KR101270718B1 - Video processing apparatus and method for detecting fire from video - Google Patents

Abstract

PURPOSE: An image processing device for detecting fire in an image and a method thereof are provided to be applied to real-environmental monitoring by obtaining candidate fire and verifying the candidate fire with a randomness test algorithm. CONSTITUTION: A color probability image generation unit(103) calculates a fire color probability for pixels included in a specific frame of an image and generates a color probability image. A motion probability image generation unit(105) calculates a motion probability for the pixels by using a background image and generates a motion probability image. A candidate fire probability image generation unit(106) generates a candidate fire probability image by using the color probability image and the motion probability image. A candidate fire feature extraction unit(107) calculates a candidate fire feature. [Reference numerals] (101) Smoothing processing unit; (102) Color distribution modeling unit; (103) Color probability image generation unit; (104) Background modeling unit; (105) Motion probability image generation unit; (106) Candidate fire probability image generation unit; (107) Candidate fire feature extraction unit; (108) Candidate fire verification unit; (109) Warning level determination unit

Description

Image processing apparatus and method for detecting fire in an image {Video processing apparatus and method for detecting fire from video}

The present invention relates to an image processing apparatus and method for detecting a fire in an image, and in particular, a candidate fire is obtained by calculating a fire color probability and a motion probability based on modeling a fire color distribution and a background of a specific frame of an image, and randomness. The present invention relates to an image processing apparatus and method for detecting a fire in an image that can be applied to a real environment monitoring outdoors by applying a test algorithm and can accurately detect a fire without depending on hardware performance.

Fire and smoke detection devices are widely used because of the economic and human damages caused by fire and smoke. In the past, many fire detection devices based on sensors have been used. Sensor-based detection devices employ thermal sensors and smoke sensors that provide a warning if the temperature is high enough or the smoke is thick enough. Sensor-based detection devices can be found in almost all buildings, especially office buildings, because they operate well and reliably provide warning in many building environments. But it is often too late to extinguish because a warning message is sent only when the fire is strong enough. In other words, the detection device can only help to reduce the damage of the fire. In addition, sensor-based detection devices are not suitable for outdoor environments such as forest fires.

In order to detect fire early and completely block the damage, rather than simply reduce the damage caused by the fire, a charge-coupled device (CCD) camera-based fire detection device has been developed. CCD camera-based fire detection devices can be applied to outdoor environments and the detected results can be remotely checked over the network using captured images.

Conventional CCD camera-based fire detection devices detect fire colored pixels and motion pixels and convert them into binary images, respectively. A candidate fire is defined by ANDing the binary image of the fire color pixel and the binary image of the motion pixel. In addition, various classifiers are applied to classify each candidate fire pixel as a fire pixel or a non-fire pixel. The classification process depends on the candidate bull pixels detected previously. However, since the fire color includes a large range of colors including white, red, orange, yellow, and the colors in between, the detection of fire color pixels is difficult. In addition, the motion detection algorithm for detecting the motion pixel is difficult to detect the motion pixel because the light blinks at the same position, and the internal area of the fire area is determined to be a static area or a background over time, and thus many internal frame pixels are detected. The loss will occur.

Therefore, in the conventional CCD camera based fire detection apparatus, the performance of the hardware is the performance of the classifier because most classifiers use the flicker characteristic of fire and flicker will have different statistics when the camera captures images at different frame rates. Will affect. Conventional CCD camera-based fire detection device has the disadvantage that it is difficult to accurately detect the fire due to the above problems, and the difference in the accuracy of the fire detection according to the camera quality is large.

An object of the present invention is to calculate the candidate fire by calculating the fire color probability and motion probability based on the model of the fire color distribution and background of a specific frame of the image, and by applying a randomness test algorithm to verify the candidate fire, the real environment of the outdoor The present invention provides an image processing apparatus and method for detecting a fire in an image that can be applied to surveillance and can accurately detect a fire without depending on hardware performance.

An image processing apparatus for detecting fire in an image according to an embodiment of the present invention uses a color Gaussian model generated by modeling a plurality of fire color distributions to determine a fire color probability for a pixel included in a specific frame of the image. A color probability image generation unit for calculating and generating a color probability image corresponding to the Boolean color probability; A motion probability image generation unit configured to calculate a motion probability for the pixels included in the specific frame by using the background image generated by modeling the background of the specific frame, and generate a motion probability image corresponding to the motion probability; And a candidate probability image generator for generating a candidate probability image on which the candidate probability of the specific frame is displayed using the color probability image and the motion probability image.

An image processing apparatus for detecting fire in the image may include: a color distribution modeling unit configured to generate a color Gaussian model by modeling a plurality of fire color distributions; And a background modeling unit configured to model the background of the specific frame to generate the background image.

The color distribution modeling unit may model the fire color distribution in a YCbCr color space.

An image processing apparatus for detecting a fire in the image comprises: an area of the candidate fire, an x coordinate and a y coordinate of the center point of the candidate fire, an average value and a standard deviation of pixels included in the candidate fire of a Cb component of the YCbCr color space; The apparatus may further include a candidate bull feature extracting unit configured to calculate a candidate bull feature including at least one of an average value and a standard deviation of pixels included in the candidate bull of the Cr component of the YCbCr color space.

The candidate fire feature extractor may determine that a fire does not exist in the specific frame when the area of the candidate fire is smaller than a predetermined threshold.

The image processing apparatus for detecting a fire in the image may further include a candidate fire verification unit for verifying a candidate fire feature of the candidate fire in a plurality of consecutive frames of the image by using a randomness test algorithm.

The image processing apparatus for detecting a fire in the image calculates a prior fire probability by averaging randomness probability that is a result of the randomness test algorithm, and calculates a posterior fire probability using the prior fire probability. The warning threshold determination unit may further include determining a fire warning level by setting a warning threshold value using the post-fire probability.

The candidate probability image generating unit may generate a bull mask image by binarizing the candidate probability image.

The image processing apparatus for detecting a fire in the image may further include a smoothing processor configured to smooth the image using a Gaussian filter.

An image processing method for detecting fire in an image according to an embodiment of the present invention uses a color Gaussian model generated by modeling a plurality of fire color distributions to determine a fire color probability for a pixel included in a specific frame of the image. Calculating; Generating a color probability image corresponding to the Boolean color probability; Calculating a motion probability of a pixel included in the specific frame by using the background image generated by modeling the background of the specific frame; Generating a motion probability image corresponding to the motion probability; And generating a candidate probability image in which the candidate probability of the specific frame is displayed using the color probability image and the motion probability image.

An image processing method for detecting fire in the image may include generating a color Gaussian model by modeling a plurality of fire color distributions; And generating the background image by modeling the background of the specific frame.

Generating the Gaussian model may include modeling the fluorine color distribution in a YCbCr color space.

An image processing method for detecting a fire in the image comprises: an area of the candidate fire, an x coordinate and a y coordinate of the center point of the candidate fire, an average value and a standard deviation of pixels included in the candidate fire of a Cb component of the YCbCr color space; The method may further include calculating a candidate bull feature including at least one of an average value and a standard deviation of pixels included in the candidate bull of the Cr component of the YCbCr color space.

The image processing method for detecting a fire in the image may further include determining that no fire exists in the specific frame when the area of the candidate fire is smaller than a predetermined threshold.

The image processing method for detecting fire in the image may further include verifying a candidate fire feature of the candidate fire in a plurality of consecutive frames of the image by using a randomness test algorithm.

An image processing method for detecting a fire in the image comprises: calculating a prior fire probability using a random probability that is a result of the randomness test algorithm; Calculating a post-fire probability using the prior probability; And determining a fire warning level by setting a warning threshold value using the post fire probability.

The image processing method for detecting a fire in the image may further include generating a fire mask image by binarizing the candidate probability image.

The image processing method for detecting a fire in the image may further include smoothing the image using a Gaussian filter.

According to an aspect of the present invention, the candidate fire is calculated by calculating the fire color probability and the motion probability based on modeling the fire color distribution and the background of a specific frame of the image, and applying the randomness test algorithm to verify the candidate fire, It is possible to provide an image processing apparatus and method for detecting a fire in an image that can be applied to a real environment monitoring of an image and can accurately detect a fire without depending on hardware performance.

1 is a view schematically showing an image processing apparatus for detecting a fire in an image according to an embodiment of the present invention.
2A to 2D are diagrams showing an example of a fire color distribution.
3A to 3D are diagrams illustrating an example of a histogram of a Cb / Cr distribution for Gaussian modeling of a color distribution modeling unit of an image processing apparatus for detecting a fire in an image according to an embodiment of the present invention.
4 is a diagram illustrating an example of a color probability image generated by a color probability image generator of an image processing apparatus for detecting a fire in an image according to an exemplary embodiment of the present invention.
5 is a diagram illustrating an example of a motion probability image generated by a motion probability image generator of an image processing apparatus for detecting a fire in an image according to an embodiment of the present invention.
6 is a diagram illustrating an example of a candidate probability image generated by a candidate probability image generator of an image processing apparatus for detecting a fire in an image according to an embodiment of the present invention.
FIG. 7 illustrates an example of a candidate fire feature generated by a candidate fire feature extracting unit of an image processing apparatus for detecting a fire in an image according to an exemplary embodiment of the present invention.
8A to 8G illustrate examples of randomness probabilities generated by a candidate fire verifier of an image processing apparatus for detecting a fire in an image according to an embodiment of the present invention.
9 is a diagram illustrating an example of a warning level determination result of a warning level determining unit of an image processing apparatus for detecting a fire in an image according to an embodiment of the present invention.
10 is a flowchart illustrating an image processing method for detecting a fire in an image according to an embodiment of the present invention.

The present invention will now be described in detail with reference to the accompanying drawings. Hereinafter, a repeated description, a known function that may obscure the gist of the present invention, and a detailed description of the configuration will be omitted. Embodiments of the present invention are provided to more fully describe the present invention to those skilled in the art. Accordingly, the shapes and sizes of the elements in the drawings and the like can be exaggerated for clarity.

1 is a view schematically showing an image processing apparatus for detecting a fire in an image according to an embodiment of the present invention.

Referring to FIG. 1, an image processing apparatus for detecting a fire in an image according to an embodiment of the present invention includes a smoothing processor 101, a color distribution modeling unit 102, and a color probability image generating unit 103. , Background modeler 104, motion probability image generator 105, candidate probability probability image generator 106, candidate probability feature extractor 107, candidate probability verifier 108, and warning level determiner 109 ) May be included. The configuration of the image processing apparatus for detecting a fire in the image shown in FIG. 1 is according to an embodiment, and in the blocks shown in FIG. 1, not all blocks are essential components. Or may be deleted. For example, in another embodiment, the image processing apparatus for detecting a fire in the image may be configured except for the smoothing processing unit 101, the color distribution modeling unit 102, and / or the background modeling unit 104, In another embodiment, it may be configured to exclude candidate candidate feature extraction unit 107, candidate candidate verification unit 108, and / or warning level determination unit 109.

The smoothing processor 101 smoothes an image. In this case, the smoothing processor 101 may smooth the image using a Gaussian filter. This is to reduce the effects of low quality images.

The color distribution modeling unit 102 generates a color Gaussian model by modeling a plurality of fluorine color distributions. That is, the color distribution modeling unit 102 may generate a color Gaussian model by obtaining a plurality of fire color distributions by using a plurality of randomly selected bull images and a plurality of mask images corresponding thereto, and modeling them. Here, the mask image refers to an image obtained by binarizing only pixels of a portion corresponding to fire in the fire image. In one embodiment, the color distribution modeling unit 102 may model the fire color distribution in the YCbCr color space. Below, the color distribution modeling unit 102 will be described a process of generating a color Gaussian model by modeling the fluorine color distribution.

2A to 2D are diagrams showing an example of a fire color distribution. 2A to 2D illustrate graphs illustrated as examples of a bulge color distribution that may be generated by the color distribution modeling unit 102. FIG. 2A shows a three-dimensional view of a fire color distribution with three axes Cb, Cr and Cb / Cr in the YCbCr color space, and FIG. 2B shows a two-dimensional view of a fire color distribution with Cb and Cr axes in the YCbCr color space. 2C shows a two-dimensional view of a fire color distribution around the Cb, Cb / Cr axis in the YCbCr color space, and FIG. 2D shows a fire color around the Cr, Cb / Cr axis in the YCbCr color space. A two-dimensional view of the distribution is shown.

3A to 3D are diagrams illustrating an example of a histogram of a Cb / Cr distribution for Gaussian modeling of a color distribution modeling unit of an image processing apparatus for detecting a fire in an image according to an embodiment of the present invention.

Instead of modeling directly based on the distribution using a polynomial based color model, in one embodiment, Cb using the fluorine color distribution as shown in FIGS. 2A-2D, as shown in FIG. 3A. A histogram of the / Cr distribution can be shown. The histogram of FIG. 3A is processed using a 3-length median filter, which can be expressed as Equation 1 to remove the pick.

Where D _i represents the i-th data of the input data sequence, D _med represents the median filtered result, and D _max and D _min are the maximum and minimum values of the input data sequence, respectively. The filtering result by the 3-length median filter is shown in FIG. 3B.

The histogram of FIG. 3B filtered as described above is processed using a 5-length average filter, which can be expressed by Equation 2.

Through Equation 2, a smooth histogram that can be used for later modeling can be obtained. A histogram smoothed using a 5-length average filter is shown in FIG. 3C.

The histogram of FIG. 3C may be normalized using Equations 3 and 4 and may be used to estimate the color Gaussian model using Equations 5 and 6.

Where T _H +1 is the bin number of the histogram. For example, in one embodiment, T _H +1 is 100. H _{Cb / Cr} (i) is the i th bin of the filtered histogram of the Cb / Cr distribution. N _{Cb / Cr} (i) is the i th value of normalized H _{Cb / Cr} , and μ and σ are the mean and standard deviation of the estimated Gaussian distribution. In FIG. 3D, the graph 301 shows a normalized histogram, and the graph 302 shows a histogram of the color Gaussian model generated by the color distribution modeling unit 102 by the above process.

Returning to FIG. 1, the color probability image generator 103 calculates a fire color probability of a pixel included in a specific frame of the image by using a color Gaussian model and generates a color probability image corresponding to the color probability. do.

In an embodiment, the process of calculating the color probability image and generating the color probability image by the color probability image generator 103 is as follows. After the color Gaussian distribution is modeled in the color distribution modeling unit 102, the violent color probability is calculated by Equation 7.

Where Cb and Cr are the values of the Cb and Cr components of the input pixel, μ and sigma represent the mean and standard deviation of the modeled Gaussian distribution, and if Cr = 0, P (Cb, Cr) = 0.

In an embodiment, equation (7) is applied to find the hue color probability for each pixel, and then to obtain the color probability image to be combined with the motion probability image to obtain a Boolean mask image.

4 is a diagram illustrating an example of a color probability image generated by a color probability image generator of an image processing apparatus for detecting a fire in an image according to an exemplary embodiment of the present invention. In FIG. 4, a color probability image 402 is shown corresponding to the original frame 401. The original frame 401 includes a fire 411, and the color probability image 402 for detecting the fire 411 is the color Gaussian model as described above in the color probability image generator 103 of FIG. 1. It can be derived by calculating the bulge color probability for the pixels included in the original frame 401, and generates a color probability image 402 corresponding to the bulge color probability. Looking at the color probability image 402, it can be seen that there is a portion other than the fire 412 that is not represented in black, and thus some noise is included.

Returning to FIG. 1, the background modeling unit 104 generates a background image by modeling a background of a specific frame of the image. In order to generate the motion probability image in the motion probability image generator 105, a background image modeling the background of the frame should be generated as a premise. In an embodiment, the process of generating the background image by the background modeling unit 104 will be described below. The background image is first updated by Equation 8 to calculate the motion probability.

Where B _t (x, y) and I _t (x, y) represent pixel values at the (x, y) position in the background image of the t frame and the original input image. The background image B ₀ is initialized with the first input image I ₀ . The original and background images used here are both black and white images.

The motion probability image generator 105 calculates a motion probability of a pixel included in a specific frame of the image by using the background image, and generates a motion probability image corresponding to the motion probability.

In an embodiment, the process of calculating the motion probability and generating the motion probability image by the motion probability image generator 105 is as follows. Based on the background image dynamically updated by the background modeling unit 104, the motion probability of each pixel is calculated by using Equations 9 and 10 below.

Here, P _m (x, y) is the motion probability of the pixel at the (x, y) position, M is a motion image calculated by removing the background, and M _max is the maximum value of M. Such a calculation can obtain a Boolean region with a high probability because the pixel value of the region is high.

5 is a diagram illustrating an example of a motion probability image generated by a motion probability image generator of an image processing apparatus for detecting a fire in an image according to an embodiment of the present invention. In FIG. 5, the motion probability image 502 corresponding to the original frame 401 is shown. As in FIG. 4, a fire 411 is included in the original frame 401, and the motion probability image 502 for detecting such a fire 411 is generated by the motion probability image generator 105 of FIG. 1. As described above, the motion probability may be calculated for a pixel included in the original frame 401 using the background image, and the motion probability image 502 corresponding to the motion probability may be derived. As a result, the fire 512 is displayed on the motion probability image 502.

Returning to FIG. 1, the candidate probability probability image generator 106 uses the color probability image generated by the color probability image generator 103 and the motion probability image generated by the motion probability image generator 105 to perform the specific frame. Create a candidate probability probability image in which candidate candidates are indicated. Here, the candidate fire refers to an area that may be detected as fire. In an embodiment, the candidate probability probability image generator 106 may generate a bull mask image by binarizing the generated candidate probability probability image.

In an embodiment, the process of generating the candidate probability image by the candidate probability probability image generating unit 106 will be described below. After the Boolean color probabilities P (Cb, Cr) and the motion probabilities P _m (x, y) are estimated, the candidate probabilities are multiplied as in Equation (11).

Here, the pixel tested at the position (x, y) is I (x, y), and (Cb, Cr) are the Cb and Cr components of I (x, y), respectively. The candidate probability probability image generator 106 may generate a candidate probability probability image corresponding to the candidate probability using the candidate probability.

And in one embodiment, the candidate probability of probability can be binarized through Equation 12 to obtain a Boolean mask image.

Where T _cf is a heuristic threshold and a value of 0.3 is used in this embodiment. Pixels with a value of 1 in a boolean mask image represent candidate Boolean pixels, and pixels with a value of 0 represent a non-bull pixel.

6 is a diagram illustrating an example of a candidate probability image generated by a candidate probability image generator of an image processing apparatus for detecting a fire in an image according to an embodiment of the present invention. First, the candidate probability probability image generator 106 of FIG. 1 generates a candidate probability probability image 601 displaying candidate candidates 611 and 612 using the color probability image 402 and the motion probability image 502. . The candidate probability image generating unit 106 may generate the bull mask image 602 by binarizing the candidate probability probability image 601.

Returning to FIG. 1, the candidate bull feature extractor 107 calculates a candidate bull feature that the candidate bull verifier 108 uses to verify the candidate bull. In one embodiment, the candidate bull feature is the candidate bull area, the x and y coordinates of the center of the candidate bull, the mean and standard deviation of the pixels contained in the candidate bull of the Cb component of the YCbCr color space, and the candidate of the Cr component of the YCbCr color space. It may include one or more of the mean and standard deviation of the pixels included in the fire.

The area of the candidate fire may be represented by Equation 13, and the x and y coordinates of the center point of the candidate fire may be represented by Equations 14 and 15.

Where F _c (x, y) represents the pixel at position (x, y) in the bull mask image F _c , X _t and Y _t are the x and y coordinates of the center point of the candidate bull for the t th frame, respectively, A _t is the area of candidate fire. The color features including the mean and standard deviation of the pixels in the candidate fire of the Cb component of the YCbCr color space and the mean and standard deviation of the pixels in the candidate fire of the Cr component of the YCbCr color space are shown in Equations 16-19. Is extracted through.

Where μ _{cb, t} and σ _{cb, t} are mean values and standard deviations of candidate bull pixels of the Cb component of the YCbCr color space, and μ _{cr, t} and σ _{cr, t} are mean values of candidate bull pixels of the Cr component and The standard deviation is Cb (x, y) and Cr (x, y) are the pixel values at the (x, y) positions of the Cb and Cr components.

는 다음의 수학식 20과 같이 정의된다.Candidate fire feature vector for t-th frame

Is defined by the following equation (20).

The candidate fire feature extractor 107 may determine that there is no fire in the specific frame when the area of the candidate fire is smaller than a predetermined threshold. In real-world applications with low quality CCTV cameras, random noise exists that can interfere with detection and give false warnings. Thus, as shown in Equation 21, the threshold value corresponding to the minimum candidate region is determined by random noise. Used to reduce the impact. This is because in the absence of fire, the area of noise is very small.

=[0,0,0,0,0,0,0]^T이다.The threshold value T _A is set to 9 in this embodiment. If Equation 21 is not satisfied, this means that no candidate boolean exists in the t frame. In other words,

= [0,0,0,0,0,0,0] ^T

FIG. 7 illustrates an example of a candidate fire feature generated by a candidate fire feature extracting unit of an image processing apparatus for detecting a fire in an image according to an exemplary embodiment of the present invention. The candidate Boolean feature extractor 107 of FIG. 1 may extract features as shown in FIG. 7 from the original frame 401.

Each drawing shows the area included in the candidate fire area 702, the x coordinate 703 of the center point of the candidate fire, the y coordinate 704 of the center point of the candidate fire, and the pixels included in the candidate fire of the Cb component of the YCbCr color space. The mean value of 705, the standard deviation 706 of the pixels contained in the candidate fire of the Cb component of the YCbCr color space, the mean value of the pixels contained in the candidate fire of the Cr component of the YCbCr color space, 707 and the Cr of the YCbCr color space. Standard deviation 708 of the pixels contained in the candidate candidates of the component.

Returning to FIG. 1, the candidate bull verification unit 108 verifies candidate bull features of candidate bulls in a plurality of consecutive frames of an image using a randomness test algorithm. In an embodiment, the process of verifying candidate candidates by the candidate bull verification unit 108 is as follows. First, in N consecutive frames, the N feature vectors are binarized using Equations 22 and 23.

는 N개의 특징 벡터의 j번째 특징 벡터의 i번째 특징을 나타낸다. μ_i는 i번째 특징의 평균값이고,

는 이진화된 값을 가지는 j번째 특징의 i번째 특징을 나타낸다. 수학식 22는 시간 영역에서 각각 7개의 특징에 대하여 산출된 7개의 평균값에 i=1,..., 7로 7번 적용된다. 그리고 수학식 23은 특징 벡터 시퀀스의 모든 특징을 이진화하기 위하여 7*N 번 적용된다. 결국,

는 수학식 24와 같은 7*N 행렬로 도출된다.here,

Denotes the i th feature of the j th feature vector of the N feature vectors. μ _i is the average value of the i th feature,

Denotes the i th feature of the j th feature with the binarized value. Equation 22 is applied seven times, i = 1, ..., 7 to seven average values calculated for seven features in the time domain, respectively. Equation 23 is applied 7 * N times to binarize all the features of the feature vector sequence. finally,

Is derived from a 7 * N matrix such as

로 이진화 되고 나면, 후보 불 검증부(108)는 영상의 복수의 연속된 프레임 내의 후보 불의 후보 불 특징을 무작위성 테스트 알고리즘을 이용하여 검증할 수 있다. 후보 불 검증부(108)는 무작위성 테스트 알고리즘으로서 월드월포위츠(Wald-Wolfowitz) 테스트 알고리즘을 이용할 수 있다. 연속적인 불의 깜박임은 불의 면적, 중심점 및 색상이 무작위로 변하도록 만들기 때문에, 월드 월포위츠 테스트 알고리즘이 N개의 연속 프레임의 특징 벡터에 기반하여 후보 불 픽셀의 무작위성 확률을 얻기 위해 적용될 수 있다.Feature vector

After binarization, the candidate bull verification unit 108 may verify candidate bull features of candidate bulls in a plurality of consecutive frames of an image using a randomness test algorithm. The candidate bull verification unit 108 may use a Wald-Wolfowitz test algorithm as the randomness test algorithm. Since the continuous fire flicker causes the area, center point, and color of the fire to change randomly, the World Wallowitz Test Algorithm can be applied to obtain a random probability of the candidate fire pixel based on the feature vectors of N consecutive frames.

는 무작위성을 테스트하기 위하여 수학식 25 내지 수학식 27과 같이 월드 월포위츠 테스트 알고리즘의 입력으로 사용될 수 있다.At this time, binarized

May be used as an input to the World Wallowitz Test Algorithm as shown in Equations 25 to 27 to test randomness.

은 0 및 1 간에 바뀐 횟수이며,

은 i번째 특징의 무작위성 테스트의 결과이다. 다시 말하면, 무작위성 테스트의 결과는 수학식 28과 같이 나타날 수 있다.Where n _i0 and n _i1 are the numbers 0 and 1 of the i th feature,

Is the number of changes between 0 and 1,

Is the result of the randomness test of the i th feature. In other words, the result of the randomness test may be represented by Equation 28.

는 수학식 29 및 수학식 30을 이용하여 산출되는 무작위성 확률

를 제공하기 위한 문턱값으로 설정된다.Once the randomness test is complete, the result

Is a randomness probability calculated using Equations 29 and 30

It is set to a threshold for providing.

8A to 8G illustrate examples of randomness probabilities generated by a candidate fire verifier of an image processing apparatus for detecting a fire in an image according to an embodiment of the present invention. The randomness probability of the feature vector in N consecutive frames is shown as in FIGS. 8A to 8F, which are randomness test results corresponding to the features shown in FIG. 7.

8A is a randomness test result 802 corresponding to the area 702 of candidate fires extracted for N consecutive frames, and FIG. 8B is a randomness test result 803 corresponding to the x coordinate 703 of the center point of the candidate fire. 8C is a randomness test result 804 corresponding to the y coordinate 704 of the center point of the candidate fire.

8D is a randomness test result 805 corresponding to an average value 705 of pixels included in candidate lights of the Cb component of the YCbCr color space, and FIG. 8E is a view of pixels included in the candidate lights of the Cb component of the YCbCr color space. Randomness test result 806 corresponding to standard deviation 706.

8F is a randomness test result 808 corresponding to an average value 707 of pixels included in candidate lights of the Cr component of the YCbCr color space, and FIG. 8G is a pixel included in the candidate lights of the Cr component of the YCbCr color space. Is a randomness test result 809 corresponding to the standard deviation 708.

1, the warning level determiner 109 calculates a prior probability by averaging the randomness probability that is the result of the randomness test algorithm, and calculates a posterior probability using the prior probability. The fire alarm level is determined by setting a warning threshold value using the post fire probability. In an embodiment, the process of determining the fire warning level by the warning level determining unit 109 is as follows.

First, the prior probability is calculated using the average of the random probability vector as shown in Equation 31.

는 현재 프레임에 대한 사전 불 확률을 나타낸다. 사전 불 확률을 산출한 후에, 시스템 신뢰성을 향상시키기 위해서 사전 불 확률에 콘볼루션(convolution)이 적용되어 시간 영역에서의 사후 불 확률(posterior)이 산출되고, N개의 연속 프레임에서의 사후 불 확률은 아래의 수학식 32 및 수학식 33에 의해 산출된다.here,

Denotes a prior fire probability for the current frame. After calculating the prior probability, convolution is applied to the prior probability to improve the system reliability, and the posterior probability in the time domain is calculated, and the posterior probability in N consecutive frames is calculated. It is calculated by the following equations (32) and (33).

는 N개의 연속 프레임에서 i번째 사전 불 확률이며,

는 몇 가지 긴급 레벨 경고 정보를 제공하기 위하여 이후에 문턱값이 설정될 수 있는 사후 불 확률을 나타낸다. 사후 불 확률은 현재 프레임에 불이 존재하는지 여부를 나타낸다.Where K (x) is the kernel function used to calculate the convolution,

Is the i-th advance fire probability in N consecutive frames,

Denotes a post-fire probability that a threshold can be set later to provide some emergency level warning information. Post-fire probability indicates whether or not fire exists in the current frame.

9 is a diagram illustrating an example of a warning level determination result of a warning level determining unit of an image processing apparatus for detecting a fire in an image according to an embodiment of the present invention. 9 shows a four-level warning level result 903 in which a pre-failure probability 901, a post-failure probability 902, and a threshold are set. The post-failure probability calculated by the above process may be set to a 4-level warning level in order to provide a 4-level warning message.

10 is a flowchart illustrating an image processing method for detecting a fire in an image according to an embodiment of the present invention.

Referring to FIG. 10, when an image processing method for detecting a fire in an image according to an embodiment of the present invention starts, first, the image is smoothed (S1001). In this case, the image may be smoothed using a Gaussian filter.

After smoothing the image, a color Gaussian model is generated by modeling a plurality of fluorine color distributions (S1002). In step S1002, the fire color distribution may be modeled in the YCbCr color space. When the color Gaussian model is generated, a fluorescence probability is calculated for pixels included in a specific frame of the image using the generated color Gaussian model (S1003). Next, a color probability image corresponding to the violent color probability is generated (S1004).

The background image is generated by modeling the background of the specific frame (S1005), and the motion probability is calculated for the pixels included in the specific frame using the background image generated by modeling the background of the specific frame (S1006). . Next, a motion probability image corresponding to the motion probability is generated (S1007).

After the color probability image and the motion probability image are generated, a candidate probability probability image displaying the candidate probability of the specific frame is generated using the color probability image and the motion probability image (S1008). In operation S1008, the candidate probability image may be binarized to generate a bull mask image.

When a candidate probability probability image is generated, the mean and standard deviation of the pixels included in the candidate bull of the area of the candidate fire, the x and y coordinates of the center point of the candidate fire, and the Cb component of the YCbCr color space and the YCbCr color space A candidate bull feature including at least one of an average value and a standard deviation of pixels included in the candidate bull of the Cr component is calculated (S1009). In operation S1009, when the area of the candidate fire is smaller than a predetermined threshold, it may be determined that no fire exists in the specific frame.

Then, the candidate bull feature of the candidate bull in the plurality of consecutive frames of the image is verified using a randomness test algorithm (S1010).

In addition, a prior probability is calculated using a random probability that is a result of the randomness test algorithm (S1011), and a posterior probability is calculated using the prior probability (S1012). The fire warning level is determined by setting a warning threshold using the probability (S1013).

Image processing method for detecting a fire in the image according to an embodiment of the present invention is similar to the image processing apparatus for detecting a fire in the image according to an embodiment of the present invention shown in FIG. Therefore, unless otherwise stated, the description in FIG. 1 is applied as it is, so detailed description thereof will be omitted. In FIG. 10, as in FIG. 1, not all steps are essential steps in the flowchart illustrated in FIG. 10, and in some embodiments, some steps may be added, changed, or deleted.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is clearly understood that the same is by way of illustration and example only and is not to be taken as limitations. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims (18)

Generate a color probability image using a color Gaussian model generated by modeling a plurality of color distributions to calculate a color probability of a pixel included in a specific frame of an image and generate a color probability image corresponding to the color probability of the color. part;
A motion probability image generation unit configured to calculate a motion probability for the pixels included in the specific frame by using the background image generated by modeling the background of the specific frame, and generate a motion probability image corresponding to the motion probability;
A candidate probability image generating unit generating a candidate probability probability image on which the candidate probability of the specific frame is displayed using the color probability image and the motion probability image; And
The pixel included in the candidate fire of the area of the candidate fire, the x- and y-coordinates of the center of the candidate fire, the mean and standard deviation of the pixels included in the candidate fire of the Cb component of the YCbCr color space, and the Cr component of the YCbCr color space. And a candidate fire feature extracting unit for calculating a candidate fire feature including at least one of an average value and a standard deviation of. The method according to claim 1,
A color distribution modeling unit configured to generate a color Gaussian model by modeling a plurality of fluorine color distributions; And
And a background modeling unit for generating the background image by modeling the background of the specific frame. The method according to claim 2,
The color distribution modeling unit,
And a fire color distribution modeling the fire color distribution in a YCbCr color space. delete The method according to claim 1,
The candidate fire feature extraction unit,
And determining that there is no fire in the specific frame when the area of the candidate fire is smaller than a predetermined threshold value. The method according to claim 1,
And a candidate bull verification unit for verifying a candidate bull feature of the candidate bull in a plurality of consecutive frames of the image by using a randomness test algorithm. The method of claim 6,
A prior probability is calculated by averaging random probability that is a result of the randomness test algorithm, a posterior probability is calculated using the prior probability, and a warning threshold is calculated using the posterior probability. And a warning level determining unit which determines a fire warning level by setting the fire warning level. The method according to claim 1,
The candidate probabilities image generator,
And a fire mask image is generated by binarizing the candidate probabilities image. The method according to claim 1,
And a smoothing processing unit for smoothing the image. Calculating fluorine color probabilities for pixels included in a specific frame of an image using a color gaussian model generated by modeling a plurality of fluorine color distributions;
Generating a color probability image corresponding to the Boolean color probability;
Calculating a motion probability of a pixel included in the specific frame by using the background image generated by modeling the background of the specific frame;
Generating a motion probability image corresponding to the motion probability;
Generating a candidate probability image in which a candidate probability of the specific frame is displayed using the color probability image and the motion probability image; And
The pixel included in the candidate fire of the area of the candidate fire, the x- and y-coordinates of the center of the candidate fire, the mean and standard deviation of the pixels included in the candidate fire of the Cb component of the YCbCr color space, and the Cr component of the YCbCr color space. And calculating a candidate fire feature comprising at least one of an average value and a standard deviation of the fire. The method of claim 10,
Modeling a plurality of fluorine color distributions to generate the color Gaussian model; And
And generating the background image by modeling the background of the specific frame. The method of claim 11,
Generating the color Gaussian model,
And modeling the fire color distribution in a YCbCr color space. delete The method of claim 10,
And determining that there is no fire in the particular frame when the area of the candidate fire is smaller than a predetermined threshold. The method of claim 10,
Verifying a candidate fire feature of the candidate fire within a plurality of consecutive frames of the image using a randomness test algorithm. The method according to claim 15,
Calculating a prior probabilities using a random probability that is a result of the randomness test algorithm;
Calculating a post-fire probability using the prior probability; And
And determining a fire alarm level by setting a warning threshold value using the post-fire probability. The method of claim 10,
And binarizing the candidate probabilities image to generate a fire mask image. The method of claim 10,
The image processing method for detecting a fire in the image further comprising the step of smoothing the image.

Images