Fire identification method and system based on video monitoring

Abstract

The application relates to the technical field of image recognition, in particular to a fire identification method and system based on video monitoring. The technical problem of insufficient reliability of a video monitoring system in fire smoke early warning is solved. The method comprises the following steps: acquiring a monitoring video stream of a monitoring area and time sequence data collected by a sensor of the monitoring area; determining an evaluation value of a to-be-detected area based on image features and morphological changes of the to-be-detected area; determining a confidence factor based on a trend of morphological changes of the to-be-detected area and a change trend of the time sequence data; determining a smoke authenticity value based on the evaluation value and the confidence factor; and adjusting an alarm threshold of a smoke alarm based on the smoke authenticity value. The application is used in a fire smoke early warning monitoring scene.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically to a fire detection method and system based on video surveillance. Background Technology

[0002] Forest fire prevention and control is a crucial issue for ecological environmental protection and public safety. Traditional manual patrols suffer from limitations such as limited coverage and delayed response. With the development of security technology, automatic fire detection systems based on video surveillance are widely used in forest areas, aiming to provide early fire warnings through real-time analysis of monitoring footage. These systems typically rely on algorithms that detect visual features such as smoke and flames in video images. However, in complex outdoor environments like forest parks, monitoring scenarios are subject to various interferences, including weather conditions and changes in lighting. Existing fire detection solutions often focus on improving the accuracy of image recognition algorithms in specific scenarios or simply linking independent smoke sensors. However, when faced with widespread natural phenomena that resemble the visual characteristics of fire smoke, their judgment criteria remain simplistic. This leads to a widespread problem of low reliability in early warning systems during actual deployment, manifested in frequent false alarms. This not only consumes significant emergency response resources but also reduces the credibility of the early warning system, impacting its practical application effectiveness. Summary of the Invention

[0003] To address the technical problem that existing methods rely on a single basis for judgment when faced with widespread natural phenomena that resemble the visual characteristics of fire smoke, resulting in low reliability of early warnings in practical deployments, the present invention aims to provide a fire identification method and system based on video surveillance. The specific technical solution adopted is as follows:

[0004] In a first aspect, the present invention provides a fire identification method based on video surveillance. The method includes: acquiring a surveillance video stream of a monitored area and time-series data collected by sensors in the monitored area; using the surveillance video stream to determine a detection area where pixel changes occur; the sensors include at least a temperature and humidity sensor and a smoke sensor; determining an evaluation value for the detection area based on image features and morphological changes; the evaluation value characterizing the probability that the detection area is smoke; determining a confidence factor based on the trend of morphological changes in the detection area and the trend of changes in the time-series data; the confidence factor used to correct the evaluation value; the consistency between the trend of morphological changes and the trend of changes in smoke sensor data is positively correlated with the confidence factor; the consistency between the trend of morphological changes and the trend of changes in temperature and humidity sensor data is negatively correlated with the confidence factor; determining a smoke authenticity value based on the evaluation value and the confidence factor; the smoke authenticity value characterizing the overall confidence that the detection area is real smoke; and adjusting the alarm threshold of the smoke alarm based on the smoke authenticity value.

[0005] In conjunction with the first aspect mentioned above, in one possible implementation, the method specifically includes: determining the intensity of the radial gradient features of the region to be detected; the radial gradient features are used to characterize the characteristic that pixel values ​​diffuse and decrease from one or more high-density points in the region to be detected to the surrounding areas; determining the morphological change rate of the region to be detected in consecutive video frames of the monitoring video stream; and determining an evaluation value based on the intensity of the radial gradient features and the morphological change rate.

[0006] In conjunction with the first aspect mentioned above, in one possible implementation, the method specifically includes: generating a temperature change trend curve, a humidity change trend curve, and a smoke concentration change trend curve based on time-series data collected by sensors; determining a first difference value between the morphological change trend and the temperature change trend curve; determining a second difference value between the morphological change trend and the humidity change trend curve; determining a third difference value between the morphological change trend and the smoke concentration change trend curve; and determining a confidence factor based on the first difference value, the second difference value, and the third difference value.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the method further includes: determining a morphological change trend curve based on the area change of the region to be detected in consecutive video frames of the monitoring video stream; the morphological change trend curve is used to characterize the trend of morphological change.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, the method specifically includes: determining at least one concentration center point within the region to be detected; the concentration center point being the point with the highest pixel value within the region to be detected; determining that the region to be detected has radial gradient features when the pixel values ​​of the concentration center point gradually decrease outward from the concentration center point in multiple directions; and determining the intensity of the radial gradient features based on the significance of the decreasing distribution.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the method specifically includes: determining the area sequence of the region to be detected in consecutive video frames; and determining the rate of morphological change based on the area difference between each adjacent image in the area sequence.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, the method specifically includes: applying a decreasing function to the first difference value and the second difference value to determine the first correction value and the second correction value; applying an increasing function to the third difference value to determine the third correction value; and determining the confidence factor based on the first correction value, the second correction value, and the third correction value.

[0011] In conjunction with the first aspect mentioned above, in one possible implementation, the method specifically includes: determining the weight adjustment coefficient of the smoke alarm based on the smoke authenticity value; lowering the original alarm threshold of the smoke alarm based on the weight adjustment coefficient; the higher the smoke authenticity value, the lower the adjusted alarm threshold.

[0012] In conjunction with the first aspect mentioned above, in one possible implementation, the method further includes: performing pixel difference processing on consecutive video frames in the monitoring video stream; marking the frame image where the first pixel change occurs as the initial frame; starting from the initial frame, determining the region in each frame image where the pixel change occurs compared to the reference template as the region to be detected where the pixel change occurs; the reference template is any frame image before the initial frame.

[0013] Secondly, this invention provides a fire detection system based on video surveillance. The system includes: a data acquisition module for acquiring a monitoring video stream of a monitored area and time-series data collected by sensors in the monitored area; the monitoring video stream for identifying areas to be detected where pixel changes occur; sensors including at least a temperature and humidity sensor and a smoke sensor; a visual evaluation module for determining an evaluation value of the area to be detected based on image features and morphological changes; the evaluation value characterizes the probability that the area to be detected is smoke; a confidence calculation module for determining a confidence factor based on the trend of morphological changes in the area to be detected and the trend of changes in time-series data; the confidence factor is used to correct the evaluation value; the consistency between the trend of morphological changes and the trend of changes in smoke sensor data is positively correlated with the confidence factor; the consistency between the trend of morphological changes and the trend of changes in temperature and humidity sensor data is negatively correlated with the confidence factor; a authenticity value determination module for determining a smoke authenticity value based on the evaluation value and the confidence factor; the smoke authenticity value characterizes the comprehensive confidence that the area to be detected is real smoke; and a threshold adjustment module for adjusting the alarm threshold of the smoke alarm based on the smoke authenticity value.

[0014] The present invention has the following beneficial effects:

[0015] This invention effectively solves the problem of high false alarm rates in traditional video fire detection in complex outdoor environments by constructing a complete logical chain: preliminary visual assessment, multi-sensor confidence correction, comprehensive authenticity determination, and dynamic threshold adjustment. First, dynamic regions are extracted from the video and a preliminary assessment is performed based on their image features. Then, synchronous time-series data from temperature, humidity, and smoke sensors are introduced. Correction factors are generated by analyzing the correlation between visual morphological change trends and the trends of various sensor data. Finally, the alarm threshold is dynamically and precisely adjusted based on the comprehensive judgment of smoke authenticity. This multi-level, multi-source information fusion decision-making mechanism significantly enhances the ability to distinguish interference such as natural fog. While ensuring a high-sensitivity response to real fires, it greatly reduces the system's false alarm rate, improving the reliability and practicality of the warning system. This solves the technical problem of existing methods relying on a single judgment basis when facing widely existing natural phenomena similar to fire smoke in visual characteristics, resulting in generally low reliability of warnings in actual deployments. Attached Figure Description

[0016] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a fire detection method based on video surveillance provided in one embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of a fire detection system architecture based on video surveillance, provided in one embodiment of the present invention. Detailed Implementation

[0019] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the video surveillance-based fire identification method and system proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] The specific solution of the fire identification method and system based on video surveillance provided by the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Please see Figure 1 The diagram illustrates a fire detection method based on video surveillance according to an embodiment of the present invention. The method includes the following steps S101-S105, which will be described in detail below.

[0023] S101. Obtain the monitoring video stream of the monitored area, as well as the time-series data collected by the sensors in the monitored area.

[0024] The monitoring video stream is used to determine the area to be detected where pixel changes occur; the sensors include at least: a temperature and humidity sensor and a smoke sensor.

[0025] In one possible implementation, while acquiring continuous video frames, the readings of temperature sensors, humidity sensors, and smoke sensors covering the same physical space are collected at that moment to form a multimodal data packet with strictly matched timestamps.

[0026] S102. Determine the evaluation value of the region to be detected based on the image features and morphological changes of the region to be detected.

[0027] The evaluation value is used to characterize the probability that the area to be detected is smoke.

[0028] In one possible implementation, for the identified region to be detected, two core analyses are performed: first, analyzing whether it exhibits a gray-level gradient distribution characteristic that radiates and diffuses from a high-concentration point in the interior to the surrounding areas, and quantifying the significance of this characteristic, denoted as gradient feature intensity; second, calculating the degree of drastic change in the area of ​​the region in consecutive video frames, denoted as morphological change rate. Finally, a single evaluation value is calculated by combining the gradient feature intensity and the morphological change rate.

[0029] For example, the gradient feature intensity can be determined by locating the brightest pixel in the region (i.e., the density center) and checking whether the pixel value in each direction shows a stable decreasing trend; the rate of morphological change can be characterized by calculating the average magnitude of the area difference of the region in consecutive frames.

[0030] S103. Determine the confidence factor based on the trend of morphological changes in the area to be detected and the trend of changes in time series data.

[0031] Among them, the confidence factor is used to correct the evaluation value; the consistency between the trend of morphological change and the trend of smoke sensor data change is positively correlated with the confidence factor; the consistency between the trend of morphological change and the trend of temperature and humidity sensor data change is negatively correlated with the confidence factor.

[0032] In one possible implementation, for the morphological change trend curve characterizing the dynamic area of ​​the region to be detected, and for three independent trend curves of temperature, humidity, and smoke concentration generated from synchronous sensor data, the degree of consistency between the morphological change trend curve and each sensor trend curve in terms of direction and amplitude of change within the same time period is calculated, resulting in three independent correlation values. Subsequently, these three correlation values ​​are used as inputs to calculate a comprehensive confidence factor.

[0033] For example, the three correlation values ​​can be obtained by calculating the rate of change sequence of the corresponding trend curves at the corresponding time points and analyzing the correlation between the sequences. The core logic of calculating the comprehensive confidence factor is to negatively process the two values ​​reflecting the correlation with temperature and humidity, positively process the value reflecting the correlation with smoke concentration, and then comprehensively calculate the processed results so that the final confidence factor can quantitatively reflect the degree of consistency between visual dynamics and fire environment characteristics.

[0034] S104. Determine the smoke authenticity value based on the evaluation value and confidence factor.

[0035] Among them, the smoke authenticity value is used to characterize the overall confidence level that the area to be detected is real smoke.

[0036] One possible implementation involves combining the visual assessment value with a confidence factor to output a comprehensive smoke authenticity value. The core of this combination operation is to use the confidence factor as a weighting factor for the initial visual assessment value, thereby obtaining a comprehensive judgment value that more closely approximates the actual situation. Specifically, the smoke authenticity value is set to be positively correlated with both the visual assessment value and the confidence factor, reaching its maximum when both are high.

[0037] For example, the combined operation can be implemented using mathematical relationships such as multiplication or weighted summation. By multiplying the assessment value representing visual probability with a confidence factor representing the reliability of multi-sensor data verification, the final authenticity value can simultaneously reflect the strength of visual evidence and its consistency with objective environmental data. This value directly quantifies the overall confidence level that the area to be detected is real fire smoke.

[0038] S105. Adjust the alarm threshold of the smoke detector based on the smoke authenticity value.

[0039] In one possible implementation, the smoke authenticity value is input into a threshold mapping function, which outputs a corresponding weight adjustment coefficient. Then, based on this coefficient, the original alarm thresholds of one or more smoke detectors closest to the detection area are adjusted in real time. The mapping function is configured such that the higher the smoke authenticity value, the larger the generated weight adjustment coefficient, and consequently, the greater the reduction in the original alarm threshold, making the detector more sensitive and more likely to trigger an alarm at the same smoke concentration.

[0040] Understandably, when the smoke accuracy value reaches or approaches its theoretical maximum value, it can be mapped to a weight adjustment coefficient close to or equal to 1. This adjusts the effective alarm threshold of the corresponding smoke alarm to an extremely low level, or even triggers the alarm directly, achieving a rapid response. Conversely, when the smoke accuracy value is low, the weight adjustment coefficient decreases accordingly, and the alarm threshold remains unchanged or is only slightly adjusted, thereby suppressing false alarms caused by visually similar objects.

[0041] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment effectively solves the problem of high false alarm rate in traditional video fire identification in complex outdoor environments by constructing a complete logical chain of visual preliminary assessment - multi-sensor confidence correction - comprehensive authenticity judgment - dynamic threshold adjustment. First, dynamic areas are extracted from the video and a preliminary assessment is performed based on their image features. Then, synchronous time-series data from temperature, humidity, and smoke sensors are introduced. Correction factors are generated by analyzing the correlation between visual morphological change trends and the trends of various sensor data. Finally, the alarm threshold is dynamically and accurately adjusted based on the comprehensively judged smoke authenticity value. This multi-level, multi-source information fusion decision-making mechanism significantly enhances the ability to distinguish interference such as natural mountain fog. While ensuring a high-sensitivity response to real fires, it greatly reduces the false alarm rate of the system, improving the reliability and practicality of the early warning system.

[0042] In one possible implementation, the process of determining the evaluation value of the region to be detected based on the image features and morphological changes of the region to be detected can be specifically implemented through the following S201-S203, which will be described in detail below.

[0043] S201. Determine the intensity of the radial gradient features in the region to be detected.

[0044] Among them, the radial gradient feature is used to characterize the characteristic that the pixel value diffuses and decreases from one or more high-concentration points in the detection area to the surrounding areas.

[0045] In one possible implementation, firstly, at least one concentration center point is located within the detection area. The concentration center point is the point with the highest pixel gray value in that area. If there are multiple adjacent high gray value points with the same gray value, the continuous area formed by these points is considered as a concentration center. Then, starting from each concentration center point, the points extend outward along multiple preset uniformly distributed directions (e.g., with the center as the origin, a direction is set every 45 degrees, for a total of 8 directions). Next, for each direction, the gray values ​​of a series of adjacent pixels from the center point outward are acquired sequentially, and it is analyzed whether these gray values ​​show a distribution trend of gradually decreasing from the center point outward. Finally, a scalar value is determined based on the significance of this decreasing trend in all directions, serving as the intensity of the radial gradient feature.

[0046] S202. Determine the rate of morphological change of the region to be detected in consecutive video frames of the monitoring video stream.

[0047] In one possible implementation, firstly, the area of ​​the region to be detected in each frame of the monitored video stream within a time period is obtained, typically quantized by the number of pixels contained in the region, forming a set of area sequences arranged in chronological order; then, the absolute difference between the area values ​​of each pair of adjacent frames in the area sequence is calculated; finally, these absolute differences are statistically processed (e.g., their average value is calculated), and the processing result is used as a numerical value characterizing the rate of morphological change.

[0048] S203. Determine the evaluation value based on the intensity and morphological change rate of the radial gradient characteristics.

[0049] In one possible implementation, the intensity of the radial gradient feature and the rate of morphological change—two quantified parameters representing spatial diffusion and temporal dynamics, respectively—are fused using a predefined evaluation rule to output a comprehensive evaluation value. The evaluation rule is configured such that the evaluation value increases with both the intensity of the radial gradient feature and the rate of morphological change. Specifically, both parameters can be used as inputs and calculated using a predefined evaluation function designed to ensure that regions possessing both high gradient feature intensity and high morphological change rate receive higher evaluation values.

[0050] For example, the evaluation value of the region to be detected as smoke in all frames of the j-th video segment. Satisfy the following formula 1:

[0051]

[0052] in, is the average value of the radial gradient feature intensity of all video frames in the j-th video segment. This value has been normalized, is dimensionless, and ranges from [0, 1]. is the number of adjacent frame pairs used to calculate the rate of morphological change within the j-th video time segment (total number of frames minus 1); x is the sequence number of consecutive video frames within the video time segment; Let be the normalized area of ​​the dynamic region to be detected in the x-th frame of the image. The area is quantized by the number of pixels and then normalized by dividing it by a reference area (such as the maximum possible area). Dimensionless, with values ​​in the range [0, 1]; It is the normalized area of ​​the dynamic region to be detected on the (x+1)th frame of the image.

[0053] It characterizes the persistence and average intensity of the typical smoke diffusion characteristics in the region during this time period; It represents the average drastic degree of regional area change (i.e., the rate of morphological change) during this period. It is a composite index that combines spatial diffusion characteristics (gradient) and temporal dynamic characteristics (rate of change), used to initially assess the possibility of a dynamic area being fire smoke from a purely visual perspective.

[0054] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment determines the evaluation value by simultaneously quantifying two key visual indicators, the intensity of the radial gradient features and the rate of morphological change of the area to be detected, providing a more robust and accurate preliminary judgment of the likelihood of smoke than relying on a single feature. The radial gradient features effectively capture the physical characteristics of smoke diffusion from the source point, while the rate of morphological change reflects the instability of the combustion process. The combination of the two makes the preliminary screening more able to distinguish areas with typical dynamic characteristics of fire smoke, laying a reliable visual evidence foundation for subsequent steps and avoiding misjudging uniformly diffused and slowly changing mountain fog as a fire.

[0055] In one possible implementation, the process of determining the confidence factor based on the trend of morphological changes in the region to be detected and the trend of changes in time series data can be specifically implemented through the following S301-S305, which will be explained in detail below.

[0056] S301. Based on the time-series data collected by the sensor, generate temperature change trend curve, humidity change trend curve and smoke concentration change trend curve.

[0057] In one possible implementation, firstly, the timestamped data sequences collected by each sensor are aligned and resampled according to the time nodes used in the video analysis (such as the moment of a video frame or the start of the analysis period) to ensure that each analysis moment corresponds to a set of synchronized environmental parameter values. Then, the resampled temperature data sequences, humidity data sequences, and smoke concentration data sequences are smoothed and filtered to suppress random noise interference and highlight their changing trends. Finally, continuous temperature change trend curves, humidity change trend curves, and smoke concentration change trend curves are generated with time as the horizontal axis and the parameter values ​​as the vertical axis.

[0058] S302. Determine the first difference value between the trend of morphological change and the trend curve of temperature change.

[0059] In one possible implementation, firstly, the morphological change trend curve and the temperature change trend curve are sampled within the same analysis time window to obtain their instantaneous rate of change sequences at the corresponding sampling time points (e.g., by calculating the slope or difference between adjacent sampling points); then, a preset difference measurement algorithm is used to calculate the overall degree of inconsistency between the two rate of change sequences; finally, the calculation result is output as the first difference value.

[0060] For example, the difference measure algorithm can be the mean absolute error algorithm. That is, it calculates the absolute value of the difference between the two change rate sequences at each sampling point, and then averages these absolute values. The first difference value output by this algorithm has the physical meaning of the average difference between morphological changes and temperature changes. The smaller this value (theoretically close to 0), the more synchronized the rate of change of visual morphology and the rate of change of temperature are in time, and the smaller the difference; the larger the value, the more asynchronous the two changes are, and the greater the difference.

[0061] S303. Determine the second difference value between the trend of morphological change and the trend of humidity change.

[0062] In one possible implementation, firstly, the morphological change trend curve and the humidity change trend curve are sampled within the same analysis time window to obtain their instantaneous rate of change sequences at the corresponding sampling time points; then, a preset difference measurement algorithm is used to calculate the overall degree of inconsistency between the two rate of change sequences; finally, the calculation result is output as a second difference value.

[0063] For example, the difference measure algorithm can be the mean absolute error algorithm. That is, the absolute value of the difference between the two change rate sequences at each sampling point is calculated, and then the average of these absolute values ​​is calculated. The second difference value output by this algorithm has the physical meaning of the average difference between morphological changes and humidity changes. The smaller this value (theoretically close to 0), the more synchronous the rate of change of visual morphology and the rate of change of humidity are in time, and the smaller the difference; the larger the value, the more asynchronous the two changes are, and the greater the difference.

[0064] S304. Determine the third degree of difference between the trend of morphological change and the trend curve of smoke concentration change.

[0065] In one possible implementation, firstly, the morphological change trend curve and the smoke concentration change trend curve are sampled within the same analysis time window to obtain their instantaneous change rate sequences at the corresponding sampling time points; then, a preset difference measurement algorithm is used to calculate the overall inconsistency between the two change rate sequences; finally, the calculation result is output as the third difference value.

[0066] For example, the difference measure algorithm can be the mean absolute error algorithm. That is, it calculates the absolute value of the difference between the two change rate sequences at each sampling point, and then averages these absolute values. The third difference value output by this algorithm has the physical meaning of the average difference between morphological changes and smoke concentration changes. The smaller this value (theoretically close to 0), the more synchronized the rate of change of visual morphology and the rate of change of smoke concentration are in time, and the smaller the difference; the larger the value, the more asynchronous the changes are, and the greater the difference.

[0067] S305. Determine the confidence factor based on the first difference value, the second difference value, and the third difference value.

[0068] In one possible implementation, the first, second, and third difference values ​​(as input) are calculated using a pre-defined confidence fusion model to output a comprehensive confidence factor. The core design logic of the fusion model is that the confidence factor should be positively correlated with the third difference value and negatively correlated with the first and second difference values. Specifically, the model processes the first and second difference values ​​using a decreasing function (e.g., by inputting them into a decreasing function or taking their complement with respect to 1) to transform their negative correlation with the confidence level; and processes the third difference value using an increasing function (e.g., by keeping the original value or by enhancing its positive influence through a function). Finally, the three processed values ​​are combined through a fusion operation (such as multiplication or weighted geometric mean) and normalized to obtain a confidence factor with a value range in the [0, 1] interval.

[0069] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment achieves a deep multi-source data cross-validation mechanism by performing correlation analysis between the visual morphological change trend and the independent temperature, humidity, and smoke concentration change trends, and determining the confidence factor based on multiple correlation values. It not only utilizes data from a directly correlated device like a smoke sensor, but also creatively introduces environmental background parameters such as temperature and humidity. By analyzing the consistency or deviation between visual dynamics and the changes in these three factors, it can effectively distinguish visually similar phenomena caused by fire (accompanied by rising smoke, possibly accompanied by some temperature change but inconsistent with the humidity change trend) and those caused by meteorological conditions (which may lead to synchronous changes in temperature and humidity), thereby greatly improving the specificity and reliability of the judgment.

[0070] In one possible implementation, it is also necessary to determine the morphological change trend curve. This process can be specifically implemented through the following S401, which will be explained in detail below.

[0071] S401. Determine the morphological change trend curve based on the area change of the region to be detected in consecutive video frames of the monitoring video stream.

[0072] Among them, the morphological change trend curve is used to characterize the trend of morphological change.

[0073] In one possible implementation, firstly, for each frame of the surveillance video stream within a selected time period, the area occupied by the region to be detected in that frame is calculated, usually quantized by the number of pixels contained in that region, thus obtaining a sequence of area values ​​arranged in chronological order; then, the original area sequence is preprocessed (such as smoothing filtering to suppress noise); finally, a continuous or discrete curve is generated with time as the horizontal axis and the processed area value as the vertical axis, which is the morphological change trend curve.

[0074] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment explicitly quantifies the trend of morphological changes into a curve based on area changes, providing a unified and computable data foundation for subsequent accurate time-series correlation analysis with sensor data. This step transforms the abstract trend into a concrete mathematical object (curve), enabling visual dynamics and sensor signals to be quantitatively compared on the same time dimension, ensuring the logical rigor and computational feasibility of the fusion analysis, which is a key prerequisite for achieving accurate data fusion in this solution.

[0075] In one possible implementation, the process of determining the intensity of the radial gradient features of the region to be detected can be specifically implemented through the following S501-S503, which will be described in detail below.

[0076] S501. Determine at least one concentration center point within the area to be detected.

[0077] In one possible implementation, the pixel values ​​(e.g., grayscale values) of all pixels in the region to be detected are traversed, and the point with the highest pixel value is identified. If there is only one such point with the highest pixel value, it is directly determined as the unique density center point. If there are multiple pixels with the same highest pixel value, it is further determined whether these points are spatially adjacent to each other (e.g., by using the 4-connectivity or 8-connectivity criterion to determine pixel adjacency). Multiple high points that meet the adjacency condition are merged into a continuous high-density region, and the geometric center of the region or any of its high points is used as a representative and marked as a density center point.

[0078] S502. When the pixel values ​​at the concentration center point gradually decrease outward from the concentration center point in multiple directions, it is determined that the area to be detected has radial gradient characteristics.

[0079] In one possible implementation, rays extend outward from the concentration center point along multiple preset directions (e.g., with the center point as the origin, an analysis direction is set every 22.5 degrees or 45 degrees). For each ray, the pixel values ​​of a series of adjacent pixels starting from the center point are sequentially acquired. Then, it is analyzed whether the sequence of these pixel values ​​on each ray shows a monotonic or overall decreasing trend from the center point outward. Finally, if the pixel value sequence is determined to satisfy the condition of decreasing distribution in all preset directions, the region to be detected is comprehensively determined to have radial gradient characteristics.

[0080] S503. Determine the intensity of the radial gradient feature based on the significance of the decreasing distribution.

[0081] In one possible implementation, for each preset analysis direction, a sub-metric characterizing the significance of the gradient in that direction is calculated based on the steepness and consistency of the pixel values ​​decreasing outward from the concentration center point. Then, the sub-metrics for all directions are aggregated (e.g., by averaging), and the aggregated result is used as the final intensity value of the radial gradient feature. Steepness reflects the rate at which pixel values ​​decrease with increasing distance; the faster the decrease, the steeper the steepness. Consistency reflects the stability and continuity of the decreasing trend; the smaller the fluctuation, the higher the consistency.

[0082] For example, the intensity of the radial gradient features of the region to be detected on the i-th frame of the video image. The following formula 2 is satisfied:

[0083]

[0084] in, and This represents the grayscale value of two adjacent pixels in the k-th direction within the region to be detected in the i-th frame of the image, where The grayscale value of the pixel closer to the marker point; It represents all directions; in the formula These are parameter tuning coefficients, and their values ​​should be extremely small positive numbers (e.g., 0.01) to avoid denominators of 0.

[0085] Radial gradient features are quantified by calculating the average of the grayscale ratios of adjacent pixels in all directions. The significance of the radial decreasing distribution pattern of pixel gray values ​​of the region to be detected in a single frame image is quantified. This feature is one of the key visual indicators for distinguishing fire smoke (which spreads from a fixed source) from natural mountain fog (which is usually evenly distributed).

[0086] In one possible implementation, for all frames within the j-th video segment... Calculate the mean to obtain , It represents the average saliency of the visual pattern of diffusion from the center outward in the region to be detected within the time window of video segment j. It is a temporal aggregate feature that is more stable than single-frame features.

[0087] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment uses the explicit definition of the concentration center point and the decreasing distribution of pixel values ​​in multiple directions as the specific criteria for determining radial gradient features, making the feature recognition process more objective, operable, and quantifiable. Determining the feature intensity based on the significance of the decreasing distribution further transforms qualitative observation into quantitative indicators, enhancing the stability and consistency of the preliminary evaluation process, reducing subjective bias, and providing a solid guarantee for generating reliable evaluation values.

[0088] In one possible implementation, the process of determining the rate of morphological change of the region to be detected in consecutive video frames of the monitoring video stream can be specifically implemented through the following S601-S602, which will be described in detail below.

[0089] S601. Determine the area sequence of the region to be detected in consecutive video frames.

[0090] In one possible implementation, according to the time sequence of the monitored video stream, for each selected consecutive multi-frame image, the total number of pixels occupied by the region to be detected in that frame image is calculated, and this value is used as the area value corresponding to that frame; then, according to the time sequence of the video frames, all the calculated area values ​​are arranged into an ordered sequence, which is the area sequence.

[0091] S602. Determine the rate of morphological change based on the area difference between each adjacent image in the area sequence.

[0092] In one possible implementation, first, an area sequence arranged in chronological order is obtained; then, the absolute difference between the area values ​​corresponding to each pair of adjacent frames (i.e., frame x and frame x+1) in the sequence is calculated; finally, all these calculated absolute differences are aggregated statistically (e.g., their arithmetic mean is calculated), and this statistical result is output as the rate of morphological change.

[0093] For example, if the area sequence contains m area values ​​(corresponding to m frames), then m-1 absolute differences between adjacent frames can be obtained. Convergence statistics can be performed by calculating the average of these differences, i.e., summing all differences and then dividing by the number of differences (m-1), thus obtaining the average area change within a unit frame interval. To handle cases where the area does not change between adjacent frames, a very small positive parameter can be introduced when calculating the differences to ensure the numerical stability and validity of the calculation. The larger this rate value, the more drastic the area fluctuation between consecutive frames.

[0094] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment quantifies the rate of morphological change by using the average absolute value of the area sequence and its adjacent differences, providing a stable and computationally efficient method to characterize the intensity of regional dynamics. This quantization method smooths the area changes between consecutive frames, effectively capturing the unstable trend of the overall regional changes, avoiding interference that may be caused by abrupt changes in a single frame, and making the parameter of morphological change rate more robustly reflect the typical dynamic characteristics of smoke.

[0095] In one possible implementation, the process of determining the confidence factor based on the first difference value, the second difference value, and the third difference value can be specifically implemented through the following S701-S703, which will be explained in detail below.

[0096] S701. Perform a decreasing function on the first difference value and the second difference value to determine the first correction value and the second correction value.

[0097] In one possible implementation, the first difference value is taken as input and mapped through a preset decreasing function to output a first correction value; similarly, the second difference value is mapped through the same or a different decreasing function to output a second correction value. The decreasing function is designed such that the output value decreases as the input value increases.

[0098] For example, the decreasing function can be a reciprocal function, a linear negative transformation function, or an exponential decay function. Regardless of the specific function used, the processed first and second correction values ​​satisfy the following: the lower the original correlation value (indicating that visual dynamics are less correlated with temperature and humidity changes), the higher the correction value; the higher the original correlation value, the lower the correction value.

[0099] S702. Perform incremental function processing on the third difference value to determine the third correction value.

[0100] In one possible implementation, the third difference value is used as input and mapped through a pre-defined incrementing function to output a third correction value. The incrementing function is designed such that the output value increases as the input value increases, and the output value tends to maximize when the input value reaches its theoretical maximum value.

[0101] For example, the incrementing function can be a simple linear function (e.g., the identity function y = x, i.e., directly using the original value) or a non-linear amplification function to produce a larger output value when the input value is high, thus highlighting the contribution of strong correlation evidence. To handle cases where the third dissimilarity value may be negative (indicating negative correlation), a lower limit can be pre-set, for example, truncating all input values ​​less than 0 to 0 before incrementing, since negative correlation does not contribute to fire detection. The processed third correction value satisfies the following: the higher the original correlation value (indicating more synchronization between visual dynamics and smoke concentration changes), the higher the correction value; the lower the original correlation value, the lower the correction value.

[0102] S703. Determine the confidence factor based on the first correction value, the second correction value, and the third correction value.

[0103] In one possible implementation, the first, second, and third correction values ​​are multiplied together to obtain a preliminary fusion value. Then, this preliminary fusion value is input into a normalization function (such as the norm function) to map its value range to a standard interval (e.g., [0, 1]). The result after mapping is the final confidence factor.

[0104] For example, the anomalous coupling between the area of ​​the region to be detected in the j-th video segment and temperature and humidity. The following formula 5 is satisfied:

[0105]

[0106] in, The first correction factor (the overall average difference between the morphological change trend and the temperature change trend) is the largest value, indicating that the two changes are more asynchronous and the correlation is weaker. The second correction factor (the overall average difference between the morphological change trend and the humidity change trend) has a larger value, indicating that the two changes are more asynchronous and the correlation is weaker. These are parameter tuning coefficients, and their values ​​should be extremely small positive numbers (e.g., 0.01) to avoid a denominator of 0. The anomalous coupling degree is positively correlated with humidity differences and negatively correlated with temperature differences. This is because fires may cause localized temperature changes, but they typically do not lead to synchronous and significant changes in ambient humidity. Therefore, This constitutes a The reciprocal transformation (a type of decreasing function processing); It is an abnormal coupling index that reflects the extent to which the visual dynamic changes of the area under test cannot be explained by the normal changes in ambient temperature and humidity. A high value suggests that the dynamic is more likely to be caused by combustion (leading to smoke) rather than ordinary weather changes (leading to mountain fog).

[0107] Understandably, abnormal coupling It is not simply proportional to the difference in temperature and humidity, but rather aims to quantify a specific anomaly pattern: that is, visual dynamics are highly uncorrelated with humidity changes, but maintain a certain correlation with temperature changes. This pattern is consistent with the physical characteristics of fire smoke, thus effectively distinguishing it from mountain fog (which is usually strongly correlated with humidity) or other disturbances without heat sources.

[0108] For example, in the j-th video segment, the confidence factor for the area to be detected as a fire smoke area. Satisfy the following formula 6:

[0109]

[0110] This represents the anomalous coupling degree between the area of ​​the region to be detected in the j-th video segment and temperature and humidity. The larger the value, the less correlated it is with temperature and humidity. Let be the slope of the morphological change trend curve in the t-th time segment within the j-th analysis time period; This represents the average rate of change of the smoke concentration trend curve within the time interval t. Used to identify parameters related to smoke concentration; These are parameter tuning coefficients, and their values ​​should be extremely small positive numbers (e.g., 0.01) to avoid a denominator of 0. This represents the total number of all time periods in the j-th video segment.

[0111] The third correction factor (the overall average difference between the morphological change trend and the smoke concentration change trend) is calculated by averaging the absolute differences of the μ segments to obtain the overall average difference between morphology and smoke. The smaller the value, the more synchronous the changes of the two are over the entire time period. Perform a reciprocal operation on the average degree of difference. Convert the degree of difference to the degree of synchronicity. The fusion process means that the final confidence level needs to meet two conditions simultaneously: it must be uncorrelated with normal environmental changes (…). (High), and also needs to be synchronized with smoke monitoring ( (High). Weak evidence from any single point will significantly reduce the product result, which is more stringent than weighted summation and effectively suppresses misjudgments caused by a single piece of evidence. This is a normalization function that normalizes the product by its maximum and minimum values, mapping the product result to a standard interval (such as [0,1]). It is a quantitative comprehensive credibility score that scientifically integrates two independent pieces of evidence: the visual dynamics being caused by non-environmental factors and the visual dynamics being highly synchronized with the monitoring of combustion products. Ultimately, it provides a probability estimate of whether the area contains real fire smoke.

[0112] It is understandable that all the intermediate variables involved in the above calculations are dimensionless values ​​that have been normalized or standardized.

[0113] The technical solution provided in the above embodiments can bring at least the following beneficial effects: This embodiment constructs an intelligent data fusion weight allocation model that conforms to the physical laws of fire by processing the correlation values ​​of temperature and humidity using a decreasing function and the correlation values ​​of smoke concentration using an increasing function, and then combining the results. It implements the logic of adding points for positive correlations with smoke trends and subtracting points for positive correlations with temperature and humidity trends, enabling the finally generated confidence factor to scientifically and quantitatively characterize the degree to which the detected area conforms to the characteristics of a fire smoke environment, significantly improving the rationality and discrimination accuracy of data fusion.

[0114] In one possible implementation, the process of adjusting the alarm threshold of the smoke detector based on the smoke authenticity value can be specifically implemented through the following S801-S802, which will be described in detail below.

[0115] S801. Determine the weight adjustment coefficient of the smoke alarm based on the smoke authenticity value.

[0116] In one possible implementation, the smoke accuracy value is input into a preset weighting mapping function, which outputs a corresponding weighting adjustment coefficient. The weighting mapping function is configured such that the output weighting adjustment coefficient is positively correlated and monotonically related to the input smoke accuracy value; that is, the higher the smoke accuracy value, the larger the determined weighting adjustment coefficient. Specifically, the weighting mapping function can be designed as a linear function or a piecewise linear function to ensure that the weighting adjustment coefficient changes continuously and stably with the smoke accuracy value, and that when the smoke accuracy value reaches or approaches its maximum value, the weighting adjustment coefficient approaches 1 (representing the maximum adjustment intensity).

[0117] For example, a linear mapping can be used: weight adjustment coefficient = smoke authenticity value. In this case, the weight adjustment coefficient and the smoke authenticity value are equal and both are in the interval [0,1].

[0118] S802. Based on the weighting adjustment coefficient, the original alarm threshold of the smoke detector is lowered.

[0119] In one possible implementation, the original alarm threshold of the smoke alarm and the weight adjustment coefficient determined by previous steps are obtained. Based on the weight adjustment coefficient, a new, reduced alarm threshold is calculated using a preset threshold adjustment function. The threshold adjustment function is configured such that the adjusted alarm threshold monotonically decreases as the weight adjustment coefficient increases. Specifically, the function can be designed to multiply the original threshold by a scaling factor related to the weight adjustment coefficient, where the scaling factor is less than or equal to 1.

[0120] For example, a linear adjustment model can be used: Adjusted threshold = Original alarm threshold × (1 - Weight adjustment coefficient). Under this model, when the weight adjustment coefficient is 0, the adjusted threshold is equal to the original threshold, i.e., no adjustment is made; when the weight adjustment coefficient is 1, the adjusted threshold becomes 0, meaning an alarm is triggered directly. To ensure system stability and prevent the threshold from being excessively lowered to negative values ​​or unreasonable ranges, a lower limit protection can be set in the function, for example: Adjusted threshold = max(Original alarm threshold × (1 - Weight adjustment coefficient), Minimum allowable threshold).

[0121] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment achieves efficient and accurate conversion from analysis conclusions to execution actions by establishing a direct mapping relationship between smoke authenticity values, weight adjustment coefficients, and the degree of alarm threshold reduction. This mechanism enables the sensitivity of the alarm system to be adjusted in real time and adaptively according to the confidence level of the threat, rapidly increasing the response level when a real fire is confirmed, and maintaining normal alert when interference is suspected. Thus, while achieving rapid emergency response, it minimizes resource waste and alarm fatigue caused by misjudgment.

[0122] In one possible implementation, it is also necessary to determine the detection area where pixel changes occur. This process can be specifically implemented through the following S901-S903, which will be described in detail below.

[0123] S901. Perform pixel difference processing on consecutive video frames in the monitoring video stream.

[0124] In one possible implementation, adjacent frames (denoted as frame N and frame N+1) are acquired sequentially according to the time sequence of the video frames. These two frames are then aligned at the pixel level (if the camera is fixed, alignment is assumed; if there is slight jitter, image stabilization can be performed first). Next, the pixel value difference between the two frames at each corresponding spatial coordinate position is calculated (for example, the grayscale difference is calculated directly for grayscale images, and for color images, it can be converted to the luminance channel or the difference between each channel can be calculated separately). Finally, a difference image with the same size as the original frame is generated, where the value of each pixel is the absolute value of the pixel difference between the two frames at the corresponding position.

[0125] S902, Mark the frame image where the first pixel change occurs as the initial frame.

[0126] In one possible implementation, during the continuous pixel difference processing of the monitoring video stream, it is determined in real time whether a pixel change region of sufficient scale that meets preset conditions has appeared. Specifically, this includes: sequentially processing the pixel difference results of each pair of adjacent video frames; when it is detected that the area of ​​the connected changed pixel region (or the total number of pixels) in the difference results of a pair of adjacent frames exceeds a preset significance threshold for the first time, the image of the next frame in this pair of adjacent frames (i.e., the frame in which the change actually occurs and is first confirmed) is marked as the initial frame.

[0127] For example, the saliency threshold is an area threshold (e.g., 0.1% of the total image area). When calculations show that the total area of ​​the connected regions formed by all the changing pixels in the differential image from frame N to frame N+1 is greater than this threshold for the first time, then frame N+1 is marked as the initial frame of this monitoring event. This threshold is used to filter out small, trivial changes caused by image noise, minor variations in illumination, etc., ensuring that the marked initial frame corresponds to the start of a noteworthy dynamic event of a certain scale (such as the beginning of smoke formation or the emergence of mountain fog).

[0128] Understandably, this step is a crucial judgment step in identifying the starting point of valid event analysis from continuous monitoring. By setting a significance threshold and identifying the first time it is exceeded, it is possible to intelligently ignore persistent and meaningless background fluctuations and accurately capture the starting moment of a new dynamic event.

[0129] S903. Starting from the initial frame, the regions in each frame of the image that have undergone pixel changes compared to the reference template are identified as the regions to be detected that have undergone pixel changes.

[0130] In one possible implementation, after marking the initial frame, dynamic region extraction based on template comparison is performed on each subsequent frame image, starting from this initial frame. Specifically, this includes: first, selecting a reference template, which is an image preceding the initial frame that has not undergone significant pixel changes in any frame (e.g., the most recent unchanged image before the initial frame, or an earlier frame representing a static background); then, for each frame image from the initial frame onwards, performing pixel-level alignment and difference operations with the reference template to calculate the pixel value difference between the current frame and the static template; finally, identifying connected regions in the difference result where pixel changes exceed an area threshold as the detection regions in the current frame where pixel changes have occurred.

[0131] The technical solution provided in the above embodiments can bring at least the following beneficial effects: This embodiment uses pixel difference, initial frame marking, and template comparison to determine the dynamic region to be detected, providing an accurate, real-time, and robust method for moving target detection against static backgrounds. This method can effectively capture newly added and changing regions in the video, locking in the accurate target range for subsequent targeted feature analysis, which is an important prerequisite for the initiation and effective execution of the entire fire identification process.

[0132] Please see Figure 2 The diagram illustrates a schematic of a video surveillance-based fire detection system architecture according to an embodiment of the present invention. The system includes: a data acquisition module 201, used to acquire a monitoring video stream of a monitored area and time-series data collected by sensors in the monitored area; the monitoring video stream is used to determine the detection area where pixel changes occur; the sensors include at least: a temperature and humidity sensor and a smoke sensor; a visual evaluation module 202, used to determine an evaluation value of the detection area based on image features and morphological changes of the detection area; the evaluation value is used to characterize the probability that the detection area is smoke; and a confidence calculation module 203, used to calculate the confidence level based on the detection area... The confidence factor is determined by the trend of morphological changes and the trend of time-series data; the confidence factor is used to correct the evaluation value; the consistency between the trend of morphological changes and the trend of smoke sensor data is positively correlated with the confidence factor; the consistency between the trend of morphological changes and the trend of temperature and humidity sensor data is negatively correlated with the confidence factor; the authenticity value determination module 204 is used to determine the smoke authenticity value based on the evaluation value and the confidence factor; the smoke authenticity value is used to characterize the comprehensive confidence that the area to be detected is real smoke; the threshold adjustment module 205 is used to adjust the alarm threshold of the smoke alarm based on the smoke authenticity value.

[0133] The technical solution provided in the above embodiments can bring at least the following beneficial effects: This embodiment integrates functional units such as data acquisition, area determination, visual assessment, trend determination, confidence calculation, authenticity value determination, and threshold control through modular design, forming a complete and collaborative hardware and software entity. This system solidifies the innovative methodology into an implementable and deployable solution, ensuring the stable and automated operation of the method in actual monitoring scenarios, and providing a directly usable system-level product implementation for early fire warning in outdoor areas such as forest parks.

[0134] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0135] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A fire detection method based on video surveillance, characterized in that, The method is applied to forest fire prevention and control, including: The system acquires a monitoring video stream of the monitored area, as well as time-series data collected by sensors in the monitored area; the monitoring video stream is used to determine the detection area where pixel changes occur; the sensors include at least: a temperature and humidity sensor and a smoke sensor. Based on the image features and morphological changes of the region to be detected, an evaluation value for the region to be detected is determined; the evaluation value is used to characterize the probability that the region to be detected is smoke. Based on the trend of morphological changes in the area to be detected and the trend of changes in the time-series data, a confidence factor is determined, including: generating a temperature change trend curve, a humidity change trend curve, and a smoke concentration change trend curve based on the time-series data collected by the sensor; determining a first difference value between the trend of morphological changes and the temperature change trend curve; determining a second difference value between the trend of morphological changes and the humidity change trend curve; determining a third difference value between the trend of morphological changes and the smoke concentration change trend curve; applying a decreasing function to the first difference value and the second difference value to determine a first correction value and a second correction value; applying an increasing function to the third difference value to determine a third correction value; and determining the confidence factor based on the first correction value, the second correction value, and the third correction value. The confidence factor is used to correct the evaluation value. The consistency between the trend of morphological changes and the trend of changes in smoke sensor data is positively correlated with the confidence factor; the consistency between the trend of morphological changes and the trend of changes in temperature and humidity sensor data is negatively correlated with the confidence factor. Based on the evaluation value and the confidence factor, a smoke authenticity value is determined; the smoke authenticity value is used to characterize the overall confidence that the area to be detected is real smoke. Adjust the alarm threshold of the smoke detector based on the smoke accuracy value; The confidence factor satisfies the following formula: Represents the confidence factor in the j-th video segment. This represents the anomalous coupling degree between the area of ​​the region to be detected in the j-th video segment and temperature and humidity. The higher the value, the less correlated it is with temperature and humidity; Let be the slope of the morphological change trend curve in the t-th time segment within the j-th video segment; This represents the average rate of change of the smoke concentration trend curve within the time interval t. Used to identify parameters related to smoke concentration; These are the parameter tuning coefficients; Represents the total number of time segments in the j-th video segment; norm represents the normalization function; This represents the third correction value in the j-th video segment; The abnormal coupling degree satisfies the following formula: in, This represents the first correction value in the j-th video segment. The larger the value, the less correlated the trend of morphological change is with the trend of temperature change. This represents the second correction value in the j-th video segment. The larger the value, the less correlated the trend of morphological change is with the trend of humidity change. These are the parameter tuning coefficients.

2. The fire detection method based on video surveillance according to claim 1, characterized in that, The step of determining the evaluation value of the region to be detected based on image features and morphological changes includes: Determine the intensity of the radial gradient feature of the region to be detected; the radial gradient feature is used to characterize the characteristic that the pixel value diffuses and decreases from one or more high-concentration points in the region to be detected to the surrounding areas. Determine the rate of morphological change of the region to be detected in consecutive video frames of the monitored video stream; The evaluation value is determined based on the intensity of the radial gradient feature and the rate of morphological change.

3. The fire detection method based on video surveillance according to claim 1, characterized in that, The method further includes: Based on the area change of the region to be detected in consecutive video frames of the monitoring video stream, a morphological change trend curve is determined; the morphological change trend curve is used to characterize the trend of the morphological change.

4. The fire identification method based on video surveillance according to claim 2, characterized in that, Determining the intensity of the radial gradient feature of the region to be detected includes: Determine at least one concentration center point within the area to be detected; the concentration center point is the point with the highest pixel value within the area to be detected. When the pixel values ​​at the concentration center point gradually decrease outward from the concentration center point in multiple directions, it is determined that the region to be detected has the radial gradient feature. The intensity of the radial gradient feature is determined based on the significance of the decreasing distribution.

5. The fire detection method based on video surveillance according to claim 2, characterized in that, Determining the rate of morphological change of the region to be detected in consecutive video frames of the monitored video stream includes: Determine the area sequence of the region to be detected in consecutive video frames; The rate of morphological change is determined based on the area difference between each adjacent image in the area sequence.

6. The fire detection method based on video surveillance according to claim 1, characterized in that, Adjusting the alarm threshold of the smoke detector based on the smoke authenticity value includes: The weight adjustment coefficient of the smoke alarm is determined based on the smoke authenticity value; Based on the weighting adjustment coefficient, the original alarm threshold of the smoke detector is lowered; the higher the smoke authenticity value, the lower the adjusted alarm threshold.

7. The fire detection method based on video surveillance according to claim 1, characterized in that, The method further includes: Pixel differential processing is performed on consecutive video frames in the monitoring video stream; The frame image where the first pixel change occurs is marked as the initial frame; Starting from the initial frame, the regions in each frame that have undergone pixel changes compared to the reference template are identified as the regions to be detected that have undergone pixel changes; the reference template is any frame image prior to the initial frame.

8. A fire detection system based on video surveillance, characterized in that, The system is configured to perform the method of claim 1, including: The data acquisition module is used to acquire the monitoring video stream of the monitored area and the time-series data collected by the sensors in the monitored area; the monitoring video stream is used to determine the detection area where pixel changes occur; the sensors include at least: a temperature and humidity sensor and a smoke sensor. A visual evaluation module is used to determine an evaluation value of the region to be detected based on its image features and morphological changes; the evaluation value is used to characterize the probability that the region to be detected is smoke. The confidence calculation module is used to determine a confidence factor based on the trend of morphological changes in the area to be detected and the trend of changes in the time-series data. This includes: generating a temperature change trend curve, a humidity change trend curve, and a smoke concentration change trend curve based on the time-series data collected by the sensor; determining a first difference value between the trend of morphological changes and the temperature change trend curve; determining a second difference value between the trend of morphological changes and the humidity change trend curve; determining a third difference value between the trend of morphological changes and the smoke concentration change trend curve; and determining the confidence factor based on the first difference value, the second difference value, and the third difference value. The confidence factor is used to correct the evaluation value. The consistency between the trend of morphological changes and the trend of changes in smoke sensor data is positively correlated with the confidence factor; the consistency between the trend of morphological changes and the trend of changes in temperature and humidity sensor data is negatively correlated with the confidence factor. The authenticity value determination module is used to determine the smoke authenticity value based on the evaluation value and the confidence factor; the smoke authenticity value is used to characterize the overall confidence that the area to be detected is real smoke; The threshold control module is used to adjust the alarm threshold of the smoke alarm based on the smoke authenticity value.

Images