KR101054649B1 - Real-time fire detection method for tunnel - Google Patents

Abstract

PURPOSE: A real time fire detection method in a tunnel is provided to monitor a fire in a tunnel with high reliability detection, minimize the damage of persons and properties because of early sensing fire, and increase stability in the tunnel by reducing tunnel maintenance cost and minimizing manpower through automatically sensing. CONSTITUTION: A fire learning step obtains covariance about RGB(Red, Green, Blue) using an equation 1 from a plurality of fire samples(S1). An image acquisition step gets digital image from image which is taken by a camera(S2). A probability distance calculation step calculates probability distance using an equation 2 in the obtained domain of image(S3). A fire territory extract step(S4) extracts fire domain using calculated probability distance.

Description

Real-time fire detection method for tunnel}

The present invention relates to a real-time fire monitoring method in a tunnel, and more particularly, to a real-time fire monitoring method in a tunnel for monitoring the occurrence of a fire using a real-time image of the tunnel.

Korea is a mountainous landscape of Gyeongdong Province. As the road speed and straightening of the road have increased recently, the number of large-scale tunnels has increased and the importance of monitoring various accidents in the tunnels has increased. In addition, there is a growing interest in preventing the loss of property and lives to the fire caused by the use of images from surveillance cameras in tunnels. In addition, with the development of intelligent surveillance systems, real-time detection technologies for various fires have been developed.In particular, due to the rapid development of computer performance and the falling price of surveillance cameras, the prevention of accidents and disasters using the video surveillance system is gradually expanded. Is being applied.

Conventional methods of monitoring fires in tunnels include direct monitoring of internal images obtained using cameras, monitoring of air flow and components by attaching cloud sensors and various chemical sensors, and periodically One way is to identify systems in tunnels.

In the case of monitoring by using the chemical sensor among the above methods, there is an advantage that can be automated by the sensor signal, but the chemical sensor has a very limited range to be recognized by its characteristics. Should be used. Therefore, the method requires excessive initial installation costs and regular inspections to confirm the correctness of the sensor operation, which incurs a lot of maintenance costs, and detects the fire until it reaches the sensor when an actual fire occurs. The delay of time is also a problem.

In addition, the rest of the methods require the staff to operate and monitor the system directly, which requires constant monitoring, especially in the midnight hours due to the nature of the road. There is a need for a relatively low cost and automatic fire monitoring system.

The present invention was devised to improve the above problems, and is based on the manner in which a person detects a fire, and combines a method of using the color and movement of the fire to provide a reliable real-time fire in a tunnel under the limited conditions inside the tunnel. Its purpose is to provide a monitoring method.

In order to achieve the above object, the present invention provides a real-time fire monitoring method in a tunnel to obtain the real-time image of the tunnel by a camera to monitor the occurrence of a fire, using the equation (1) from a plurality of fire samples A fire learning step of obtaining a covariance; An image acquisition step of acquiring a digital image of an image photographed by a camera; A probability distance calculation step of calculating a probability distance with respect to the area of the obtained image using Equation 2; And a fire zone extraction step of extracting a fire station using the calculated probability distance.

Preferably, the fire area extracting step may include a VOV calculation step of obtaining a variance of variance (VOV) using Equations 3, 4, and 5 on the obtained image; And a fire region extraction step using the VOV which extracts the fire region from the image using the VOV calculated in the above step.

More preferably, the step of extracting the fire zone may be performed by using a plurality of selected cross windows, and calculating the crossness of the brightness of the image obtained by Equation 6 for each cross window, and the cross brightness of the entire light may be determined by each window. Calculating a crossover of brightnesses which averages the crossovers of the brightnesses; And a fire zone extraction step using the crossover of brightness to extract the fire zone by using the crossover of the brightness calculated in the above step.

More preferably, the crossing window is divided into three sections in which the bright and dark windows are divided diagonally, the bright and dark windows are divided up and down, and the bright and dark windows are divided left and right.

Preferably, the fire zone extraction step further includes a final fire zone extraction step of extracting a final fire zone by calculating a common area of a fire zone using a probability distance, a fire zone using a VOV, and a fire zone using a crossover of brightness. Characterized in that.

Preferably, in the probability distance calculation step, the probability distance calculation is characterized by using a LUT that has previously calculated a square image for each coordinate.

Preferably, the crossover of brightness in the step of calculating the crossover of brightness is characterized by using an incremental image.

Through the above problem solving means, the present invention can monitor with high reliability for the fire occurring in the tunnel, can minimize the damage of life and property due to early fire detection, furthermore through the automatic detection It has the effect of increasing the stability in the tunnel by minimizing manpower and reducing the cost of tunnel maintenance.

1 is a flow chart illustrating a procedure of a real-time fire monitoring method in a tunnel according to the present invention,
2 is a sample image for the fire learning of the present invention,
3 is a diagram showing a standard fire in a probability space,
4 is an image showing a case of a fire image in the tunnel,
FIG. 5 is a graph illustrating a change in color of an area of flame in the image of FIG.
FIG. 6 is a graph illustrating a change in color per hour of an area without flame in the image of FIG. 4.
7 is an example of a table showing a value calculated by variance for each unit time.
8 is an example showing a cross window for calculating a brightness cross ratio,
9 is an example of an area for a gradual image calculation,
FIG. 10 is an image showing a fire zone from which the video image of FIG. 4 is extracted by the method according to the present invention. FIG.

Hereinafter, with reference to the accompanying drawings, the configuration of the present invention will be described in detail.

The present invention is characterized by using the color and movement of the fire based on the method of recognizing the fire by the human visual information.

If we first look at how a person perceives fire, it can be divided into two types.

First, fire has a red color compared to other objects and has a very bright characteristic, and when it is given as a single light source, it makes the object's illuminance brighter.

And secondly, the fire is basically burning and burning, and at the same time, the surrounding haze occurs, and incomplete combustion and complete combustion occur at the same time.

The present invention detects the occurrence of fire using the characteristics of the overall color of the fire, the variable color change based on the above recognition method.

In particular, in the present invention, after the fire occurs, the location of the fire is constant and the fire is recognized by using the invariability of the fire location where the smoke continuously occurs around the point where the fire occurs.

The fire monitoring method proposed in the present invention consists of nine steps as shown in FIG. 1, and a detailed configuration of each step will be described.

1. Fire Learning Steps ( S1 )

In this step, the average and the covariance according to Equation 1 are obtained using the tunnel internal fire sample image as shown in FIG. 2.

Equation 1:

Where R _m is the mean value of R, n is the number of pixels of the trained standard fire sample, and Σ _RG is the covariance between RGs. Covariance between the remaining GB and RB is also obtained in the same manner as in Equation 1 above.

As a result, a covariance Σ with mean μ = [R _m , G _m , B _m ] and a 3 × 3 diagonal matrix with diagonal values Σ _RG , Σ _GB , Σ _BR is obtained from the standard fire sample. Where G _m and R _m are the average of G and R, Σ _RG , Σ _GB , Σ _BR Is the covariance value of RG, GB, and BR.

2. Image Acquisition Step ( S2 )

The image acquisition step is a procedure of acquiring a digital image using a digital camera, and includes an image photographed by the digital camera.

The digital camera may have a suitable lens brightness and may be a device capable of photographing a relatively dark tunnel in color, and the output image is continuously stored in a computer.

3. Calculate the probability distance S3 )

This step is a procedure for calculating the probability distance using the digital image provided in the image acquisition step.

When the RGB value is expressed as x (i, j) = [R, G, B] in the coordinates (i, j) of the image obtained in the previous step, the probability distance D _M is calculated using Equation 2 below.

Equation 2:

Meanwhile, if necessary for the calculation of the probability distance, the following square image calculation method may be used.

The square image is pre-calculated x _sqr (i, j) = x (i, j) ² by reducing the number of calculations by squaring the brightness value of each coordinate in advance. In this case, when the LUT (Look Up Table) is constructed for each coordinate, the square values of the corresponding positions are directly corresponded on the table, thereby increasing the efficiency of calculation.

4. Extraction of fire zone by probability distance ( S4 )

If the present step, when the probability of distance D _m value as compared with the D _lim of selecting the probability distances of the image coordinate (i, j) calculated in the previous step is less than D _lim, proceed to the next step is greater than D _lim is It is judged that no fire has occurred and processes other images.

The D _m is rapidly increased as it does not cluster in a standard fire, as shown in FIG. 3, so that an appropriate D _lim value is selected using the characteristic.

5. Variance of Variance ( VOV ) Calculation step ( S5 )

In this step, the change of R is calculated using the characteristics of the flame in the image. The flame in the image is usually red in color and constantly swinging. The flame is perceived as a light source in the image, so the center is close to white, but the edge of the flame is constantly changing irregularly. In addition, as a light source, the flame has a characteristic of constantly changing the brightness of the surrounding object surface.

As shown in FIG. 4, the color change of the portion where the flame is located and the color change of the flameless point were analyzed using the fire image generated in the tunnel. FIG. 5 shows the RGB color change over time of the portion where the flame is located for 10 seconds in the image of 10 fps, and FIG. 5 shows the color change according to the traffic of a general vehicle without the flame at the same time. As shown in FIG. 5, the part with the flame has a very bright R of RGB and shows irregular movement in all colors, while the color change of the part of the vehicle outside of the light of FIG. 5 has a constant color. Only brightness changes appear. In this way, the flame has a noticeable difference in motion compared to the background or other objects, and a continuous color change occurs over time.

In order to detect the presence of fire using these characteristics, first calculate the brightness dispersion using Equation 3.

Equation 3:

Where Var (st, ed) is the variance of the brightness values from frame _st to frame _ed at the same location, frame _i is the brightness in i image, sd-st is the number of images used, μ _{st - ed} is frame _st To the frame _ed is the average brightness of the same location in the image.

7 is a result of calculating a dispersion value of R in RGB according to a unit time. The active and persistent brightness change of the flame keeps Var (st, ed) high and the var (st, ed) of the flameless point has a high value in some intervals as shown in FIG. It is not persistent.

The Var (st, ed) value can be obtained as shown in Equation 4 for each section.

Equation 4:

Here, μ _var (st, f, nm) is the average for each interval, nm is the number of intervals used for the average, and f is the number of images used to calculate Var (st, ed).

Using Equation 4, a VOV (variance of variance) as shown in Equation 5 can be obtained.

Equation 5:

If necessary, a square image calculation method may be applied to the VOV calculation.

6. VOV Fire zone extraction step by S6 )

In this step, the fire area is extracted by preparing the VOV calculated in the previous step in the image.

The VOV value has the effect of amplifying the value very large when the variance value constantly changes. When the VOV value is calculated using Equation 5, a flame having a constantly irregular high dispersion value has a high VOV value, whereas a surrounding background or an object has a relatively low VOV value. Fire can be detected by comparing this value to the variance of the brightness values in the whole image. In the case of calculating the VOV value with respect to the value used in FIG. 7, the VOV of the fire zone is VOV _fire (1,10,10) = 6,083,306, and the difference in the background portion VOV _non (1,10,10) = 0.245 is significantly different. Seems.

Accordingly, when the VOV value is compared with a specific value of VOV _lim , a fire may extract an area, and the VOV _lim value may be appropriately selected by learning.

7. Brightness Crossness  Calculation step ( S7 )

In this step, the brightness intersection of the region is calculated to distinguish between moving and non-moving objects in the image.

Backgrounds such as tunnels and roads are still images that do not move over time when viewed from a camera. Vehicles and people are classified as moving objects on the road. The outward appearance of the flame does not move in place as a whole, but the details are constantly changing. Thus, a flame can be characterized as a thing that changes shape at a fixed location in a tunnel. In other words, the fire in the tunnel is a moving object, but its position does not change. It is a motion that is different from the background of a stationary, non-moving background and a vehicle moving its position along the road, and this can be used to search for fire, and this step uses this property to calculate the crossover of brightness.

FIG. 8 is three cross windows for calculating the crossover of brightness, and the crossover is defined by Equation 6 below.

Equation 6:

Here, P _cross is an intersection value of brightness with respect to the window of FIG. 8A, and Σblack-Σwhite are values obtained by subtracting the sum of the brightnesses of the bright parts from the sum of the brightnesses of the dark parts. In the case of FIG. 8A, the sum of the brightnesses of the dark parts is greater than the sum of the brightnesses of the light parts, and a negative value. If the color changes in place like a flame, this value repeats negative and positive at the same position.In a unit time, the value of Σblack-Σwhite intersects 0 and zerocross (the number of times the positive and negative values are repeated). Σ black-Σ white).

In the same way, the results of FIGS. 8B and 8C are called P _ver and P _hor , respectively, and the average of the results of the three extractors is P _result .

On the other hand, if the color of the dark and light areas inside the cross extractor window is similar, the light and light cross each other, like a fire, even with a small light change. Can be reduced.

On the other hand, the present invention can reduce the calculation time when performing a calculation method such as the following a lot of iterative calculation in the step of obtaining the dispersion of brightness over time and the P _result .

An incremental image may be used, and the integrative image calculates x _int (i, j) by the following equation.

Equation 7:

Where x _int (i, j) is the sum of the brightness values from (0,0) to (i, j).

For example, if the sum of the shaded areas to obtain the brightness value I in Fig. 9 can quickly absorbs the value by the following formula is effective in the calculation _result of the P value.

8. Brightness In crossness  Fire zone extraction step (S8)

In this step, the fire area is extracted using the P _result value calculated in the previous step. The area where the P _result value is equal to or greater than the P _lim is determined to be a fire area.

The _Plim value can also be appropriately selected by learning the previous fire pattern.

9. Final fire zone extraction step ( S9 )

Extracting the areas intersecting each other among the three fire zones by the above steps, the fire zone where the fire occurred is extracted.

The fire zone is an intersection point of the area extracted at each step, and the intersection point can be easily obtained by the coordinates of the area.

Meanwhile, the three topical extraction steps may be changed in order. For example, the fire zone detection may be applied first to the brightness crossover, and the remaining two methods may be applied, and the fire zone by the VOV may be detected first and the rest may be performed. You can also perform only some of the detections as needed. For example, it can only be performed by monitoring by VOV and probability distance, or by VOV and brightness crossover.

As a result of performing the method according to the present invention on the video screen of FIG. As shown in FIG. 10, it may be confirmed that a plurality of windows have been extracted at a fire occurrence point, and it may be confirmed that a fire is generated in the window.

While the invention has been shown and described with respect to certain preferred embodiments thereof, the invention is not limited to these embodiments, and has been claimed by those of ordinary skill in the art to which the invention pertains. It includes all the various forms of embodiments that can be carried out without departing from the spirit.

Claims (7)

In the real-time fire monitoring method in the tunnel to obtain the real-time image of the tunnel by the camera to monitor the fire occurrence,
A fire learning step of obtaining covariance for RGB using Equation 1 from a plurality of fire samples;
An image acquisition step of acquiring a digital image of an image photographed by a camera;
A probability distance calculation step of calculating a probability distance with respect to the area of the obtained image using Equation 2; And
And a fire zone extraction step of extracting a fire station using the calculated probability distance.
Equation 1 is

Where R _m is the mean value of R, n is the number of pixels of the standard fire sample _trained , Σ _RG is the covariance between RG,
Equation 2 is

Where μ = [R _m , G _m , B _m ], G _m and R _m are the mean values of G and R, Σ is a 3 × 3 diagonal matrix with diagonal values Σ _RG , Σ _GB , Σ _BR to be.
The method of claim 1, wherein the fire zone extraction step
Calculating a VOV of variance (VOV) of the obtained image using equations (3), (4) and (5); And
And a fire zone extraction step using a VOV for extracting a fire zone from the image using the VOV calculated in the step.
However, Equation 3 is

Where Var (st, ed) is the variance of the brightness values from frame _st to frame _ed at the same location, frame _i is the brightness in the i image, sd-st is the number of images used, μ _{st - ed} is the average of brightness at the same position in the image from frame _st to frame _ed ,
Equation 4 is

Μ _var (st, f, nm) is the average for each interval, nm is the number of intervals used for the average, f is the number of images used to calculate Var (st, ed),
Equation 5 is

to be.
The method of claim 2, wherein the fire zone extraction step
Calculate the crossover of the brightness of the image obtained by Equation 6 for each cross window using the selected multiple cross windows, and the overall cross brightness of brightness is the average of the crossover of the brightness of each window. Crossover calculation step; And
And a fire zone extraction step using a crossover of brightness to extract a fire zone by using the crossover of brightness calculated in the step.
However, Equation 6 is

Where P _cross is the intersection of brightness for each cross window, and Σblack-Σwhite is the sum of the brightness of the dark parts minus the sum of the brightness of the light areas.
The real time in a tunnel of claim 3, wherein the crossing window comprises three bright and dark windows divided diagonally, bright and dark windows vertically divided, and bright and dark windows divided left and right. Topical monitoring methods.
The method of claim 3, wherein the fire zone extraction step
The real-time topic in a tunnel further comprising a final fire zone extraction step of extracting a final fire zone by calculating a common area of a fire zone using a probability distance, a fire zone using a VOV, and a fire zone using a crossover of brightness. Surveillance Method.
The method according to claim 1, wherein the probability distance calculation in the probability distance calculation step, the real-time topic monitoring method in a tunnel, characterized in that using the LUT pre-calculated square image for each coordinate.
4. The method of claim 3, wherein in the step of calculating the crossover of brightness, the crossover of brightness uses an incremental image.

Images