A high-speed flame detection method and device based on a HVS visual model

Abstract

The application discloses a high-speed flame detection method and device based on an HVS visual model, relates to the technical field of fire safety monitoring, and comprises the following steps: converting an image to HSV space, and performing color preliminary screening based on a Gaussian mixture model and local contrast modulation; adopting a chessboard sampling and neighborhood weighted voting to perform motion detection, and generating a candidate region; calculating a circularity and boundary roughness to remove interference; storing a plurality of frames of luminance sequences by using a ring buffer, statistically obtaining three time domain descriptors based on a JND threshold, and calculating flicker confidence; and finally outputting a result by means of softmax dynamic weighting and nonlinear entropy weight fusion decision. The high-speed flame detection method and device based on the HVS visual model have the advantages of high detection rate, low false alarm rate, strong real-time performance, good self-adaptability, convenient deployment and the like.

Description

Technical Field

[0001] This invention relates to the field of fire safety monitoring technology, and in particular to a high-speed flame detection method and device based on the HVS visual model. Background Technology

[0002] With the acceleration of urbanization and the emergence of numerous new building structures (such as high-rise buildings, complexes, and transportation hubs), traditional contact-type fire detectors (such as smoke and heat detectors) are failing to meet the demand for reliable early alarms due to the severe attenuation of fire characteristic signals (smoke and temperature) during long-distance transmission. Video image-based fire detection technology, with its advantages of fast response, rich information, and non-contact monitoring, has become a research hotspot and development direction in this field.

[0003] However, existing video flame detection technology still has the following significant drawbacks, which severely restrict its large-scale practical application: First, feature description is disconnected from the HVS perception mechanism. Most algorithms start from a purely mathematical or signal processing perspective, using objective indicators such as mean squared error and fixed threshold segmentation for feature extraction and discrimination. These methods fail to effectively simulate the sensitivity of the human visual system to color saturation, brightness contrast, edge texture roughness, and irregular high-frequency jitter, resulting in poor consistency between the detection results and human subjective judgment. For example, some algorithms rely solely on simple threshold segmentation of the RGB channels (such as R>150 and R>G>B), making them highly susceptible to interference from bright white light, red flags, orange pipe paint, or the afterglow of sunset, generating a large number of false alarms.

[0004] Second, the contradiction between computational complexity and real-time performance is prominent. To achieve high-precision detection, some advanced algorithms have introduced three-dimensional convolutional neural networks (3D-CNN) or optical flow methods to capture the spatiotemporal dynamic features of flames. Although these methods have high accuracy, the huge number of parameters and computational load makes it difficult to achieve real-time processing requirements (≥25fps) on embedded devices or conventional security servers. While simple frame difference methods or full-pixel background subtraction methods are fast, they often can only extract the outline of moving targets or reveal internal holes, failing to provide stable and complete regional information for subsequent identification.

[0005] Third, single-feature decision-making leads to weak generalization ability. Many methods tend to rely on a single "strong feature" (such as using only color or only strobe) for the final decision. This single-point decision-making mechanism may be effective in specific scenarios, but once the environment changes (such as sudden changes in lighting or slight detector jitter), the system performance will drop sharply, manifested as a significant increase in the false alarm rate and the missed detection rate.

[0006] Fourth, it has low sensitivity to early-stage, small-scale flames. In the early stages of a fire, flames often occupy only a very small number of pixels in an image (usually less than 3% of the total number of pixels). Traditional target detection methods based on sliding windows or fixed anchor boxes are designed with receptive fields and feature pyramids that are mainly aimed at medium to large-sized targets. They are unable to effectively capture the features of such small targets, causing them to be ignored by the network downsampling layer or background model, thus delaying the best time to extinguish the fire. Summary of the Invention

[0007] The purpose of this invention is to provide a high-speed flame detection method and device based on the HVS visual model to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, this invention provides a high-speed flame detection method based on an HVS visual model, comprising: S1. Color space conversion and preprocessing steps: Convert the acquired raw video image from the device-dependent color space to the HSV perceptual color space, and apply median filtering to the luminance component; S2. Color saliency screening step: In the HSV space, based on the multi-component Gaussian mixture color model with illumination adaptive weights, combined with the local contrast modulation factor and dynamic saturation-brightness constraint interval, the color saliency determination function is calculated pixel by pixel to generate a binarized saliency color map. S3. Moving target confirmation step: The image pixels are divided into active pixels and non-active pixels using a checkerboard sampling strategy, and the sampling density is adaptively adjusted according to the scene motion state; Gaussian mixture background model is built only for active pixels to perform foreground detection, and the state of non-active pixels is determined by a neighborhood weighted voting mechanism; The salient color map and the moving foreground map are logically overlapped to generate a joint motion-color candidate region. S4. Morphological and shape analysis steps: Perform morphological opening and closing operations and connective region marking on candidate regions, calculate the circularity and boundary roughness of each region, and eliminate regions that do not meet the shape determination criteria. S5. Temporal flicker feature extraction steps: Extract the brightness change time series in the candidate region, calculate the inter-frame change intensity based on the human eye's just distinguishable difference threshold, and construct three temporal statistics within the sliding time window: average activation rate, change stability index, and cumulative fluctuation energy, and calculate the flicker confidence based on these. S6. Multi-feature fusion decision steps: Construct a four-dimensional feature confidence vector, dynamically allocate the weights of each feature according to the scene state parameters through the softmax function, calculate the global confidence using a nonlinear entropy weight fusion function, and output the final detection result based on a dual-threshold three-state decision mechanism.

[0009] Preferably, step S2 specifically includes: The multi-component Gaussian mixture color model is as follows: ; Component weights An adaptive function for pixel position and local brightness: ; For pixels The local brightness normalized value, The saturation and brightness components are determined by the dynamic constraints of the illumination influence factor. ; Saturation lower bound The color saliency decreases monotonically with increasing brightness; the color saliency determination function is: ; in Local contrast modulation factor, For pixels The local brightness contrast with its eight neighbors; output 1 if all conditions are met, otherwise output 0, generating a significant color map. .

[0010] Preferably, the checkerboard sampling mask in step S3 is defined as follows: ; Sampling density is based on the proportion of the motion region in the previous frame. Adaptive adjustment: ; in Based on the base sampling rate; a Gaussian mixture background model is established only for active pixels with a sampling mask value of 1, and its probability density function is: ; If the current observation value is the same as the first If there are Gaussian distributions matching, then according to the learning rate Update the mean and variance of the distribution; if no distribution matches, initialize a new Gaussian distribution with the current observations and replace the one with the smallest weight.

[0011] Preferably, in step S3, the motion state of the non-active pixels is determined by a weighted vote of the eight neighboring active pixels: ; Domain weight In other words, the smaller the variance of the Gaussian mixture model, the higher the weight of the active pixel in the judgment result. The voting threshold is used; the motion-color joint candidate region generation method is as follows: ; Only pixels that simultaneously meet the conditions of color salience and moving foreground are retained.

[0012] Preferably, a 3×3 square structural element is used in step S4. Performing morphological operations, the erosion operation is defined as follows: The expansion operation is defined as follows: morphologically filtered connected regions Calculate roundness and boundary roughness ,in For the area, The perimeter of the region. Let be the perimeter of the convex hull of the region; the criterion is: if or If so, the corresponding region will be removed from the candidate set. Take a value between 0.65 and 0.75. Take a value between 1.10 and 1.20.

[0013] Preferably, in step S5, the candidate region At any moment The intensity of the inter-frame brightness change is: ; The exactly resolvable difference JND is presented in a piecewise nonlinear form: ; ; in the sliding time window ( Define three time-domain statistics within the range of 25 to 32: Average activation rate ; Stability index of change ; Accumulated fluctuation energy .

[0014] Preferably, the confidence level of the flash gravel in step S5 The calculation function is: ; in: ; ; and This is an empirical reference value for flame samples. A sinusoidal window constraint is applied to the average activation rate so that the spark confidence level outputs a valid value only when the rate of change is within a reasonable range.

[0015] Preferably, in step S6, the four-dimensional confidence vector Scene state vector ,in This is the global brightness normalized value. The proportion of the dynamic region to the entire map. This is the signal-to-noise ratio estimate; the weights of each feature are dynamically calculated using the softmax function. ; in and The weight mapping parameters are for offline learning; the global confidence score for nonlinear entropy weight fusion is: ; in It is a nonlinear modulation index. Information entropy of the confidence vector: ; and The entropy penalty intensity coefficient is used; the dual-threshold three-state decision mechanism is as follows: if If the value is between 0.70 and 0.80, a fire is detected and an alarm is triggered; if If the value is between 0.50 and 0.60, it is considered a suspicious state, and cumulative confidence tracking is introduced. If continuous Frame satisfies ( If the value is 5 to 8, it will be upgraded to an alarm; otherwise, it will be reset. It was determined that there was no fire.

[0016] A high-speed flame detection device based on an HVS visual model includes: The front-end image acquisition module is used to acquire video image signals of the monitored area in real time. The data buffer unit is used to store the current processing frame and historical image data of multiple consecutive frames; The core heterogeneous computing unit comprises a Field Programmable Gate Array (FPGA) subsystem and a Digital Signal Processor (DSP) or ARM processor subsystem. The FPGA subsystem performs pixel-level parallel processing, including color space conversion, image filtering, and morphological operations. The DSP or ARM processor subsystem performs Gaussian mixture model parameter maintenance, connected component labeling, temporal flicker feature statistics, and multi-feature fusion decision-making. After completing pixel-level preprocessing of the entire image, the FPGA subsystem transmits the candidate region coordinates to the DSP or ARM processor subsystem, which then performs background modeling updates and flicker feature calculations only for the candidate regions. The program and parameter storage unit is used to store the startup program, flame statistical model parameters, and dynamic threshold configuration; The alarm output interface is used to output an alarm signal when a fire is detected.

[0017] Therefore, the high-speed flame detection method and device based on the HVS visual model described above have the following beneficial effects: (1) A multi-level serial verification structure of "color statistical model → motion spatiotemporal filtering → shape texture analysis → flicker feature verification" is constructed. Each level filters different types of interference sources in sequence: color screening is based on statistical models rather than fixed thresholds to adaptively distinguish between flames and fire-like interference objects; motion detection filters out static fire-like objects; shape analysis uses roundness and boundary roughness to eliminate regular shape interference; flicker analysis distinguishes between high-frequency irregular flickering of flames and stable light sources or periodic flickering in the time domain. The four-fold verification gradually narrows the judgment conditions, and a single feature's occasional high value cannot independently trigger an alarm, ensuring a low false alarm rate in complex scenarios.

[0018] (2) A heterogeneous computing architecture of "FPGA hardware acceleration + ARM software algorithm" is adopted. The FPGA uses hardware parallelism to complete pixel-level preprocessing, while the ARM only performs calculations on the candidate regions selected by the FPGA; the key pixel sampling strategy reduces the computational load of motion detection to less than one-quarter of that of the traditional full-pixel Gaussian mixture model; the flicker analysis uses a circular buffer incremental update mechanism, so that the updates of the three time-domain statistics are constant-time operations in each frame, completely avoiding the recalculation of sliding window traversal. The above optimizations make the end-to-end processing time of a single frame much smaller than the frame period, and real-time full analysis can be completed with only the embedded chip.

[0019] (3) The color judgment threshold is adaptively shifted with the global brightness of the image, and the lower bound of saturation decreases monotonically with the increase of pixel brightness; the sampling density is automatically adjusted according to the intensity of scene motion; the feature fusion weight is dynamically mapped from the scene state vector through softmax, and the contribution of flicker and motion is automatically increased under low illumination. The entire system does not need to preset multiple sets of parameters for different environments, and maintains consistent detection performance under various lighting and scene complexity.

[0020] (4) The color statistical model is updated online using false alarm and correct alarm samples in actual operation. The learning rate decreases as the feedback samples accumulate - "rapid adaptation in the early stage and stable convergence in the later stage", so that the system can gradually gain a stronger ability to identify scene-specific interference sources in long-term operation, and the false alarm rate continues to decrease as the deployment time extends.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1This is a schematic diagram of the overall process of an embodiment of a high-speed flame detection method based on an HVS vision model according to the present invention; Figure 2 This is a schematic diagram of an embodiment of a high-speed flame detection device based on an HVS visual model according to the present invention. Figure 3 This is a schematic diagram of the specific application structure of the device according to an embodiment of the present invention; Figure label: 1. Front-end image acquisition module; 2. Housing; 3. Core heterogeneous computing unit; 4. Data cache unit; 5. Program and parameter storage unit; 6. RJ45 network interface; 7. Aviation connector; 8. Battery. Detailed Implementation

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0025] Example Please see Figure 1 This invention provides a high-speed flame detection method based on the HVS visual model, comprising: S1, Video image acquisition and adaptive color space mapping.

[0026] The raw video stream of the monitored scene is acquired through an image sensor (CCD or CMOS), supporting YUV, RGB, or RAW formats. To eliminate the interference of ambient light intensity on subsequent color determination and to simulate the human eye's independent perception of hue, saturation, and brightness, the raw image is converted from the device-dependent RGB color space to the perceptually uniform HSV color space. During the conversion, a parallel lookup table (LUT) in the FPGA hardware acceleration unit is used for fast calculation. After conversion, a median filter (window size of 3×3 or 5×5) is applied to the luminance component V to suppress sensor noise while maintaining the sharpness of the flame edges, simulating the lateral suppression mechanism in the human visual system.

[0027] S2. Rapid initial screening of visual saliency based on statistical color models.

[0028] In the offline phase, a priori statistical model of flame color is constructed. At least 2,000 flame image samples covering various fuel types (liquid hydrocarbon fuels and solid combustibles) and various lighting conditions (strong daylight, cloudy days, nighttime, and indoor lighting) are collected and pixel-level statistical analysis is performed in HSV space.

[0029] A mixture probability density model with K Gaussian components (K=3~5) is established in the hue H dimension to cover the complete chromaticity range of the flame from the inner flame (bright white-yellow) to the outer flame (red-orange). The weights of each Gaussian component are designed as an adaptive function of the local brightness of the pixel—the contribution weight of the low-saturation yellow component is automatically reduced in high-brightness areas, simulating the nonlinear shift in color perception of the human eye under strong light.

[0030] In the dimensions of saturation (S) and brightness (V), dynamic constraint intervals are established based on the Weber-Fechner Law of the human eye. The lower bound of saturation decreases monotonically with increasing brightness—flames can still be captured even when their saturation decreases under strong light. The upper and lower bounds of brightness adaptively shift with the global average brightness of the image, enabling the algorithm to maintain stable filtering capabilities in both low-light (dark red flames at night) and high-light (bright yellow flames during the day) scenarios.

[0031] During the online determination phase, the hue probability density is calculated for each pixel and compared with a base threshold modulated by local contrast. The local contrast modulation factor uses a sigmoid function to automatically lower the determination threshold for areas with abrupt changes in local brightness (i.e., flame edges) in the image, thereby accurately simulating the enhanced sensitivity of the human eye's lateral inhibition mechanism to edge regions. When all three determination conditions (hue probability, saturation range, and brightness range) are simultaneously met, the pixel is marked as "color saliency," and a binarized saliency color map is output.

[0032] S3. Rapid moving target identification based on spatial redundancy optimization.

[0033] By leveraging pixel spatial correlation, image pixels are evenly divided into "active pixels" and "non-active pixels" in a checkerboard pattern. The sampling density is not a fixed 50%, but rather adaptively adjusted based on the proportion of the moving region in the previous frame—sampling tends to be denser during intense motion (up to 75%), while remaining sparse in static scenes (down to 25%), ensuring that computing resources are precisely allocated to dynamic regions.

[0034] A Gaussian Mixture Background Model (GMM) is built only for active pixels. Model updates employ a standard online EM strategy: for matching pixels, the mean and variance are updated according to the learning rate; for mismatches, a new distribution is created to replace the one with the smallest weight. The foreground / background determination of active pixels is based on whether the matching Gaussian distribution belongs to the background subset determined by the weight-variance ratio ranking.

[0035] Non-active pixels are not modeled; their motion state is determined by a weighted vote among active pixels in their eight neighborhoods. The voting weight is inversely proportional to the variance of the Gaussian model of each active pixel—pixels with more stable models (smaller variance) enjoy higher decision weights. This ensures the reliability of the decision result even under sparse sampling.

[0036] A pixel-by-pixel logical AND operation is performed between the salient color map and the moving foreground map. Only pixels that simultaneously satisfy both color salience and motion attributes constitute "motion-color joint candidate regions." This overlap strategy reduces the computational load of subsequent fine-grained analysis by 60% to 70%, and is one of the keys to achieving high-speed processing.

[0037] S4. Disturbance removal based on morphology and shape complexity.

[0038] Morphological opening operations (erosion + dilation) are performed on the candidate region to remove isolated noise points, and then closing operations (dilation + erosion) are performed to connect the small breaks caused by color or motion segmentation, simulating the human eye's perception of the complete closed contour of an object based on Gestalt psychology principles.

[0039] Eight-connected region labeling is performed on the morphologically filtered binary image. For each connected region, two shape descriptors are calculated: circularity (the ratio of area to the square of the perimeter) and boundary roughness (the ratio of the convex hull perimeter to the actual perimeter). Real flames are highly irregular in shape due to turbulent entrainment: circularity is typically below 0.6 (far less than the 1.0 of a circle), and boundary roughness is typically above 1.25 (jagged edges). Based on this, a judgment criterion is set: regions with a circularity above 0.7 (close to a circle) or a boundary roughness below 1.15 (overly smooth edges) are judged as interference and removed. This step effectively eliminates fire-like interference sources such as vehicle headlights, streetlights, circular signs, and human silhouettes.

[0040] S5. High-frequency flicker feature extraction based on cumulative change in the time domain.

[0041] For the retained candidate regions, their minimum bounding rectangles are extracted as Regions of Interest (ROIs), and a circular buffer with a length of 25-32 frames is maintained in the memory to store the luminance sequence. The design of the circular buffer enables the statistic to be updated in only O(1) time when a new frame arrives, without recalculating the entire window.

[0042] Define the inter-frame luminance change intensity: count the proportion of pixels in the ROI whose luminance change exceeds the Just Noticeable Difference (JND) threshold of the human eye. The JND threshold adopts a piecewise nonlinear function: the threshold is lower in the low luminance interval (V≤40) (the human eye is more sensitive after dark adaptation), the threshold is moderate in the middle luminance interval (40<V≤180) (the Weber fraction is about 0.07), and the threshold is higher in the high luminance interval (V>180) (the sensitivity of the human eye decreases under strong light). This design accurately simulates the perceptual characteristic that the human eye is most sensitive to changes in the middle luminance region.

[0043] Calculate three time-domain statistics within the sliding time window: (1) Average activation rate — reflecting the overall dynamic activity level of the region; (2) Change stability index — reflecting the fluctuation intensity of the flicker amplitude; (3) Cumulative fluctuation energy — performing weighted summation on the difference between the change amounts of adjacent frames, with higher weights for recent changes than for distant changes, so as to enhance the real-time response capability.

[0044] The three statistics are input into a flicker confidence function. This function is composed of the product of three sub-functions: (1) Variance saturation activation function — the confidence level saturates rapidly after the stability index reaches a reference value; (2) Energy accumulation function — presenting an exponential response to the cumulative fluctuation energy; (3) Sine window constraint function — outputting a valid value only when the average activation rate is within a reasonable interval (neither static nor drastically mutated), otherwise it is directly set to zero. The product of the three finally generates a flicker confidence in the interval of [0, 1], which effectively characterizes the unique "continuous and irregular" high-frequency flicker behavior of flame without requiring computationally complex Fourier transform.

[0045] S6. Multi-feature decision based on nonlinear entropy weight fusion.

[0046] Construct a four-dimensional confidence vector: color confidence (the proportion of pixels satisfying the color model), motion confidence (the proportion of foreground pixels), shape confidence (the comprehensive score of circularity and roughness), and flicker confidence (calculated in step S5).

[0047] The scene state vector consists of three components: a global brightness normalized value (determining lighting conditions), a dynamic region occupancy (determining scene motion complexity), and a signal-to-noise ratio estimate (determining image quality). The weights of each feature are mapped using a softmax function—automatically increasing the mapping coefficients for flicker and motion weights under low-light conditions, and increasing the mapping coefficients for shape weights under strong interference backgrounds, thus achieving true scene adaptation.

[0048] The core innovation lies in the nonlinear entropy-weighted fusion function: the numerator is the weighted nonlinear power sum of the confidence levels of each feature (the power is not equal to 1), and the denominator adds an entropy penalty term to the weighted power sum. The smaller the information entropy (confidence highly concentrated on a single feature), the closer the exponent term is to 1, the larger the denominator becomes, and the compressed fusion output is achieved; the larger the information entropy (confidence uniformly distributed), the closer the exponent term is to 0, and the denominator is reduced to a weighted power sum. This design mathematically ensures that the global confidence level can only reach the alarm threshold when multiple features simultaneously provide medium-to-high confidence evidence; any random high value of a single feature cannot independently trigger an alarm.

[0049] The decision-making process employs a dual-threshold, three-state mechanism: high confidence (≥0.75) triggers an immediate alarm, low confidence (<0.55) allows for safe passage, and intermediate states trigger cumulative tracking—if the confidence level remains at a medium to high level for multiple consecutive frames, the alarm is escalated; otherwise, it is reset. This mechanism balances response speed (immediate alarm upon instantaneous high confidence) and anti-interference capability (suspicious states require confirmation over multiple frames).

[0050] like Figure 2 A high-speed flame detection device based on the HVS visual model includes: The system includes a front-end image acquisition module 1 (wide dynamic range CMOS camera), a data cache unit 4 (DDR3 / DDR4 memory with a capacity of not less than 1GB), a core heterogeneous computing unit 3 (FPGA+DSP / ARM dual-core architecture), a program and parameter storage unit 5 (Flash memory), and an alarm output interface (dry contact / Ethernet / RS-485).

[0051] The FPGA subsystem is responsible for pixel-level parallel processing (color conversion, filtering, morphology); the DSP / ARM subsystem is responsible for logic control (GMM parameter maintenance, region marking, flicker statistics, fusion decision). After the FPGA completes the full-image preprocessing, it transmits the candidate region coordinates to the DSP / ARM, which then performs subsequent calculations only on the candidate regions. This division of labor—"FPGA full-image coarse screening + DSP fixed-point fine calculation"—is the hardware guarantee for the device to achieve a real-time processing speed of ≥25fps.

[0052] Example 1: Flame detection process in a typical outdoor scenario.

[0053] This embodiment takes the monitoring of an outdoor unit area in a petrochemical industrial park as an example. An explosion-proof camera installed on the pipe rack captures 704×576 resolution PAL video images at 25fps. The original image is in YUV format, converted to HSV color space by an FPGA parallel lookup table circuit. The on-site illuminance is approximately 8500 lux, and the system accordingly stretches the dynamic range of the luminance V channel to [40, 235].

[0054] Color selection stage: Pixel-by-pixel judgment is performed based on a color model containing four Gaussian components (corresponding to bright white inner flame, yellow-orange flame core, red-orange outer flame, and transition region, respectively). In strong sunlight, the system automatically lowers the saturation lower bound from 0.25 to 0.18 to tolerate saturation decay of the flame color under strong light. A local contrast modulation factor lowers the judgment threshold for flame edge areas by approximately 15%, ensuring that edge pixels are not missed. This stage marks the actual flame area (oil pan fire, approximately 450 pixels) and part of the orange pipe paint surface (approximately 1200 pixels).

[0055] Motion confirmation phase: Sampling density GMM background matching was performed only on 54% of the key pixels. The orange pipe paint surface has no motion properties, and its non-key pixels were determined to be background through neighborhood voting. After logical overlap, the pipe paint surface was filtered out, reducing the candidate area to approximately 380 pixels.

[0056] Shape analysis phase: Circularity of the actual flame area Boundary roughness 1.274. A red warning flag fluttering in the distance. , .set up , The flag because Removed.

[0057] Flicker analysis phase: Maintain a 30-frame circular buffer. Flame area. Flickering confidence .

[0058] Fusion Judgment Phase: Four-Dimensional Confidence Scene status , The weight vector is approximately equal to... Nonlinear fusion This triggered an alarm.

[0059] like Figure 3The high-speed flame detection device based on the HVS vision model provided in this embodiment has an overall appearance resembling a gun-shaped camera. It consists of five main functional modules: a front-end image acquisition module 1, a core heterogeneous computing unit 3, a data cache unit 4, a program and parameter storage unit 5, and an alarm output interface. Each module is integrated within a waterproof and dustproof housing 2, and data interaction is achieved through an internal high-speed data bus, forming a complete signal processing link from front-end sensing to back-end alarm. Externally, the device only has one RJ45 network interface 6 for power supply and data transmission, and one aviation connector 7 for connecting an external alarm. The overall installation method is exactly the same as that of a regular network camera, allowing for rapid deployment at various monitoring points. An internal battery 8 is also included for emergency backup during power outages.

[0060] When the device is in operation, the front-end image acquisition module 1 continuously acquires video images of the monitored area at a rate of 25 to 30 frames per second, and transmits the image data to the core heterogeneous computing unit 3 via the MIPI interface. The core heterogeneous computing unit 3 runs the built-in flame detection algorithm program, performing six steps sequentially on each frame of the image: color space conversion, initial color saliency screening, moving target confirmation, morphological and shape analysis, temporal flicker feature extraction, and multi-feature fusion decision. When the determination result is a fire, the core heterogeneous computing unit 3 drives the alarm output interface to send an alarm signal, and simultaneously uploads the alarm information (including a compressed image of the flame location marking box) to the monitoring center or fire management platform via the RJ45 network interface 6.

[0061] The front-end image acquisition module 1 is fixedly installed at the front of the device housing 2, used to convert the optical signals of the monitored scene into digital video signals, and is the signal source of the entire device. The data cache unit 4 stores the current frame image data to be processed and historical brightness data of multiple consecutive frames, providing inter-frame data support for motion detection and temporal flicker analysis. The data cache unit 4 consists of a DDR3 SDRAM main cache chip, a circular buffer, and a cache controller. The core heterogeneous computing unit 3 is the data processing center of the device, adopting a heterogeneous dual-core architecture of a field-programmable gate array (FPGA) subsystem and an ARM processor subsystem. The FPGA subsystem is responsible for low-level pixel-level parallel processing, while the ARM processor subsystem is responsible for high-level logic control and algorithm stream processing. The two subsystems communicate and exchange data through an AXI-DMA high-speed data bridge. The program and parameter storage unit 5 is used to persistently store the program code, algorithm model parameters, and system configuration information required for system operation. The alarm output interface is used to convert the fire judgment results of the fusion decision processor into various output signal forms conforming to fire protection industry standards, enabling interface with external fire alarm systems and monitoring management platforms.

[0062] Therefore, this invention employs the aforementioned high-speed flame detection method and device based on the HVS visual model. Regarding detection reliability, a four-level cascaded verification structure (color, motion, shape, and flicker) filters out fire-related interference at each level, preventing a single feature from independently triggering an alarm. This effectively solves the false alarm problem caused by common interference sources such as sunlight, vehicle headlights, and red objects. In terms of early warning timeliness, color screening based on a visual attention mechanism allows small-scale flames to be locked in the first step, without relying on large-area features. Combined with local contrast modulation and morphological closing operations, it ensures that weak early flame signals are not missed. Regarding real-time processing performance, FPGA hardware acceleration and ARM software algorithms work together. A key pixel sampling strategy reduces computation to less than a quarter of traditional methods. A circular buffer enables constant-time updates to flicker statistics, and the end-to-end processing time per frame is much shorter than the frame period, allowing real-time full analysis to be completed using only an embedded chip. Regarding environmental adaptability, the color threshold adaptively shifts with illumination, the sampling density automatically adjusts according to the intensity of motion, and the fusion weight is dynamically mapped from the scene state. The entire system can operate stably in various environments without requiring manual parameter presets. Regarding continuous system evolution, the color model is incrementally updated online using feedback from actual operation, resulting in a continuous decrease in the false alarm rate over time. In terms of engineering applications, the device's appearance is identical to that of a regular camera, supports PoE power supply and the ONVIF standard, allowing for seamless integration into existing security systems. It can also be spawned into various product forms such as edge computing boxes and drone modules, demonstrating strong practicality and promotional value.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A high-speed flame detection method based on the HVS visual model, characterized in that, include: S1. Convert the acquired raw video image from the device-dependent color space to the HSV perceptual color space, and apply median filtering to the luminance component; S2. In the HSV space, based on the multi-component Gaussian mixture color model with illumination adaptive weights, combined with the local contrast modulation factor and dynamic saturation-brightness constraint interval, the color saliency determination function is calculated pixel by pixel to generate a binarized salient color map. S3. A checkerboard sampling strategy is used to divide image pixels into active pixels and non-active pixels, and the sampling density is adaptively adjusted according to the scene motion state. A Gaussian mixture background model is built for active pixels to perform foreground detection, and the state of inactive pixels is determined by a neighborhood weighted voting mechanism; the salient color map and the moving foreground map are logically overlapped to generate a joint motion-color candidate region. S4. Perform morphological opening and closing operations and connective region marking on the candidate regions, calculate the circularity and boundary roughness of each region, and remove regions that do not meet the shape determination criteria. S5. Extract the brightness change time series in the candidate region, calculate the inter-frame change intensity based on the human eye's just distinguishable difference threshold, construct three time-domain statistics within the sliding time window: average activation rate, change stability index, and cumulative fluctuation energy, and calculate the flicker confidence. S6. Construct a four-dimensional feature confidence vector, dynamically allocate the weights of each feature according to the scene state parameters using the softmax function, calculate the global confidence using a nonlinear entropy weight fusion function, and output the final detection result based on a dual-threshold three-state decision mechanism.

2. The high-speed flame detection method based on the HVS visual model according to claim 1, characterized in that, Step S2 specifically includes: The multi-component Gaussian mixture color model is as follows: ； Component weights An adaptive function for pixel position and local brightness: ； For pixels The local brightness normalized value, The saturation and brightness components are determined by the dynamic constraints of the illumination influence factor. ； Saturation lower bound The color saliency decreases monotonically with increasing brightness; the color saliency determination function is: ； in Local contrast modulation factor, For pixels The local brightness contrast with its eight neighbors; output 1 if all conditions are met, otherwise output 0, generating a significant color map. .

3. The high-speed flame detection method based on the HVS visual model according to claim 2, characterized in that, In step S3, the checkerboard sampling mask is defined as follows: ； Sampling density is based on the proportion of the motion region in the previous frame. Adaptive adjustment: ； in Based on the base sampling rate, a Gaussian mixture background model is established for active pixels with a sampling mask value of 1, and its probability density function is: ； If the current observation value is the same as the first If there are Gaussian distributions matching, then according to the learning rate Update the mean and variance of the distribution; If no distribution matches, a new Gaussian distribution is initialized with the current observations to replace the one with the smallest weight.

4. The high-speed flame detection method based on the HVS visual model according to claim 3, characterized in that, In step S3, the motion state of the non-active pixels is determined by a weighted vote of the eight neighboring active pixels: ； Domain weight In a Gaussian mixture model, the smaller the variance of an active pixel, the higher its weight in the determination result. The voting threshold is used; the motion-color joint candidate region generation method is as follows: ； Pixels that simultaneously satisfy the conditions of color salience and moving foreground are retained.

5. The high-speed flame detection method based on the HVS visual model according to claim 4, characterized in that: Step S4 uses a 3×3 square structural element. Performing morphological operations, the erosion operation is defined as follows: The expansion operation is defined as follows: ; for connected regions after morphological filtering Calculate roundness and boundary roughness ,in For the area, The perimeter of the region. Let be the perimeter of the convex hull of the region; the criterion is: if or If so, the corresponding region will be removed from the candidate set. Take a value between 0.65 and 0.

75. Take a value between 1.10 and 1.

20.

6. The high-speed flame detection method based on the HVS visual model according to claim 5, characterized in that: In step S5, candidate regions At any moment The intensity of the inter-frame brightness change is: ； The exactly resolvable difference JND is presented in a piecewise nonlinear form: ； in, ; in the sliding time window Three time-domain statistics are defined internally: Average activation rate ; Stability index of change ; Accumulated fluctuation energy .

7. The high-speed flame detection method based on the HVS visual model according to claim 6, characterized in that: Confidence level of sharding in step S5 The calculation function is: ； in: ； ； and This is an empirical reference value for flame samples. Apply a sinusoidal window constraint to the average activation rate so that the confidence level of the flash gravel outputs an effective value when the rate of change is within a reasonable range.

8. The high-speed flame detection method based on the HVS visual model according to claim 7, characterized in that: In step S6, the four-dimensional confidence vector Scene state vector ,in This is the global brightness normalized value. The proportion of the dynamic region to the entire map. This is the signal-to-noise ratio estimate; The weights of each feature are dynamically calculated using the softmax function: ； in and The weight mapping parameters are for offline learning; the global confidence score for nonlinear entropy weight fusion is: ； in It is a nonlinear modulation index. Information entropy of the confidence vector: ； and The entropy penalty intensity coefficient is used; the dual-threshold three-state decision mechanism is as follows: if If the value is between 0.70 and 0.80, a fire is detected and an alarm is triggered; if If the value is between 0.50 and 0.60, it is considered a suspicious state, and cumulative confidence tracking is introduced. If continuous Frame satisfies Then escalate to an alarm; otherwise reset. It was determined that there was no fire.

9. A high-speed flame detection device based on an HVS visual model, used to implement the high-speed flame detection method based on an HVS visual model as described in any one of claims 1-8, characterized in that, include: The front-end image acquisition module is used to acquire video image signals of the monitored area in real time. The data buffer unit is used to store the current processing frame and historical image data of multiple consecutive frames; The core heterogeneous computing unit includes a field-programmable gate array (FPGA) subsystem and a digital signal processor (DSP) or ARM processor subsystem. The FPGA subsystem is used to perform pixel-level parallel processing, including color space conversion, image filtering, and morphological operations. The DSP or ARM processor subsystem is used to perform Gaussian mixture model parameter maintenance, connected component labeling, temporal flicker feature statistics, and multi-feature fusion decision. After completing pixel-level preprocessing of the entire image, the FPGA subsystem transmits the candidate region coordinates to the DSP or ARM processor subsystem. The DSP or ARM processor subsystem only performs background modeling updates and flicker feature calculations for the candidate regions. The program and parameter storage unit is used to store the startup program, flame statistical model parameters, and dynamic threshold configuration; The alarm output interface is used to output an alarm signal when a fire is detected.

Images