An optimized early fire detection method and early fire detection device

Abstract

The present invention discloses an optimized early fire detection method and an early fire detection device thereof. The optimized early fire detection method includes the following steps: acquiring a picture stream in real time; calculating the global energy of each picture in the global; judging whether the global energy difference between the current picture and the previous frame picture is greater than a preset global energy difference; if so, checking whether the current picture is interfered with; if not, increasing the shooting cycle to start high-speed shooting, calculating the displacement field of the current picture relative to the previous frame picture in the picture stream after high-speed shooting; judging whether the displacement field is within the preset displacement field range; if so, starting a fire alarm. The present invention optimizes the corresponding global energy model and designs a matching penalty function to improve the accuracy of the global energy difference, thereby solving the technical problem that the alarm accuracy of the fire early detection method based on the flow field schlieren display technology of the existing fire protection system needs to be improved.

Description

Technical Field

[0001] The present invention relates to a real-time detection method and a detection device for early-stage fires in large public buildings in the field of Internet of Things devices, and in particular to an optimized early-stage fire detection method and an optimized early-stage fire detection device. Background Art

[0002] Schlieren technology dates back to the 17th century. Hooke, while studying optically inhomogeneous media, first proposed a prototype device for schlieren imaging. Using the human pupil as a knife edge, he visualized the transparent convection plume of a candle flame through variations in light density. In the 1850s and 1860s, scholars used knife-edge schlieren to observe images of natural convection, including those of human plumes and ultrasonic stationary fields. High-speed camera technology emerged in the late 19th century. After decades of exploration and effort by engineers and scientists, and with the development and application of electronic technology in the first half of the 20th century, high-speed cameras gradually adopted electronic shutters to capture shorter time intervals. In the 1860s and 1970s, high-speed cameras were widely used in scientific research, military testing, engineering testing, and motion analysis, including the application of high speed in schlieren systems to analyze moving flow fields.

[0003] In fire detection technology, the first heat detector was developed in Britain in 1890. Heat detectors dominated until the 1940s. Heat detectors are suitable for locations generating large amounts of heat, but their sensitivity is relatively low, making them slow to detect fires. By the time an alarm is generated, the fire has already grown to an uncontrollable level. They are also unresponsive to small and weak fires, especially smoldering ones. In the 1950s, Swiss physicist Ernst Meili developed the prototype of the modern ionization smoke detector. Ionization detectors offer excellent response to smoldering smoke, enabling early detection of fires. In the late 1970s, breakthroughs in long-life photoelectric element technology led to the emergence of photoelectric smoke detectors. Both ionization and photoelectric smoke detectors have their advantages. Ionization smoke detectors are slow to respond to smoldering fires, while photoelectric smoke detectors have poor response to invisible smoke particles smaller than 0.4 μm. After the 1980s, with the development of computer technology, artificial intelligence, and single-chip microcomputers, fire detection technology began to intersect and integrate more widely with other technologies, moving towards independence, intelligence, and information technology. Researchers developed heat detectors using new semiconductors and thermosensitive materials, cable-type heat detectors for special locations like subways and tunnels, standalone smoke detectors for residential use, combined smoke and temperature detectors, intelligent detectors with built-in single-chip microcomputers, and wireless fire detectors using wireless networking technology.

[0004] With the increasing diversification and functionalization of buildings, their structures and installations are becoming increasingly complex. Fire hazards in large buildings are not only caused by problems with electrical circuits, gas pipelines, and heating and cooking equipment, but also by improper use of flammable items and tobacco products, or by false alarms. Existing technologies for early fire detection and control in large public buildings are inadequate for the complex and changing internal environments and suffer from various technical deficiencies. For example, heat detectors rely on smoke to heat the probe, often triggering an alarm only after a fire has developed significantly, delaying the optimal firefighting time. Smoke detectors, on the other hand, rely on the presence of smoke to trigger an alarm, making them ineffective for flammable alcohols, which produce little smoke when burned. Gas detectors can effectively detect the early stages of a fire, but they typically monitor carbon monoxide or carbon dioxide, limiting their detection targets and failing to detect early signs of fire or explosion, such as ruptured gas pipelines, leaks, or tank leaks. Infrared detection of flames in fire detection technology can detect the early stages of fire, but it is prone to false alarms for high-temperature objects. It relies on the generation of obvious flames. For flammable objects with unclear flames (methanol and ethanol flames are light blue or even colorless), fire image recognition technology cannot work well.

[0005] In schlieren images, fire plumes differ from plumes of general interference sources in terms of velocity, pulsation, edge entrainment, and flow pattern variations, all of which can be used to distinguish fires from non-fires. Applying an optical flow velocity measurement algorithm to the image can identify flame velocity pulsation. Speed can be used to determine fire severity. Using convolutional neural network architectures, training and simulation can be used to distinguish fires from non-fires.

[0006] Schlieren technology is applicable to fire detection because its imaging principle enables observation of temperature fields with density differences in the air. The very early stages of a fire are often accompanied by the release of heat energy, leading to the leakage of flammable and explosive gases, forming a transparent plume with a temperature and density difference from the surrounding environment. These plumes are invisible in traditional fire images, but can be observed with schlieren, and the plume's rising velocity is closely related to the heat release rate. Optimizing the optical flow algorithm used in schlieren velocity measurement enables real-time plume velocity measurement, combining it with the fire plume velocity to determine the fire's size and dynamics. This overcomes the inability of traditional fire detectors to issue fire alarm signals in the earliest stages of a fire, combining velocity characteristics to determine the fire's dynamics. Furthermore, it can respond to the heat plume generated by smoldering fires and detect the early signs of gas leaks, overcoming the limited functionality of traditional fire detection technology for detecting flammable and explosive gas leaks (which can only monitor leaks of a few or specific gases). Compared with modern fire detection technology, it has obvious advantages. The advantages are that it can detect the very early stage of fire and some smoldering fire phenomena. At the same time, it also has good image detection effects on the schlieren for various flammable and explosive gas leaks.

[0007] However, the alarm accuracy of the traditional early fire detection method based on flow field schlieren display technology still needs to be improved. In the final analysis, it is too sensitive and prone to false alarms, while insensitive and prone to missed alarms, thus losing credibility and ultimately reducing customer trust. Summary of the Invention

[0008] In order to solve the technical problem that the alarm accuracy of the existing fire early detection method of the fire protection system needs to be improved, the present invention provides an optimized fire early detection method and an optimized fire early detection device.

[0009] The present invention is implemented by the following technical solution: an optimized fire early detection method, which includes the following steps:

[0010] Acquire an image stream in real time, where each image in the image stream is a signal matrix consisting of M×N pixel units;

[0011] Calculate the global energy of each image in the world;

[0012] Determine whether the global energy difference between the current image and the previous frame is greater than the preset global energy difference. If yes, check whether the current image is interfered with. If not, increase the shooting cycle to start high-speed shooting.

[0013] Calculate the displacement field of the current image relative to the previous frame in the image stream after high-speed shooting;

[0014] Determine whether the displacement field is within a preset displacement field range, and if so, activate a fire alarm;

[0015] Among them, the corresponding global energy model optimization is:

[0016]

[0017] Where E(u,v) is the global energy of the image, I1(i,j) is the brightness of the first image at the pixel position (i,j) in the two adjacent frames, and I2(i+u i，j ,j+v i，j ) is the displacement of the second picture in two adjacent frames (u i，j ,v i，j ) is the brightness at the pixel position (i, j), λ is the regularization weight coefficient, ρ d and ρ s are data penalty function and space penalty function respectively; u i+1，j and v i+1，j are the vertical displacement component and horizontal displacement component at the pixel position (i+1, j), u i，j+1 and v i，j+1 are the vertical displacement component and the horizontal displacement component at the pixel position (i, j+1) respectively;

[0018] Among them, ρ d and ρ s The same penalty function is used: ρ(x) = (x 2 +∈ 2 ) a , x is the brightness residual or displacement error, ∈ is a small positive number taken to avoid the denominator being zero during calculation, and a is the exponent.

[0019] As a further improvement to the above solution, the method for checking whether the current image is interfered with is as follows:

[0020] Between multiple consecutive frames of adjacent images: Is the global energy difference between the current image and the previous frame amplified in a local area? If so, it can be considered that there is foreign interference in the local area. If the global energy difference is amplified and there is no increase in the global energy difference, it can be considered that the image is interfered with by other factors. If the global energy difference changes continuously and is greater than a preset value, it can be considered that the image is not interfered with.

[0021] As a further improvement of the above scheme, the displacement field (u i，j ,v i，j ) is to perform multi-layer Gaussian smoothing and downsampling on the two adjacent frames I1 and I2 to obtain the displacement field (u i，j ,v i，j ).

[0022] Furthermore, the displacement field (ui，j ,v i，j ) includes the following steps:

[0023] Decompose two adjacent frames of images I1 and I2 into k layers of images with different resolutions according to the image size, with the lowest resolution belonging to the highest layer and the original resolution belonging to the lowest layer;

[0024] A single displacement field (u is applied to the top layer images of two adjacent frames I1 and I2 at the pixel position (i, j) i，j ,v i，j )1 single energy E i，j (u i，j ,v i，j ) calculation, the corresponding single energy model is;

[0025]

[0026] E i，j (u i，j ,v i，j ) is interpolated downward to the initial result of the second layer image with higher resolution for each frame;

[0027] Warp the second layer image of the second frame image I2 to have the same scale as the second layer image of the first frame image I1 to obtain a warped second layer image of the second frame image I2;

[0028] Calculate the corresponding single displacement field (u) at the pixel position (i, j) of the distorted second layer image and the second layer image of the first frame image I1. i，j ,v i，j )2, and the corresponding single displacement field (u i，j ,v i，j )2 single energy Ei, j (u i，j ,v i，j ) is calculated, and the corresponding calculation results are interpolated downward to the initial results of the adjacent next layer image with higher resolution of each frame image, and the loop is iterated until the corresponding single displacement field (u i，j ,v i，j ) k , as the displacement field (u i，j ,v i，j ).

[0029] Furthermore, Gaussian filtering is used when decomposing each layer of the image.

[0030] Preferably, the Gaussian filter uses a 5×5 Gaussian distribution filter to perform filtering processing on each layer of image.

[0031] Furthermore, the displacement field (u i，j ,v i，j ) includes the following steps:

[0032] (1) Initialization: The initial result of the displacement field of each layer of image is set to (u i,j (0) ,v i,j (0) );

[0033] (2) Displacement field of the highest layer image: According to the initial displacement field result, the final result of the displacement field of the current layer (u i-1,j-1 (n) ,v i-1,j-1 (n) );

[0034] (3) From the next layer of the highest layer image to the lowest layer image:

[0035] Interpolate downward the displacement field:

[0036] The displacement field calculation results of the previous layer (u i,j (n) ,v i,j (n) ), interpolate and expand it to adapt to the size of the current layer image, and obtain the estimated result of the same size as the current layer image (u i-1,j-1 (n) ,v i-1,j-1 (n) );

[0037] (4) Iteratively optimize the current layer image:

[0038] a) According to the displacement field estimation value (u i-1,j-1 (n) ,v i-1,j-1 (n) ), calculate the energy residual of the current layer image:

[0039]

[0040]

[0041] b) For each pixel position (i, j), calculate the gradient of its regularization term:

[0042]

[0043]

[0044]

[0045]

[0046] c) Update the displacement field estimate (u) of each pixel position (i, j) of the current layer image i-1,j-1 (n+1) ,v i-1,j-1 (n+1) ):

[0047]

[0048]

[0049] Among them, (n) represents the nth iteration, (n+1) represents the n+1th iteration, α1 is the learning rate, E is the energy residual term, and R is the regularization term:

[0050]

[0051]

[0052] e) Repeat step (4) on the next layer of images until the iteration reaches the lowest layer of images;

[0053] (5) Return the displacement field estimation result u of the bottom layer image i,j (n+1) and v i,j (n+1) ;

[0054] (6) Finally, we get the vectors u and v of the entire displacement field, which is the bottom-level image result u i,j (n+1) and v i,j (n+1) , and as the displacement field (u i，j ,v i，j ).

[0055] As a further improvement to the above solution, the optimized fire early detection method further includes the following steps:

[0056] When the fire alarm is activated, the fire source size is calculated and the corresponding fire source size model is designed as follows:

[0057]

[0058] Where α is the ratio of the convective heat release rate to the total heat release rate (usually 0.6 to 0.8, depending on the specific type of fire source material), g is the acceleration of gravity, and c is the acceleration of gravity. p is the specific heat capacity of air, T ∞ is the ambient air temperature, ρ ∞ is the ambient air density, z is the height from the speed measurement point to the fire source, is the plume velocity of the current image relative to the previous frame.

[0059] Further, The calculation method is designed as follows: take 95% to 98% of the maximum value of the velocity field of all pixel positions in the current image, The corresponding plume velocity model is designed as:

[0060]

[0061] Where t is the time between two adjacent frames, D r is the actual distance represented by the pixel position (i, j).

[0062] The present invention also provides an optimized early fire detection device, which adopts the above-mentioned optimized early fire detection method.

[0063] Compared with the existing technology, the advantages of the present invention are: it can detect gas tank and gas pipeline leakage, and ignite the early characteristics of fire. At the same time, it optimizes the optical flow algorithm of the schlieren image, calculates the plume flow field more quickly, and through empirical formulas and speed measurement results, it can dynamically reflect the size of the fire in real time on the terminal. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a flow chart of the optimized early fire detection method of the present invention.

[0065] Figure 2 To adopt Figure 1 Schematic diagram of the displacement field obtained by the early fire detection method.

[0066] Figure 3 To adopt Figure 1 Schematic diagram of the velocity field obtained by the early fire detection method. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0068] See also Figure 1The optimized early fire detection method of the present invention performs ordinary shooting first and then high-speed shooting when the early fire detection system is started. The so-called ordinary shooting and high-speed shooting in this embodiment refer to different shooting cycles, that is, different speeds. High-speed shooting is faster than ordinary shooting. For example, ordinary shooting can be set to 20 frames per second, and high-speed shooting can be set to 200 frames per second. The specific setting can refer to the type and scale of the fire source that may occur on the scene. If it is a place where fuel is stored or there are dense accumulations of flammable and explosive materials, there is a gas pipeline passing through, and the fire spreads quickly, real-time monitoring should be used to prevent the early fire from evolving into a major safety accident in a short time. If there is no obvious fire source or fire safety hazard in the venue, interval shooting can be considered, shooting every 1 minute, or even every few minutes. When an abnormality is detected, high-speed shooting can be started, such as 200 frames per second.

[0069] The optimized early fire detection method of the present invention includes the following steps: acquiring a picture stream in real time in a normal shooting mode. Calculating the global energy of each picture in the global image, judging whether the global energy difference between the current picture and the previous frame is greater than the preset global energy difference, if yes, checking whether the current picture is disturbed, and continuing with normal shooting if the global energy difference is not greater than the preset global energy difference. If there is no interference, increasing the shooting cycle to start high-speed shooting, and if there is interference, continuing with normal shooting. Calculating the displacement field of the current picture relative to the previous frame in the picture stream after high-speed shooting, judging whether the displacement field is within the preset displacement field range, if yes, starting a fire alarm, otherwise continuing to calculate the displacement field of the next frame. When the fire alarm is activated, the size of the fire source is calculated.

[0070] Global energy cannot be used to determine whether a fire has occurred because other abnormalities, such as obstructions by foreign objects, flying birds, and insects, can cause changes in the brightness of the two images, resulting in a larger residual brightness difference. In this case, the fire alarm cannot be triggered. The abnormality must be eliminated and the image confirmed to be a thermal plume before the calculation can continue. Furthermore, considering the lifespan of high-speed photography, its transmission resource consumption, and its lifespan, it is best to activate high-speed photography mode after a fire has been detected. In this embodiment, the method for checking whether the current image has been interfered with is to check whether the global energy difference between the current image and the previous image is amplified in a local area between multiple consecutive adjacent frames. If so, it can be considered that foreign object interference has occurred in the local area. If the global energy difference is amplified and does not increase, it can be considered that the image has been interfered with by other factors. If the global energy difference is continuously changing and is greater than a preset value, it can be considered that the image has not been interfered with.

[0071] In the specific implementation process, the optimized early fire detection method of the present invention can be set in the form of software in application, such as being designed as an independent APP, or embedded software that can be called at any time, and applied in a computer terminal (such as a fire smoke alarm with an imaging function). The computer terminal includes a memory, a processor, and a computer program stored in the memory and run on the processor. The computer terminal can also be a smart phone, tablet computer, laptop computer, etc. that can execute programs. In some embodiments, the processor can be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is generally used to control the overall operation of the computer device. In this embodiment, the processor is used to run the program code stored in the memory or process data. When the processor executes the program, the steps of the optimized early fire detection method of the present invention can be implemented.

[0072] The optimized early fire detection method of the present invention can also be designed to store computer program instructions in a readable storage medium, such as a USB-shield. When the various drivers stored in the USB-shield are read and executed by a processor, the steps of the optimized early fire detection method of the present invention can be implemented. When the USB-shield is configured to be electronically plugged into the data port of a conventional early fire detector, the CPU of the fire detector can read and execute the computer program instructions in the USB-shield, thereby improving the sensitivity and accuracy of existing early fire detectors. Therefore, the present invention can also facilitate software upgrades for existing early fire detectors, thereby facilitating the promotion and application of the present invention.

[0073] Whether non-embedded or embedded, they can be summarized as corresponding early fire detection devices. The early fire detection device includes an image acquisition module, a global energy calculation module, a displacement field calculation module, a fire alarm module, and a fire source size calculation module, which respectively perform the above steps.

[0074] Each picture in the picture stream is a signal matrix composed of M×N pixel units. In this embodiment, the traditional global energy model is optimized and designed as follows:

[0075]

[0076] Where E(u,v) is the global energy of the image, I1(i,j) is the brightness of the first image at the pixel position (i,j) in the two adjacent frames, and I1(i+u i，j ,j+v i，j ) is the displacement field (u) of the second picture in two adjacent framesi，j ,v i，j ) is the brightness at the pixel position (i, j), λ is the regularization weight coefficient, ρ d and ρ s are data penalty function and space penalty function respectively; u i+1，j and v i+1，j are the vertical displacement component and horizontal displacement component at the pixel position (i+1, j), u i，j+1 and v i，j+1 are the vertical displacement component and the horizontal displacement component at the pixel position (i, j+1) respectively;

[0077] Where ρ(x)=(x 2 +∈ 2 ) a , x is the brightness residual or displacement error, ∈ is a small positive number taken to avoid the denominator being zero during calculation, which is 0.001, and a is the exponent, which is 0.40.

[0078] When calculating the two-dimensional velocity field of a high-speed schlieren image, the schlieren image is a signal matrix composed of M×N pixel units, where the grayscale (pixel value) at the pixel position (i, j) is considered to be about the displacement field V in adjacent frames. i,j function, V i,j =u i,j +v i,j ,u i,j and v i,j It is the vertical and horizontal displacement components corresponding to the pixel position (i, j). We use V to represent the calculated motion field of the entire image, that is, the displacement field.

[0079] The problem of solving the motion field of two adjacent Schlieren images is an optimization problem. The optimization process is to solve the motion field corresponding to the minimum value of the energy functional.

[0080]

[0081] E(i,j) is the energy functional at position (i,j), I1(i,j) is the brightness at the first image (i,j), I2(i+u i,j ,j+v i,j ) is the distance between (i,j) and (u i,j ,v i,j ) brightness after displacement, u i,j and v i,j is the vertical and horizontal displacement component corresponding to (i, j), λ is the corresponding regularization weight coefficient, ρ D and ρ s is the corresponding data penalty function and space penalty function. In this embodiment, ρD and ρ s The same penalty function is adopted.

[0082] The energy functional is also the displacement (u i,j ,v i,j ) function:

[0083]

[0084] The global energy function is

[0085]

[0086] The optimal pixel displacement field is solved by minimizing the energy functional. In simple terms, in two adjacent images, if the pixel at position (i, j) in the first image is at position (i+u i,j ,j+v i,j ), then the pixel displacement can be considered as (u i,j ,v i,j ), then the displacement is Excessive brightness differences or non-smooth motion fields will cause E to not converge or diverge, which will lead to inaccurate process motion estimation.

[0087] Note: In this invention, all subscripts such as V i,j 、u i,j 、v i,j 、v i+1,j 、u i+1,j etc. represent the characteristic quantities of a certain pixel position in a specific image, while V, (u, v), and E represent global variables, such as global velocity field, displacement field, and global energy.

[0088] Introducing the data penalty term (ρ D The purpose of ) is to optimize the optical flow estimation by using the brightness consistency constraint in the image. The brightness consistency constraint means that the brightness value of the pixels of the same object should remain unchanged in time between consecutive frames. By minimizing the brightness difference between the pixels of the current frame and the pixels of the next frame, the optical flow estimation result can be made more accurate. The spatial penalty term (ρ s) aims to optimize optical flow estimation by utilizing the smoothness of the image space. Between consecutive frames, the motion of adjacent pixels is often similar, because the motion of the object changes continuously. By introducing a spatial penalty term in the optical flow calculation, the optical flow field can be made smooth in space, that is, the difference in optical flow between adjacent pixels is small. This can improve the robustness of the optical flow estimation and reduce the impact of noise and outliers on the results. The regularization parameter is a parameter used to control the weight balance between the data term and the regularization term in the energy function. The energy function is the optimization target of the optical flow calculation, and the optical flow field is solved by minimizing the energy function. The regularization parameter is usually expressed as λ, which can adjust the relative importance of the regularization term in the energy function. The regularization term is to introduce additional prior knowledge or constraints to improve the accuracy and robustness of the optical flow estimation.

[0089] In this embodiment, the global energy model is optimized by designing a penalty function, ρ d and ρ s The same penalty function is used: ρ(x) = (x 2 +∈ 2 ) a This penalty function is asymptotically nonconvex, with a = 0.40. Compared to the general quadratic penalty function, this function has better stability and is easier to converge to the global optimal solution in iterative calculations. The regularization parameters λ and ε are set to 5 and 0.001, respectively.

[0090] In order to further improve the accuracy of the fire early detection method of the present invention, the displacement field (u i，j ,v i，j ) is optimized, and the corresponding optimization design is the displacement field (u i，j ,v i，j ) is to perform multi-layer Gaussian smoothing and downsampling on the two adjacent frames I1 and I2 to obtain the displacement field (u i，j ,v i，j ).

[0091] Specifically, the displacement field (u i，j ,v i，j ) includes the following steps:

[0092] Decompose two adjacent frames of images I1 and I2 into k layers of images with different resolutions according to the image size, with the lowest resolution belonging to the highest layer and the original resolution belonging to the lowest layer;

[0093] A single displacement field (u is applied to the top layer images of two adjacent frames I1 and I2 at the pixel position (i, j) i，j ,v i，j )1 single energy E i，j (ui，j ,v i，j ) calculation, the corresponding single energy model is;

[0094]

[0095] E i，j (u i，j ,v i，j ) is interpolated downward to the initial result of the second layer image with higher resolution for each frame;

[0096] Warp the second layer image of the second frame image I2 to have the same scale as the second layer image of the first frame image I1 to obtain a warped second layer image of the second frame image I2;

[0097] Calculate the corresponding single displacement field (u) at the pixel position (i, j) of the distorted second layer image and the second layer image of the first frame image I1. i，j ,v i，j )2, and the corresponding single displacement field (u i，j ,v i，j )2 single energy Ei, j (u i，j ,v i，j ) is calculated, and the corresponding calculation results are interpolated downward to the initial results of the adjacent next layer image with higher resolution of each frame image, and the loop is iterated until the corresponding single displacement field (u i，j ,v i，j ) k , as the displacement field (u i，j ,v i，j ).

[0098] Among them, Gaussian filtering is used when decomposing each layer of image, and Gaussian filtering uses a 5×5 Gaussian distributed filter to filter each layer of image.

[0099] In this embodiment, for the displacement field (u i，j ,v i，j ) is given in detail with examples of methods to improve the accuracy.

[0100] Displacement field (u i，j ,v i，j ) includes the following steps:

[0101] (1) Initialization: The initial result of the displacement field of each layer of image is set to (u i,j (0) ,v i,j (0) );

[0102] (2) Displacement field of the highest layer image: According to the initial displacement field result, the final result of the displacement field of the current layer (u i-1,j-1 (n) ,v i-1,j-1 (n) );

[0103] (3) From the next layer of the highest layer image to the lowest layer image:

[0104] Interpolate downward the displacement field:

[0105] The displacement field calculation results of the previous layer (u i,j (n) ,v i,j (n) ), interpolate and expand it to adapt to the size of the current layer image, and obtain the estimated result of the same size as the current layer image (u i-1,j-1 (n) ,v i-1,j-1 (n) );

[0106] (4) Iteratively optimize the current layer image:

[0107] a) According to the displacement field estimation value (u i-1,j-1 (n) ,v i-1,j-1 (n) ), calculate the energy residual of the current layer image:

[0108]

[0109]

[0110] b) For each pixel position (i, j), calculate the gradient of its regularization term:

[0111]

[0112]

[0113]

[0114]

[0115] c) Update the displacement field estimate (u) of each pixel position (i, j) of the current layer image i-1,j-1 (n+1) ,v i-1,j-1 (n+1) ):

[0116]

[0117]

[0118] Among them, (n) represents the nth iteration, (n+1) represents the n+1th iteration, α1 is the learning rate, E is the energy residual term, and R is the regularization term:

[0119]

[0120]

[0121] d) Repeat step (4) on the next layer of images until the iteration reaches the lowest layer of images;

[0122] (5) Return the displacement field estimation result u of the bottom layer image i,j (n+1) and v i,j (n+1) ;

[0123] (6) Finally, we get the vectors u and v of the entire displacement field, which is the bottom-level image result u i,j (n+1) and v i,j (n+1) , and as the displacement field (u i，j ,v i，j ).

[0124] The displacement field in the image is obtained by iteratively obtaining the flow vector of the entire field, such as Figure 2 As shown in , the velocity field of the Schlieren image in two adjacent images can be obtained by the actual size and the interval time of adjacent frames, as shown in Figure 3 As shown. That is:

[0125]

[0126] Apply globally:

[0127]

[0128] Where V i,j represents the velocity of point (i, j) (m / s), u i,j Represents the horizontal displacement component (piexl) of point (i, j), v i,j Indicates the vertical displacement component (piexl) of the (i, j) point motion, D r Represents the actual distance between pixels (m / piexl)) represents the actual distance represented by each unit pixel width in the Schlieren image, and t represents the time between two adjacent frames (s).

[0129] In very early fires, we usually believe that the maximum velocity of the fire plume at the same height is often at the center line. Therefore, in this invention, 95% to 98% of the maximum value in the velocity field is used as the true velocity of the plume center line. Based on this, the fire plume centerline velocity formula is used to infer the convective heat release rate. Normally The value of α is generally 0.6 to 0.8, depending on the specific fire source material, which can be obtained by checking the corresponding manual.

[0130] In the empirical formula of fire plume velocity, due to the diameter of the fire source, the distance from the virtual point source to the measuring point is usually used instead of the actual height to reduce the error.

[0131] Virtual point source height formula:

[0132]

[0133] Where z0 is the height of the virtual point source (m);

[0134] is the total heat release rate (kW); In the following it refers to the size of the fire source.

[0135] D is the diameter of the fire source (m).

[0136] The formula for the centerline velocity of the fire plume is:

[0137]

[0138] Where u0 is the plume velocity at the height z of the plume centerline from the point source (m / s);

[0139] g is the acceleration due to gravity (m / s 2 );

[0140] c p is the specific heat capacity of air (kJ / (kg·K));

[0141] T ∞ is the ambient air temperature (K);

[0142] ρ ∞ is the ambient air density (kg / m 3 );

[0143] is the convective heat release rate (kW);

[0144] z is the height from the speed measurement point to the fire source (m);

[0145] z0 is the position of the virtual point source (m).

[0146] In the early stages of a fire, the fire is very small. and D are both small series, so their linear combination can be considered as a small quantity, that is, the position of the virtual point source can be considered to coincide with the actual position of the fire source. Therefore, when other environmental parameters and gravitational acceleration are known, the centerline velocity of the fire plume can be considered as the height z and the convective heat release rate. function, when the height is known, the fire size can be inverted by measuring the velocity Right now:

[0147]

[0148] According to the above formula, the fire intensity of the fire source can be calculated, where α, g, c p , ρ ∞ , z are constants, and the fire source size can be obtained by measuring the plume velocity

[0149] The early detection method of fire detection of the present invention can dynamically capture the thermal plume schlieren in the air above a closed place by using an optimized optical flow algorithm, and can detect weak characteristic signals in the early stage of a fire.

[0150] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An optimized fire early detection method comprising the following steps: Get the picture stream in real time, each picture in the picture stream is M × N A signal matrix composed of pixel units; Calculate the global energy of each image in the world; Determine whether the global energy difference between the current image and the previous frame is greater than the preset global energy difference. If yes, check whether the current image is interfered with. If not, increase the shooting cycle to start high-speed shooting. Calculate the displacement field of the current image relative to the previous frame in the image stream after high-speed shooting; Determine whether the displacement field is within a preset displacement field range, and if so, activate a fire alarm; It is characterized in that the corresponding global energy model is optimized as follows: ； Where, E ( u , v ) is the global energy of the image in the world, I 1( i , j ) is the pixel position of the first picture in two adjacent frames ( i , j ) at a brightness of 1, I 1( i+u i，j , j + v i，j ) is the displacement field of the second picture in two adjacent frames ( u i，j , v i，j ) and then at pixel position ( i , j ) at brightness 2, λ is the regularization weight coefficient, ρ d and ρ s They are data penalty function and space penalty function respectively; u i+1，j and v i+1，j At pixel locations ( i +1, j ), the vertical and horizontal displacement components at u i，j+1 and v i，j+1 At pixel locations ( i , j +1) vertical displacement component and horizontal displacement component; in, ρ d and ρ s Using the same penalty function: , x is the brightness residual or displacement error, A small positive number used to avoid the denominator being zero during calculations. a is the index.

2. The optimized fire early detection method according to claim 1, characterized in that: The way to check whether the current image is interfered with is: Between multiple consecutive frames of adjacent pictures, there exists the following condition: the global energy difference between the current picture and the previous picture is greater than the preset global energy difference, and the global energy difference does not increase.

3. The optimized fire early detection method according to claim 1, characterized in that: Displacement field ( u i，j , v i，j ) is to improve the accuracy of two adjacent frames of pictures I 1 and I 2 Perform multi-layer Gaussian smoothing and downsampling to obtain two adjacent frames of images I 1 and I 2 displacement field ( u i，j , v i，j ).

4. The optimized fire early detection method according to claim 3, characterized in that: Displacement field ( u i，j , v i，j ) includes the following steps: Two adjacent frames of pictures I 1 and I 2. Decompose the image into different resolutions according to the image size k Layer images, the lowest resolution belongs to the highest layer image, and the original resolution belongs to the lowest layer image; For two adjacent frames I 1 and I 2's top-level image at pixel position ( i , j ) for a single displacement field ( u i，j , v i，j )1 single energy E i，j ( u i，j , v i，j ) calculation, the corresponding single energy model is; ； Will E i，j ( u i，j , v i，j ) is interpolated downward to the initial result of the second layer image with higher resolution for each frame; For the second frame I The second layer of image 2 is distorted to match the first frame image I The second layer image scale of 1 is the same, and the second frame image is obtained I 2 is the distorted second layer image; The distorted second layer image and the first frame image I 1's second layer image at pixel location ( i , j ) to calculate the corresponding single displacement field ( u i，j , v i，j )2, and perform the corresponding single displacement field ( u i，j , v i，j )2 single energy E i，j ( u i，j , v i，j ) calculation, and the corresponding calculation results are interpolated downward to the initial results of the adjacent next layer image with higher resolution of each frame image, and the loop is iterated until the adjacent two frames of images are obtained I 1 and I 2The corresponding single displacement field of the bottom image ( u i，j , v i，j ) k , as the displacement field ( u i，j , v i，j ).

5. The optimized fire early detection method according to claim 4, characterized in that: Gaussian filtering is used when decomposing each layer of the image.

6. The optimized fire early detection method according to claim 5, characterized in that: Gaussian filtering uses a 5×5 Gaussian distributed filter to filter each layer of image.

7. The optimized early fire detection method according to claim 4, characterized in that: Displacement field ( u i，j , v i，j ) includes the following steps: (1) Initialization: The initial result of the displacement field of each layer of image is set to ; (2) Displacement field of the highest layer image: Based on the initial displacement field result, the final result of the displacement field of the current layer is obtained through a finite number of iterations of the HS displacement field algorithm. ; (3) From the next layer of the highest layer image to the lowest layer image: Interpolate downward the displacement field: Calculation results of the displacement field of the previous layer , interpolate and expand it to adapt to the size of the current layer image, and obtain an estimated result of the same size as the current layer image ; (4) Iteratively optimize the current layer image: a ) According to the displacement field estimation value of the current layer image , calculate the energy residual of the current layer image: ； b ) for each pixel position ( i , j ), calculate the gradient of its regularization term: ； c ) Update each pixel position of the current layer image ( i , j ) displacement field estimate : ； in,( n ) indicates the n iterations, ( n +1) indicates the n +1 iteration, α 1 is the learning rate, E is the energy residual term, R is the regularization term: ； Repeat step (4) on the next layer of image until it iterates to the lowest layer of image; (5) Return the single displacement estimation result of the bottom layer image and ; (6) Finally, we get the vector sum of the entire displacement field, that is, all the bottom layers. and and as the displacement field ( u i，j , v i，j ).

8. The optimized fire early detection method according to claim 1, characterized in that: The fire early detection method further comprises the following steps: When the fire alarm is activated, the fire source size is calculated and the corresponding fire source size model is designed as follows: ； Where, α is the ratio of the convective heat release rate to the total heat release rate, g is the acceleration due to gravity, c p is the specific heat capacity of air, T ∞ is the ambient air temperature, ρ ∞ is the ambient air density, z is the height from the speed measurement point to the fire source, is the plume velocity of the current image relative to the previous frame.

9. The optimized fire early detection method according to claim 8, characterized in that: The calculation method is designed as follows: take 95%~98% of the maximum value of the velocity field of all pixel positions in the current image, The corresponding plume velocity model is designed as: ； Where, t is the time between two adjacent frames of pictures, D r The actual distance represented by the unit pixel width.

10. An optimized fire early detection device, characterized in that: It adopts the optimized early fire detection method as claimed in any one of claims 1 to 9.

Images