Fire smoke and fire source accurate positioning method based on image recognition

Abstract

The invention discloses a fire smoke and fire source accurate positioning method based on image recognition, and the method comprises the following steps: S1, collecting multi-modal data, and carrying out the graying, denoising, histogram equalization and shielding correction preprocessing; s2, inputting the preprocessed data into a vortex sensing network to generate a preliminary detection result of flame and smoke; s3, dividing the scene into dynamic space blocks by using a dynamic space division algorithm and allocating priorities; s4, eliminating non-fire feature interference through an environment feature filter, and generating an optimization candidate region; s5, inputting the candidate area into an event-driven time sequence analysis module, and extracting a key frame to analyze flame and smoke diffusion and time trend; s6, verifying position consistency by adopting a flame and smoke joint detection method, and generating a three-dimensional space positioning result; and S7, generating a fire three-dimensional thermodynamic diagram according to three-dimensional positioning and triggering an alarm mechanism. According to the invention, accurate detection and positioning of fire smoke and a fire source are realized by using a vortex sensing network, an event-driven mechanism and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition and fire detection, and in particular to a method for accurately locating fire smoke and fire sources based on image recognition. Background Art

[0002] In recent years, with the development of social economy and the acceleration of urbanization, the frequent occurrence of building fires has become an important issue of global concern. The rapid spread of fire not only poses a huge threat to life and property, but may also have long-term adverse effects on the ecological environment. Therefore, how to quickly and accurately detect fires, especially to effectively identify and locate flames and smoke in the early stages of a fire, has become a research focus in the field of fire warning and prevention.

[0003] Traditional fire detection technology mainly relies on hardware devices such as smoke detectors, temperature sensors and infrared sensors. These devices show high sensitivity and reliability in many scenarios, but they still have many shortcomings in dynamic and complex environments. For example, smoke detectors usually require a high concentration of smoke to trigger the alarm, so it is difficult to respond in time at the beginning of a fire; although temperature sensors can detect the increase in ambient temperature, their detection effect may be seriously disturbed when the temperature difference in the scene is large or the flame radiation is weak. In addition, these single sensor devices cannot provide spatial distribution information of flames and smoke, and it is difficult to meet the needs of real-time monitoring and positioning.

[0004] With the rapid development of artificial intelligence and image processing technology, fire detection methods based on image recognition have gradually become a research hotspot. Compared with traditional sensor equipment, image recognition technology can more comprehensively capture fire characteristics, such as flame color, shape, brightness, and smoke diffusion pattern. Especially with the support of multispectral image fusion technology, by simultaneously collecting visible light and infrared light image data, the limitations of a single spectrum under complex lighting conditions can be effectively overcome. However, the existing image recognition technology still has the following problems in practical applications.

[0005] First, existing methods have limitations in extracting flame and smoke features. As dynamically changing physical phenomena, flames and smoke have highly irregular and unstable features, such as the jumping nature of flames and the diffusivity of smoke. Although traditional convolutional neural networks can extract edge features to a certain extent, they are difficult to capture the fine-grained changes in the dynamic characteristics of flames and smoke. This deficiency makes the detection results easily affected by environmental noise, such as reflections and shadows under complex lighting conditions. In addition, traditional feature extraction methods often process flames and smoke separately, lacking a comprehensive analysis of the spatial relationship between the two.

[0006] Secondly, in terms of regional division and optimization of fire scenes, existing technologies mostly use fixed segmentation or global analysis methods. Although this method is simple to calculate, it cannot adapt to the changes in local characteristics in complex fire scenes. For example, in small-scale fire scenes (such as homes or offices), flames and smoke are usually distributed in a localized manner, and the global analysis method will lead to waste of resource allocation and reduced detection accuracy. At the same time, for the processing of background noise, existing technologies are often limited to simple filtering operations, lacking a precise suppression mechanism for light source reflection and environmental background noise, resulting in a high false alarm rate.

[0007] In addition, existing methods also have obvious deficiencies in time series dynamic analysis. Traditional time series analysis methods are usually based on recurrent neural networks (RNN) or long short-term memory networks (LSTM). Although they can capture certain time dependencies, these methods often need to process a large amount of irrelevant information, have high computational complexity, and poor real-time performance. At the same time, existing time series analysis technology has limited ability to predict the spread trend of flames and smoke, and cannot provide accurate prediction information for the spread path of fire.

[0008] Finally, in terms of the joint detection and three-dimensional positioning of flames and smoke, the existing technology has limited research. Most methods can only provide the plane location of the fire, but it is difficult to generate three-dimensional spatial distribution information of the fire source and smoke. This deficiency is particularly obvious in high-rise buildings or multi-story scenes. The lack of three-dimensional positioning results will make it difficult to accurately determine the location of the fire source during the emergency response process. In addition, the existing technology lacks a consistency verification mechanism for flame and smoke characteristics, which is prone to false alarms or missed alarms.

[0009] Therefore, how to provide a method for accurately locating fire smoke and fire sources based on image recognition is an urgent problem that technicians in this field need to solve. Summary of the invention

[0010] One purpose of the present invention is to propose a method for accurately locating fire smoke and fire sources based on image recognition. The present invention utilizes multispectral image recognition technology, dynamic time series analysis and three-dimensional positioning methods, and realizes accurate detection and real-time three-dimensional positioning of flames and smoke in fire scenes through multimodal data fusion, vortex perception network and event-driven mechanism. Through dynamic feature extraction and spatial optimization, the present invention significantly improves the accuracy of fire detection in complex environments, and effectively suppresses the interference of light source reflection and background noise. At the same time, the present invention has the function of predicting the dynamic spread trend of fire, providing efficient and reliable technical support for fire warning and emergency treatment, and has the advantages of strong real-time performance, high accuracy and good robustness.

[0011] According to an embodiment of the present invention, a method for accurately locating fire smoke and fire source based on image recognition includes the following steps: S1. Deploy multispectral cameras to collect visible light and infrared multimodal data of fire scenes, obtain flame texture, smoke diffusion characteristics and thermal signals, and perform grayscale, denoising, histogram equalization and correction of occluded areas to generate pre-processed multimodal image data; S2, inputting the preprocessed multimodal image data into the vortex perception network, extracting the highlight features of the flame and the diffusion boundary features of the smoke by adjusting the dynamic vortex kernel, and generating preliminary detection results of the flame and smoke; S3. Based on the preliminary detection results of flames and smoke, the fire scene is divided into multiple dynamic space blocks using a dynamic space segmentation algorithm, and the dynamic space blocks are prioritized according to the layout of the fire scene to generate preliminary candidate detection areas; S4, combining the environmental feature filter to eliminate the interference of non-fire features, including light source reflection and background noise, and generate an optimized candidate detection area; S5. Input the optimized candidate detection area into the event-driven timing analysis module, trigger the event point by monitoring the significant changes of flame and smoke characteristics, extract key frames to analyze the dynamic diffusion position of flame and smoke, and predict the time trend; S6. Based on the output results of the event-driven timing analysis module, the joint detection method of flame and smoke is used to verify the position consistency of the fire source and the smoke, and generate a three-dimensional spatial positioning result; S7. Use the three-dimensional spatial positioning results to generate a three-dimensional heat map of the fire, and trigger the alarm mechanism in combination with the diffusion trend of flames and smoke, providing the location of the fire source, the range of smoke diffusion and fire dynamics information through sound and light alarms and remote notifications.

[0012] Optionally, the S2 specifically includes: S21, construct the input layer of the vortex perception network, and record the preprocessed multimodal image data as ,in represents a visible light image, Represents infrared images, multimodal image data The convolutional backbone structure that is used as input into the vortex perception network is used to capture the characteristic differences of flames and smoke in different bands; S22. Set a dynamic vortex kernel in the core convolutional layer of the vortex perception network to simulate the irregular diffusion and edge jumping of flames and smoke: ; in, represents the dynamic vortex core, represents the convolution kernel coordinates, represents the diffusion scale of the dynamic vortex core in the horizontal and vertical directions, represents the phase offset, represents the angular frequency, represents the exponential function, represents the covariance matrix; The dynamic vortex kernel is dynamically adjusted according to the local features of flame and smoke in the input multimodal image. , , and ; S23, convolution extraction of the highlight features of the flame, using a dynamic vortex kernel designed for flame thermal radiation Working with multimodal image data , and get the flame highlight feature map: ; in, Represents the flame highlight feature map, represents the scaling factor of the flame characteristics, represents the hyperbolic tangent function, Represents the convolution operation; S24. Perform convolution extraction on the diffusion boundary of smoke, using a dynamic vortex kernel targeting the low transparency diffusion characteristics of smoke. Working with multimodal image data , and obtain the smoke diffusion boundary feature map: ; in, represents the smoke diffusion boundary feature map, Scaling factor representing smoke characteristics; S25, input the flame highlight feature map and the smoke diffusion boundary feature map into the feature fusion module, and generate a comprehensive feature map by nonlinear combination: ; in, represents the comprehensive feature map, Represents the weighting coefficient of the flame highlight feature map, represents the weighting coefficient of the smoke diffusion boundary feature map, represents the differential boundary enhancement coefficient, Represents the gradient operation, which is used to enhance the difference between the flame and smoke edges; S26, based on the comprehensive feature map, perform threshold determination to generate preliminary detection results of flame and smoke, set the detection threshold, and obtain a set of detection results : ; in, Represents the detected fire and smoke pixel coordinate set, Representing coordinates The comprehensive feature map of .

[0013] Optionally, the S3 specifically includes: S31, reconstructing the two-dimensional spatial features of the preliminary detection results of the generated flame and smoke, and defining the joint feature matrix of the flame and smoke: ; in, Represents the distribution matrix of flame and smoke in the scene space, Indicates the flame highlight feature map at coordinates The pixel value at Indicates the smoke diffusion boundary feature map in coordinates The pixel value at Represents the maximum value of all pixel values ​​in the flame highlight feature map, Represents the maximum value of all pixel values ​​in the smoke diffusion boundary feature map, and represents the weight coefficient; S32, according to Based on the spatial characteristics of the scene, a dynamic space segmentation algorithm is used to divide the scene into multiple dynamic space blocks and determine the boundary of each dynamic space block: ; in, represents a dynamic space block, represents the size of the spatial block, Indicates the vertical index of the dynamic space block. Indicates the horizontal index of the dynamic space block; S33. In each dynamic space block, calculate the cumulative weight of fire characteristics: ; in, represents the cumulative weight of fire characteristics of dynamic spatial blocks; S34. Setting a priority allocation function according to the location and risk level of the space block: ; in, Indicates the priority of the dynamic space block, Represents the maximum value of the cumulative weight of fire characteristics of all dynamic space blocks, represents the fire scene layout weight of the dynamic space block, represents the maximum value of the fire layout weight, The adjustment factor representing the weight of the fire scene layout; S35. Sort the dynamic space blocks according to the calculation result of the priority allocation function to form a priority set: ; in, represents a set of dynamic space blocks that meet the priority requirements, Indicates the priority threshold; S36: Dynamic space blocks that meet the priority requirements are collected The high-priority spatial blocks in are output as preliminary candidate detection regions.

[0014] Optionally, the S4 specifically includes: S41. Construct an environmental feature filter and define the interference feature matrix using the light source reflection intensity and background noise intensity: ; in, Representing coordinates The interference characteristic value at Representing coordinates The reflected intensity of the light source at Representing coordinates The background noise intensity at Indicates the maximum value of the light source reflection intensity. represents the maximum value of the background noise intensity, Represents the weight factor of light source reflection, represents the weight factor of background noise; S42, filtering the dynamic space blocks of the preliminary candidate detection area by interference feature matrix, and calculating the optimized credible fire feature value: ; in, Representing coordinates The optimized fire eigenvalue matrix is Represents the coordinates in the preliminary candidate detection area Fire response value, represents the interference feature suppression coefficient; S43, for the optimized fire eigenvalue matrix , set the filtering threshold and generate the filtered fire feature matrix ; S44, performing dynamic spatial block accumulation on the filtered fire feature matrix, and calculating the fire feature weight of each dynamic spatial block: ; in, represents the fire feature weight of the dynamic spatial block, represents the pixel range within the dynamic space block, represents a dynamic space block, Indicates the vertical index of the dynamic space block. Indicates the horizontal index of the dynamic space block; S45. According to the optimized spatial block weight, set the minimum fire weight threshold , remove low-weight dynamic space blocks and retain high-weight dynamic space blocks as optimized candidate detection areas: ; in, represents the optimized candidate detection area; S46. Output optimized candidate detection area , as the input of the event-driven timing analysis module.

[0015] Optionally, the S5 specifically includes: S51, the optimized candidate detection area is input into an event-driven timing analysis module, and the event-driven timing analysis module includes: A change detection unit, used to calculate the change value of the flame and smoke feature matrix between consecutive frames; An event trigger unit, used to generate an event trigger signal by comparing the change value with a set threshold; A key frame extraction unit, used to extract a key frame set within a time window according to an event trigger point; A diffusion analysis unit, used to perform differential operations on a set of key frames to generate a diffusion velocity matrix; A trend prediction unit, which is used to combine the diffusion velocity matrix, calculate the dynamic diffusion position, and generate a time trend graph; S52, the change detection unit performs differentiation on the continuous frames of the flame and smoke feature matrices: ; in, Indicates Frame at coordinates The fire characteristic change value at Indicates Frame at coordinates The fire eigenvalue matrix at Indicates Frame at coordinates The fire eigenvalue matrix at ; S53, define the event triggering condition, the event triggering unit compares With setting threshold , generate a trigger signal: ; in, Indicates Frame at coordinates The event trigger signal at when When the current frame time point is recorded is the event point; S54, key frame extraction unit at event point Set a time window before and after , extract the fire feature value frames within the time window to form a sparse time series key frame set: ; in, represents a set of key frames, Indicates the time window size; S55, the diffusion analysis unit performs time difference calculation on the key frame set to generate a diffusion speed matrix of flame and smoke: ; in, Indicated in coordinates The average diffusion velocity at Indicates Frame at coordinates The fire eigenvalue matrix at ; S56. The trend prediction unit predicts the dynamic diffusion position of the flame and smoke using the diffusion velocity matrix and calculates the prediction position matrix: ; in, Indicated in coordinates The predicted diffusion value at represents the prediction time interval; S57. Combining the dynamic diffusion position matrix with the key frame difference result, a time trend diagram of flame and smoke is generated: ; in, Indicated in coordinates The flame and smoke diffusion tendency value at the location.

[0016] Optionally, the S6 specifically includes: S61. Perform joint feature analysis on the dynamic diffusion position matrix of flame and smoke, and calculate the spatial consistency score of flame and smoke features: ; in, Indicated in coordinates The flame and smoke consistency score at represents the weight factor, Indicated in coordinates The predicted diffusion value at represents the smoke diffusion intensity matrix; S62: Set a consistency score threshold, screen the joint detection area of ​​flame and smoke, and only retain the positions with consistency scores higher than the consistency score threshold to obtain the effective area feature matrix ; S63, effective area feature matrix after joint detection Perform triangulation and combine multi-view fire feature data to calculate the three-dimensional spatial coordinates of flames and smoke: ; in, represents the three-dimensional space coordinates, , , represents the coordinate index, represents the triangulation function based on multi-view features, Represents a collection of deep data matrices; S64, performing regional aggregation on the three-dimensional spatial coordinate data, aggregating adjacent flame and smoke positions into a single fire source region, and obtaining a three-dimensional spatial positioning result: ; in, represents the aggregated fire source area set, represents the coordinates of the aggregation center, Indicates the aggregation threshold.

[0017] The beneficial effects of the present invention are: First, the present invention uses a multi-spectral camera to simultaneously collect visible light and infrared light image data, integrating the color characteristics of the flame, the diffusion pattern of the smoke, and the thermal signal, overcoming the limitations of single spectral data under complex light conditions and effectively improving the adaptability of detection. In response to the noise and contrast problems of image data, the present invention designs pre-processing operations such as graying, denoising, histogram equalization, and occlusion area correction, which not only improves the quality of the image, but also provides a highly reliable data foundation for subsequent feature extraction.

[0018] Secondly, in terms of feature extraction, the present invention introduces a vortex perception network (VSNet). By dynamically adjusting the shape and size of the vortex core, the network can accurately extract the highlight features of the flame and the diffusion boundary features of the smoke, effectively capture the jumping of the flame and the diffusion pattern of the smoke, and significantly improve the recognition ability of complex dynamic features. Compared with traditional convolutional networks, the vortex perception network shows higher applicability in the fine-grained perception of feature extraction. In addition, through the nonlinear weighted combination of feature fusion modules, the present invention realizes the comprehensive analysis of flame and smoke features, and improves the integrity and consistency of the detection results.

[0019] Thirdly, in terms of regional division and optimization of fire scenes, the present invention adopts a dynamic space segmentation algorithm to divide the fire scene into multiple dynamic space blocks, and combines the scene layout to allocate priorities, thereby achieving reasonable allocation of resources and accurate detection of key areas. In addition, combined with the environmental feature filter, the present invention can effectively suppress the interference of light source reflection and background noise, significantly reducing the false alarm rate. Especially in complex indoor environments, the method's ability to suppress background interference is fully demonstrated, and the detection results are more stable and reliable.

[0020] In addition, the present invention breaks through the limitations of traditional time series analysis methods through an event-driven time series analysis module. Through the event-driven mechanism, the present invention only triggers analysis when the fire characteristics change significantly, avoiding redundant processing of irrelevant frames, and greatly improving the real-time performance and computational efficiency of time series analysis. Combined with key frame extraction and differential operation, the present invention can accurately capture the dynamic diffusion position and speed changes of flames and smoke, providing a reliable basis for predicting the spread path of fire.

[0021] Finally, in terms of the joint detection and three-dimensional positioning of flames and smoke, the present invention effectively reduces the problems of false alarms and missed alarms in fire detection through a consistency verification method. By integrating depth data with triangulation, the present invention generates a three-dimensional spatial distribution map of flames and smoke, providing accurate fire source location information and dynamic analysis capabilities of smoke diffusion range. This method is particularly suitable for multi-story buildings or complex scenes, and provides important technical support for emergency responders to quickly determine the location and spread trend of fires. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is an overall flow chart of a method for accurately locating fire smoke and fire sources based on image recognition proposed by the present invention; Figure 2 This is a flow chart of the event-driven timing analysis module of the fire smoke and fire source precise positioning method based on image recognition proposed by the present invention. DETAILED DESCRIPTION

[0023] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0024] refer to Figure 1 and Figure 2 , a method for accurately locating fire smoke and fire source based on image recognition, comprising the following steps: S1. Deploy multispectral cameras to collect visible light and infrared multimodal data of fire scenes, obtain flame texture, smoke diffusion characteristics and thermal signals, and perform grayscale, denoising, histogram equalization and correction of occluded areas to generate pre-processed multimodal image data; S2, inputting the preprocessed multimodal image data into the vortex perception network, extracting the highlight features of the flame and the diffusion boundary features of the smoke by adjusting the dynamic vortex kernel, and generating preliminary detection results of the flame and smoke; S3. Based on the preliminary detection results of flames and smoke, the fire scene is divided into multiple dynamic space blocks using a dynamic space segmentation algorithm, and the dynamic space blocks are prioritized according to the layout of the fire scene to generate preliminary candidate detection areas; S4, combining the environmental feature filter to eliminate the interference of non-fire features, including light source reflection and background noise, and generate an optimized candidate detection area; S5. Input the optimized candidate detection area into the event-driven timing analysis module, trigger the event point by monitoring the significant changes of flame and smoke characteristics, extract key frames to analyze the dynamic diffusion position of flame and smoke, and predict the time trend; S6. Based on the output results of the event-driven timing analysis module, the joint detection method of flame and smoke is used to verify the position consistency of the fire source and the smoke, and generate a three-dimensional spatial positioning result; S7. Use the three-dimensional spatial positioning results to generate a three-dimensional heat map of the fire, and trigger the alarm mechanism in combination with the diffusion trend of flames and smoke, providing the location of the fire source, the range of smoke diffusion and fire dynamics information through sound and light alarms and remote notifications.

[0025] In this implementation, S2 specifically includes: S21, construct the input layer of the vortex perception network, and record the preprocessed multimodal image data as ,in represents a visible light image, Represents infrared images, multimodal image data The convolutional backbone structure that is used as input into the vortex perception network is used to capture the characteristic differences of flames and smoke in different bands; S22. Set a dynamic vortex kernel in the core convolutional layer of the vortex perception network to simulate the irregular diffusion and edge jumping of flames and smoke: ; in, represents the dynamic vortex core, represents the convolution kernel coordinates, represents the diffusion scale of the dynamic vortex core in the horizontal and vertical directions, represents the phase offset, represents the angular frequency, represents the exponential function, represents the covariance matrix; The dynamic vortex kernel is dynamically adjusted according to the local features of flame and smoke in the input multimodal image. , , and ; S23, convolution extraction of the highlight features of the flame, using a dynamic vortex kernel designed for flame thermal radiation Working with multimodal image data , get the flame highlight feature map: ; in, Represents the flame highlight feature map, represents the scaling factor of the flame characteristics, represents the hyperbolic tangent function, Represents the convolution operation; S24. Perform convolution extraction on the diffusion boundary of smoke, using a dynamic vortex kernel targeting the low transparency diffusion characteristics of smoke. Working with multimodal image data , and obtain the smoke diffusion boundary feature map: ; in, represents the smoke diffusion boundary feature map, Scaling factor representing smoke characteristics; S25, input the flame highlight feature map and the smoke diffusion boundary feature map into the feature fusion module, and generate a comprehensive feature map by nonlinear combination: ; in, represents the comprehensive feature map, Represents the weighting coefficient of the flame highlight feature map, represents the weighting coefficient of the smoke diffusion boundary feature map, represents the differential boundary enhancement coefficient, Represents the gradient operation, which is used to enhance the difference between the flame and smoke edges; S26, based on the comprehensive feature map, perform threshold determination to generate preliminary detection results of flame and smoke, set the detection threshold, and obtain a set of detection results : ; in, Represents the detected fire and smoke pixel coordinate set, Representing coordinates The comprehensive feature map of .

[0026] In this implementation, S3 specifically includes: S31, reconstructing the two-dimensional spatial features of the preliminary detection results of the generated flame and smoke, and defining the joint feature matrix of the flame and smoke: ; in, Represents the distribution matrix of flame and smoke in the scene space, Indicates the flame highlight feature map at coordinates The pixel value at Indicates the smoke diffusion boundary feature map in coordinates The pixel value at Represents the maximum value of all pixel values ​​in the flame highlight feature map, Represents the maximum value of all pixel values ​​in the smoke diffusion boundary feature map, and represents the weight coefficient; S32, according to Based on the spatial characteristics of the scene, a dynamic space segmentation algorithm is used to divide the scene into multiple dynamic space blocks and determine the boundary of each dynamic space block: ; in, represents a dynamic space block, represents the size of the spatial block, Indicates the vertical index of the dynamic space block. Indicates the horizontal index of the dynamic space block; S33. In each dynamic space block, calculate the cumulative weight of fire characteristics: ; in, represents the cumulative weight of fire characteristics of dynamic spatial blocks; S34. Setting a priority allocation function according to the location and risk level of the space block: ; in, Indicates the priority of the dynamic space block, Represents the maximum value of the cumulative weight of fire characteristics of all dynamic space blocks, represents the fire scene layout weight of the dynamic space block, represents the maximum value of the fire layout weight, The adjustment factor representing the weight of the fire scene layout; S35. Sort the dynamic space blocks according to the calculation result of the priority allocation function to form a priority set: ; in, represents a set of dynamic space blocks that meet the priority requirements, Indicates the priority threshold; S36: Dynamic space blocks that meet the priority requirements are collected The high-priority spatial blocks in are output as preliminary candidate detection regions.

[0027] In this implementation, S4 specifically includes: S41. Construct an environmental feature filter and define the interference feature matrix using the light source reflection intensity and background noise intensity: ; in, Representing coordinates The interference characteristic value at Representing coordinates The reflected intensity of the light source at Representing coordinates The background noise intensity at Indicates the maximum value of the light source reflection intensity. represents the maximum value of the background noise intensity, Represents the weight factor of light source reflection, represents the weight factor of background noise; S42, filtering the dynamic space blocks of the preliminary candidate detection area by interference feature matrix, and calculating the optimized credible fire feature value: ; in, Representing coordinates The optimized fire eigenvalue matrix is Represents the coordinates in the preliminary candidate detection area Fire response value, represents the interference feature suppression coefficient; S43, for the optimized fire eigenvalue matrix , set the filtering threshold and generate the filtered fire feature matrix ; S44, performing dynamic spatial block accumulation on the filtered fire feature matrix, and calculating the fire feature weight of each dynamic spatial block: ; in, represents the fire feature weight of the dynamic spatial block, represents the pixel range within the dynamic space block, represents a dynamic space block, Indicates the vertical index of the dynamic space block. Indicates the horizontal index of the dynamic space block; S45. Set the minimum fire weight threshold according to the optimized spatial block weight , remove low-weight dynamic space blocks and retain high-weight dynamic space blocks as optimized candidate detection areas: ; in, represents the optimized candidate detection area; S46. Output optimized candidate detection area , as the input of the event-driven timing analysis module.

[0028] In this implementation manner, S5 specifically includes: S51, the optimized candidate detection area is input into an event-driven timing analysis module, and the event-driven timing analysis module includes: A change detection unit, used to calculate the change value of the flame and smoke feature matrix between consecutive frames; An event trigger unit, used to generate an event trigger signal by comparing the change value with a set threshold; A key frame extraction unit, used to extract a key frame set within a time window according to an event trigger point; A diffusion analysis unit, used to perform differential operations on a set of key frames to generate a diffusion velocity matrix; A trend prediction unit, which is used to combine the diffusion velocity matrix, calculate the dynamic diffusion position, and generate a time trend graph; S52, the change detection unit performs differentiation on the continuous frames of the flame and smoke feature matrices: ; in, Indicates Frame at coordinates The fire characteristic change value at Indicates Frame at coordinates The fire eigenvalue matrix at Indicates Frame at coordinates The fire eigenvalue matrix at ; S53, define the event triggering condition, the event triggering unit compares With setting threshold , generate a trigger signal: ; in, Indicates Frame at coordinates The event trigger signal at when When the current frame time point is recorded is the event point; S54, key frame extraction unit at event point Set a time window before and after , extract the fire feature value frames within the time window to form a sparse time series key frame set: ; in, represents a set of key frames, Indicates the time window size; S55, the diffusion analysis unit performs time difference calculation on the key frame set to generate a diffusion speed matrix of flame and smoke: ; in, Indicated in coordinates The average diffusion velocity at Indicates Frame at coordinates The fire eigenvalue matrix at ; S56. The trend prediction unit predicts the dynamic diffusion position of the flame and smoke using the diffusion velocity matrix and calculates the prediction position matrix: ; in, Indicated in coordinates The predicted diffusion value at represents the prediction time interval; S57. Combining the dynamic diffusion position matrix with the key frame difference result, a time trend diagram of flame and smoke is generated: ; in, Indicated in coordinates The flame and smoke diffusion trend value at the location.

[0029] In this implementation manner, S6 specifically includes: S61. Perform joint feature analysis on the dynamic diffusion position matrix of flame and smoke, and calculate the spatial consistency score of flame and smoke features: ; in, Indicated in coordinates The flame and smoke consistency score at represents the weight factor, Indicated in coordinates The predicted diffusion value at represents the smoke diffusion intensity matrix; S62: Set a consistency score threshold, screen the joint detection area of ​​flame and smoke, and only retain the positions with consistency scores higher than the consistency score threshold to obtain the effective area feature matrix ; S63, effective area feature matrix after joint detection Perform triangulation and combine multi-view fire feature data to calculate the three-dimensional spatial coordinates of flames and smoke: ; in, represents the three-dimensional space coordinates, , , represents the coordinate index, represents the triangulation function based on multi-view features, Represents a collection of deep data matrices; S64, performing regional aggregation on the three-dimensional spatial coordinate data, aggregating adjacent flame and smoke positions into a single fire source region, and obtaining a three-dimensional spatial positioning result: ; in, represents the aggregated fire source area set, represents the coordinates of the aggregation center, Indicates the aggregation threshold.

[0030] Embodiment 1: In order to verify the feasibility of the present invention in implementation, the present invention is applied to the indoor fire monitoring scene of a large shopping mall. The mall covers an area of ​​about 2,000 square meters, including multiple functional areas such as supermarkets, dining areas, clothing areas, and entertainment facilities. Due to the dense flow of people, complex environment, and changeable light conditions in this scene, traditional fire detection methods have problems such as delayed response, high false alarm rate, and difficulty in locating the source of fire in this environment. In response to these problems, the present invention significantly improves the accuracy and real-time performance of fire monitoring through the application of key technologies such as multispectral image recognition technology, vortex perception network, and event-driven timing analysis module.

[0031] In this embodiment, 12 multispectral cameras are installed, which are deployed at high points in the main channel, dining area and storage area of ​​the mall to ensure that the camera's field of view covers the entire mall. The camera can collect visible light and infrared light data at the same time, where visible light is used to capture the color and shape characteristics of the flame, and infrared light is used to identify the thermal radiation characteristics of the smoke. In the collected image data, the distinguishability of the flame and smoke features is improved through grayscale, denoising, histogram equalization and occlusion correction operations.

[0032] The collected data is preprocessed and then input into the vortex perception network. The network dynamically adjusts the shape and size of the vortex core, extracts the highlight features of the flame and the diffusion boundary features of the smoke, and generates preliminary detection results. In one experiment, the system detected abnormal heat signals and slight smoke diffusion features at a stacking point in the supermarket area. After optimization of the dynamic space segmentation algorithm, the system accurately locked the abnormal feature blocks in the area and eliminated the interference of light source reflection and background noise through the environmental feature filter.

[0033] Subsequently, the optimized detection area was input into the event-driven timing analysis module. This module triggers event analysis when the flame characteristics change significantly, extracts key frame data, and calculates the diffusion trend of flames and smoke. In this experiment, the system completed the extraction of key frames and the calculation of diffusion speed within 3 seconds after the triggering event, successfully predicted the direction and range of possible flame spread, and provided accurate decision support for emergency response.

[0034] Through the joint detection method of flame and smoke, the system verified the consistency of the position of the fire source and the smoke diffusion area, and used the triangulation method to generate a three-dimensional spatial distribution map of the flame and smoke. In this experiment, the system located the fire source with an accuracy of 97.5% and generated a clear three-dimensional thermal map showing the center position of the flame and the range of smoke diffusion, providing firefighters with real-time and effective fire scene information.

[0035] In order to verify the performance of the present invention in complex scenarios, the actual monitoring data of the system within 30 days was recorded and compared with the traditional fire detection method. The specific data is shown in Table 1.

[0036] Table 1 Comparison of fire detection effects between traditional methods and the method of the present invention

[0037] From the data analysis of Table 1 above, it can be seen that the present invention is significantly superior to traditional methods in multiple key performance indicators of fire detection. First, in terms of response time, the traditional method takes 15-25 seconds to make a detection response, while the response time of the present invention is shortened to 3-5 seconds, indicating that the present invention has higher real-time performance. This rapid response capability can provide timely warnings in the early stages of a fire, effectively reduce the risk of fire spread, and win valuable time for fire emergency response.

[0038] Secondly, in terms of false alarm rate, the false alarm rate of the traditional method is 8.7%, while the false alarm rate of the present invention is only 1.2%. This significant reduction reflects the present invention's ability to accurately suppress light source reflection and background noise in complex environments, effectively avoiding unnecessary waste of firefighting resources. At the same time, the reduction in the false alarm rate of the present invention further enhances the reliability of the fire warning system, making it more suitable for actual scenarios.

[0039] In terms of the accuracy of fire source positioning, the present invention has shown obvious advantages. The accuracy of fire source positioning of the traditional method is 85.3%, while the accuracy of the present invention is improved to 97.5%. This improvement is due to the precise extraction of flame characteristics by the vortex sensing network and the application of three-dimensional positioning methods. The present invention can not only identify the location of the flame, but also accurately generate the three-dimensional spatial coordinates of the fire source, providing firefighters with clear and intuitive fire distribution information.

[0040] In addition, the present invention has also significantly improved the accuracy of smoke diffusion range detection. The accuracy of the traditional method is 78.6%, while the accuracy of the present invention reaches 95.8%. This progress reflects the technical advantages of the present invention in smoke diffusion feature extraction and trend prediction, which can more comprehensively monitor the dynamics of smoke diffusion in fire scenes and provide a reliable basis for the judgment of fire development trends.

[0041] Finally, from the perspective of real-time scoring, the traditional method scored 6, while the present invention scored 9.5, which fully demonstrates the excellent performance of the present invention in terms of real-time performance. Through the event-driven timing analysis module, the present invention can accurately capture the dynamic diffusion information of flames and smoke, and quickly generate diffusion trend diagrams to provide real-time support for firefighting decisions.

[0042] This embodiment verifies the efficiency and reliability of the fire smoke and fire source precise positioning method of the present invention in a complex environment by applying it in a large shopping mall. Experimental results show that the present invention successfully realizes the rapid detection, precise positioning and dynamic diffusion trend prediction of flames and smoke by using core technologies such as multi-spectral image acquisition, vortex perception network and event-driven timing analysis module. Compared with traditional methods, the present invention has significant improvements in response time, false alarm rate, fire source positioning accuracy and smoke diffusion detection accuracy. Through real-time monitoring and three-dimensional positioning of fire characteristics in shopping mall scenes, the present invention effectively solves the problems of delayed response, high false alarm rate and low positioning accuracy in traditional methods, provides comprehensive and intelligent technical support for fire warning and emergency response, and fully reflects its practicality and innovation.

[0043] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for accurately locating fire smoke and fire sources based on image recognition, characterized in that: The steps include: S1. Deploy multispectral cameras to collect visible light and infrared multimodal data of fire scenes, obtain flame texture, smoke diffusion characteristics and thermal signals, and perform grayscale, denoising, histogram equalization and correction of occluded areas to generate pre-processed multimodal image data; S2, inputting the preprocessed multimodal image data into the vortex perception network, extracting the highlight features of the flame and the diffusion boundary features of the smoke by adjusting the dynamic vortex kernel, and generating preliminary detection results of the flame and smoke; S3. Based on the preliminary detection results of flames and smoke, the fire scene is divided into multiple dynamic space blocks using a dynamic space segmentation algorithm, and the dynamic space blocks are prioritized according to the layout of the fire scene to generate preliminary candidate detection areas; S4, combining the environmental feature filter to eliminate the interference of non-fire features, including light source reflection and background noise, and generate an optimized candidate detection area; S5. Input the optimized candidate detection area into the event-driven timing analysis module, trigger the event point by monitoring the significant changes of flame and smoke characteristics, extract key frames to analyze the dynamic diffusion position of flame and smoke, and predict the time trend; S6. Based on the output results of the event-driven timing analysis module, the joint detection method of flame and smoke is used to verify the position consistency of the fire source and the smoke, and generate a three-dimensional spatial positioning result; S7. Use the three-dimensional spatial positioning results to generate a three-dimensional heat map of the fire, and trigger the alarm mechanism in combination with the diffusion trend of flames and smoke, providing the location of the fire source, the range of smoke diffusion and fire dynamics information through sound and light alarms and remote notifications.

2. The method for accurately locating fire smoke and fire source based on image recognition according to claim 1, characterized in that: The S2 specifically includes: S21, construct the input layer of the vortex perception network, and record the preprocessed multimodal image data as ,in represents a visible light image, Represents infrared images, multimodal image data The convolutional backbone structure that is used as input into the vortex perception network is used to capture the characteristic differences of flames and smoke in different bands; S22. Set a dynamic vortex kernel in the core convolutional layer of the vortex perception network to simulate the irregular diffusion and edge jumping of flames and smoke: ； in, represents the dynamic vortex core, represents the convolution kernel coordinates, represents the diffusion scale of the dynamic vortex core in the horizontal and vertical directions, represents the phase offset, represents the angular frequency, represents the exponential function, represents the covariance matrix; The dynamic vortex kernel is dynamically adjusted according to the local features of flame and smoke in the input multimodal image. , , and ; S23, convolution extraction of the highlight features of the flame, using a dynamic vortex kernel designed for flame thermal radiation Working with multimodal image data , get the flame highlight feature map: ； in, Represents the flame highlight feature map, represents the scaling factor of the flame characteristics, represents the hyperbolic tangent function, Represents the convolution operation; S24. Perform convolution extraction on the diffusion boundary of smoke, using a dynamic vortex kernel targeting the low transparency diffusion characteristics of smoke. Working with multimodal image data , and obtain the smoke diffusion boundary feature map: ； in, represents the smoke diffusion boundary feature map, Scaling factor representing smoke characteristics; S25, input the flame highlight feature map and the smoke diffusion boundary feature map into the feature fusion module, and generate a comprehensive feature map by nonlinear combination: ； in, represents the comprehensive feature map, Represents the weighting coefficient of the flame highlight feature map, represents the weighting coefficient of the smoke diffusion boundary feature map, represents the differential boundary enhancement coefficient, Represents the gradient operation, which is used to enhance the difference between the flame and smoke edges; S26, based on the comprehensive feature map, perform threshold determination to generate preliminary detection results of flame and smoke, set the detection threshold, and obtain a set of detection results : ； in, Represents the detected fire and smoke pixel coordinate set, Representing coordinates The comprehensive feature map of .

3. The method for accurately locating fire smoke and fire source based on image recognition according to claim 1, characterized in that: The S3 specifically includes: S31, reconstructing the two-dimensional spatial features of the preliminary detection results of the generated flame and smoke, and defining the joint feature matrix of the flame and smoke: ； in, Represents the distribution matrix of flame and smoke in the scene space, Indicates the flame highlight feature map at coordinates The pixel value at Indicates the smoke diffusion boundary feature map in coordinates The pixel value at Represents the maximum value of all pixel values ​​in the flame highlight feature map, Represents the maximum value of all pixel values ​​in the smoke diffusion boundary feature map, and represents the weight coefficient; S32, according to Based on the spatial characteristics of the scene, a dynamic space segmentation algorithm is used to divide the scene into multiple dynamic space blocks and determine the boundary of each dynamic space block: ； in, represents a dynamic space block, represents the size of the spatial block, Indicates the vertical index of the dynamic space block. Indicates the horizontal index of the dynamic space block; S33. In each dynamic space block, calculate the cumulative weight of fire characteristics: ； in, represents the cumulative weight of fire characteristics of dynamic spatial blocks; S34. Setting a priority allocation function according to the location and risk level of the space block: ； in, Indicates the priority of the dynamic space block, Represents the maximum value of the cumulative weight of fire characteristics of all dynamic space blocks, represents the fire scene layout weight of the dynamic space block, represents the maximum value of the fire layout weight, The adjustment factor representing the weight of the fire scene layout; S35. Sort the dynamic space blocks according to the calculation result of the priority allocation function to form a priority set: ； in, represents a set of dynamic space blocks that meet the priority requirements, Indicates the priority threshold; S36: Dynamic space blocks that meet the priority requirements are collected The high-priority spatial blocks in are output as preliminary candidate detection regions.

4. The method for accurately locating fire smoke and fire source based on image recognition according to claim 1, characterized in that: The S4 specifically includes: S41. Construct an environmental feature filter and define the interference feature matrix using the light source reflection intensity and background noise intensity: ； in, Representing coordinates The interference characteristic value at Representing coordinates The reflected intensity of the light source at Representing coordinates The background noise intensity at Indicates the maximum value of the light source reflection intensity. represents the maximum value of the background noise intensity, Represents the weight factor of light source reflection, represents the weight factor of background noise; S42, filtering the dynamic space blocks of the preliminary candidate detection area by interference feature matrix, and calculating the optimized credible fire feature value: ； in, Representing coordinates The optimized fire eigenvalue matrix is Represents the coordinates in the preliminary candidate detection area Fire response value, represents the interference feature suppression coefficient; S43, for the optimized fire eigenvalue matrix , set the filtering threshold and generate the filtered fire feature matrix ; S44, performing dynamic spatial block accumulation on the filtered fire feature matrix, and calculating the fire feature weight of each dynamic spatial block: ； in, represents the fire feature weight of the dynamic spatial block, represents the pixel range within the dynamic space block, represents a dynamic space block, Indicates the vertical index of the dynamic space block. Indicates the horizontal index of the dynamic space block; S45. According to the optimized spatial block weight, set the minimum fire weight threshold , remove low-weight dynamic space blocks and retain high-weight dynamic space blocks as optimized candidate detection areas: ； in, represents the optimized candidate detection area; S46. Output optimized candidate detection area , as the input of the event-driven timing analysis module.

5. The method for accurately locating fire smoke and fire source based on image recognition according to claim 1, characterized in that: The S5 specifically includes: S51, the optimized candidate detection area is input into an event-driven timing analysis module, and the event-driven timing analysis module includes: A change detection unit, used to calculate the change value of the flame and smoke feature matrix between consecutive frames; An event trigger unit, used to generate an event trigger signal by comparing the change value with a set threshold; A key frame extraction unit, used to extract a key frame set within a time window according to an event trigger point; A diffusion analysis unit, used to perform differential operations on a set of key frames to generate a diffusion velocity matrix; A trend prediction unit, which is used to combine the diffusion velocity matrix, calculate the dynamic diffusion position, and generate a time trend graph; S52, the change detection unit performs differentiation on the continuous frames of the flame and smoke feature matrices: ； in, Indicates Frame at coordinates The fire characteristic change value at Indicates Frame at coordinates The fire eigenvalue matrix at Indicates Frame at coordinates The fire eigenvalue matrix at ; S53, define the event triggering condition, the event triggering unit compares With setting threshold , generate a trigger signal: ； in, Indicates Frame at coordinates The event trigger signal at when When the current frame time point is recorded is the event point; S54, key frame extraction unit at event point Set a time window before and after , extract the fire feature value frames within the time window to form a sparse time series key frame set: ； in, represents a set of key frames, Indicates the time window size; S55. The diffusion analysis unit performs time difference calculation on the key frame set to generate a diffusion speed matrix of flame and smoke: ； in, Indicated in coordinates The average diffusion velocity at Indicates Frame at coordinates The fire eigenvalue matrix at ; S56. The trend prediction unit predicts the dynamic diffusion position of the flame and smoke using the diffusion velocity matrix and calculates the prediction position matrix: ； in, Indicated in coordinates The predicted diffusion value at represents the prediction time interval; S57. Combining the dynamic diffusion position matrix with the key frame difference result, a time trend diagram of flame and smoke is generated: ； in, Indicated in coordinates The flame and smoke diffusion tendency value at the location.

6. The method for accurately locating fire smoke and fire source based on image recognition according to claim 1, characterized in that: The S6 specifically includes: S61. Perform joint feature analysis on the dynamic diffusion position matrix of flame and smoke, and calculate the spatial consistency score of flame and smoke features: ； in, Indicated in coordinates The flame and smoke consistency score at represents the weight factor, Indicated in coordinates The predicted diffusion value at represents the smoke diffusion intensity matrix; S62: Set a consistency score threshold, screen the joint detection area of ​​flame and smoke, and only retain the positions with consistency scores higher than the consistency score threshold to obtain the effective area feature matrix ; S63, effective area feature matrix after joint detection Perform triangulation and combine multi-view fire feature data to calculate the three-dimensional spatial coordinates of flames and smoke: ； in, represents the three-dimensional space coordinates, , , represents the coordinate index, represents the triangulation function based on multi-view features, Represents a collection of deep data matrices; S64, performing regional aggregation on the three-dimensional spatial coordinate data, aggregating adjacent flame and smoke positions into a single fire source region, and obtaining a three-dimensional spatial positioning result: ； in, represents the aggregated fire source area set, represents the coordinates of the aggregation center, Indicates the aggregation threshold.

Images