicates the pixel value of the pixel l∈Ω. The average image at time t is represented by / (/ indicates the average) Y
_t = {/ y _{l, t} | _l {Ω}, the root mean square image is calculated as / Y ² _t =
{/ Y ² _{l, t} | _l {Ω}. Here, each pixel value is estimated as follows.

[0016]

(Equation 3) Further, the moving area average image at time t is represented by / A _t = ｛/
a _{l, t} | _l {Ω}, the root-mean-square image is _expressed as / A ² _t = ｛/ a ²
_{l, t} | _l {Ω}. Each pixel value is estimated as follows using the above-mentioned moving area mask image M _t.

[0017]

(Equation 4) Here, n ^{l, t} _a represents the number of pixels in the moving area obtained at the position 1 until the time t.

[0018]

(Equation 5) Actually, since a moving image area can be obtained only for a part of an image, it is not practical to obtain an average value for each pixel. Therefore, updating is performed using an average value in a square area of a × b around each pixel. Here, the actual update processing is adaptively performed based on rough knowledge of the target.

FIG. 2 shows this state. Results of classification by rough knowledge, if it is determined that the entire image is moving, the entire distribution / Y, and updates only / Y ^2. If it is determined that the moving area is the target object, the moving area distribution / A, / A
^2, if the entire distribution / Y, / Y ² are both image updating is stationary does not perform the update process. No. (3),
Equations (4), (6), and (8) are expressions of sequential processing, and these updating processing can be performed sequentially for each frame.

As described above, the average value / y _{l, t} , the root mean square value / y ² _{l, t} at each position l∈Ω at time t, the average value for the moving region / a _{l, t} , the root mean square value / A ² _{l, t} is obtained. Based on these values, the luminance distribution is represented by a beta distribution represented by the following equation.

[0021]

(Equation 6) Since the beta distribution has a finite domain and can cope with a steep distribution, it is considered to be suitable for expressing a pixel distribution. The parameters α and β of the beta distribution are represented by the following equation using the average value / μ and the variance σ ² .

[0022]

(Equation 7) However, σ ² <μ (1−μ) must be satisfied.

When expressed using the root mean square / μ ² instead of σ ^2, the following equation is obtained.

[0024]

(Equation 8) Probability P pixel value of y _l for a pixel l (y _l)
Is given by

[0025]

(Equation 9) Here, y ′ _l is obtained by normalizing the luminance value y _l to be [0, 1]. Similarly, the pixel l probability P (y _l pixel value is y _l as occupied by the target
| T) is expressed by the following equation.

[0026]

(Equation 10) From these, when the pixel value _yl is observed, the probability P (T | _yl ) that the pixel l is occupied by the target can be calculated as follows.

[0027]

[Equation 11] Here, P (T) indicates the probability that the pixel 1 is occupied by the target. Obtained probability image Z = {P (T | yl) | _l }
By binarizing Ω｝ by appropriate threshold processing, a silhouette image of the object can be obtained.

FIGS. 4 and 5 are diagrams for explaining the effect of the embodiment of the present invention. FIG. 4A shows an input image, which is an image of a first frame, a fifth frame, a twentieth frame, and a fifty frame when the hand is moved left and right in a state where the background luminance is different. In such an input image, since the brightness in the image is not uniform, it is not possible to cut out the image of the hand, which is the target area, with a common threshold value.

FIG. 4B shows an average image of all pixels. Since the average image of the first frame is the average of only one frame, it matches the input image. However, since the image of the fifth frame is the average of the images of the five frames, the average image of the moving region (hand) is obtained. Influence is thin, 20th frame, 5th
This tendency is even stronger in the average image of the 0th frame.

FIG. 5A shows an average image of a moving area, and FIG. 5B shows a probability image. As shown in FIG.
As the number of frames increases, the brightness distribution of the moving area average image is updated. FIG. 5B shows a higher probability as the luminance is higher. As described above, the probability of the target region increases with the update of the pixel value distribution, and the region can be easily divided.

[0031]

As described above, according to the present invention, a moving object region is extracted, and a pixel value distribution for each of a target object and a background is obtained based on rough knowledge about the extracted moving object region and the target. By generating a probability image indicating the probability of the existence of the target object for each pixel on the image obtained by expressing the pixel value distribution by the beta distribution, by cutting out the image with a high probability, The body region and the background region can be easily divided.

[Brief description of the drawings]

FIG. 1 is a schematic block diagram of an apparatus for realizing an embodiment of the present invention.

FIG. 2 is a flowchart for explaining the operation of the embodiment of the present invention;

FIG. 3 is a diagram for explaining a method of cutting out a part of a moving object in an image by an inter-frame difference.

FIG. 4 is a diagram for explaining an effect of the embodiment of the present invention, showing an input image and an average image of all pixels.

FIG. 5 is a diagram for explaining an effect of the embodiment of the present invention, showing a moving region average image and a probability image.

[Explanation of symbols]

 1 camera 2 image processing device

Continuation of the front page (56) References Overseas, "Hand region extraction based on pixel value distribution estimated from time-series images", March 24, 1997, Vol. 1997, Pt. 7, p. 361, Overseas, "Estimation of Hand Posture Using Multiple Cameras Using Distance Conversion Values," IEICE Technical Report, 1996, Vol. 82 (MVE 96 15-33), p. 13-18 (58) Field surveyed (Int.Cl. ⁶ , DB name) G06T 7/20 G06T 1/00 INSPEC (DIALOG) JICST file (JOIS)

Claims (2)

(57) [Claims] 1. A moving object region dividing method for dividing a moving object region and a background region in an image photographed by a camera, the method comprising the steps of: Extracting the moving object region, obtaining a pixel value distribution for each of the target object and the background, based on the rough knowledge about the extracted moving object region and the target, and converting the obtained pixel value distribution into a beta distribution. A moving object region dividing method, comprising the step of generating a probability image indicating the probability of the presence of a target object for each pixel on a newly obtained image by expressing. 2. The moving object region dividing method according to claim 1, wherein the rough knowledge is a movement and a size of the target object.