JP7503174B2 - Smoke detection device and method - Google Patents

Description

The present invention relates to a smoke detection device and a smoke detection method that detect smoke caused by a fire from an image of a monitored area captured by a camera.

Conventionally, various devices and systems have been proposed that detect fires by applying image processing to images of a monitored area captured by a surveillance camera.

With such smoke detection devices, early detection of fires is important from the perspective of early fire extinguishing and evacuation guidance.

For this reason, the conventional device (Patent Document 1) derives from an image the following phenomena caused by smoke associated with a fire: a decrease in transmittance or contrast, convergence of brightness values to a specific value, a decrease in brightness variance due to a narrowing of the brightness distribution range, a change in the average brightness value due to smoke, a decrease in the sum of edges, and an increase in intensity in the low-frequency band, and makes a comprehensive judgment of these to detect smoke.

However, conventional fire detection systems that detect fires from images of smoke associated with such fires process the entire image captured by a surveillance camera to detect characteristic changes caused by smoke and determine whether or not there is a fire, which increases the processing load required to determine whether there is a fire from the entire image, resulting in a problem of increased processing time.

To solve this problem, in Patent Document 2, the smoke rising from a fire contains areas of high and low density, and since there is a phenomenon in which areas of high density smoke appear one after another and move upward, classes are placed in the areas of high density smoke from the captured image, for example according to the rules of a particle filter, and each class is tracked for each image to obtain position information, and the characteristic upward movement caused by the smoke is detected, making it possible to reliably identify a fire.

JP 2008-046916 A JP 2017-191544 A Junpei Amano and 4 others, Study on visualization and detection of smoke using features of spatio-temporal domain, IEEJ Research Papers, Perception Information Research Session PI-18-049-060, Next Generation Industrial Systems Research Session IIS-18-030-041, July 30, 2018, pp.37-41

However, in such conventional methods of using particle filters to place classes in smoke regions and track them, when scattering particles in the target image to place classes, there is a problem that the particles concentrate in one place, making it impossible to place and track multiple classes. Therefore, by shining multiple spot lights onto the monitored area and extracting and tracking images of areas irradiated by spot lights scattered in the image of the monitored area captured by the surveillance camera, it is possible to generate and track multiple classes.

However, a special spot irradiation device is required to shine multiple spot lights onto the surveillance area, and an infrared light source is used to make it difficult to tell that the spot lights are hitting the area, meaning that it cannot be used in a lighting environment with normal visible light, resulting in a lack of versatility.

The present invention aims to provide a smoke detection device and method that targets images captured in a visible light lighting environment, prevents particles from concentrating in one place when tracking classes using a particle filter, and arranges multiple classes that indicate smoke regions in a time series to capture movements specific to smoke, enabling early detection of fires.

(Smoke detection device)
The present invention provides a smoke detection device that assigns a class corresponding to a smoke region in each image captured by an imaging means and detects a specific characteristic change due to smoke from a movement trajectory of the class, comprising:
When rearranging each class arranged in the first image to correspond to the second image in a second image that follows in chronological order from the first image in which the arrangement of the classes has been completed, a predetermined number of particles are scattered for each class with a scattering center set to a predetermined position within an area above the position of the center of gravity of the class in the first image, and the class is rearranged with the center of gravity of the scattered predetermined particles set to the center of gravity of the class in the second image;
The distance from the center of gravity of the class in the first image to the scattering center is changed in accordance with the movement speed corresponding to the color of the smoke.

(Smoke detection method)
Another aspect of the present invention is a smoke detection method for arranging a class corresponding to a smoke region in each image captured by an imaging means, and detecting a specific characteristic change due to smoke from a movement trajectory of the class, the method comprising:
When rearranging each class arranged in the first image to correspond to the second image in a second image that follows in chronological order from the first image in which the arrangement of the classes has been completed, a predetermined number of particles are scattered for each class with a scattering center set to a predetermined position within an area above the position of the center of gravity of the class in the first image, and the class is rearranged with the center of gravity of the scattered predetermined particles set to the center of gravity of the class in the second image;
The distance from the center of gravity of the class in the first image to the scattering center is changed in accordance with the movement speed corresponding to the color of the smoke.

(Effectiveness of smoke detection devices)
The present invention is a smoke detection device that, for each image captured by an imaging means, arranges a class corresponding to a smoke region in the image and detects a specific characteristic change due to smoke from the movement trajectory of the class. When rearranging each class arranged in a first image to correspond to a second image that follows in chronological order from a first image in which class arrangement has been completed, a specific number of particles are scattered for each class, with a specific position in an area above the center of gravity of the class in the first image as the scattering center, and the center of gravity of the scattered specific particles is set as the center of gravity of the class in the second image to determine the class. Since the movement of smoke between images changes depending on the speed at which the smoke rises, when the smoke moves quickly, the distance from the center of gravity of the class to the center for scattering particles is increased during class rearrangement, and the distance is shortened when the smoke moves slowly. This enables particles to be scattered in an optimal area for class rearrangement in accordance with the speed of movement of the smoke, improving the accuracy of class tracking through class rearrangement and enabling more accurate detection of class movement trajectories.

The effect of the smoke detection method is the same as that of the smoke detection device described above.

FIG. 1 is an explanatory diagram showing a monitoring area in which a smoke detection device according to the present invention is installed. A block diagram showing an outline of the functional configuration of a smoke detection device. An explanatory diagram showing a frame image captured by a surveillance camera and its difference image. An explanatory diagram showing the procedure for placing new classes by scattering initial particles in the first smoke region that appears. An explanatory diagram showing the particle scattering area when class rearrangement is performed. FIG. 5 is an explanatory diagram showing a new class arrangement following the class rearrangement of FIG. 4. A diagram showing the time series changes in class placement and class movement trajectories for differential images that change over time. Diagram showing rearrangement of stuck classes 3 is a flowchart showing a smoke detection process performed by the smoke detection device of FIG. 2. A flowchart showing details of class tracking using a particle filter in step S2 of FIG.

[Basic Concept of the Embodiment]
FIG. 1 is an explanatory diagram showing a monitoring area in which a smoke detection device of the present invention is installed, and FIG. 2 is a block diagram showing an outline of the functional configuration of the smoke detection device.

The basic concept of this embodiment is to input moving images captured by a surveillance camera 10, which functions as an imaging means, into a smoke detection device 12, and to detect a fire at an early stage. From the images input into the smoke detection device 12, a difference image is generated for each of the multiple images by a difference image generation unit 24, which functions as a difference image generation unit, and a class tracking unit 26, which functions as a class tracking unit using a particle filter, scatters multiple particles in the image according to a predetermined rule for each difference image, and weights each particle according to the likelihood that it is smoke-like, and determines the center of gravity of the smoke region, repeating the process of arranging classes by tracking the classes and detecting their movement trajectories. Next, a fire judgment unit 28, which functions as a fire judgment unit, detects a characteristic change due to smoke from the movement trajectory of the class detected by the class tracking unit 26, and judges a fire.

In particular, in the case of a difference image in which no classes have been placed, the class tracking unit 26 randomly scatters multiple particles in the image as an initialization process to place a new class; in the case of a difference image in which a class has been placed, the class is rearranged by scattering multiple particles according to a predetermined probability density distribution centered on a predetermined position above the position of each particle when the class was previously placed; and after rearranging all classes, the new class is placed by randomly scattering multiple particles in the area excluding the class.

In this way, when the class tracking unit 26 rearranges classes, it scatters particles mainly above the position of each particle when the class was previously placed, and then, once rearrangement of all classes is complete, scatters particles in areas other than the class placement area to place new classes. This makes it possible to continuously place and rearrange classes without mutual interference in a mass of smoke of a certain density that rises as a smoke vortex (smoke area), and to reliably determine that it is a fire by capturing the unique characteristics of smoke from the movement trajectories of multiple classes.

In addition, because smoke generally does not easily descend, when rearranging a class, particles are scattered in an area slightly above the position of each particle when the class was previously placed, which reliably prevents particles of different classes from overlapping, and makes it possible to detect the movement trajectory of a class that accurately tracks the rising of the smoke cloud. This is explained in detail below.

[Overview of Smoke Detection Device]
1, a surveillance camera 10 is installed in a surveillance area 14, and the surveillance camera 10 sequentially captures the surveillance area illuminated by lighting fixtures at a frame rate of, for example, 30 frames per second or 60 frames per second, and outputs a video consisting of a series of color frame images. The surveillance camera 10 is installed, for example, in the center of an upper corner of a surveillance area (surveillance space) 14 that is partitioned into top and bottom, left and right, and front and back, and is positioned with its imaging optical axis facing diagonally downward toward the opposing wall surface, enabling it to capture the entire surveillance area 14.

The surveillance camera 10 captures video of visible images of the surveillance area 14, and the video captured by the surveillance camera 10 is transmitted via a transmission line to a smoke detection device 12 installed in a caretaker's room or the like, which uses image processing with a particle filter to detect smoke 16 rising from a fire source 15 such as a trash can, determines that there is a fire, and transmits a fire detection signal to a fire alarm system 18 to output a fire alarm. The size of the video captured by the surveillance camera 10 is, for example, 1280 x 720 pixels.

[Functional configuration of the smoke detection device]
2, the smoke detection device 12 is composed of hardware such as a computer circuit equipped with a CPU, memory, various input/output ports, etc., and is provided with a control unit 20, an image input unit 22, a differential image generation unit 24, a class tracking unit 26, and a fire determination unit 28 as functions realized by the CPU executing a program. In addition, an alarm display unit 30 and an audio alarm unit 32 are provided to issue an alarm when a fire is determined.

The control unit 20 performs overall control of each function provided in the surveillance camera 10 and the smoke detection device 12. The surveillance camera 10 operates upon receiving instructions from the control unit 20, and transmits image data of the monitored area as moving image data, for example at 30 frames per second, through transmission control. The image data is received by the image input unit 22 of the smoke detection device 12 and stored in a memory (not shown). Under instructions from the control unit 20, the image input unit 22 reads out the stored moving images frame by frame and outputs them to the difference image generation unit 24.

[Difference image generation unit]
FIG. 3 is an explanatory diagram showing a frame image captured by a surveillance camera and a difference image thereof. FIG. 3(A) shows an image of smoke, and FIG. 3(B) shows a difference image of smoke.

The difference image generating unit 24 generates a difference image according to the following procedure.
(Step 1) The frame image is converted to grayscale and smoothed using a Gaussian filter to reduce noise.
(Step 2) The maximum and minimum values of each pixel in the most recent frames, for example 30 frames, including the current frame image, are monitored, and maximum and minimum value images for the 30 frames are created.
(Step 3) A difference image is created from the maximum value image and the minimum value image generated in step 2.

Note that the generation of the difference image is not limited to steps 1 to 3, and any appropriate method for generating the difference image can be applied.

The length of 30 frames used to calculate the difference is the processing length for images of smoke with a slow flow velocity. For smoke with a fast flow velocity, for example, six frames are required. In actual processing, this is handled by parallel processing that creates a difference image using multiple frames corresponding to the flow velocity.

As shown in FIG. 3(A), in smoke image 16a of frame 34, which captures smoke 16 rising from the fire source 15 shown in FIG. 1, the smoke has the characteristic of rising while fluctuating. When focusing on a point in the image where smoke is present, the presence or absence of smoke changes on the line connecting the viewpoint and the background, and it is considered that there are a state where the image density is the original background and a state where the density is the smoke. Using such characteristics in the time-space domain, smoke is visualized by differential processing.

These subtraction processes of steps 1 to 3 generate a smoke subtraction image 16b shown in the subtraction frame 36 of FIG. 3(B), which visualizes the appearance of multiple smoke regions, each of which is a mass of smoke with a certain density, rising in a time series. Note that in the subtraction frame 36, the fire source image 15a disappears as indicated by the dotted line.

[Class Tracking Department]
FIG. 4 is an explanatory diagram showing the procedure for placing new classes by scattering initial particles in the first-appearing smoke region, FIG. 5 is an explanatory diagram showing the particle scattering region during class rearrangement, and FIG. 6 is an explanatory diagram showing the placement of new classes following the class rearrangement of FIG. 4.

(Class Tracking Division Overview)
The class tracking unit 26 uses a particle filter to arrange smoke regions that are masses of smoke as classes in the difference image, and detects the movement trajectories of the classes by tracking them in time series.

The class tracking unit 26 repeats the processes of sampling, prediction, and observation, which are the processing steps of a particle filter. That is, for each difference image generated by the difference image generation unit 24, the class tracking unit 26 scatters multiple particles in the image according to a predetermined rule, weights each particle according to the likelihood that it is smoky, finds the center of gravity of the smoky region, and repeats the process of arranging classes, tracking the classes and detecting the movement trajectory. Here, one particle corresponds to one pixel, for example, and the particle's density value is used as the likelihood, and weighting is performed according to the pixel value.

The size of a particle may be a predetermined number of pixels, in which case the sum of the pixel values of the pixels contained in the particle is weighted as the likelihood.

(Initial class placement)
If no class has been placed in the previous processing, the class tracking unit 26 performs processing to place a new class in the read differential image by randomly scattering a predetermined number of particles, for example, 40 particles, in the image.

For example, as shown in FIG. 4, for processing frame 38-1, where a smoke area 41, which is surrounded by a dotted line and is the first mass of smoke, appears, 40 particles 40 are randomly scattered throughout the entire image.

Next, as shown in processing frame 38-2, the pixel value of each particle is weighted as a likelihood, with low-likelihood particles represented as white circles and high-likelihood particles represented as black circles. As shown in processing frame 38-3, 34 low-likelihood particles are eliminated, leaving high-likelihood smoke-estimated particles 42. A center of gravity calculation is then performed on the weight value of the remaining smoke-estimated particles 42, resulting in the placement of class 44-1 with center of gravity 46-1, as shown in processing frame 38-4.

Here, class 44-1 indicates a smoke region that is a mass of smoke, but in reality it is a collection of smoke-estimated particles 42, and for ease of explanation, it is shown as a circular region.

(Class rearrangement)
4, after arranging the class in the first smoke region, when the next difference image is read, the class tracking unit 26 performs a process of scattering particles at coordinate positions generated according to a predetermined probability density distribution, for example, a normal distribution random number, for the six particles remaining in the previous class arrangement, with a center above the position (center of gravity) 46-1 of each particle at the previous arrangement, determined by the likelihood of the particle serving as the position reference, so that the total number of rearranged particles is 40, and rearranges the class with a total of 40 particles scattered. At this time, the higher the likelihood of the particle at the previous arrangement, the greater the number of particles to be rearranged based on that position.

Here, the distance L from the previously placed position (center of gravity) 46-1 of each particle to the scattering center 48 is changed according to the movement speed corresponding to the color of the smoke, or the movement speed of the past class. This is because the movement of smoke between images changes depending on the speed at which the smoke rises, so if the movement speed of the smoke is high, the distance L from the center of gravity 46-1 of the class to the scattering center 48 of the particles is lengthened when rearranging the class, and if the movement speed of the smoke is low, the distance L is shortened, so that the particles can be scattered in an area optimal for class rearrangement according to the movement speed of the smoke.

For example, black smoke from a combustion fire rises quickly, so by lengthening the distance L to the particle scattering center, and white smoke from a smoldering fire rises slowly, by shortening the distance L to the particle scattering center, accurate class reassignment can be achieved in line with the movement of the smoke.

Furthermore, the class tracking unit 26 designates the area below the horizontal line 52 passing through the position (centre of gravity) 46-1 of each particle previously placed in the scattering area 50 shown in FIG. 5, which is the area for the relocation of the class, as a non-scattering area 54 as shown by the dotted line, excludes the particles present in the non-scattering area 54 from the calculation of the centroid position of the class to be relocated, and performs the process of finding the centroid from the likelihood of the particles present in the scattering area 50 present above the horizontal line 52 and relocating them.

Particles present in the non-scattering region 54 can be excluded by setting the particle weight to zero. Alternatively, the center of gravity of the class may be calculated as particles in the scattering region 50 by moving all particles present in the non-scattering region 54 onto the horizontal line 52.

When rearranging classes in this way, particles located below the center of gravity of the class in the previous image are excluded, or the class is rearranged by moving to the center of gravity of the class. Since smoke does not normally move downwards, by rearranging the class so that it does not move downwards, it is possible to detect the movement trajectory of the class specific to smoke.

Processing frames 38-5 and 38-6 in Figure 6 show the relocation of classes for the next differential image following processing frame 38-4 in Figure 4. Processing frame 38-5 shows a state in which particles are scattered so that the total number is 40, centered on a specified position above the position (centre of gravity) 46-1 of each particle at the time of previous placement. As shown in the next processing frame 38-6, particles with low likelihood are eliminated, particles with high likelihood shown by black circles are left, and the centre of gravity 46-2 of the class of particles with high likelihood is found, resulting in class 44-2 being relocated.

When rearranging the classes, the distance L from the previously placed position (center of gravity) 46-1 of each particle shown in Figure 5 to the scattering center 48 is changed according to the speed of movement of the smoke. However, the distance L can be made approximately constant by changing the number of frames from which the difference is taken in the difference image generation unit 24 shown in Figure 2 according to the speed of movement of the smoke.

For example, if the differential image generating unit 24 generates differential images in units of 30 frames for smoke with a slow flow velocity and in units of 6 frames for smoke with a fast flow velocity, the amount of movement of the smoke region in the differential image in units of 30 frames with a slow flow velocity and in the differential image in units of, for example, 6 frames with a fast flow velocity can be made approximately the same, simplifying the processing of the class tracking unit 26.

(Initial class placement after class reassignment)
When the class tracking section 26 has finished rearranging all the classes in an image in which classes have been arranged in the previous image, it performs class initialization arrangement in which 40 particles are randomly scattered in areas other than the rearranged class areas to arrange new classes.

In the class rearrangement shown in processing frames 38-5 and 38-6 in Figure 6, one class was placed in the immediately preceding image, and class rearrangement is completed in one processing run. When class rearrangement is completed in this manner, as shown in processing frame 38-7, 40 particles 40 are randomly scattered in the area excluding class 44-2, which has the center of gravity 46-2 of the rearranged class, and the particles are weighted according to the likelihood of their pixel values in the same manner as in the class initialization arrangement shown in Figure 4 to calculate the center of gravity, and if a new smoke area appears, a new class can be placed there.

(Class movement trajectory detection)
The class tracking unit 26 performs processing to detect the movement trajectory of classes moving in a time series by repeating class allocation and class rearrangement on the difference images read in a time series.

Figure 7 (A) shows processing frames 60-1 to 60-5 that were processed in chronological order at times t1 to t5, with classes C1 to C5 being sequentially arranged and rising due to rearrangement.

As a result, for example, the class tracking unit 26 can detect the movement trajectories of the class centers of gravity Gt1 to Gt5 in class C1 shown in FIG. 7(B) for class C1. The movement trajectories of the class centers of gravity are similarly detected for the remaining classes C2 to C5.

(Removal of stuck classes)
As shown in the processing frame 60-5 of FIG. 7A, class C1, which has been moved due to the rising of smoke, is not rearranged when it reaches the top end of the image and remains there.

When the class tracking unit 26 finishes the class relocation process, if it detects a stuck class C1, as shown in processing frame 60-5 in Figure 8, it erases (resets) class C1, as shown by the dotted line, and makes it available for allocation to a new class.

Regarding reuse after erasing a class that has stayed, for example, assuming the maximum number of classes is 5, as shown in FIG. 8, when class C1 is detected as staying and erased in processing frame 60-5, as shown in processing frame 60-6 for the next sub-image, the area below the smoke area of classes C2 to C5 that exist at that time, for example the area below the horizontal line 62 that passes below class C5, is set as a scattering area 64, and 40 particles 40 are randomly scattered to generate a new class, and the newly generated class is placed as class C1 that was erased due to staying, as shown in processing frame 60-6 for the next sub-image.

In addition, when class C1 remains and is deleted, if there are no other classes, the area below the horizontal line that passes under the deleted class C1 is designated as the scattering area 64, and initialization processing is performed to place a new class.

In this embodiment, the maximum number of classes that can be placed by the class tracking unit 26 is set to, for example, 16 classes, but by deleting and rearranging such stagnant classes and arranging classes in a cyclical manner, it is possible to arrange classes without being restricted by the maximum number of classes.

(Effect of relocation based on class priority)
As shown in processing frames 60-1 to 60-5 in FIG. 7, when the class tracking unit 26 repeatedly places and rearranges classes for smoke regions that appear one after another, it sets a different priority for each class and rearranges them in order from the highest class, and performs a process of finding the center of gravity from the likelihood of the remaining particles, excluding particles that overlap with the smoke regions of the highest classes that have already been rearranged, from among the multiple particles scattered for class rearrangement, and rearranging them.

For example, if a higher priority is set for a class that has been created earlier, taking the example of relocation from processing frame 60-4 to processing frame 60-5 in Figure 7, the classes C1 to C4 will be assigned higher priorities in that order.

At this time, the class tracking unit 26 first rearranges the highest priority class C1, but the particle scattering range is not limited by the other lower priority classes C2 to C4. Next, the class tracking unit 26 rearranges the next highest priority class C2. At this time, the smoke region of the high priority class C1 that has been rearranged is masked, and the particles scattered for the rearrangement of class C2 are removed from the smoke region of class C1, so that the smoke regions of classes C1 and C2 do not overlap.

This allows classes to be relocated without overlapping in response to continuously rising smoke regions, allowing the movement trajectories of the classes to be detected accurately.

[Fire Judgment Department]
The fire determination unit 28 detects a specific characteristic change caused by smoke based on the movement trajectory of the center of gravity position of each class detected by the class tracking unit 26, and determines the occurrence of smoke caused by a fire.

The fire judgment unit 28 also sets accumulation conditions for fire judgment based on the movement trajectory of the class, and if, for example, the movement trajectory of the class continues for a predetermined period of time or more, it judges that smoke has been generated due to a fire and outputs a fire alarm while also transmitting a fire detection signal to the outside.

[Processing operation of the smoke detection device]
FIG. 9 is a flowchart showing an outline of smoke detection processing by the smoke detection device of FIG. 2, and FIG. 10 is a flowchart showing the details of class tracking by the particle filter in step S2 of FIG.

(Overview of smoke detection processing operation)
As shown in Figure 9, the control unit 20 of the smoke detection device 12 shown in Figure 2 operates the differential image generation unit 24 in step S1, detects the maximum and minimum values of each pixel for a predetermined number of images captured by the surveillance camera 10 from the image input unit 22, creates a maximum value image and a minimum value image across multiple frames, and generates a differential image from the maximum value image and the minimum value image.

The control unit 20 then proceeds to step S2, where it operates the class tracking unit 26, and for each difference image generated by the difference image generation unit 24, scatters multiple particles in the image according to a predetermined rule, weights each particle according to its likelihood of being smoke, and repeats the process of locating classes by finding the center of gravity of the smoke region by weighting each particle according to its likelihood of being smoke, and tracks the classes to detect a movement trajectory indicating the movement of the center of gravity of the class.

Then, the control unit 20 proceeds to step S3, operates the fire determination unit 28, and if the movement trajectory of the center of gravity of the class detected by the class tracking unit 26 continues for a predetermined period of time or more, determines that the fire is smoke caused by a fire.

Then, the control unit 20 proceeds to step S4, where it determines whether the fire judgment in step S3 satisfies a predetermined accumulation condition. If the accumulation condition is met, for example, a predetermined number of consecutive fire judgments, it proceeds to step S5, where it operates the alarm display unit 30 and the audible alarm unit 32 to output a fire alarm, and also transmits a fire detection signal to the fire alarm system 18 shown in FIG. 1 to output a fire alarm.

(Class Tracking Processing)
The class tracking unit 26, which has been operated in response to an instruction from the control unit 20 in step S2 of Fig. 9, performs the class tracking process shown in Fig. 10. The class tracking unit 26 reads the difference image generated by the difference image generation unit 24 in step S11, and if a smoke region appears for the first time in the difference image, it determines in step S12 that there is no existing class since no class has been allocated in the processing up to that point, and proceeds to step S13 to perform class allocation as initialization processing.

In the class placement as initialization processing by the class tracking unit 26, in step S13, for example, 40 particles are randomly scattered in the difference image as shown in FIG. 4, in step S14, the density value of each particle is set as a likelihood and a weight is set, particles with low likelihood are deleted, and particles with high likelihood are left, in step S15, the center of gravity of the weights of the remaining particles is calculated, a new class with the found center of gravity is placed, in step S16 the center of gravity (coordinate position) of the placed class is stored, and the process returns to the main routine in FIG. 10.

On the other hand, if the class tracking unit 26 determines in step S12 that an existing class exists, the process proceeds to step S17, where the class rearrangement process is performed. In the class rearrangement process by the class tracking unit 26, the center of gravity of the existing class is extracted in step S17. Here, if multiple classes have already been placed, the class priority is set in order of oldest to newest, so the class tracking unit 26 extracts the center of gravity of the class with the highest priority (the class that appeared the oldest).

Then, the class tracking unit 26 proceeds to step S18 to determine whether the class from which the center of gravity was extracted is a non-smoke-like class, and if it is not a smoke-like class, proceeds to step S19 to delete the class and make it available for new class placement. For example, as shown in processing frame 60-5 in Figure 7, a class that has reached and remained at the top of the screen is deleted as a non-smoke-like class. Also, if the width of the distribution of particles used to calculate the class center of gravity is wider than a specified range (if the standard deviation exceeds a specified value), it is deleted as a non-smoke-like class.

When the class tracking unit 26 determines in step S18 that the class is likely to be smoke, it proceeds to step S20, where, as shown in FIG. 5, in addition to the particles left in the previous process, it sets a scattering center 48 at a predetermined distance L above the center of gravity 46-1 of the class, generates coordinates using normal distribution random numbers for the number of particles deleted in the previous process, and scatters them. In step S21, it sets weights using the concentration value of each particle as likelihood. In step S22, it deletes particles with low likelihood, leaving particles with high likelihood, calculates the center of gravity of the weights of the remaining particles, and rearranges the class. It then proceeds to step S23 and stores the center of gravity (coordinate position) of the rearranged class.

The class tracking unit 26 then proceeds to step S24 to determine whether rearrangement of all classes has been completed, and if not, returns to step S17, extracts the center of gravity of the next highest priority class, and rearranges the classes using steps S20 to S23.

When the class tracking unit 26 determines in step S24 that rearrangement of all classes has been completed, it proceeds to step S26, and when it determines that a class has been deleted in the class rearrangement processing of steps S17 to S25, it proceeds to step S27, where it randomly scatters 40 particles below the existing classes as shown in processing frame 60-6 in Figure 8, weights the particle concentration values as likelihood in steps S14 to S16, deletes particles with low likelihood, leaves particles with high likelihood, and performs initialization processing to determine the center of gravity position to place a new class.

When there are no other classes other than the deleted class, the scattering of particles in step S27 involves randomly scattering 40 particles below the deleted class, weighting the particle concentration values as likelihood in steps S14 to S16, eliminating particles with low likelihood, leaving particles with high likelihood, and determining the center of gravity to place the new class.

If the class tracking unit 26 determines in step S26 that no classes have been deleted, it proceeds to step S28, where it randomly scatters 40 particles into the empty space of the existing classes that have been rearranged, as shown in processing frame 38-7 in FIG. 6, and performs initialization processing to place new classes using steps S14 to S16. Note that the scattering of particles in step S28 can be done by randomly scattering 40 particles while masking the smoke areas of the existing classes.

[Modifications of the present invention]
(Class placement)
In the above embodiment, when a stuck class is deleted and a new class is placed, the class is placed by randomly scattering particles below the area in which other classes exist, or below the deleted class if no other classes exist; however, this is not limited to this, and for example, a predetermined number of particles may be scattered uniformly and randomly over the entire screen.

(Fire detection)
In addition, in the above embodiment, a particle filter is used to place classes in characteristic regions to track characteristic time series changes caused by smoke, but this is not limited to this, and other appropriate filter processing may be used to track time series changes that indicate characteristic movements caused by smoke and determine whether a fire has occurred.

(Image processing in normal and smoke-generated conditions)
Also, the number of images to be processed may be reduced under normal circumstances (for example, one image per five captures), and when it is determined that there is a high possibility of smoke, the number of images to be processed may be increased (all captured images). Here, the number of images specifically means the number of frames.

In addition, if it is determined that there is a high possibility of smoke, all images may be processed again, starting from a time slightly before it was determined that there was a high possibility of smoke, including images that were not subject to processing (for example, the remaining four images out of five images taken).

Furthermore, as another type of image processing, a process of simply changing the imaging interval may be used, for example, lengthening the imaging interval under normal circumstances and shortening the imaging interval when it is determined that there is a high possibility of smoke.

(others)
Furthermore, the present invention is not limited to the above-described embodiment, but includes appropriate modifications that do not impair the objects and advantages of the present invention, and is not limited to the numerical values shown in the above-described embodiment.

10: Surveillance camera 12: Smoke detection device 14: Surveillance area 15: Fire source 16: Smoke 16a: Smoke image 16b: Smoke difference image 18: Fire alarm system 20: Control unit 22: Image input unit 24: Difference image generation unit 26: Class tracking unit 28: Fire judgment unit 30: Alarm display unit 32: Sound alarm unit 34: Frame 36: Difference frame 40: Particle 42: Smoke estimated particle 44-1, 44-2: Class 48: Scattering center 50, 58, 64: Scattering area 54: Non-scattering area

Claims (2)

A smoke detection device that arranges a class corresponding to a smoke region in each image captured by an imaging means and detects a specific characteristic change due to smoke from a movement trajectory of the class, comprising:
When rearranging each class arranged in the first image in a second image chronologically following the first image in which the arrangement of the classes has been completed, so as to correspond to the second image, for each class, scattering a predetermined number of particles with a scattering center set to a predetermined position within an area above a center of gravity of the class in the first image, and rearranging the class with the center of gravity of the scattered predetermined particles set to the center of gravity of the class in the second image;
A smoke detection device characterized in that the distance from the center of gravity of a class in a first image to the scattering center is changed in accordance with a moving speed corresponding to the color of the smoke.
A smoke detection method for detecting smoke, the method comprising: arranging a class corresponding to a smoke region in each image captured by an imaging means; and detecting a specific characteristic change due to smoke from a movement trajectory of the class, the method comprising:
When rearranging each class arranged in the first image in a second image chronologically following the first image in which the arrangement of the classes has been completed, so as to correspond to the second image, for each class, scattering a predetermined number of particles with a scattering center set to a predetermined position within an area above a center of gravity of the class in the first image, and rearranging the class with the center of gravity of the scattered predetermined particles set to the center of gravity of the class in the second image;
A method for detecting smoke, comprising changing a distance from a center of gravity of a class in a first image to the scattering center in accordance with a moving speed corresponding to the color of the smoke.

Images