KR100918436B1 - Fire detection system and method basedon visual data - Google Patents

Abstract

The present invention relates to a vision-based fire detection system and method for detecting a fire zone from the image information.

The fire detection system includes: a motion detection module for initializing a gray image of a previous frame to a background and updating and outputting a previous background frame to reflect the current frame; A color modeling module configured to detect a first fire candidate region by applying a color model to a frame output by the motion detection module; A luminance map processing module configured to apply a luminance map to the current frame to detect a second fire candidate region brighter than a surrounding; And a SVM processing module for extracting a final fire detection region for the first and second fire candidate regions according to a variation model of temporal wavelet transform energy using a support vector machine (SVM).

Fire Detection, Imaging, Color Modeling, Motion Detection, Luminance Map, SVM (Support Vector Machine)

Description

Vision-based fire detection system and method {FIRE DETECTION SYSTEM AND METHOD BASEDON VISUAL DATA}

The present invention relates to a fire detection technology, and more particularly to a vision-based fire detection system and method for detecting a fire zone from the image information.

The size and complexity of buildings increase the convenience of real life, while causing enormous property and human damage in safety accidents such as fire. Accordingly, the necessity of establishing a fire alarm system is emerging.

Early fire suppression is the most important, but current fire alarm systems are mostly sensor-based detectors, which can be detected only after a certain time after the fire has spread heat or smoke.

As a result, early evolution is difficult and its performance is degraded in large spaces and open spaces. On the other hand, the following advantages exist in the fire detection system using a camera.

First, a fire detection system using a camera can be implemented by using a surveillance camera installed inside and outside the building, and thus additional costs are low.

Second, in the event of a fire or smoke, the camera can immediately detect a remote location without waiting for heat or smoke to spread.

Third, in the event of a fire alarm, the worker can see immediately at a remote location without having to go directly to the place, so immediate response is possible.

Despite the advantages of the fire detection system using the camera described above until now, the research is still insufficient.

For example, the fire monitoring technique presented by W. Phillips extracts fire zones from an experimental data image and uses a Gaussian distribution of the color histograms of the extracted zones to create a color look-up table based on fire color. . The fire candidate region is extracted by using the color lookup table, the fire is discriminated by the amount of temporal change of the region, and the fire is finally detected by using an additional erosion technique and an area enlargement technique. However, due to the heuristic characteristics of the experimental data, there is a disadvantage in that the performance decreases when the input image is different.

The fire monitoring technique presented by C. B. Liu created a color model focusing on the shape of the fire among the characteristics of the fire. In the color image, the fire zone is in the form of a core, the inner central circle of which is white or yellow and orange or red when away from the central circle. In gray, the central circle is brighter than the nucleus periphery. According to this characteristic, if the color area that shows high contrast or the change of white or yellow to orange or red color is overlaid in the form of ring, it is regarded as the form of fire. Fires are detected by using these areas as a fire model and analyzing the external changes in the areas. However, it is difficult to distinguish an object having a color similar to a real fire zone, and thus accurate fire detection performance is deteriorated.

The fire monitoring technique presented by TH Chen is generally red in color, and is characterized by the RGB / HSI color model of the fire zone, depending on the nature of the flame from red to yellow at low temperatures and white at high temperatures. Detect and distinguish fires by analyzing the change in size for the detected flame zone, assuming that the fire zone is gradually expanded. However, because the difference between two adjacent frames detects a change in the size of the flame region, there is a lot of small noise or errors about moving objects.

The fire monitoring technique proposed by B. U. Toreyin distinguishes between fire pixels or fire zones by comparing them with predefined color models in moving pixels or regions, and discriminates fires using temporal / spatial wavelet analysis of the zones identified as fire zones. In order to detect the motion area, the motion area and the background are divided using the difference between the current frame and the previous frame, and the motion area is compared with the fire color pixels obtained according to experimental results on a pixel-by-pixel basis. Fire color pixels were generated using a Gaussian Mixture Model from the RGB color model of the fire zone in the experimental data image and waveleted to detect more precise temporal changes in the detected fire candidate area. We proposed a method of detecting fire by analyzing the variation of high frequency components using transformation. This method has a good fire detection capability, but also uses too many heuristic thresholds, which reduces the ability to distinguish fires from various images.

The present invention has been made to solve the above problems, by applying a fire color model by marking the movement of the image information input from the image pickup device, and analyzes the amount of energy change through the luminance map and wavelet transform and using the SVM It is an object of the present invention to provide a vision-based fire detection system and method for effectively discriminating fire zones from non-fire zones.

A fire detection system for achieving the above object and to solve the problems of the prior art, the motion detection module for initializing the gray image of the previous frame to the background after updating the previous background frame to reflect the current frame; A color modeling module configured to detect a first fire candidate region by applying a color model to a frame output by the motion detection module; A luminance map processing module configured to apply a luminance map to the current frame to detect a second fire candidate region brighter than a surrounding; And a SVM processing module for extracting a final fire detection region for the first and second fire candidate regions according to a variation model of temporal wavelet transform energy using a support vector machine (SVM).

In particular, the motion detection module is configured to assign a motion pixel value to a corresponding frame when a difference between gray image values of a previous frame with respect to a current frame is greater than a threshold.

The color modeling module may be configured to determine the first fire candidate region if each of the R channel value and the saturation value belong to a predetermined threshold range for each pixel of the image information.

In addition, the threshold range for the R channel value and the threshold range for saturation are determined based on the color model for the flame.

The luminance map processing module may include a gray image converter configured to convert input image information into gray image information; A down sampling module for down sampling the gray image information; A plurality of filters for filtering the down-sampled image information to different sizes; An average calculator configured to calculate an average of output values of the plurality of filters; A restoring unit generating a luminance map by restoring an output of the average calculating unit to an original size; A Gaussian filter for Gaussian filtering the luminance map; A change amount detector configured to analyze a temporal change amount of the Gaussian filtered luminance map; The non-fire region removal unit may be configured to determine a second fire candidate region according to the change amount and to remove the remaining region.

In addition, a morphological opening operation is performed on the result of the SVM processing module to remove a noise region for color and a change amount and finally to discriminate fire.

The present invention detects a fire candidate region by temporal variation of color information, motion information, and luminance map with respect to an image input using an imaging device, and finally discriminates a fire by using an SVM. There are advantages in overcoming the shortcomings of existing vision-based fire detection methods in which a number of problems exist, such as the occurrence of detection, and obtaining more accurate results.

In addition, the present invention analyzes the temporal variation of fire candidate pixels in the luminance map in successive frames as well as analyzing simple color information and the amount of change in motion to detect the fire region, and then effectively removes the non-fire region. We apply the SVM, which is a binary class classifier, to learn the fire / non-fire model and finally detect the fire zone using the learned data.

Thus, the present invention has a robust effect on noise caused by smoke or subtle differences between frames.

The configuration of a fire detection system according to a preferred embodiment of the present invention will be described in detail with reference to FIG. 1.

The fire detection system initializes the gray image of the previous frame to the background and then updates the previous background frame by reflecting the current frame and outputs the motion detection module 100 and applies a color model to the frame output by the motion detection module. A color modeling module 102 for detecting a first fire candidate region, a luminescence map processing module 104 for detecting a second fire candidate region brighter than a surrounding by applying a luminance map to the current frame, and the first And an SVM processing module 106 for extracting the final fire detection region according to the variation model of temporal wavelet transform energy using a support vector machine (SVM) for the second fire candidate region.

<Motion detection module 100>

An operation of the motion detection module 100 of the fire detection system will be described.

The motion detection module 100 detects motion pixels using a background initialization and update method. The motion detection module 100 initializes the gray image of the first frame to the background according to the characteristics of a flame whose shape changes from time to time according to air flow, material, or fuel. After that, the previous frame is updated to reflect the current frame.

는 현재 프레임에서 x위치에 대한 그레이 영상이고,

는 현재 프레임에서 배경을 의미한다. 상기 수학식 1의 배경과 현재 프레임의 차분이 임계치 값

의 값보다 크다면 해당 픽셀을 움직임 픽셀 값이 된다. 상기 전경 혹은 배경 영역에 포함된 각 픽셀 값은 수학식 2에 의해서 업데이트되고 임계값은 수학식 3에 의해서 업데이트된다.here,

Is the gray image for x position in the current frame,

Means background in the current frame. The difference between the background and the current frame of Equation 1 is a threshold value

If the value is greater than, the pixel becomes a moving pixel value. Each pixel value included in the foreground or background region is updated by Equation 2, and the threshold is updated by Equation 3.

Here, α, β, and γ are weight values and represent the degree of reflection of the current frame and the previous background frame.

FIG. 2 is a result screen of detecting a motion pixel according to the present invention, in which movement regions of the input fire image (a) and the non-fire image (b) appear in red in (c) and (d), respectively. You can check it.

<Color Modeling Module 102>

The operation of the color modeling module 102 of the fire detection system will be described.

The color modeling module 102 applies Celik's algorithm to construct a probabilistic model for an RGB channel and to detect fire candidate pixels therefrom. Assuming that the distributions of the respective RGB color channels are independent of each other and show Gaussian distribution, a fire color probability model is generated using Equation 4 after extracting fire color pixels from fire images in 100 training fire frames.

의 i채널 확률 값을 수학식 5를 이용하여 획득한다. In Equation 4, μ denotes an average of i-color channels obtained from the training data, and σ denotes a variance of the i-color channel. New input pixels from the probability model Pi of each channel

An i-channel probability value of is obtained by using Equation 5.

In Equation 5, τ denotes a threshold value, and the probability value of the entire color channel is calculated using Equation 5 of each color channel, and when the value is greater than the threshold value τ, the pixel is a fire color. Pixel.

A preferred embodiment of the fire color probability graph assumed to be Gaussian distribution according to the present invention will be described with reference to the accompanying drawings.

3 is a graph of a fire color model for detecting a fire candidate pixel using fire color information according to the present invention, each of which is a Gaussian probability graph for an RGB channel having a value from 0 to 255 obtained from training data. Is shown.

4 illustrates a screen of a result of detecting a fire candidate pixel by using motion pixel detection and color information according to the present invention. FIG. 4 shows fire images (a), (b) and non-fire images (c), (d). As a result, it can be seen that the areas detected as fire candidate pixels are shown as red areas. (A) and (b) it can be seen that the fire area is detected correctly, but some areas included in the moving object in (c) and (d) are incorrectly detected as the fire area.

<Luminance map processing module 104>

The operation of the luminance map processing module 104 of the fire detection system will be described.

The luminance map processing module 104 is used to remove non-fire pixels using temporal luminance variation. The luminance map processing module 104 generates and uses a luminance map from a color image by using a fact that a fire region is brighter than a periphery in a gray image. The fire zone on the luminance map is prominent and the non-fire zone can be effectively removed by using the change amount of the luminance map within several frames for the fire zone having a large change in time.

Referring to the accompanying drawings, a preferred embodiment of a method for detecting a fire area using a temporal change amount in a luminance map according to the present invention will be described below.

5 shows a detailed configuration of the luminance map processing module 104.

The luminance map generator includes a gray image converter 200, a down sampling module 202, a first filter 204, a second filter 206, an average calculation unit 208, a reconstruction unit 210, and a Gaussian filter ( 212), the change amount detector 214, and the non-fire region removal unit 216.

The luminance map processing module 104 converts the original color image to a gray image to generate the luminance map L, downsamples the size to 1/2, and then filters s∈ {11 × 11) having different sizes. , 13 × 13}.

Thereafter, after generating the feature map by taking the average value of the different luminance maps obtained by applying the 11 × 11 and 13 × 13 filters, the image is restored to its original size and the Gaussian filter is applied to the image to remove noise. To obtain the final luminance map.

This process can be represented by Equation 6.

는 1/2의 크기로 다운 샘플링된 회색도 영상에 대해 11×11과 13×13 필터를 적용한 서로 다른 루미넌스 맵이고, 이들의 평균값이 특징 맵

이 된다. 상기

은 다시 원래의 크기로 복원되고, 이 영상에 대해 가우시안 필터를 적용해서 노이즈를 제거하고 루미넌스 맵을 생성한다. 상기 최종 루미넌스 맵은 화재 영역이 주변에 비해 두드러지는 형태를 띄는 루미넌스 영상을 얻어낼 수 있다. In Equation 6

Are different luminance maps with 11 × 11 and 13 × 13 filters for grayscale images downsampled to a size of 1/2, and their mean values are feature maps.

Becomes remind

Is restored to its original size, and a Gaussian filter is applied to the image to remove noise and generate a luminance map. The final luminance map may obtain a luminance image in which the fire region is prominent compared to the surroundings.

In order to analyze the temporal variation of the final luminance map generated as described above, Equation 7 is used to measure the variation of the luminance map with respect to the pixels of x and y positions in 10 consecutive frames.

At this time, the luminance variation of the pixels included in the fire zone reflects the characteristics of the flame and shows a high variation. Accordingly, if the amount of change of the pixel at the x and y positions in the luminance map is smaller than the threshold L _T , the pixel is removed from the fire candidate region from the previous step.

In Equation 8, H _{x , y, L} ( u ) represents a luminance histogram of x , y pixels within 10 consecutive frames.

6 is a diagram illustrating an input image, a motion region detection, a fire color model application, a luminance map generation result, and a fire detection result for the image.

Referring to FIG. 6, the motion region detection result of the input image (a) is shown in (b), the result of applying the fire color model is shown in (c), and the luminance map is shown in (d). It can be seen from the luminance map that the fire zone is brighter than the surroundings. And (e) shows the result of detecting the fire by using the color information, the motion information and the luminance variation. You can see that most of the moving objects and the non-fire areas have been removed, but the area still in the light remains.

<SVM Processing Module 106>

The configuration and operation of the SVM processing module 106 of the fire detection system will now be described.

The present invention uses a temporal variation of the pixels in the luminance map to remove the non-fire region and then to refine the classifier for the pixels remaining in the fire candidate region instead of the heuristic threshold values used in previous papers. Finally, the SVM is used to discriminate fire pixels.

In order to learn the SVM, wavelet transform of fire image and non-fire image is performed and wavelet energy is calculated. The wavelet energy is expressed by Equation 9 by modifying the equation of T.

In Equation 9, e _n ( x , y ) is the wavelet energy of the horizontal, vertical and diagonal high frequency components of the nth frame and extracts 150 energy from fire pixels and 150 energy from non-fire pixels. . Since only the fixed energy extracted from one frame cannot be used to distinguish fire and non-fire pixels, it is judged as energy in several consecutive frames.

7 and 8 are graphs of temporal wavelet energy changes for fire zones and non-fire zones.

Referring to FIGS. 7 and 8, when the pixels at positions x and y are fire pixels in Equation 9, the amount of change in wavelet energy is large, and in the case of non-fire pixels, the amount of change is little or relatively low. Also, if the object is not a fire, there will be only one or two changes, and there will be little change in the overall frame. The present invention extracts the wavelet energy for the pixel at the same position in 10 frames and equalizes it to have a value of 0 to 1 because the variation of the wavelet energy is large. The temporal wavelet energy change in the fire pixel appears in the form of a similar pattern with a wide range of fluctuations, but the non-fire pixels show different patterns such as little change or a large change in energy over a short period of time due to passing objects. Appears in the form of. In order to train these SVMs, they appear in a similar pattern, so the energy values are arranged in ascending order.

9 is a graph arranged in ascending order after leveling wavelet energies of fire and non-fire pixels.

 The present invention extracted feature vectors from ten dimensions. In order to map input vectors in high-dimensional feature space, we used the RBF (Radial Basis Function) kernel function, which is superior to other methods.

In Equation 10, x is an input vector, y is an input pattern, and the standard deviation σ is 3.5, and the exchange parameter C between the largest margin value and the minimum value of training error is set to 5. The standard deviation σ represents the width of the base function. When the fire pixel detection is finally made using the SVM, the morphology opening operation removes the noise region for the color and the amount of change and finally discriminates the fire.

10 is a table comparing the results for the detection of fire according to the present invention with the results of experiments using the algorithm of T, which is the most recent research. Method 1 is the fire detection result according to the present invention and Method 2 is the fire according to the algorithm of T. The detection result.

Referring to FIG. 10, the average false detection rate (False Positive) is 9.2% when using Method 1 and 13.7% when using Method 2, but the present invention is higher than when the algorithm of T is used, but the average false detection rate is Method 1 and Method. In case of using 2, it is rather low, 12.9% and 6.8%, respectively. However, the average detection rate is 84.2% for Method 1 and 79.5% for Method 2, which shows overall good results. In the case of Movie 1, both methods have a low average detection rate, because Movie 1 has a low average bit rate, sunlight is reflected on the branches around the fire, and has a similar color. Since the fire area in Movie 3 and Movie 4 is very small compared to the surroundings, it was often removed when noise was removed using the morphology open operation. In case of detecting a fire using Method 2, many false detections occur in Movie 3, Movie 5 and Movie 7, which is similar to the fire color and movement of a person wearing a red shirt included in the video. This is because the smoke generated in the fire zone has not been properly analyzed for the variation of high frequency components through wavelet transformation. However, the proposed algorithm showed relatively high retrieval rate of 84.2% and 79.5% compared to fire detection using Method 2 using luminance map and SVM.

11 is a view showing a fire detection results according to the present invention.

1 is a block diagram of a fire detection system according to the present invention.

2 is a diagram illustrating a motion region detection process using background initialization and update according to the present invention.

3 illustrates a fire color model using candidate fire colors in accordance with the present invention.

4 is a diagram illustrating a fire candidate region to which a fire color model is applied after detecting a motion pixel according to the present invention;

5 is a configuration diagram of a luminance map processing module according to the present invention;

6 is a diagram illustrating a candidate fire region removal result screen according to an input image and a luminance map generation result of the image according to the present invention;

FIG. 7 is a graph illustrating a temporal wavelet energy change graph for a fire zone and a fire zone according to the present invention. FIG.

8 illustrates a graph of temporal wavelet energy variation for non-fire regions and non-fire regions according to the present invention.

9 is a graph illustrating the wavelet energy of fire pixels and non-fire pixels according to the present invention after being leveled and arranged in ascending order.

10 is a table showing a comparison of experimental results using the fire detection results and the algorithm of T according to the present invention.

11 is a view showing a fire detection result screen according to the present invention.

Claims (12)

In the fire detection system, A motion detection module for initializing the gray image of the previous frame to the background and updating and outputting the previous background frame to reflect the current frame; A color modeling module configured to detect a first fire candidate region by applying a color model to a frame output by the motion detection module; A luminance map processing module configured to apply a luminance map to the current frame to detect a second fire candidate region brighter than a surrounding; And a SVM processing module for extracting a final fire detection region for the first and second fire candidate regions according to a variation model of temporal wavelet transform energy using a support vector machine (SVM), The motion detection module, And if the difference between gray image values of the previous frame with respect to the current frame is greater than a threshold, a motion pixel value is assigned to the frame. delete The method of claim 1, The color modeling module, A fire detection system generating a probability model of fire colors after extracting fire colors from fire images in a learning fire frame, assuming that the distribution of each RGB color channel is independent of each other and having a Gaussian distribution in the fire region. The method of claim 1, The luminance map processing module, A gray image converting unit converting input image information into gray image information; A down sampling module for down sampling the gray image information; A plurality of filters for filtering the down-sampled image information to different sizes; An average calculator configured to calculate an average of output values of the plurality of filters; A restoring unit generating a luminance map by restoring an output of the average calculating unit to an original size; A Gaussian filter for Gaussian filtering the luminance map; A change amount detector configured to analyze a temporal change amount of the Gaussian filtered luminance map; And a non-fire zone removal unit for determining a second fire candidate zone and removing the remaining zones according to the change amount. The method of claim 1, The wavelet energy detection equation of the SVM processing module according to Equation (11).

The method of claim 1, Fire detection system, characterized in that to remove the noise area for the color and the amount of change through the morphology opening operation on the result of the SVM processing module to finally discriminate the fire. In the fire detection method, After initializing the gray image of the previous frame to the background, updating the previous background frame to reflect the current frame and outputting a frame indicating a motion; Detecting a first fire candidate region by applying a color model to the frame marked with the motion; Detecting a second fire candidate area brighter than a periphery by applying a luminance map to the current frame; Extracting a final fire detection zone for the first and second fire candidate zones according to a variation model of temporal wavelet transform energy using a support vector machine (SVM), The mark of the movement, And if the difference between the gray image values of the previous frame with respect to the current frame is larger than the threshold, the motion pixel value is assigned to the frame. delete The method of claim 7, wherein Application of the color model, And calculating a probability value of the entire color channel using the fire color model, and when the value is larger than the threshold value, discriminates the first fire candidate region. The method of claim 7, wherein Application of the luminance map, Converting the input image information into gray image information; Down sampling the gray image information; Filtering the down-sampled image information to different sizes; Calculating an average of output values of the plurality of filters; Generating a luminance map by restoring an output of the average calculator to an original size; Gaussian filtering the luminance map; Analyzing a temporal change amount of the Gaussian filtered luminance map; Determining a second fire candidate region according to the change amount and removing the remaining region. The method of claim 7, wherein The wavelet energy detection equation is according to the equation (12).

The method of claim 7, wherein Fire detection method, characterized in that to remove the noise area for the color and the amount of change and finally to discriminate the fire through the morphology opening operation for the SVM processing result.

Images