Fire object information tracing method and system based on flame portraying technology

Abstract

The invention relates to a fire object information tracing method and system based on a flame portrait technology, and belongs to the technical field of information tracing. Comprising the following steps: collecting and sorting fire video data, laboratory simulation data and network platform public data in an existing fire event, and constructing a fire video database; on the basis of a deep learning technology, a fire-related video analysis large model system based on a flame portrait technology is developed, and the matching problem of fire flames, smoke characteristics and fire objects in fire-related case event fire videos is solved; establishing a feature matching and fire object information traceability model, and quickly judging whether a fire is triggered by a combustion improver or not; and carrying out model optimization and updating. According to the invention, a plurality of links such as data acquisition, fire flame detection, feature extraction, matching traceability, model optimization, system application and actual combat inspection are covered, the accuracy and actual combat value of the technology are ensured, and intelligent research and judgment of fire-related cases are powerfully promoted.

Description

Technical Field

[0001] The present invention relates to a method and system for tracing information of a fire object based on flame portrait technology, belonging to the technical field of information tracing. Background Art

[0002] In the investigation of fire-related incidents, quickly and accurately characterizing the fire is the primary task. The flame and smoke characteristics of the fire video of "fire accidents" caused by general combustible materials and "arson cases" carried out using accelerants are obviously different. Some typical fire sources contain characteristic elements in their chemical composition, which causes the smoke to have a special color. The fire characteristics of lithium batteries and electrical fault fires are even more typical. These are all derived from the close correlation between the chemical composition of the fire source and its flame and smoke characteristics.

[0003] However, the application of existing video analysis in fire investigation still faces technical bottlenecks: (1) Traditional fire video analysis is mainly used for the spatial positioning of fire time, fire location or fire point, lacks systematic quantification and scientific analysis of fire flame and smoke characteristics, and is difficult to accurately judge the changes in flame and smoke characteristics. The analysis of video flame and smoke characteristics mainly relies on the experience of analysts, and the analysis results can only be simply described qualitatively. Therefore, the technology for associating and tracing the information of fire objects / fire sources is lacking. Accurately identifying fire objects / fire sources or discovering clues to fire object / fire source information is the key to fire investigation, especially the identification of arson cases related to ignition of accelerants.

[0004] Among existing patent technologies, CN114998822A discloses a fire detection method, device, equipment, and storage medium, comprising obtaining a target frame image captured by a target camera; performing fire recognition on the target frame image through a preset fire recognition model to obtain a fire recognition result; and sending the fire recognition result to a client for display. This invention performs fire recognition on the image captured by the camera through a preset fire recognition model. CN119445802A relates to the technical field of fire source tracing and positioning, and particularly to a method and device for accurately locating a fire source based on deep learning, wherein the method comprises: obtaining a sensor data set and a fire source information data set of a fire area, and generating a feature vector based on the sensor data set to obtain a feature matrix, obtaining fire source information corresponding to each feature matrix in the fire source information data set based on the distribution of the sensor data to generate a training set, a validation set, and a test set, and using a preset deep learning strategy based on the training set, validation set, and test set to obtain fire source information of the predicted fire area, and obtaining fire source positioning information based on the predicted fire source information. Although the above-mentioned patented technology has solved certain technical problems to a certain extent, it still lacks systematic quantification and precise analysis of the morphological characteristics of fire flames, making it difficult to scientifically judge changes in flame and smoke characteristics, and difficult to effectively trace the type of fire source, especially the identification of arson cases related to ignition of accelerants. Summary of the Invention

[0005] To address the aforementioned shortcomings of the existing technologies, the present invention proposes a method and system for tracing the source of fire-related information based on flame profiling technology. This method overcomes the limitations of traditional video analysis in fire investigations and promotes the transformation of fire video analysis technology from empirical judgment to a data-driven, intelligent analysis model. This method addresses the issues of systematic quantification and precise analysis, as well as the effective tracing of fire-related sources.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A method for tracing the source of fire-causing objects based on flame profiling technology comprises the following steps:

[0008] S 1. Collect and organize fire video data from existing fire-related incidents, laboratory simulation data, and public data from online platforms to build a fire video database;

[0009] S2. Develop a large-scale fire video analysis model system based on deep learning technology and flame profiling technology to solve the problem of matching the characteristics of flames and smoke with the source of fire in fire incident videos. Establish a feature matching and source tracing model for source information to quickly determine whether the fire was caused by an accelerant.

[0010] S3. Model optimization and updating: using incremental learning technology to dynamically adjust parameters, improve recognition accuracy, and achieve adaptive optimization, so that the model can continuously adapt to different environments and fire types, and improve the ability to trace information on fire sources or fire sources in complex scenarios.

[0011] Furthermore, the step S1 includes data collection and preprocessing, specifically including the following steps:

[0012] S 11. Build a multi-source fire video database, the data sources include:

[0013] Fire-related incident video: Collect actual fire video data from public security and fire-related incidents;

[0014] Laboratory simulated fire data: Fire video data obtained from laboratory controlled combustion experiments to obtain accurate flame and smoke characteristic data. This involves burning different fire materials in a controlled environment and obtaining fire videos for flame and smoke characteristic analysis. This includes fire accelerants, paper, wood, plastic, gasoline, and electrical equipment.

[0015] Public online videos: Crawl fire videos on public online platforms;

[0016] The video dataset in the multi-source fire video database is divided into training and test sets in proportion, which are used for model construction and optimization respectively to improve its generalization ability;

[0017] S 12. Data cleaning and denoising

[0018] After data collection is completed, video processing software combined with manual review is used for data preprocessing: cleaning, removing fuzzy, incomplete or interfering fragments to ensure data quality; establishing a video classification method, labeling the video according to the fire location, fire object, and fire source elements, and building a database for subsequent model training and analysis. FFmpeg and Adobe Premiere tools are used to remove low-quality, fuzzy, repeated, and irrelevant fragments to ensure data accuracy; Gaussian filtering and bilateral filtering image denoising methods are used to reduce video noise interference and improve the accuracy of flame and smoke detection.

[0019] Furthermore, the combustion aid in S11 is gasoline, diesel, alcohol or paint thinner; the fire videos on the public network platform include news reports and videos uploaded by users of self-media platforms such as Tik Tok and Bilibili.

[0020] Furthermore, S2 includes image positioning and feature extraction, namely, studying the accurate recognition technology of flames and smoke under unfavorable shooting conditions such as "overexposure", "reflection", and "low illumination", and deeply learning the unique physical and dynamic characteristics of flames and smoke, vectorizing the characteristics of fire flames and smoke, and intelligently building the association between feature extraction and the flame and smoke characteristics of fire videos;

[0021] The specific steps include:

[0022] S21. Flame area detection

[0023] A deep learning target detection algorithm is used to automatically detect flame areas. Edge detection, color segmentation, and optical flow analysis methods are used to assist and improve the model's adaptability to different environments.

[0024] S22. Flame and smoke feature extraction

[0025] Extract multi-dimensional features of flames and smoke, including:

[0026] Color features: Use HSV color space conversion and flame color histogram analysis to extract flame color distribution information;

[0027] Morphological features: Use HOG oriented gradient histogram and morphological processing methods to extract the contour features of the flame;

[0028] Dynamic features: Through optical flow, frame difference, and LSTM time series analysis, the flame combustion pattern and motion trajectory are extracted, mainly including flame spread rate, area change rate, volume change rate, flicker frequency, illumination change, number of sharp corners, and their probability density;

[0029] Temperature characteristics: Use thermal imaging analysis to extract the temperature distribution characteristics of the flame and combine it with the blackbody radiation model for correction; the temperature characteristics are based on infrared video;

[0030] Smoke characteristics: Gray-level co-occurrence matrix (GLCM) is used to analyze the texture characteristics of smoke and reflect the distribution of smoke particles. Fourier transform is used to analyze the frequency domain characteristics of smoke and capture periodic changes.

[0031] S23. Feature Vectorization

[0032] The above features of S22 are converted into vector form and input into the deep learning network for further analysis, and the flame and smoke features are mapped to high-dimensional space to improve the classification accuracy.

[0033] Furthermore, in S21, the deep learning target detection algorithm includes the YOLO algorithm and the FasterR-CNN algorithm; the different environments include illumination changes, smoke interference, and reflections; in S23, the deep learning network includes the ResNet and EfficientNet networks.

[0034] Furthermore, in S2, after image positioning and feature extraction, it also includes flame and smoke feature matching and information tracing of the source of the fire. That is, by building a flame and smoke feature matching model, using intelligent comparison technology, optimizing feature vector calculation, and improving the matching accuracy; using flame and smoke features to reversely infer the source of the fire, an efficient video case query system is established to support rapid comparison and analysis of fire-related cases;

[0035] The specific steps include:

[0036] S24. Constructing a fire flame and smoke feature matching model

[0037] Using the fire video database built in S1, a fire flame and smoke feature library is established, covering the flame and smoke features of different fire materials, combustion accelerants, and combustion modes; a flame and smoke feature matching model is constructed, and the intelligent comparison technology K-NN is used to optimize the feature vector calculation to improve the matching accuracy; feature indexing technology is used to improve the retrieval speed of flame and smoke feature matching;

[0038] S25. Fire flame and smoke feature matching method

[0039] The similarity of flame and smoke features is calculated using Euclidean distance, cosine similarity, and Mahalanobis distance methods. Hash index + KD-Tree is combined to accelerate large-scale data matching and improve the efficiency of flame tracing. The SiameseNetwork is introduced to optimize the matching capabilities of flame and smoke features through metric learning.

[0040] S26. Analysis and determination of fire species

[0041] Through the classification network, the fire object classification model is trained to realize the intelligent identification of fire objects; combined with laboratory data, the flame and smoke characteristics of different combustion objects are analyzed in a fine-grained manner to improve the accuracy of the fire object information traceability.

[0042] Furthermore, in S26, the clustering logic judgment sequence is as follows: first, determine whether an accelerant was involved in the combustion. If so, investigate the case as suspected arson, and further determine the type of fire based on video analysis. If an accelerant is not the cause of the fire, further determine whether the ignition source is a hazardous chemical. If not, determine that the fire was caused by ordinary combustible materials. Based on the smoke and flame characteristics, further trace the ignition source / fire source information.

[0043] Furthermore, in S24, the feature indexing technology includes FAISS technology; in S26, the classification network is ResNet or VGG.

[0044] Furthermore, in S3, the model optimization and update is the optimization and update of the fire flame and smoke recognition model, which is specifically as follows:

[0045] S31. Model Optimization

[0046] Experiments were conducted using test set video data to evaluate the model's recognition accuracy and generalization capabilities. Incremental learning techniques were used to dynamically adjust model parameters, improve recognition accuracy, and achieve adaptive optimization. By continuously introducing new fire types and environmental data, the model was able to adapt to different environments and fire types, enhancing its traceability capabilities in complex scenarios.

[0047] S32. Model update

[0048] Establish a dynamic model update mechanism, regularly introduce new video data and feature information, and continuously optimize model performance; iteratively optimize the model based on feedback from actual applications to ensure the accuracy and practicality of the model.

[0049] A fire source information tracing system based on flame profiling technology adopts the above-mentioned fire source information tracing method. The fire source information tracing system integrates flame and smoke detection, feature extraction, and fire source information tracing functions. It uses PyTorch+Flask / Django to build the backend and combines it with a web front-end visualization interface to implement interactive analysis functions. Specifically, it includes the following modules:

[0050] Fire video database module: It is used to collect and organize fire video data from existing fire-related incidents, laboratory simulation data, and public data from online platforms to build a fire video database;

[0051] Fire-related video analysis module: Based on deep learning technology, it is used to develop a large-scale fire video analysis model system based on flame profiling technology to solve the problem of matching the characteristics of fire and smoke with the burning objects in fire-related incident videos;

[0052] Information tracing module: This module is used to establish a feature matching and fire source information tracing model to quickly determine whether the fire was caused by an accelerant;

[0053] Model optimization and update module: This module uses incremental learning technology to dynamically adjust parameters, improve recognition accuracy, and achieve adaptive optimization, enabling the model to continuously adapt to different environments and fire types, thereby improving the ability to trace fire sources or fire sources in complex scenarios.

[0054] Practical application test module: It is used to combine fire case data to test the actual application effect of the fire flame recognition system, evaluate the accuracy, recall rate, processing speed, and optimize system performance.

[0055] After adopting the above technical solution, the present invention has at least one of the following beneficial effects compared with the prior art:

[0056] The present invention draws on the underlying logic of "face / gait recognition technology" in image recognition, feature extraction and intelligent comparison, and expands the application of this technology to the field of fire flame and smoke feature recognition. On the basis of building a fire video database, the static and dynamic characteristics of fire flames and smoke are associated with the information of fire objects / fire sources based on deep learning technology, and an intelligent fire video analysis system based on multimodal intelligent comparison is developed to infer "what kind of fire" through fire videos.

[0057] This invention utilizes deep learning, computer vision, and data mining technologies to construct a comprehensive intelligent fire video analysis system focused on identifying flames and smoke and tracing the source of the ignited object. The research methodology encompasses data acquisition, flame detection, feature extraction, matching and tracing, model optimization, system application, and field testing to ensure the accuracy and practical value of the technology. Ultimately, the research results will provide efficient and accurate technical support for public security agencies and promote the intelligent analysis and judgment of fire-related cases.

[0058] This invention aims to break through the limitations of video analysis in traditional fire-related incident investigations and promote the transformation of fire video analysis technology from experience-based judgment to a data-driven, intelligent analysis model.

[0059] The implementation of this invention will meet the public security department's urgent need for intelligent analysis technology for fire-related video, providing accurate and efficient technical support. It will improve the accuracy of fire case investigations, promote intelligent and information-based investigations, enhance public security technical capabilities, and achieve a new level of public security combat effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 The invention discloses a fire source information tracing model based on flame image technology. DETAILED DESCRIPTION

[0061] The following is combined with Figure 1 The present invention will be further described in detail with specific implementations to facilitate a clear understanding of the present invention, but they do not constitute a limitation to the present invention.

[0062] In the description of the present invention, it should be noted that the terms "upper", "lower", "front", "back", "left", "right", "vertical", "inside", "outside", etc., indicating directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, they cannot be understood as limiting the present invention.

[0063] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood broadly. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; and direct or indirect connections through an intermediary. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0064] Example 1

[0065] As attached Figure 1 As shown, a method for tracing the source of fire-causing objects based on flame profiling technology in this embodiment includes the following steps:

[0066] S 1. Collect and organize fire video data from existing fire-related incidents, laboratory simulation data, and public data from online platforms to build a fire video database.

[0067] S1 includes data collection and preprocessing, which specifically includes the following steps:

[0068] S 11. Build a multi-source fire video database, the data sources include:

[0069] Fire-related incidents video: Collect actual fire video data from fire-related incidents by public security and fire departments.

[0070] Laboratory simulated fire data: Fire video data obtained from controlled laboratory combustion experiments, capturing accurate flame and smoke characteristics. This involves burning various fire materials in a controlled environment, capturing fire videos for flame and smoke characterization analysis. This includes fire accelerants, paper, wood, plastic, gasoline, and electrical equipment. In this example, the accelerant is gasoline, diesel, alcohol, or paint thinner.

[0071] Public online videos: Crawl fire videos from public online platforms, including news reports and user-uploaded videos on social media platforms such as Douyin and Bilibili.

[0072] The video dataset in the multi-source fire video database is divided into training set and test set in proportion, which are used for model construction and optimization respectively to improve its generalization ability.

[0073] S 12. Data cleaning and denoising

[0074] After data collection, video processing software combined with manual review was used for data preprocessing: cleaning and removing blurry, incomplete, or interfering clips to ensure data quality. A video classification method was established, labeling videos based on fire location, burning objects, and fire source. A database was constructed for subsequent model training and analysis. FFmpeg and Adobe Premiere were used to remove low-quality, blurry, repetitive, and irrelevant clips to ensure data accuracy. Gaussian and bilateral filtering image denoising methods were used to reduce video noise interference and improve the accuracy of flame and smoke detection.

[0075] S2. Develop a large-scale fire video analysis model system based on deep learning technology and flame profiling technology to solve the problem of matching the characteristics of flames and smoke in fire incident videos with the source of the fire. Establish a feature matching and source tracing model to quickly determine whether the fire was caused by an accelerant.

[0076] S2 includes image positioning and feature extraction, that is, studying the accurate identification technology of flames and smoke under unfavorable shooting conditions such as "overexposure", "reflection" and "low illumination", and deeply learning the unique physical and dynamic characteristics of flames and smoke, vectorizing the characteristics of fire flames and smoke, and intelligently constructing the association between feature extraction and flame and smoke characteristics of fire videos.

[0077] The specific steps include:

[0078] S21. Flame area detection

[0079] A deep learning target detection algorithm is used to automatically detect flame areas. Edge detection, color segmentation, and optical flow analysis are used to assist in improving the model's adaptability to diverse environments. In this embodiment, the deep learning target detection algorithms include the YOLO algorithm and the Faster R-CNN algorithm. Different environments include lighting changes, smoke interference, and reflections.

[0080] S22. Flame and smoke feature extraction

[0081] Extract multi-dimensional features of flames and smoke, including:

[0082] Color features: Use HSV color space conversion and flame color histogram analysis to extract flame color distribution information.

[0083] Morphological features: HOG oriented gradient histogram and morphological processing methods are used to extract the contour features of the flame.

[0084] Dynamic features: Through optical flow method, frame difference method, and LSTM time series analysis, flame combustion pattern and motion trajectory are extracted, mainly including flame spread rate, area change rate, volume change rate, flicker frequency, illumination change, number of sharp corners and their probability density.

[0085] Temperature characteristics: Thermal imaging analysis is used to extract the temperature distribution characteristics of the flame, and correction is performed in combination with the blackbody radiation model. The temperature characteristics are based on infrared video.

[0086] Smoke characteristics: Gray-level co-occurrence matrix (GLCM) is used to analyze the texture characteristics of smoke and reflect the distribution of smoke particles. Fourier transform is used to analyze the frequency domain characteristics of smoke and capture periodic changes.

[0087] S23. Feature Vectorization

[0088] The above features of S22 are converted into vector form and input into a deep learning network for further analysis. This maps the flame and smoke features into a high-dimensional space to improve classification accuracy. Deep learning networks include ResNet and EfficientNet.

[0089] In S2, after image positioning and feature extraction, the system also includes flame and smoke feature matching and tracing the source of the fire. This involves building a flame and smoke feature matching model, applying intelligent comparison technology, optimizing feature vector calculation, and improving matching accuracy. Using flame and smoke features, the source of the fire can be inferred, establishing an efficient video case query system to support rapid comparison and analysis of fire-related cases.

[0090] The specific steps include:

[0091] S24. Constructing a fire flame and smoke feature matching model

[0092] Using the fire video database constructed in S1, a flame and smoke feature library was established, covering flame and smoke characteristics for different igniters, accelerants, and combustion modes. A flame and smoke feature matching model was constructed, and the intelligent K-NN matching technology was used to optimize feature vector calculation and improve matching accuracy. Feature indexing technology was used to increase the search speed of flame and smoke feature matching. Feature indexing technologies include FAISS. In S26, the classification networks used were ResNet and VGG.

[0093] S25. Fire flame and smoke feature matching method

[0094] Euclidean distance, cosine similarity, and Mahalanobis distance are used to calculate the similarity of flame and smoke features. A hash index combined with a KD-Tree algorithm accelerates large-scale data matching and improves the efficiency of flame tracing. The SiameseNetwork is introduced to optimize flame and smoke feature matching through metric learning.

[0095] S26. Analysis and determination of fire species

[0096] Using a classification network, a fire source classification model is trained to intelligently identify fire sources. Combined with laboratory data, a fine-grained analysis of the flame and smoke characteristics of different combustible materials is performed to improve the accuracy of fire source information tracing. The clustering logic for identification is as follows: First, determine whether an accelerant contributed to the combustion. If so, investigate the case as suspected arson, and further identify its type based on video analysis. If an accelerant is ruled out, further determination is made as to whether the fire source is a hazardous chemical. If not, the fire is determined to be caused by ordinary combustible materials. Smoke and flame characteristics are then used to further trace the source of the fire source.

[0097] S3. Model optimization and updating: using incremental learning technology to dynamically adjust parameters, improve recognition accuracy, and achieve adaptive optimization, so that the model can continuously adapt to different environments and fire types, and improve the ability to trace information on fire sources or fire sources in complex scenarios.

[0098] In S3, model optimization and update are for the fire flame and smoke recognition models, as follows:

[0099] S31. Model Optimization

[0100] Experiments were conducted using test set video data to evaluate the model's recognition accuracy and generalization capabilities. Incremental learning techniques were used to dynamically adjust model parameters, improve recognition accuracy, and achieve adaptive optimization. By continuously introducing new fire types and environmental data, the model was able to adapt to different environments and fire types, enhancing its ability to trace fire sources in complex scenarios.

[0101] Specifically, in this embodiment, an online learning framework is used to enable the smoke recognition model to continuously adjust parameters when receiving new fire data without the need for complete retraining. Knowledge distillation technology is combined to reduce computing costs. Pre-trained models, such as ResNet on the ImageNet and COCO datasets, are used to fine-tune the fire flame recognition task to improve recognition capabilities under small sample data. The network structure is adjusted for different fire scenarios, such as forest fires and indoor fires, to improve adaptability. Automatic hyperparameter optimization methods, such as WieOptuna and Hyperopt, are used to optimize model parameters and improve accuracy. In combination with attention mechanisms, such as SE-Net and Transformer, the model's ability to capture fire flame and smoke characteristics is enhanced.

[0102] S32. Model update

[0103] Establish a dynamic model update mechanism, regularly introduce new video data and feature information, and continuously optimize model performance. Based on feedback from actual applications, iteratively optimize the model to ensure its accuracy and practicality.

[0104] Example 2

[0105] This embodiment provides a fire source information tracing system based on flame profiling technology. It adopts the fire source information tracing method of the above-mentioned embodiment 1. The fire source information tracing system integrates flame and smoke detection, feature extraction, and fire source information tracing functions. It uses PyTorch+Flask / Django to build the backend and combines it with a Web front-end visualization interface to implement interactive analysis functions. Specifically, it includes the following modules:

[0106] Fire video database module: It is used to collect and organize fire video data from existing fire-related incidents, laboratory simulation data, and public data from online platforms to build a fire video database.

[0107] Fire-related video submodule: Based on deep learning technology, it is used to develop a large model system for fire-related video analysis based on flame profiling technology to solve the problem of matching the fire flames, smoke characteristics and burning objects in fire-related incident videos.

[0108] Information tracing module: It is used to establish feature matching and fire source information tracing models to quickly determine whether the fire was caused by an accelerant.

[0109] Model optimization and update module: Used to dynamically adjust parameters with the help of incremental learning technology, improve recognition accuracy, achieve adaptive optimization, enable the model to continuously adapt to different environments and fire types, and improve the ability to trace the source of fire or fire information in complex scenarios.

[0110] Practical application test module: It is used to combine fire case data to test the actual application effect of the fire flame recognition system, evaluate the accuracy, recall rate, processing speed, and optimize system performance.

[0111] The above is merely a preferred embodiment of the present invention and does not constitute any formal limitation on the structure of the present invention. The layout and number of the present invention are not limited to this example and can be optimized according to actual engineering practices. Any modifications, equivalent changes, and decorations to the above embodiment based on the technical principles of the present invention that do not depart from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A method for tracing the source of fire-causing objects based on flame profiling technology, characterized in that: The steps include: S1. Collect and organize fire video data from existing fire incidents, laboratory simulation data, and public data from online platforms to build a fire video database; S2. Develop a large-scale fire video analysis model system based on deep learning technology and flame profiling technology to solve the problem of matching the characteristics of flames and smoke with the source of fire in fire incident videos. Establish a feature matching and source tracing model for source information to quickly determine whether the fire was caused by an accelerant. S3. Model optimization and updating: using incremental learning technology to dynamically adjust parameters, improve recognition accuracy, and achieve adaptive optimization, so that the model can continuously adapt to different environments and fire types, and improve the ability to trace information on fire sources or fire sources in complex scenarios.

2. The method for tracing the source of fire-causing objects based on flame profiling technology according to claim 1, characterized in that: S1 includes data collection and preprocessing, which specifically includes the following steps: S11. Build a multi-source fire video database. Data sources include: Fire-related incident video: Collect actual fire video data from public security and fire-related incidents; Laboratory simulated fire data: Fire video data obtained from laboratory controlled combustion experiments to obtain accurate flame and smoke characteristic data. This involves burning different fire materials in a controlled environment and obtaining fire videos for flame and smoke characteristic analysis. This includes fire accelerants, paper, wood, plastic, gasoline, and electrical equipment. Public online videos: Crawl fire videos on public online platforms; The video dataset in the multi-source fire video database is divided into training and test sets in proportion, which are used for model construction and optimization respectively to improve its generalization ability; S12. Data cleaning and denoising After data collection is completed, video processing software combined with manual review is used for data preprocessing: cleaning, removing fuzzy, incomplete or interfering fragments to ensure data quality; establishing a video classification method, labeling the video according to the fire location, fire object, and fire source elements, and building a database for subsequent model training and analysis. FFmpeg and Adobe Premiere tools are used to remove low-quality, fuzzy, repeated, and irrelevant fragments in S11 to ensure data accuracy; Gaussian filtering and bilateral filtering image denoising methods are used to reduce video noise interference and improve the accuracy of flame and smoke detection.

3. The method for tracing the source of fire-causing objects based on flame profiling technology according to claim 2, characterized in that: The combustion aid in S11 is gasoline, diesel, alcohol or paint thinner; the fire videos on the public network platform include news reports, videos uploaded by users of self-media platforms such as Tik Tok and Bilibili.

4. The method for tracing the source of fire-causing objects based on flame profiling technology according to claim 3, characterized in that: S2 includes image positioning and feature extraction, namely, studying the accurate recognition technology of flames and smoke under adverse shooting conditions such as "overexposure", "reflection", and "low illumination", and deeply learning the unique physical and dynamic characteristics of flames and smoke, vectorizing the characteristics of fire flames and smoke, and intelligently building the association between feature extraction and flame and smoke characteristics of fire videos; The specific steps include: S21. Flame area detection A deep learning target detection algorithm is used to automatically detect flame areas. Edge detection, color segmentation, and optical flow analysis methods are used to assist and improve the model's adaptability to different environments. S22. Flame and smoke feature extraction Extract multi-dimensional features of flames and smoke, including: Color features: Use HSV color space conversion and flame color histogram analysis to extract flame color distribution information; Morphological features: Use HOG oriented gradient histogram and morphological processing methods to extract the contour features of the flame; Dynamic features: Through optical flow, frame difference, and LSTM time series analysis, the flame combustion pattern and motion trajectory are extracted, mainly including flame spread rate, area change rate, volume change rate, flicker frequency, illumination change, number of sharp corners, and their probability density; Temperature characteristics: Use thermal imaging analysis to extract the temperature distribution characteristics of the flame and combine it with the blackbody radiation model for correction; the temperature characteristics are based on infrared video; Smoke characteristics: Gray-level co-occurrence matrix (GLCM) is used to analyze the texture characteristics of smoke and reflect the distribution of smoke particles. Fourier transform is used to analyze the frequency domain characteristics of smoke and capture periodic changes. S23. Feature Vectorization The above features of S22 are converted into vector form and input into the deep learning network for further analysis, and the flame and smoke features are mapped to high-dimensional space to improve the classification accuracy.

5. The method for tracing the source of fire-causing objects based on flame profiling technology according to claim 4, characterized in that: In S21, the deep learning target detection algorithm includes the YOLO algorithm and the Faster R-CNN algorithm; the different environments include illumination changes, smoke interference, and reflections; in S23, the deep learning network includes the ResNet and EfficientNet networks.

6. The method for tracing the source of fire-causing objects based on flame profiling technology according to claim 5, characterized in that: In S2, after image positioning and feature extraction, it also includes flame and smoke feature matching and information tracing of the source of the fire. That is, by building a flame and smoke feature matching model, using intelligent comparison technology, optimizing feature vector calculation, and improving matching accuracy; using flame and smoke features to reversely infer the source of the fire, an efficient video case query system is established to support rapid comparison and analysis of fire-related cases; The specific steps include: S24. Constructing a fire flame and smoke feature matching model Using the fire video database built in S1, a fire flame and smoke feature library is established, covering the flame and smoke features of different fire starters, combustion aids, and combustion modes; a flame and smoke feature matching model is constructed, and the intelligent comparison technology K-NN is used to optimize feature vector calculation to improve the matching accuracy; feature indexing technology is used to improve the retrieval speed of flame and smoke feature matching; S25. Fire flame and smoke feature matching method The similarity of flame and smoke characteristics is calculated using Euclidean distance, cosine similarity, and Mahalanobis distance methods; Combining hash index + KD-Tree to accelerate large-scale data matching and improve the efficiency of flame tracing; introducing the Siamese Network to optimize the matching capabilities of flame and smoke characteristics through metric learning; S26. Analysis and determination of fire species Through the classification network, the fire object classification model is trained to realize the intelligent identification of fire objects; combined with laboratory data, the flame and smoke characteristics of different combustion objects are analyzed in a fine-grained manner to improve the accuracy of the fire object information traceability.

7. The method for tracing the source of fire-causing objects based on flame profiling technology according to claim 6, characterized in that: In S26, the cluster logic judgment sequence is as follows: first, determine whether an accelerant is involved in the combustion. If so, investigate the case as an arson suspect, and further determine its type based on video analysis; If the fire is not caused by an accelerant, the fire is further determined to be caused by hazardous chemicals. If not, the fire is determined to be caused by ordinary combustible materials. Based on the smoke and flame characteristics, the fire source information is further traced.

8. The method for tracing the source of fire-causing objects based on flame profiling technology according to claim 7, characterized in that: In S24, the feature indexing technology includes FAISS technology; in S26, the classification network is ResNet or VGG.

9. The method for tracing the source of fire-causing objects based on flame profiling technology according to claim 8, characterized in that: In S3, the model optimization and update is the optimization and update of the fire flame and smoke recognition model, which is as follows: S31. Model Optimization Experiments were conducted using test set video data to evaluate the model's recognition accuracy and generalization capabilities. Incremental learning techniques were used to dynamically adjust model parameters, improve recognition accuracy, and achieve adaptive optimization. By continuously introducing new fire types and environmental data, the model was able to adapt to different environments and fire types, enhancing its traceability capabilities in complex scenarios. S32. Model update Establish a dynamic model update mechanism, regularly introduce new video data and feature information, and continuously optimize model performance; iteratively optimize the model based on feedback from actual applications to ensure the accuracy and practicality of the model.

10. A fire-causing object information tracing system based on flame profiling technology, which adopts the fire-causing object information tracing method according to claim 9, characterized in that: The fire source information tracing system integrates flame and smoke detection, feature extraction, and fire source information tracing functions. It uses PyTorch+Flask / Django to build the backend and combines it with a Web front-end visualization interface to implement interactive analysis functions. Specifically, it includes the following modules: Fire video database module: It is used to collect and organize fire video data from existing fire-related incidents, laboratory simulation data, and public data from online platforms to build a fire video database; Fire-related video analysis module: Based on deep learning technology, it is used to develop a large-scale fire video analysis model system based on flame profiling technology to solve the problem of matching the characteristics of fire and smoke with the burning objects in fire-related incident videos; Information tracing module: This module is used to establish a feature matching and fire source information tracing model to quickly determine whether the fire was caused by an accelerant; Model optimization and update module: This module uses incremental learning technology to dynamically adjust parameters, improve recognition accuracy, and achieve adaptive optimization, enabling the model to continuously adapt to different environments and fire types, thereby improving the ability to trace fire sources or fire sources in complex scenarios. Practical application test module: It is used to combine fire case data to test the actual application effect of the fire flame recognition system, evaluate the accuracy, recall rate, processing speed, and optimize system performance.

Images