Data outlier elimination method and system for fire monitoring

Abstract

The application relates to the image recognition technical field and discloses a data abnormal value elimination method and system for fire monitoring. The method comprises the following steps: acquiring an original video stream of a fire monitoring area; extracting an image frame of a current time from the original video stream; extracting a foreground motion area of the image frame from the image frame; extracting a fire feature of the foreground motion area; identifying a fire area from the foreground motion area according to the fire feature of the foreground motion area; calculating a feature mutation trend of each pixel in the fire area; identifying an interference pixel from the fire area according to the feature mutation trend of each pixel in the fire area; and performing elimination processing on the interference pixel. The application has made significant progress in reducing the false alarm rate, improving the adaptability to complex scenes, retaining the dynamic information of key features and the like, and provides effective technical support for the reliable operation of a fire monitoring system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, and more specifically, to a data outlier elimination method and system for fire monitoring. BACKGROUND

[0002] As a high-occurrence disaster threatening public safety, early monitoring and early warning of fire has a key significance for reducing casualties and property losses. With the popularization of video monitoring technology, computer vision-based fire monitoring systems have become an important technical means in the field of intelligent fire fighting due to their strong intuitiveness, wide coverage and other advantages. Such systems usually collect the original video stream of the fire monitoring area, use image processing and pattern recognition techniques to automatically identify fire characteristics (such as the color, brightness, flicker frequency of fire, and the diffusion pattern of smoke, etc.), thereby realizing real-time monitoring and alarm of fire.

[0003] However, in practical applications, a core challenge faced by fire monitoring systems is how to accurately distinguish between real fire characteristics and pseudo-fire signals generated by various interference factors. Due to the complexity of the monitoring scene, there are often a large number of non-fire related dynamic changes in the original video stream, such as periodic changes in natural light (such as sunrise and sunset, cloud cover), noise fluctuations of monitoring equipment, random motion of small targets such as flying insects or birds, light and shadow disturbances of human activities (such as light switches, vehicle lights), and even other types of aerosols (such as dust, water mist), etc. These interference factors may appear as pixel value mutations or continuous changes in local regions in the image, and if not effectively eliminated, they may cause the system to mistakenly identify interference regions as fire regions, resulting in a large number of false alarms; on the contrary, if over-filtered, it may miss the real early fire characteristics, causing missed alarms.

[0004] Therefore, how to accurately eliminate interference pixels from video data and retain real fire characteristics is a key technical difficulty in improving the reliability of fire monitoring systems. SUMMARY

[0005] In order to solve the above technical problems, the present application is proposed to provide a data outlier elimination method and system for fire monitoring, which can accurately eliminate interference pixel data in video data and retain real fire characteristics.

[0006] In a first aspect, the present invention provides a method for removing outliers in fire monitoring data, comprising: acquiring an original video stream of a fire monitoring area; extracting an image frame of the current time from the original video stream; extracting a foreground motion region from the image frame; extracting fire features from the foreground motion region; identifying a fire region from the foreground motion region based on the fire features; calculating the feature mutation trend of each pixel in the fire region; identifying interfering pixels from the fire region based on the feature mutation trend of each pixel in the fire region; and removing the interfering pixels.

[0007] Preferably, in the aforementioned method for removing outlier data in fire monitoring, extracting the foreground motion region of the image frame includes: taking any pixel in the image frame as a target pixel, obtaining the color histogram of the target pixel, and calculating the dense optical flow field of the target pixel; extracting the color features of the target pixel from the color histogram of the target pixel; calculating the optical flow direction features of the target pixel based on the dense optical flow field of the target pixel; concatenating the color features and optical flow direction features of the target pixel to form the comprehensive features of the target pixel; generating a foreground mask based on the comprehensive features of each pixel in the image frame; and processing the image frame using the foreground mask to obtain the foreground motion region of the image frame.

[0008] Preferably, in the aforementioned method for removing outlier data in fire monitoring, generating a foreground mask based on the comprehensive features of each pixel in the image frame includes: setting a pixel property function for the target pixel. ,in For the preset color weights, This represents the dot product. The function is a natural exponential function. Indicates the position of the target pixel. The comprehensive features of the target pixel, It is the mean of the comprehensive features of all pixels in the image frame. The variance of the combined features of all pixels in the image frame is given by . The dense optical flow field of the target pixel. The mean of the dense optical flow field of all pixels in the image frame; a foreground indicator value is set for the target pixel. Set a pixel property function for the target pixel. At that time, the foreground indication value of the target pixel Select 1 otherwise select 0, where The preset segmentation threshold is used; the foreground mask is generated based on the foreground indication value of each pixel in the image frame.

[0009] Preferably, in the aforementioned method for removing outlier data in fire monitoring, extracting fire features from the foreground moving region includes: obtaining historical flame regions from historical image frames from the previous time; selecting candidate flame regions from the foreground moving region based on the historical flame regions and a preset magnification ratio; identifying the flame regions from the candidate flame regions using a preset U-Net network; and calculating the area change rate and edge jaggedness of the flame regions as fire features of the foreground moving region.

[0010] Preferably, in the aforementioned method for removing outlier data in fire monitoring, calculating the edge jaggedness of the flame region includes: setting a grid size according to a preset fire monitoring accuracy; dividing the flame region into K grids of the same size according to the grid size; and counting at least one pixel within the grid whose color gradient exceeds a preset color threshold. Number of grids ,in, The function is a conditional function; it takes the value 1 if the condition is met, and 0 otherwise. The function is used to find the maximum value. For the i-th grid in K grids, for Any pixel in, for Color gradient; calculate the fractal dimension of the edge curve of the flame region. ; Calculate the edge serration of the flame region. .

[0011] Preferably, in the aforementioned method for removing outlier data in fire monitoring, extracting the fire features of the foreground moving area further includes: extracting the color and texture features of the foreground moving area; concatenating the color and texture features of the foreground moving area into a fusion feature of the foreground moving area; inputting the fusion feature of the foreground moving area into a preset neural network, which then identifies a smoke area from the foreground moving area; and extracting smoke diffusion features from the smoke area as the fire features of the foreground moving area.

[0012] Preferably, in the aforementioned method for removing outlier data in fire monitoring, extracting smoke diffusion features from the smoke area includes: obtaining a texture histogram of the smoke area at the current time. And the historical texture histogram of the previous time period of the current time. ; Calculate the texture histogram probability distribution The historical texture histogram a probability distribution of and the texture histogram and the historical texture histogram a joint probability distribution of ; calculating a texture continuity degree of smoke features ; calculating an energy standard deviation of the texture histogram ; calculating an optical flow vector field of the smoke region at the current time ; calculating the optical flow vector field a velocity variance , a velocity standard deviation and a velocity gradient ; calculating a motion smoothness degree of the smoke region wherein, denotes a dot product; calculating a physical continuity degree of the smoke region

[0013] as the smoke diffusion features.

[0014] Preferably, in the foregoing data outlier elimination method for fire monitoring, wherein the interference pixels are identified from the fire region according to the feature mutation trend of each pixel in the fire region, comprising: generating an anomaly probability matrix according to the feature mutation trend of each pixel in the fire region; inputting the anomaly probability matrix into a trained neural network to output an environmental noise mask corresponding to the fire region; using the environmental noise mask to identify the interference pixels from the fire region.

[0015] Preferably, in the foregoing data outlier elimination method for fire monitoring, wherein before the interference pixels are eliminated, it further comprises: obtaining an abnormal confidence of the interference pixels from a preset sensor feedback; detecting whether the abnormal confidence of the interference pixels is higher than a preset confidence threshold; when the judgment result is yes, entering the step of eliminating the interference pixels.

[0016] In a second aspect, the present application provides a data outlier elimination system for fire monitoring, comprising: a video stream acquisition module, which acquires an original video stream of a fire monitoring area; an image frame extraction module, which extracts an image frame of a current time from the original video stream; a moving area extraction module, which extracts a foreground moving area of the image frame from the image frame; a fire feature extraction module, which extracts a fire feature of the foreground moving area; a fire area extraction module, which identifies a fire area from the foreground moving area according to the fire feature of the foreground moving area; a feature mutation calculation module, which calculates a feature mutation trend of each pixel in the fire area; an interference pixel identification module, which identifies an interference pixel from the fire area according to the feature mutation trend of each pixel in the fire area; and an elimination processing module, which performs elimination processing on the interference pixel.

[0017] The above one or more technical solutions of the present application have at least one or more of the following beneficial effects:

[0018] According to the technical solution of the present application, through the progressive processing flow of "foreground moving area extraction-fire feature focusing-pixel-level mutation trend analysis-interference pixel elimination", significant progress has been made in reducing false positive rate, improving adaptability to complex scenes, and retaining dynamic information of key features, thereby providing effective technical support for reliable operation of a fire monitoring system. BRIEF DESCRIPTION OF DRAWINGS

[0019] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application, when taken in conjunction with the accompanying drawings. The drawings provided in the present application are used to provide further understanding of the embodiments of the present application, and constitute a part of the specification, and are used to explain the present application together with the embodiments of the present application, and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0020] Figure 1 A flowchart of a data outlier elimination method for fire monitoring according to an embodiment of the present application;

[0021] Figure 2 A partial flowchart of a data outlier elimination method for fire monitoring according to an embodiment of the present application;

[0022] Figure 3 Another layout flowchart of a data outlier elimination method for fire monitoring according to an embodiment of the present application;

[0023] Figure 4 Still another partial flowchart of a data outlier elimination method for fire monitoring according to an embodiment of the present application;

[0024] Figure 5A block diagram of a data outlier elimination system for fire monitoring according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] Some embodiments of the present application will be described below with reference to the accompanying drawings. Those skilled in the art will understand that these embodiments are only used to explain the technical principles of the present application, and are not intended to limit the protection scope of the present application.

[0026] As shown in FIG. 1, in one embodiment of the present application, a data outlier elimination method for fire monitoring is provided, comprising: Figure 1

[0027] Step S110: acquiring an original video stream of a fire monitoring area.

[0028] Step S120: extracting an image frame of a current time from the original video stream.

[0029] Step S130: extracting a foreground motion area of the image frame from the image frame.

[0030] In the present embodiment, the foreground motion area of the image frame is extracted through step S130, effectively excluding the interference of static backgrounds (such as buildings, vegetation) and irrelevant dynamic backgrounds, and only processing dynamic targets (such as flames, smoke) that may have fire risks. On the one hand, the data volume of subsequent feature extraction and mutation trend analysis is greatly reduced, improving the real-time performance of the algorithm. On the other hand, the interference of static background noise on fire features is avoided, ensuring that the extracted fire features (such as flame color, smoke shape) are more concentrated and accurate, providing high-quality input data for subsequent identification.

[0031] Step S140: extracting fire features of the foreground motion area.

[0032] Step S150: identifying a fire area from the foreground motion area according to the fire features of the foreground motion area.

[0033] Step S160: calculating a feature mutation trend of each pixel in the fire area.

[0034] In the present embodiment, unlike traditional methods that rely on fixed thresholds and machine learning methods that require a large number of labeled samples, the present embodiment directly mines the dynamic rules of real fire features through pixel-level mutation trend analysis, without the need for artificial setting of fixed thresholds or reliance on training data of specific scenes. Experimental verification shows that this method can still maintain a high accuracy of fire area identification in weak light environments on rainy days, long-range fuzzy scenes, and multi-light source mixed scenes (such as the presence of streetlights and car lights in the monitoring area), significantly improving the adaptability of traditional methods in complex scenes.

[0035] ​Step S170, identifying the interference pixels from the fire area according to the feature mutation trend of each pixel in the fire area.

[0036] Step S180, performing the elimination processing on the interference pixels.

[0037] In the embodiment, only the identified interference pixels are eliminated, the original pixel value and the mutation trend of the real fire area are retained, the key dynamic information such as flame color and smoke diffusion is completely retained, and richer data support is provided for subsequent fire grade evaluation, fire spread prediction and other advanced analysis.

[0038] According to the technical scheme of the embodiment, through the progressive processing flow of 'foreground motion area extraction-fire feature focusing-pixel-level mutation trend analysis-interference pixel elimination', significant progress has been made in reducing the false positive rate, improving the adaptability to complex scenes, retaining key feature dynamic information, and providing effective technical support for the reliable operation of the fire monitoring system.

[0039] As shown in Figure 2 one embodiment of the present application provides a data outlier elimination method for fire monitoring. Compared with the foregoing embodiment, the data outlier elimination method for fire monitoring of the embodiment, wherein step S130 comprises:

[0040] Step S210, taking any pixel in the image frame as a target pixel, obtaining a color histogram of the target pixel, and calculating a dense optical flow field of the target pixel.

[0041] Step S220, extracting a color feature of the target pixel from the color histogram of the target pixel.

[0042] Step S230, calculating an optical flow direction feature of the target pixel according to the dense optical flow field of the target pixel.

[0043] In the embodiment, static background interference (such as a moving curtain, bird flight, etc. short-term moving target) may be misjudged as a fire area, and dynamic background interference (such as shaking of a monitoring camera, flickering of vehicle lights, etc. global motion) may damage the foreground segmentation effect. The embodiment constructs a comprehensive feature for each pixel through steps S210-S240: the color histogram captures the chroma distribution characteristics of the flame (such as orange-red tone, high brightness area), and the dense optical flow field reflects the dynamic motion mode of the pixel (such as upward spread of the flame, horizontal diffusion of the smoke). The splicing of the two forms a more comprehensive feature expression, effectively distinguishing between real fire movement (color and motion direction consistent with the burning law) and interference movement (color mutation but motion direction random).

[0044] Step S240, splicing the color feature and the optical flow direction feature of the target pixel to form a comprehensive feature of the target pixel.

[0045] Step S250, generating a foreground mask according to the comprehensive feature of each pixel in the image frame.

[0046] Step S260, processing the image frame using the foreground mask to obtain a foreground motion region of the image frame.

[0047] In step S250, the method comprises:

[0048] (1) setting a pixel property function for the target pixel , wherein, is a preset color weight, represents point multiplication, the function is a natural exponential function, represents the position of the target pixel, is the comprehensive feature of the target pixel, is the mean value of the comprehensive feature of all pixels in the image frame, is the variance of the comprehensive feature of all pixels in the image frame, is the dense optical flow field of the target pixel, is the mean value of the dense optical flow field of all pixels in the image frame.

[0049] In the embodiment, the pixel property function innovatively combines the statistical characteristics of color distribution and motion mode: through the Gaussian kernel function , the color deviation from the mean value is punished to suppress background noise (such as local brightness fluctuation caused by light change); through the point multiplication term , the consistency of the pixel motion direction and the global motion trend is measured to exclude reverse motion interference (such as reverse displacement caused by camera shaking). The function dynamically balances the weights of color and motion features through the parameter , compared with the segmentation method with fixed threshold, it can adapt to the feature distribution difference in different scenes.

[0050] (2) setting a foreground indication value for the target pixel , when the pixel property function of the target pixel is set, the foreground indication value of the target pixel takes 1, otherwise it takes 0, wherein, is a preset segmentation threshold.

[0051] In the embodiment, through the binary processing of the foreground indication value , combined with the setting of the adaptive threshold T, it is beneficial to filter various interferences such as transient interference and global interference.

[0052] (3) generating a foreground mask according to the foreground indication value of each pixel in the image frame.

[0053] According to the technical solution of this embodiment, through feature fusion, adaptive statistical modeling and efficient computational design, more accurate foreground segmentation and interference suppression are achieved in fire monitoring scenarios, laying a solid foundation for subsequent outlier removal and fire determination.

[0054] like Figure 3 As shown, one embodiment of the present invention provides a method for removing outlier data in fire monitoring. Compared with the previous embodiments, the method for removing outlier data in fire monitoring in this embodiment includes step S140:

[0055] Step S310: Obtain the historical flame region from the previous historical image frame at the current time.

[0056] Step S320: Select candidate flame regions in the foreground motion region based on the historical flame regions and the preset magnification ratio.

[0057] Step S330: Identify the flame region from the candidate flame region using a preset U-Net network.

[0058] In this embodiment, through steps S310-S330, the historical flame area of ​​the previous frame is used as prior information, and the candidate region range is dynamically adjusted by combining a preset magnification ratio, which effectively reduces the search space and significantly reduces the subsequent computational load. The U-Net network is used to perform semantic segmentation on the candidate region, and its encoder-decoder structure can accurately capture the texture and morphological features of the flame (such as irregular edges and dynamic spread patterns).

[0059] Step S340: Calculate the area change rate and edge jaggedness of the flame region as fire characteristics of the foreground moving region.

[0060] The step S340, which calculates the edge jaggedness of the flame region, includes:

[0061] (1) Set the grid size according to the preset fire monitoring accuracy.

[0062] (2) Divide the flame area into K grids of the same size according to the grid size.

[0063] (3) The color gradient of at least one pixel within the grid exceeds a preset color threshold. Number of grids ,in, The function is a conditional function; it takes the value 1 if the condition is met, and 0 otherwise. The function is used to find the maximum value. For the i-th grid in K grids, for Any pixel in, for color gradient,

[0064] (4) Calculate the fractal dimension of the edge curve of the flame region. ,

[0065] (5) Calculate the edge serration of the flame area. .

[0066] In this embodiment, the dynamic geometric features of flames are incorporated into the outlier determination system for the first time through the area change rate and edge jaggedness indices proposed in step S340. The area change rate reflects the spread / contraction trend of flames. The flame area of ​​a real fire usually shows a continuous growth or fluctuating upward trend, while interference from flying insects or flickering lights can cause abrupt changes in area. This embodiment innovates the calculation of edge jaggedness by quantifying the complexity of the flame edge through grid division and color gradient statistics. The edge of a real flame exhibits fractal characteristics due to combustion instability, while the edges of artificial light sources or reflective objects are usually regular geometric shapes. Therefore, the fractal dimension can be transformed into a quantifiable feature. In scenarios with strong illumination (such as direct headlights) or low visibility (such as rainy or foggy weather), fractal dimension analysis can effectively suppress edge noise caused by light reflection, significantly improving the accuracy of flame recognition.

[0067] According to the technical solution of this embodiment, the extracted area change rate and edge jaggedness features provide high-confidence input for subsequent pixel-level mutation trend analysis: background pixels are eliminated by filtering out flame regions, allowing mutation trend analysis to focus on the real flame region; fractal dimension is used to assist in judging the dynamic stability of the flame edge, providing new dimensional features for fire level assessment. In summary, this embodiment, through a spatiotemporally correlated flame recognition mechanism, fractal dimension feature quantification, and multi-feature fusion decision-making, achieves more accurate fire feature extraction and interference suppression in complex scenarios, providing key technical support for improving the reliability of fire monitoring systems.

[0068] like Figure 4 As shown, one embodiment of the present invention provides a method for removing outlier data in fire monitoring. Compared with the previous embodiments, the method for removing outlier data in fire monitoring in this embodiment further includes step S140 as follows:

[0069] Step S410: Extract the color and texture features of the foreground motion region.

[0070] Step S420: The color features and texture features of the foreground motion region are spliced ​​together to form the fused features of the foreground motion region.

[0071] Step S430: Input the fused features of the foreground motion region into a preset neural network, and the neural network identifies the smoke region from the foreground motion region.

[0072] In this embodiment, color features can reflect the color distribution of captured smoke (such as grayish-white tones and low saturation characteristics); texture features can reflect the diffusion pattern of smoke (such as irregular edges and blurred gradients). The combination of the two forms a more comprehensive feature representation, which, combined with the semantic segmentation capability of neural networks, significantly improves the accuracy of smoke region recognition compared to traditional methods.

[0073] Step S440: Extract smoke diffusion features from the smoke area as fire features of the foreground moving area.

[0074] Step S440 includes:

[0075] (1) Obtain the texture histogram of the smoke region at the current time. And the historical texture histogram of the previous time. .

[0076] (2) Calculate the texture histogram probability distribution Historical texture histogram probability distribution and texture histogram Histogram of historical textures joint probability distribution .

[0077] (3) Calculate the texture continuity of the smoke features.

[0078] ,

[0079] In this embodiment, the proposed texture continuity index Calculating the mutual information between the texture distribution of the current frame and historical frames helps to quantify the spatiotemporal continuity of smoke diffusion: due to the continuous combustion process, the texture distribution of real smoke has a time correlation; interference noise, such as flying insects, can easily cause abrupt changes in texture in a single frame. This index can effectively filter out instantaneous interference while preserving the real smoke area to the greatest extent.

[0080] (4) Calculate the energy standard deviation of the texture histogram. .

[0081] (5) Calculate the optical flow vector field of the smoke region at the current time. .

[0082] (6) Calculate the optical flow vector field velocity variance Speed ​​standard deviation and velocity gradient .

[0083] (7) Calculate the motion smoothness of the smoke region wherein, denotes the dot product.

[0084] (8) Calculate the physical continuity of the smoke region as a smoke diffusion feature.

[0085] In this embodiment, the physical continuity index combines the physical law of smoke diffusion with the motion feature for the first time: the low-speed diffusion feature of real smoke due to airflow disturbance is embodied by , the smooth gradient of the smoke edge due to the change in density makes small, and the dispersion degree of the smoke texture energy distribution can be embodied by . The index formed by the combination of the three can effectively distinguish smoke from artificial light sources (such as flashing car lights).

[0086] According to the technical solution of the embodiment, by introducing the physical continuity feature analysis of the smoke region and the multi-modal feature fusion mechanism, the accuracy and anti-interference ability of smoke recognition in fire monitoring are significantly improved, high-precision smoke recognition and interference suppression are realized in complex scenes, and a key technical breakthrough is provided for the reliability improvement of the fire monitoring system.

[0087] In one embodiment of the present application, a data outlier elimination method for fire monitoring is provided. Compared with the foregoing embodiment, the data outlier elimination method for fire monitoring of the present embodiment, step S170 comprises:

[0088] (1) Generating an anomaly probability matrix according to the feature mutation trend of each pixel in the fire region.

[0089] (2) Inputting the anomaly probability matrix into a trained neural network to output an environmental noise mask corresponding to the fire region.

[0090] (3) Using the environmental noise mask to identify interference pixels from the fire region.

[0091] In this embodiment, the pixel-level anomaly probability distribution is generated by the feature mutation trend (such as brightness gradient, color diffusion speed), which reflects the anomaly confidence of each pixel. The anomaly probability of the real fire region usually presents spatial continuity (such as the gradual distribution of the flame edge), while the interference noise (such as flying insects, light flickering) shows local discrete high-probability points. A lightweight convolutional neural network is used for semantic segmentation of the anomaly probability matrix, and its spatial attention mechanism can effectively suppress isolated noise points.

[0092] Before step S180, the method further comprises:

[0093] (1) Obtain the abnormal confidence level of the interference pixel fed back by the preset sensor.

[0094] (2) Detect whether the abnormal confidence level of the interfering pixel is higher than the preset confidence threshold.

[0095] (3) If the judgment result is yes, proceed to step S180.

[0096] According to the technical solution of this embodiment, the physical anomaly confidence level of interfering pixels is calculated by combining multi-source data such as temperature, smoke concentration, and gas sensor data. For example, insect interference may only trigger visual sensor anomalies, while a real fire will be accompanied by a simultaneous increase in temperature and smoke concentration. The preset confidence threshold is automatically adjusted according to the environmental background noise level (such as light scattering caused by rain or fog). For example, in strong interference scenarios, the threshold can be dynamically lowered to ensure that real fire signals are not missed; in stable environments, the threshold can be increased to reduce false alarms.

[0097] like Figure 5 As shown, one embodiment of the present invention provides a data outlier removal system for fire monitoring, comprising:

[0098] The video stream acquisition module 510 acquires the original video stream of the fire monitoring area.

[0099] The image frame extraction module 520 extracts the image frame at the current time from the original video stream.

[0100] The motion region extraction module 530 extracts the foreground motion region of the image frame from the image frame.

[0101] In this embodiment, by extracting the foreground motion region of the image frame in step S130, interference from static backgrounds (such as buildings and vegetation) and irrelevant dynamic backgrounds is effectively eliminated. Only dynamic targets (such as flames and smoke) that may pose a fire risk are processed. On the one hand, this significantly reduces the amount of data for subsequent feature extraction and mutation trend analysis, improving the real-time performance of the algorithm. On the other hand, it avoids interference from static background noise on fire features, ensuring that the extracted fire features (such as flame color and smoke pattern) are more concentrated and accurate, providing high-quality input data for subsequent identification.

[0102] Fire feature extraction module 540 extracts fire features from the foreground moving area.

[0103] The fire area extraction module 550 identifies the fire area from the foreground moving area based on the fire characteristics of the foreground moving area.

[0104] The feature mutation calculation module 560 calculates the feature mutation trend of each pixel in the fire area.

[0105] In this embodiment, unlike the traditional method relying on fixed threshold and the machine learning method requiring a large number of labeled samples, the embodiment directly mines the dynamic law of the real fire characteristics through pixel-level mutation trend analysis, without the need for manual setting of fixed threshold and without the need for training data of specific scenes. Experiments verify that the method can still maintain a high fire area recognition accuracy in weak light environment on rainy days, long-range fuzzy scenes, and multi-light source mixed scenes (such as the presence of street lamps and car lights in the monitored area), and the adaptability of the method in complex scenes is significantly improved compared with the traditional method.

[0106] The interference pixel recognition module 570 identifies interference pixels from the fire area according to the feature mutation trend of each pixel in the fire area.

[0107] The elimination processing module 580 performs elimination processing on the interference pixels.

[0108] In this embodiment, only the identified interference pixels are eliminated, the original pixel values and their mutation trends of the real fire area are preserved, the key dynamic information such as flame color and smoke diffusion is completely preserved, and richer data support is provided for subsequent advanced analysis such as fire rating assessment and fire spread prediction.

[0109] According to the technical solution of this embodiment, through the progressive processing flow of “foreground motion area extraction-fire feature focusing-pixel-level mutation trend analysis-interference pixel elimination”, significant progress has been made in reducing false positive rate, improving adaptability in complex scenes, and preserving key feature dynamic information, thereby providing effective technical support for reliable operation of the fire monitoring system.

[0110] The above describes the basic principles of the application in conjunction with specific embodiments, but it should be noted that the advantages, advantages, effects and the like mentioned in the application are only examples and are not limiting, and these advantages, advantages, effects and the like cannot be considered as the application must have. In addition, the above specific details are only for the purpose of example and for the purpose of understanding, and are not limited to the application which must use the above specific details to realize.

[0111] The block diagrams of the devices, apparatuses, equipment, systems involved in the application are only illustrative examples and are not intended to require or imply that the connection, arrangement, configuration shown in the block diagram must be connected, arranged, configured. As those skilled in the art will recognize, these devices, apparatuses, equipment, systems can be connected, arranged, configured in any way. Words such as "include", "contain", "have" and the like are open-ended words, which mean "include but not limited to", and can be used interchangeably. The words "or" and "and" used herein mean the word "and / or", and can be used interchangeably unless the context clearly indicates otherwise. The word "such as" used herein means the phrase "such as but not limited to", and can be used interchangeably.

[0112] It is also important to note that the devices, apparatuses and methods described in the present application can be embodied in a variety of other forms, modi fications and alt ernatives, some of which have been discussed above and some of which are gathe red as wi ll be apparent to those reasonably skilled in the art. The described aspects and embodiments are to be considered in a descriptive sense only and not for purposes of limitation. Therefore, the scope of the present application is not to be determined strictly by the description in the specification but by the appended claims, and all changes that come within the meaning and range of equivalency of the claims are intended to be embraced therein.

[0113] The above description of disclosed aspects is intended to be illustrative, and not restrictive. Other aspects, including modifications and alternative forms, will be apparent upon reading the disclosure. The scope of the disclosure is not limited to the aspects described herein, but is defined by the appended claims.

[0114] The above description has been presented for the purpose of illustration and description. It is not intended to be exhaustive or to limit the embodiments of the present application to the precise forms disclosed. Although various example aspects and embodiments have been discussed above, those of ordinary skill in the art will appreciate a variety of modifications, alternatives, permutations, additions, and subcombinations.

Claims

1. Methods for removing outliers in fire monitoring data, including: Acquire the raw video stream of the fire monitoring area; Extract the image frame at the current time from the original video stream; Extract the foreground motion region of the image frame from the image frame; Extract the fire features of the foreground motion region; Based on the fire characteristics of the foreground moving area, identify the fire area from the foreground moving area; Calculate the feature mutation trend of each pixel in the fire area; Based on the characteristic mutation trend of each pixel in the fire area, interfering pixels are identified from the fire area; The interfering pixels are then removed. The extraction of fire features from the foreground motion region includes: Obtain the historical flame region from the previous historical image frame at the current time; Based on the historical flame region and the preset magnification ratio, candidate flame regions are selected in the foreground motion region; The flame region is identified from the candidate flame region using a preset U-Net network; The area change rate and edge jaggedness of the flame region are calculated as fire characteristics of the foreground moving area. The calculation of the edge jaggedness of the flame region includes: Set the grid size according to the preset fire monitoring accuracy; According to the stated grid size, the flame region is divided into K grids of the same size; The color gradient of at least one pixel within the statistical grid exceeds a preset color threshold. Number of grids ,in, The function is a conditional function; it takes the value 1 if the condition is met, and 0 otherwise. The function is used to find the maximum value. For the i-th grid in K grids, for Any pixel in, for Color gradient; Calculate the fractal dimension of the edge curve of the flame region. ; Calculate the edge serration of the flame region. , The process of identifying interfering pixels in the fire area based on the characteristic mutation trend of each pixel in the fire area includes: An anomaly probability matrix is ​​generated based on the characteristic mutation trend of each pixel in the fire area; The anomaly probability matrix is ​​input into a trained neural network, which outputs an environmental noise mask corresponding to the fire area. Using the ambient noise mask, interfering pixels are identified from the fire area.

2. The method for removing outlier data in fire monitoring according to claim 1, wherein, Extracting the foreground motion region of the image frame from the image frame includes: Taking any pixel in the image frame as the target pixel, obtain the color histogram of the target pixel, and calculate the dense optical flow field of the target pixel; Extract the color features of the target pixel from its color histogram; Based on the dense optical flow field of the target pixel, calculate the optical flow direction characteristics of the target pixel; The color features and optical flow direction features of the target pixel are combined to form the comprehensive features of the target pixel; A foreground mask is generated based on the combined features of each pixel in the image frame; The image frame is processed using the foreground mask to obtain the foreground motion region of the image frame.

3. The method for removing outlier data in fire monitoring according to claim 2, wherein, Generate a foreground mask based on the combined features of each pixel in the image frame, including: Set the pixel property function for the target pixel ,in, For the preset color weights, This represents the dot product. The function is a natural exponential function. Indicates the position of the target pixel. The comprehensive features of the target pixel, It is the mean of the comprehensive features of all pixels in the image frame. The variance of the combined features of all pixels in the image frame is given by . The dense optical flow field of the target pixel. It is the mean of the dense optical flow field of all pixels in the image frame; Set a foreground indicator value for the target pixel. Set a pixel property function for the target pixel. At that time, the foreground indication value of the target pixel Select 1 otherwise select 0, where, The preset segmentation threshold; The foreground mask is generated based on the foreground indication value of each pixel in the image frame.

4. The method for removing outlier data in fire monitoring according to claim 1, wherein, Extracting the fire features of the foreground motion region also includes: Extract the color and texture features of the foreground motion region; The color and texture features of the foreground motion region are combined to form the fusion feature of the foreground motion region; The fused features of the foreground motion region are input into a preset neural network, which then identifies the smoke region from the foreground motion region. Smoke diffusion features are extracted from the smoke area and used as fire features in the foreground moving area.

5. The method for removing outlier data in fire monitoring according to claim 4, wherein, Extracting smoke diffusion features from the smoke region includes: Obtain the texture histogram of the smoke region at the current time. And the historical texture histogram of the previous time period of the current time. ; Calculate the texture histogram probability distribution The historical texture histogram probability distribution and the texture histogram With the historical texture histogram joint probability distribution ; Calculate the texture continuity of smoke features ; Calculate the energy standard deviation of the texture histogram. ; Calculate the optical flow vector field of the smoke region at the current time. ; Calculate the optical flow vector field velocity variance Speed ​​standard deviation and velocity gradient ; Calculate the motion smoothness of the smoke region ,in, Dot product; Calculate the physical continuity of the smoke area This serves as a characteristic of the smoke diffusion.

6. The method for removing outlier data in fire monitoring according to claim 1, wherein, Before removing the interfering pixels, the method for removing outlier data in fire monitoring further includes: Obtain the abnormal confidence level of the interfering pixel fed back by the preset sensor; Detect whether the abnormal confidence level of the interfering pixel is higher than a preset confidence threshold; If the determination result is yes, proceed to the step of removing the interfering pixels.

7. For a fire monitoring data outlier removal system, the method for removing data outliers for fire monitoring according to any one of claims 1-6 is executed, comprising: The video stream acquisition module acquires the raw video stream of the fire monitoring area; The image frame extraction module extracts the image frame at the current time from the original video stream; The motion region extraction module extracts the foreground motion region of the image frame from the image frame; The fire feature extraction module extracts the fire features of the foreground moving area; The fire area extraction module identifies the fire area from the foreground moving area based on the fire characteristics of the foreground moving area; The feature mutation calculation module calculates the feature mutation trend of each pixel in the fire area; An interference pixel identification module identifies interference pixels from the fire area based on the characteristic mutation trend of each pixel in the fire area. The rejection processing module performs rejection processing on the interfering pixels.

Images