JP7691063B2 - Smoke detection devices - Google Patents

Description

This disclosure relates to a technology for detecting the occurrence of smoke by applying image processing to images captured by a surveillance camera.

There are conventional devices that use image processing technology to extract only the smoke areas from within a monitored area captured by a surveillance camera (see, for example, Patent Document 1). Such conventional devices extract features caused by smoke by applying image processing to the image captured by the surveillance camera, and detect a state in which smoke has been generated as an abnormal state that differs from a steady state.

JP 2010-97265 A

However, the conventional techniques have the following problems.
Conventional smoke detection methods that apply image processing technology aim to improve the accuracy of smoke detection by distinguishing between "smoke" and "non-smoke" based on a discrimination boundary learned from training data in a multidimensional feature space consisting of multiple feature values.

However, the types of features used in conventional smoke detection methods are lacking in those that evaluate the luminance fluctuations that are characteristic of smoke. For example, there is a problem in that luminance changes due to lighting or sunlight, or luminance changes due to a person passing by, are easily mistaken for smoke. Therefore, it is desirable to use features that focus on the luminance fluctuations that are characteristic of smoke, in order to prevent erroneous recognition of smoke.

The present disclosure has been made to solve the problems described above, and aims to provide a smoke detection device that can detect smoke occurring within a monitored area using features focusing on the luminance fluctuations that are characteristic of smoke, and can reduce false recognition of smoke.

The smoke detection device according to the present disclosure is a smoke detection device that detects the occurrence of smoke by applying image processing to a smoke candidate region in a monitoring target image captured by a surveillance camera, and includes a feature extraction unit that calculates a co-occurrence matrix for the monitoring target image captured by the surveillance camera using the luminance values of each pixel in the smoke candidate region and calculates a texture feature based on the calculated co-occurrence matrix, and a smoke occurrence detection unit that determines whether smoke has occurred in the smoke candidate region in the monitoring target image based on the texture feature calculated by the feature extraction unit, and further includes an image memory that stores the monitoring target images captured by the surveillance camera as time-series monitoring target images at a predetermined sampling period, and the feature extraction unit calculates the co-occurrence matrix using the luminance values of each pixel in the smoke candidate region in two or more chronologically adjacent monitoring target images included in the time-series monitoring target images, taking into account the adjacency relationship between a pixel of interest in one image and adjacent pixels relative to the pixel of interest, and further taking into account the adjacency relationship with pixels at the same position as the pixel of interest in previous and subsequent frames in the chronological order .

The smoke detection device according to the present disclosure is a smoke detection device that detects the occurrence of smoke by performing image processing on a smoke candidate region in a monitoring target image captured by a surveillance camera, and includes an image memory that stores the monitoring target image captured by the surveillance camera at each predetermined sampling period during surveillance as a time-series monitoring target image, a pre-processing unit that sequentially calculates an EMA image consisting of pixels that are the result of performing an exponential moving average using the time-series monitoring target image for each pixel, and compares the luminance value of the EMA image with the luminance value of the monitoring target image captured when the EMA image was calculated or in the next sampling period, for each pixel, to generate a binary image based on the magnitude of the luminance value, and a smoke occurrence detection unit that determines whether smoke has occurred in the smoke candidate region in the monitoring target image based on the binary image generated in the pre-processing unit within the smoke candidate region.

According to the present disclosure, it is possible to obtain a smoke detection device that can detect smoke occurring within a monitored area using features that focus on the brightness fluctuations that are characteristic of smoke, and can reduce the false recognition of smoke.

1 is a configuration diagram of a smoke detection device according to a first embodiment of the present disclosure. 11A and 11B are diagrams for explaining differences in EMA difference images depending on the type of moving object in the first embodiment of the present disclosure. FIG. 11 is an image diagram for explaining a technique for creating a GLCM from a single image input by a feature extraction unit according to the first embodiment of the present disclosure. FIG. 11 is an image diagram for explaining a method for creating a GLCM from multiple image inputs arranged in chronological order by a feature extraction unit according to the first embodiment of the present disclosure. 11 is an image that visualizes the highs and lows of entropy calculated for a smoky scene and a scene with moving people in the first embodiment of the present disclosure. 1 is a diagram showing the probability density distribution of entropy in a smoke region and a non-smoke region in embodiment 1 of the present disclosure. FIG. 11 is an image that visualizes the highs and lows when non-uniformity is calculated for a smoky scene and a scene in which people are moving in the first embodiment of the present disclosure. 11 is a diagram showing a probability density distribution of non-uniformity in a smoke region and a non-smoke region in the first embodiment of the present disclosure. FIG. 11 is an image that visualizes the highs and lows of energy obtained for a smoky scene and a scene in which people are moving in the first embodiment of the present disclosure. 1 is a diagram showing the energy probability density distribution in a smoke region and a non-smoke region in embodiment 1 of the present disclosure. FIG.

The following describes a preferred embodiment of the smoke detection device of the present disclosure with reference to the drawings. The present disclosure has a technical feature of calculating features that focus on the luminance fluctuations that are characteristic of smoke using GLCM (Gray Level Co-occurrence Matrix), EMA (Exponential Moving Average), or a combination of GLCM and EMA.

Therefore, we will first explain a smoke detection method based on a combination of GLCM and EMA, and then explain smoke detection methods based on GLCM alone or EMA alone.

Embodiment 1.
1 is a configuration diagram of a smoke detection device according to a first embodiment of the present disclosure. The smoke detection device according to the first embodiment includes an image memory 10 and a detection processing unit 20. The image memory 10 is configured as an image memory for a plurality of frames so that images of a monitoring target captured at a predetermined sampling period by a monitoring camera 1 corresponding to an image sensor can be stored as time-series monitoring target images for a certain period of time in the past. One frame of an image is composed of a plurality of pixels.

The detection processing unit 20 is also configured to include a preprocessing unit 21, a feature extraction unit 22, and a smoke generation detection unit 23, and detects the generation of smoke by performing image processing on a smoke candidate area in the monitored image captured by the surveillance camera 1. First, an overview of the series of processes performed by the detection processing unit 20 will be described.

The pre-processing unit 21 sequentially calculates EMA images consisting of pixels that are the result of performing exponential moving averages using the time-series images of the monitored object. Furthermore, for each pixel, the pre-processing unit 21 compares the brightness value of the monitored object image captured in the sampling period when the EMA image was calculated or the brightness value of the monitored object image captured in the next sampling period with the brightness value of the EMA image, and generates a binary image based on the magnitude of the comparison.

The feature extraction unit 22 calculates a co-occurrence matrix using the binary values of each pixel in the binary images sequentially generated within the smoke candidate region by the pre-processing unit 21. Furthermore, the feature extraction unit 22 calculates texture features within the smoke candidate region by performing statistical processing on the calculated co-occurrence matrix to identify the presence or absence of smoke.

The smoke generation detection unit 23 determines whether smoke has been generated in a smoke candidate region in the monitored image based on the texture features calculated by the feature extraction unit 22.

Next, we will provide a detailed explanation of the computational process to obtain the EMA difference image and the computational process to obtain the GLCM, which are preliminary steps to obtaining texture features that focus on the luminance fluctuations that are unique to smoke.

<About EMA difference images>
As a preprocessing for calculating a feature amount having time-series information, a calculation process of an EMA difference image is performed by the preprocessing unit 21. First, a method for creating an EMA image will be described, and then a method for creating an EMA difference image will be described.
(1) Method for Creating EMA Image The preprocessing unit 21 calculates an exponential moving average (hereinafter referred to as EMA) in the time series direction for each luminance in each image of the time series monitoring target image corresponding to a series of time series images. In this disclosure, an image in which each calculated EMA value is set as a luminance value is defined as an EMA image.

The EMA image is an image that is updated sequentially over time, and if there is no change in the monitored image, there will be no change in the EMA image either. On the other hand, if a moving object moves across the field of view, the trajectory of the moving object's movement will remain as an afterimage on the EMA image. Also, if the illuminance of the entire screen changes gradually, the EMA image will adapt sequentially and its brightness will change gradually.

The calculation formula for EMA for each pixel is shown as the following formula (1).
S _t = α×Y _t + (1-α)×S _t-1 (1)
where S _t is the exponential moving average at time t, which is the current sampling period.
S _t-1 : exponential moving average at time t-1, which is the previous sampling period
Y _t : The brightness value of the image of the target captured by the surveillance camera at time t
α: A smoothing coefficient set as a value between 0 and 1

By appropriately setting the value of the smoothing coefficient α according to the monitored environment, the speed of brightness change due to the influence of disturbing elements, etc., it is possible to change the degree of weighting of the EMA value at the previous sampling time when calculating the EMA value at the current sampling time in an EMA image that is continuously updated over time.

(2) Creation of EMA Difference Image Next, the pre-processing unit 21 compares, for each pixel, a first luminance value, which is the luminance value of the monitored image captured in the sampling period when the EMA image was calculated or in the next sampling period, with a second luminance value, which is the luminance value of the EMA image. Then, the pre-processing unit 21 creates an EMA difference image according to the magnitude relationship between the first luminance value and the second luminance value.

As an example, the pre-processing unit 21 performs binarization processing on each pixel to make it white if the first luminance value is higher than the second luminance value, and to make it black if the first luminance value is equal to or lower than the second luminance value, thereby creating an EMA difference image. Here, if the first luminance value and the second luminance value each have 256 gradations from 0 to 255, then white corresponds to 255 and black corresponds to 0.

Note that, for the sake of simplicity, the EMA difference image is a binary image here, but it is also possible to create an EMA difference image as a grayscale image with three or more values in which the difference between the first luminance value and the second luminance value is divided into multiple levels.

The characteristics of the EMA difference image created in this manner will be described. FIG. 2 is a diagram for explaining the difference in EMA difference images depending on the type of moving object in embodiment 1 of the present disclosure. FIG. 2(a) shows an EMA difference image created for a scene in which there is no movement within the monitored area, FIG. 2(b) shows an EMA difference image created when a smoke generation area is included within the monitored area, and FIG. 2(c) shows an EMA difference image created when a person, which is a moving object, is included within the monitored area.

The EMA difference image appears differently depending on factors such as the movement of the subject within the monitored area and the periodicity of changes in the surface texture. When no changes occur within the screen, random noise caused by minute fluctuations in pixel values due to the scanning lines or AGC (auto gain control) appears across the entire screen in the EMA difference image, as shown in Figure 2(a).

In addition, when a change in brightness occurs within the monitored area due to a moving object such as a person, a pattern corresponding to the brightness level of the object and an afterimage on the moving trajectory appear in the moving object area on the EMA difference image, as shown in Figure 2(c).

In addition, when the brightness of the monitored area changes significantly due to changes in illuminance caused by room lighting or external light, the EMA difference image shows objects on the screen in silhouette according to their shades while the difference between the first luminance and the second luminance immediately after the change is large. After that, as the EMA image is applied to the screen after the illuminance change and the difference between the first luminance value and the second luminance value becomes smaller, the EMA difference image returns to random noise.

On the other hand, when a smoke generation area, which is the actual detection target, appears within the monitored area, a unique shading pattern caused by the flow and periodicity of the smoke appears within the smoke generation area on the EMA difference image, as shown in Figure 2 (b).

About GLCM
In the EMA difference image calculated as described above, different textures appear depending on the presence or absence of a moving object in the screen or the difference in the moving direction of the moving object. Therefore, by quantitatively identifying these differences using texture features, it becomes possible to identify whether or not there is an area where smoke has occurred. Here, a method for creating a GLCM used to extract texture features focusing on the luminance fluctuations specific to smoke will be described.

Regarding GLCM, there are cases where the GLCM is derived from a single image, and cases where the GLCM is derived from multiple images arranged in time series. In this disclosure, the former is referred to as a two-dimensional GLCM, and the latter is referred to as a three-dimensional GLCM. The specific method of creating a GLCM performed by the feature extraction unit 22, and the calculation of texture features suitable for smoke detection based on GLCM, will be explained below for two dimensions and three dimensions.

(1) Creation of a Two-Dimensional GLCM Fig. 3 is an image diagram for explaining a method for creating a GLCM from one image input by the feature extraction unit 22 in the first embodiment of the present disclosure. Fig. 3(a) shows a binary image of 5 x 5 pixels as an example of an input image. Fig. 3(b) shows a matrix A corresponding to the GLCM created by the feature extraction unit 22 based on the input image.

The GLCM corresponds to matrix A, which quantitatively expresses the degree of luminance variation based on the positional relationship between pixels. When creating a two-dimensional GLCM, the luminance value of the pixel of interest is taken as the row, and the luminance values of the surrounding pixels of the pixel of interest are taken as the column. In other words, when each element of matrix A shown in FIG. 3B is A _i,j , i means the density value of the pixel of interest, and j means the density value of the neighboring pixel.

The feature extraction unit 22 calculates the matrix A by obtaining the cumulative sum of each element of A _i,j . Specifically, each element of A _0,0 , A _0,1 , A _1,0 , and A _1,1 is counted up when the following condition is met:
A _0,0 : If the luminance value of the pixel of interest is 0 and the luminance value of the neighboring pixel is 0, then +1 is added.
A _0,1 : If the luminance value of the pixel of interest is 0 and the luminance value of the neighboring pixel is 1, then +1 is added.
A _1,0 : If the luminance value of the pixel of interest is 1 and the luminance value of the neighboring pixel is 0, +1 is added.
A _1,1 : If the luminance value of the pixel of interest is 1 and the luminance value of the neighboring pixel is 1, then +1 is added.

For example, let us consider a case where the pixel surrounded by a double line in the binary input image shown in Fig. 3(a) is the pixel of interest, and the pixels on the right and left are neighboring pixels. In this case, the density value of the pixel of interest is "0", the density value of the neighboring pixel on the right is "1", and the density value of the neighboring pixel below is "0". Therefore, of the four elements of matrix A, 1 is added to _A0,0 and _A0,1 .

By repeatedly performing the same process for all pixels in the block while moving the pixel of interest in sequence, the feature extraction unit 22 can obtain the cumulative sum result for each element. Furthermore, the feature extraction unit 22 can obtain a normalized matrix A by dividing the cumulative sum result for each element of matrix A by the sum of the cumulative sum results for all elements.

By performing normalization, each element A _i,j represents the probability that the luminance values of pixels are adjacent to each other as (0,0), (0,1), (1,0), and (1,1).

In the specific example described in FIG. 3, the brightness values of the EMA difference image are binary, 0 and 1, so matrix A equivalent to the GLCM is a 2x2 matrix. However, this is just one example. For example, if the EMA difference image has four values, 0, 1, 2, and 3, matrix A equivalent to the GLCM will be a 4x4 matrix.

In the specific example described in FIG. 3, the two pixels on the right and bottom are considered to be adjacent images, but it is also possible to use pixels other than the right and bottom as adjacent pixels. For example, it is also possible to use the two pixels on the left and top as adjacent images. It is also possible to use diagonal pixels as adjacent pixels.

The feature extraction unit 22 can calculate the matrix A that corresponds to the two-dimensional GLCM by performing the above-mentioned calculation process on the input image.

(2) Creation of a 3D GLCM The above-mentioned 2D GLCM is calculated by using one difference image created at regular intervals in accordance with the period for calculating the feature amount as an input image. In this case, there may be a problem that the feature amount is not stable due to the bias of random noise. Alternatively, there may be a problem that a sufficient number of features cannot be obtained because feature amount extraction is not performed for most of the frames for which a difference image is not created.

Therefore, in order to improve smoke detection accuracy, it is desirable to create a GLCM that can extract statistically stable features. In order to solve these problems, instead of a two-dimensional GLCM that quantitatively expresses the degree of luminance variation based on the positional relationship between pixels in a single input image, it is possible to create a three-dimensional GLCM that also takes into account the relationship between adjacent pixels in the forward and backward directions in a chronological order.

As an example, it is possible to accumulate 32 frames of EMA difference images and perform three-dimensional GLCM creation and texture analysis for each cubic voxel consisting of 32 pixels x 32 pixels x 32 frames.

Figure 4 is an image diagram for explaining a method for creating a GLCM from multiple image inputs arranged in time series by the feature extraction unit 22 in embodiment 1 of the present disclosure. Figure 4(a) shows a cubic voxel as a time-series binary image in which 32 frames of EMA difference images are accumulated.

Figure 4(b) also shows an example of a case where a three-dimensional GLCM is calculated using a total of three input images, taking into consideration the input image at sampling time t, the input image at the previous sampling time t-1, and the input image at the next sampling time t+1.

When creating the two-dimensional GLCM described above in Figure 3, the adjacent relationships between the pixel of interest and the two adjacent pixels to the right and below in a single image were taken into consideration. In contrast, as shown in Figure 4, when creating a three-dimensional GLCM, the adjacent relationships between pixels in the same position as the pixel of interest in the previous and following frames in the chronological order are also taken into consideration, and the degree of luminance variation based on the positional relationships between pixels is quantitatively evaluated.

The feature extraction unit 22 creates a matrix A using four adjacent pixels for a pixel of interest in the input image at time t, taking into consideration two adjacent pixels (in the x and y directions) to the right and below in the input image at time t, as well as two adjacent pixels in the input images at times t-1 and t+1 that are in the same position as the pixel of interest in the input image at time t.

In this way, the feature extraction unit 22 can create a three-dimensional GLCM using four adjacent pixels based on three input images arranged in chronological order. As a result, compared to creating a two-dimensional GLCM, it is possible to create a GLCM that takes into account images to be monitored over a greater number of sampling periods and calculates more stable features.

(3) Extraction Processing of Texture Features Based on GLCM Next, a specific example of statistical processing will be described in which, after matrix A is created as a two-dimensional GLCM or a three-dimensional GLCM, the feature extraction unit 22 extracts texture features from matrix A that are used to determine smoke detection.

In this disclosure, a specific example will be described below in which three texture features, entropy, heterogeneity, and energy, are extracted by performing statistical processing on GLCM as texture features. Note that other texture features can also be used to extract the texture features.

(3-1) Entropy When GLCM is P(A), the entropy can be calculated as the following formula (2).

Taking the difference image shown in Figure 2 above as an example, the more random the arrangement of black and white in the difference image is, such as in the "motionless scene" in Figure 2(a), the higher the entropy value. On the other hand, if the arrangement of black and white is regular, or if the white or black is uniform, such as the afterimage behind the person in the "moving object (person)" in Figure 2(c), the entropy value will be low.

Figure 5 is an image that visualizes the entropy obtained for a smoky scene and a scene in which people are moving in the first embodiment of the present disclosure. Specifically, Figure 5(a) is an image that visualizes the entropy in a smoky scene, and Figure 5(b) is an image that visualizes the entropy in a scene in which people are moving.

Furthermore, FIG. 6 shows the probability density distribution of entropy in smoke and non-smoke regions in embodiment 1 of the present disclosure. Specifically, FIG. 6(a) shows the probability density distribution of entropy calculated for a two-dimensional GLCM, and FIG. 6(b) shows the probability density distribution of entropy calculated for a three-dimensional GLCM.

In addition, in each of Figures 6(a) and 6(b), the probability density distribution of entropy in the smoke region is shown as "distribution 1," and the probability density distribution of entropy in the non-smoke region is shown as "distribution 2."

As shown in Figure 6, distribution 1, which is the probability density distribution of entropy within a smoke region, tends to have small entropy. For this reason, in each of Figures 6(a) and 6(b), it can be seen that distribution 1 has a lower peak and the peak is shifted to the left compared to distribution 2. Furthermore, a comparison of Figures 6(a) and 6(b) shows that the difference between distribution 1 and distribution 2 is more clearly evident in the texture features calculated based on the three-dimensional GLCM.

Therefore, the smoke generation detection unit 23 can quantitatively determine whether or not smoke has occurred in the monitored area from the probability density distribution of the texture features related to the entropy calculated by the feature extraction unit 22.

(3-2) Regarding heterogeneity When GLCM is P(A), the heterogeneity can be calculated as the following formula (3).

The non-uniformity value is high when the probability of consecutive white or black pixels in the difference image is low. Thus, non-uniformity values tend to be low inside smoke and high in areas where random noise is present.

Figure 7 shows images that visualize the level of non-uniformity obtained for a smoky scene and a scene in which people are moving in the first embodiment of the present disclosure. Specifically, Figure 7(a) shows an image that visualizes the level of non-uniformity in a smoky scene, and Figure 7(b) shows an image that visualizes the level of non-uniformity in a scene in which people are moving.

Furthermore, FIG. 8 shows the probability density distribution of inhomogeneity in smoke and non-smoke regions in embodiment 1 of the present disclosure. Specifically, FIG. 8(a) shows the probability density distribution of inhomogeneity calculated for a two-dimensional GLCM, and FIG. 8(b) shows the probability density distribution of inhomogeneity calculated for a three-dimensional GLCM.

In addition, in each of Figures 8(a) and 8(b), the probability density distribution of non-uniformity within the smoke region is shown as "distribution 1," and the probability density distribution of non-uniformity within the non-smoke region is shown as "distribution 2."

As shown in Figure 8, distribution 1, which is the probability density distribution of heterogeneity within a smoke region, tends to have smaller heterogeneity. For this reason, it can be seen in both Figures 8(a) and 8(b) that the peak of distribution 1 is shifted to the left compared to distribution 2. Furthermore, a comparison of Figures 8(a) and 8(b) shows that the difference between distribution 1 and distribution 2 is more clearly shown in the texture features calculated based on the three-dimensional GLCM.

Therefore, the smoke generation detection unit 23 can quantitatively determine whether or not smoke has occurred in the monitored area from the probability density distribution of the texture features related to non-uniformity calculated by the feature extraction unit 22.

(3-2) Regarding Energy When GLCM is P(A), the energy can be calculated as the following formula (4).

Energy has high values when the difference image is solidly white or black, or when certain patterns are repeated. Thus, energy values tend to be high inside smoke and low in areas where random noise is present.

Figure 9 is an image that visualizes the energy levels obtained for a smoky scene and a scene in which people are moving in the first embodiment of the present disclosure. Specifically, Figure 9(a) is an image that visualizes the energy levels in a smoky scene, and Figure 9(b) is an image that visualizes the energy levels in a scene in which people are moving.

Furthermore, FIG. 10 is a diagram showing the energy probability density distribution in the smoke region and the non-smoke region in the first embodiment of the present disclosure. Specifically, FIG. 10(a) is the energy probability density distribution calculated for a two-dimensional GLCM, and FIG. 10(b) is the energy probability density distribution calculated for a three-dimensional GLCM.

In addition, in each of Figures 10(a) and 10(b), the probability density distribution of non-uniformity within the smoke region is shown as "distribution 1," and the probability density distribution of non-uniformity within the non-smoke region is shown as "distribution 2."

As shown in Figure 10, distribution 1, which is the probability density distribution of energy within a smoke region, tends to have large energy. For this reason, it can be seen in both Figures 10(a) and 10(b) that the peak of distribution 1 is shifted to the right compared to distribution 2. Furthermore, a comparison of Figures 10(a) and 10(b) shows that the texture features calculated based on the three-dimensional GLCM more clearly show the difference between distribution 1 and distribution 2.

Therefore, the smoke generation detection unit 23 can quantitatively determine whether or not smoke has occurred in the monitored area from the probability density distribution of the texture features related to energy calculated by the feature extraction unit 22.

In the first embodiment, a detailed description has been given of a case in which a GLCM is created using an EMA difference image as an input image to calculate a feature quantity focusing on the luminance fluctuations specific to smoke, and a statistical process is then performed to quantitatively determine whether or not smoke has occurred in a monitored area. However, the smoke detection device according to the present disclosure is not limited to such a configuration.

Therefore, we provide additional explanation below on a configuration that uses an EMA difference image but does not use GLCM, and a configuration that uses GLCM without using an EMA difference image.

<Configuration using EMA difference image without using GLCM>
Such a configuration can be realized by configuring the detection processing unit 20 as a pre-processing unit 21 and a smoke generation detection unit 23 as described below.

During monitoring, images of the monitored object captured by the monitoring camera 1 at a predetermined sampling period are stored in the image memory 10 as time-series images of the monitored object.

The pre-processing unit 21 sequentially calculates an EMA image consisting of pixels that are the result of performing an exponential moving average on the time-series monitored image for each pixel. Furthermore, the pre-processing unit 21 compares the brightness value of the EMA image for each pixel with the brightness value of the monitored image captured when the EMA image was calculated or in the next sampling period, and generates a binary image based on the magnitude of the brightness value.

Then, the smoke generation detection unit 23 determines whether or not smoke has been generated in the smoke candidate region in the monitored image based on the binary image generated in the smoke candidate region by the preprocessing unit 21.

In other words, the detection processing unit 20 performs this type of calculation processing, and without calculating the GLCM for the EMA difference image as shown in Figure 2 above, calculates features from the EMA difference image that focus on the brightness fluctuations that are unique to smoke, thereby being able to identify the state shown in Figure 2 (b), and preventing erroneous recognition of smoke.

<Configuration using GLCM without using EMA difference image>
Such a configuration can be realized by configuring the detection processing unit 20 as a feature extraction unit 22 and a smoke generation detection unit 23 as described below.

The feature extraction unit 22 creates a co-occurrence matrix using the luminance values of each pixel in the smoke candidate region for the monitored image captured by the surveillance camera 1. Furthermore, the feature extraction unit 22 calculates texture features based on the created co-occurrence matrix.

Specifically, the feature extraction unit 22 uses the monitored image captured by the surveillance camera 1 instead of the EMA difference image, converts the monitored image, for example an image with 256 gradations of luminance, into a binary or higher grayscale image with 256 gradations or less, and then creates a co-occurrence matrix.

Then, the smoke generation detection unit 23 determines whether smoke has been generated in the smoke candidate region in the monitored image based on the texture features calculated by the feature extraction unit.

In other words, even by performing this type of calculation processing and creating a GLCM without using an EMA difference image, the detection processing unit 20 can determine whether smoke has occurred by calculating features that focus on the luminance fluctuations that are unique to smoke, thereby preventing erroneous recognition of smoke.

1 Surveillance camera, 10 Image memory, 20 Detection processing unit, 21 Preprocessing unit, 22 Feature extraction unit, 23 Smoke generation detection unit.

Claims (4)

A smoke detection device that detects smoke generation by performing image processing on a smoke candidate region in a monitoring target image captured by a monitoring camera,
a feature extraction unit that calculates a co-occurrence matrix using a luminance value of each pixel in the smoke candidate region for the monitoring target image captured by the monitoring camera, and calculates a texture feature based on the calculated co-occurrence matrix;
a smoke generation detection unit that determines whether or not smoke has been generated in the smoke candidate region in the monitoring target image based on the texture feature calculated by the feature extraction unit,
The image memory stores the images of the object to be monitored captured by the monitoring camera as time-series images of the object to be monitored at each predetermined sampling period,
The feature extraction unit calculates the co-occurrence matrix by using the luminance values of each pixel in the smoke candidate region in two or more chronologically adjacent monitoring target images included in the time-series monitoring target images, taking into consideration an adjacency relationship between a pixel of interest in one image and adjacent pixels to the pixel of interest, and further taking into consideration an adjacency relationship with pixels at the same position as the pixel of interest in previous and subsequent frames.
Smoke detection devices. A smoke detection device that detects smoke generation by performing image processing on a smoke candidate region in a monitoring target image captured by a monitoring camera,
an image memory for storing, during monitoring, the images of the object to be monitored captured by the monitoring camera at each predetermined sampling period as time-series images of the object to be monitored;
a pre-processing unit which sequentially calculates, for each pixel, an EMA image consisting of pixels that is a result of performing an exponential moving average using the time-series monitoring target image, and compares, for each pixel, a luminance value of the EMA image with a luminance value of the monitoring target image captured when the EMA image was calculated or in the next sampling period, and generates a binary image based on the magnitude of the luminance value of the EMA image;
a smoke generation detection unit that determines whether smoke has been generated in the smoke candidate region in the monitored image based on the binary image generated within the smoke candidate region by the preprocessing unit. A feature extraction unit is further provided which calculates a co-occurrence matrix using the binary values of each pixel in the binary images sequentially generated within the smoke candidate region by the pre-processing unit, and calculates a texture feature based on the calculated co-occurrence matrix,
The smoke detection device according to claim 2 , wherein the smoke generation detection unit determines whether or not smoke has been generated in the smoke candidate region in the monitored image based on the texture feature calculated by the feature extraction unit, instead of using the binary image. the preprocessing unit stores the binary images sequentially generated for each sampling period in the image memory as time-series binary images;
The smoke detection device according to claim 3 , wherein the feature extraction unit calculates the co-occurrence matrix using binary values of each pixel in the smoke candidate region of two or more chronologically adjacent binary images included in the time- series binary image.