Fire reason analysis method and analysis system based on visual features

Abstract

The invention provides a fire reason analysis method and analysis system based on visual features, and relates to the technical field of fire fighting, and the analysis method comprises the steps: obtaining a video stream of a current fire, and extracting a plurality of fire features from the video stream of the current fire; based on the plurality of fire features, identifying a fire development stage by using a machine learning model; distributing a weight for each fire feature according to the fire development stage; performing weighted fusion on the plurality of fire features based on the weights to obtain a multi-modal feature vector, and establishing a multi-modal feature sequence of the multi-modal feature vector by taking the video frame number of the video stream of the current fire as a time axis; the multi-modal feature sequence is compared with a pre-constructed fire type feature library, and the fire type of the current fire is obtained.According to the fire type identification method and device, the multiple visual features of the fire scene are extracted, and the feature weight is dynamically adjusted according to the development stage of the fire, so that the fire type identification accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of fire fighting technology, and in particular to a fire cause analysis method and analysis system based on visual features. Background Art

[0002] In the field of fire monitoring, surveillance video is the key evidence for fire accident investigation and prevention, and its importance is self-evident. However, traditional video monitoring methods are often limited to simple screen playback, and cannot intuitively and quantitatively display the entire process of fire occurrence and development. Even if observers watch the video, it is difficult to accurately capture and calculate the data changes of the key characteristics of the fire in the visual characteristics of the monitoring screen during the initial stage of the fire. Therefore, there is a lack of quantitative data as an important scientific basis for the analysis and identification of the cause of the fire, which greatly limits the efficiency and accuracy of subsequent investigations and fire accident prevention work.

[0003] Therefore, it is necessary to provide a fire cause analysis method and analysis system based on visual features to solve the above technical problems. Summary of the invention

[0004] To solve the above technical problems, the present invention provides a fire cause analysis method and analysis system based on visual features, which extracts multiple visual features of the fire scene and dynamically adjusts the feature weights according to the development stage of the fire to improve the accuracy of fire type identification.

[0005] The present invention provides a method for analyzing the cause of fire based on visual features, the method comprising the following steps:

[0006] S1: Obtain a video stream of a current fire, and extract multiple fire features from the video stream of the current fire;

[0007] S2: Based on the multiple fire characteristics, identifying the fire development stage using a machine learning model;

[0008] S3: assigning weights to the various fire characteristics according to the fire development stage;

[0009] S4: performing weighted fusion on the multiple fire features based on the weights to obtain a multimodal feature vector, and establishing a multimodal feature sequence of the multimodal feature vector using the video frame number of the video stream of the current fire as the time axis;

[0010] S5: Compare the multimodal feature sequence with a pre-constructed fire type feature library to obtain the fire type of the current fire, wherein the fire type feature library includes fire events of multiple known fire types and corresponding known multimodal feature sequences.

[0011] Preferably, the multiple fire characteristics include brightness characteristics, color characteristics, texture characteristics, shape characteristics and motion characteristics.

[0012] Preferably, step S1 comprises the following steps:

[0013] S11: receiving a video stream from the fire scene and preprocessing the received video stream;

[0014] S12: Separate the fire area from the background using image segmentation technology from the preprocessed video stream;

[0015] S13: In the separated fire area, brightness features, color features, texture features, shape features and motion features are extracted respectively.

[0016] Preferably, step S2 comprises the following steps:

[0017] S21: training the constructed machine learning model using the collected training data, wherein the training data includes video stream data of a plurality of known fire events and corresponding annotated fire development stages;

[0018] S22: Input the extracted multiple fire features into a trained machine learning model, and the machine learning model outputs the fire development stage of the current fire, wherein the fire development stage includes an initial stage, a rising stage, and a stable stage.

[0019] Preferably, step S3 comprises the following steps:

[0020] S31: Establish a weight distribution formula, wherein the weight distribution formula is:

[0021] ω i =α i ·f(D)

[0022] Among them, ω i represents the weight of the i-th fire feature, α i represents the basic weight of the ith fire characteristic, f(D) represents the influencing factor of the fire development stage D on the weight of the fire characteristic;

[0023] S32: dynamically adjusting the impact factor according to the fire development stage of the current fire;

[0024] S33: Based on the adjusted influencing factors, the weights of the various fire characteristics are dynamically calculated using the weight allocation formula.

[0025] Preferably, step S4 comprises the following steps:

[0026] S41: performing weighted fusion on the multiple fire features based on the weights to obtain a multimodal feature vector, wherein the calculation formula of the multimodal feature vector is:

[0027]

[0028] Among them, V t represents the multimodal feature vector of the tth frame, ω i (t) represents the weight of the i-th fire feature in the t-th frame, v i (t) represents the feature vector of the i-th fire feature in the t-th frame;

[0029] S42: Using the video frame number of the video stream of the current fire as the time axis, establishing a multimodal feature sequence of the multimodal feature vector, wherein the multimodal feature sequence is expressed as:

[0030] {V1, V2, ..., V T}

[0031] Where T represents the total number of frames in the video stream of the current fire.

[0032] Preferably, step S5 comprises the following steps:

[0033] S51: Calculate the similarity between the multimodal feature sequence and the known multimodal feature sequence of each known fire type in the fire type feature library, wherein the similarity is represented by Euclidean distance, and the calculation formula of the Euclidean distance is:

[0034]

[0035] Where a and b represent the multimodal feature sequence of the current fire and the known multimodal feature sequence, respectively, T represents the total number of frames in the video stream of the current fire, d(a, b) represents the Euclidean distance, and a t represents the multimodal features of the current fire at frame t, b t The known multimodal features representing the known fire type at the t-th frame;

[0036] S52: Selecting a multimodal feature sequence of a known fire type that has the smallest Euclidean distance to the multimodal feature sequence, and using the fire type corresponding to the multimodal feature sequence as the fire type of the current fire.

[0037] The present invention also provides a fire cause analysis system based on visual features, which is used to execute the fire cause analysis method based on visual features. The analysis system comprises:

[0038] A feature extraction module, used for acquiring a video stream of a current fire and extracting a plurality of fire features from the video stream of the current fire;

[0039] A fire development stage identification module, configured to identify the fire development stage using a machine learning model based on the plurality of fire characteristics;

[0040] A weight allocation module, used for allocating weights to various fire characteristics according to the fire development stage;

[0041] A feature sequence construction module, used for weighted fusion of the plurality of fire features based on the weights to obtain a multimodal feature vector, and establishing a multimodal feature sequence of the multimodal feature vector using the video frame number of the video stream of the current fire as the time axis;

[0042] The fire type determination module is used to compare the multimodal feature sequence with a pre-built fire type feature library to obtain the fire type of the current fire, wherein the fire type feature library includes fire events of multiple known fire types and corresponding known multimodal feature sequences.

[0043] Compared with the related art, the fire cause analysis method and analysis system based on visual features provided by the present invention have the following beneficial effects:

[0044] The present invention extracts multimodal features including brightness features, color features, texture features, shape features and motion features, and dynamically adjusts the weights of these features according to the fire development stage, so that feature selection is more precise and the accuracy of fire type identification is improved.

[0045] At the same time, using machine learning models to identify the development stage of a fire can more accurately determine the development stage of the fire, and then adjust the feature weights in a targeted manner, thereby improving the accuracy of the overall analysis.

[0046] The analysis method of the present invention has a high degree of automation. The user only needs to provide a video stream of the fire scene to automatically complete the identification of the fire type, thereby simplifying the operation process and lowering the user's usage threshold. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A flow chart of a method for analyzing fire causes based on visual features provided by the present invention;

[0048] Figure 2 A module structure diagram of a fire cause analysis system based on visual features provided by the present invention. DETAILED DESCRIPTION

[0049] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only the parts related to the present invention, rather than all structures, are shown in the accompanying drawings. In addition, the embodiments of the present invention and the features in the embodiments may be combined with each other without conflict.

[0050] It should also be noted that, for ease of description, only the part relevant to the present invention but not all content is shown in the accompanying drawings. It should be mentioned before discussing exemplary embodiments in more detail that some exemplary embodiments are described as processing or methods depicted as flow charts. Although the flow chart describes each operation (or step) as sequential processing, many operations therein can be implemented in parallel, concurrently or simultaneously. In addition, the order of each operation can be rearranged. When its operation is completed, the processing can be terminated, but it can also have additional steps not included in the accompanying drawings. The processing can correspond to methods, functions, procedures, subroutines, subprograms, etc.

[0051] Embodiment 1

[0052] The present invention provides a method for analyzing the cause of fire based on visual features. Figure 1 As shown, the analysis method includes the following steps:

[0053] S1: Obtain a video stream of the current fire, and extract multiple fire features from the video stream of the current fire.

[0054] Specifically, step S1 includes the following steps:

[0055] S11: Receive a video stream from the fire scene and pre-process the received video stream.

[0056] In this embodiment, a video stream of the fire scene is acquired in real time through a camera or other video acquisition equipment, and then the video stream is preprocessed, and the preprocessing includes:

[0057] Brightness and contrast adjustment: Adjust the brightness and contrast according to the actual situation of the video stream to ensure image quality.

[0058] Noise removal: Remove noise interference in the video stream through filtering technology to improve image quality.

[0059] The preprocessing steps can effectively improve the image quality and lay a good foundation for subsequent feature extraction. Among them, operations such as dehazing and noise removal can reduce interference factors in the image and improve the accuracy of feature extraction.

[0060] S12: From the preprocessed video stream, the fire area is separated from the background using image segmentation technology.

[0061] In this embodiment, a threshold-based image segmentation method is used to separate the fire area from the background, and then the location and range of the fire area are determined so that subsequent feature extraction is focused on the fire area.

[0062] Accurately locating the fire area can avoid interference from the background area, improve the pertinence of feature extraction, and only extract features in the fire area, which can significantly improve the speed and efficiency of feature extraction.

[0063] S13: In the separated fire area, brightness features, color features, texture features, shape features and motion features are extracted respectively.

[0064] In this embodiment, feature extraction is performed in the fire area, specifically:

[0065] Brightness feature: Calculate the average brightness value of the fire area.

[0066] Color feature: Extract the color distribution information of the fire area and represent it using a color histogram.

[0067] Texture features: The gray level co-occurrence matrix method is used to extract the texture features of the fire area.

[0068] Shape features: Analyze the contour shape of the fire area and extract shape descriptors including but not limited to perimeter and area.

[0069] Motion features: Use the optical flow method to track the motion trajectory of the fire area and extract the motion direction and speed.

[0070] By extracting multiple features, different aspects of a fire can be fully captured, providing rich information for subsequent identification of fire types. The combination of multiple features can improve the accuracy of fire type identification, which is particularly important when distinguishing different types of fire.

[0071] S2: Using machine learning models to identify fire development stages based on multiple fire characteristics.

[0072] Specifically, step S2 includes the following steps:

[0073] S21: Train the constructed machine learning model using the collected training data, wherein the training data includes video stream data of multiple known fire events and corresponding labeled fire development stages.

[0074] In this embodiment, video stream data of multiple known fire events and corresponding annotated fire development stages are collected as training data.

[0075] Then, the training data is preprocessed as necessary, including but not limited to data cleaning, feature selection, data enhancement, etc., to improve the effect of model training.

[0076] Select appropriate machine learning algorithms, including but not limited to support vector machine (SVM), random forest and neural network, to build a machine learning model.

[0077] The constructed machine learning model is trained using the training dataset, and the model parameters are optimized through the back-propagation method so that it can accurately identify the development stage of the fire.

[0078] By using a large amount of labeled data for training, the model can learn how to accurately identify fires at different stages of development and improve recognition accuracy. At the same time, a fully trained model has strong generalization capabilities and can better cope with new data that has not been seen, which is very important for practical applications.

[0079] Once the model is trained, it can automatically identify the development stages of a fire, greatly reducing the need for human intervention

[0080] S22: Input the extracted multiple fire features into the trained machine learning model, and the machine learning model outputs the fire development stage of the current fire, where the fire development stage includes the initial stage, the rising stage and the stable stage.

[0081] In this embodiment, the extracted multiple fire features are input into a trained machine learning model, and the machine learning model makes predictions based on the input features and outputs the current fire development stage, including the initial stage, rising stage, and stable stage.

[0082] S3: Assign weights to individual fire characteristics based on the fire development stage.

[0083] Specifically, step S3 includes the following steps:

[0084] S31: Establish a weight distribution formula, where the weight distribution formula is:

[0085] ω i =α i ·f(D)

[0086] Among them, ω i represents the weight of the i-th fire feature, α i represents the basic weight of the ith fire characteristic, and f(D) represents the influencing factor of the fire development stage D on the weight of the fire characteristic.

[0087] In this embodiment, it is first necessary to construct a reasonable weight distribution formula to dynamically adjust the weights of various fire features according to the fire development stage. The weight distribution formula takes into account the basic weight of each fire feature and the influencing factor of the fire development stage on the feature weight. This formula is intended to adjust the relative importance of features according to different stages of the fire (such as the initial stage, the rising stage, and the stable stage). For example, in the early stage of a fire, the color and brightness characteristics of the flame may be more important because these characteristics can quickly indicate the occurrence of a fire; while in the stable stage, the shape and texture characteristics of the flame may be more critical because they help determine the type and scale of the fire.

[0088] By introducing the weight distribution formula, this method can flexibly adjust the weight of each feature according to the characteristics of different fire development stages, thereby improving the accuracy and robustness of fire type identification. This dynamic weight adjustment mechanism makes the feature extraction and fusion process more consistent with the actual development of the fire, and improves the reliability of the analysis results.

[0089] S32: Dynamically adjust the impact factor according to the fire development stage of the current fire.

[0090] In this embodiment, a mechanism is designed to dynamically adjust the impact factor according to the specific development stage of the current fire. Specifically, it is: identification of the fire development stage and a set of predefined mapping tables that map the fire development stage to different impact factor values. For example, in the early stage of the fire, the impact factor value can be set to a higher value to increase the weight of color and brightness features; while in the stable period, the impact factor value is adjusted to a lower value to balance the weight of shape and texture features.

[0091] By dynamically adjusting the impact factors, the importance changes of different features at different stages of the fire can be more accurately reflected, thereby further improving the accuracy of fire type identification. This dynamic adjustment mechanism enables the feature weights to be automatically adjusted as the fire develops, enhancing the adaptability and intelligence of the analysis process.

[0092] S33: Based on the adjusted influencing factors, the weights of each fire characteristic are dynamically calculated using the weight allocation formula.

[0093] In this embodiment, the actual weight of each fire feature at the current fire development stage is calculated using a weight distribution formula based on the adjusted influencing factors and the basic weight of each fire feature. This calculation process will take into account all relevant factors, including the fire development stage, the basic importance of the feature and the influencing factors, to obtain a weight distribution scheme that best suits the current situation.

[0094] By dynamically calculating feature weights based on the adjusted influencing factors, this method can ensure that the weight distribution of each feature is both reasonable and accurate during the fire type identification process. This dynamic weight calculation mechanism not only improves the efficiency of feature extraction and fusion, but also enhances the accuracy and reliability of fire type identification.

[0095] S4: Perform weighted fusion on multiple fire features based on weights to obtain a multimodal feature vector, and establish a multimodal feature sequence of the multimodal feature vector using the video frame number of the current fire video stream as the time axis.

[0096] Specifically, step S4 includes the following steps:

[0097] S41: performing weighted fusion on multiple fire features based on weights to obtain a multimodal feature vector, wherein the calculation formula of the multimodal feature vector is:

[0098]

[0099] Among them, V t represents the multimodal feature vector of the tth frame, ω i (t) represents the weight of the i-th fire feature in the t-th frame, v i (t) represents the feature vector of the i-th fire feature in the t-th frame.

[0100] In this embodiment, in step S41, weighted fusion of multiple fire features based on weights is mainly performed to generate a multi-modal feature vector for each frame. The specific steps are as follows:

[0101] According to the weights of each fire feature at the current fire development stage calculated in step S3, these weights are applied to the corresponding fire feature vectors. The size of the weight reflects the importance of the feature for fire type identification at the current fire development stage.

[0102] Feature fusion: Through weighted summation, the weighted fire feature vectors are fused into a single multimodal feature vector. This vector integrates information on brightness, color, texture, shape, and motion, providing comprehensive data support for subsequent fire type identification.

[0103] By assigning dynamic weights to each fire feature and performing feature fusion based on these weights, the actual situation of the fire can be reflected more accurately, thereby improving the accuracy of fire type identification; the construction of multimodal feature vectors takes into account a variety of fire characteristics, enabling the system to cope with fire situations under different environmental conditions, thereby enhancing the robustness of the system.

[0104] S42: Using the video frame number of the current fire video stream as the time axis, a multimodal feature sequence of the multimodal feature vector is established, wherein the multimodal feature sequence is expressed as:

[0105] {V1, V2, ..., V T}

[0106] Where T represents the total number of frames in the video stream of the current fire.

[0107] In this embodiment, in step S42, the main process is to construct a multimodal feature sequence using the video frame number as the time axis. The specific steps are as follows:

[0108] Timeline establishment: Based on the video frame number of the current fire video stream, a timeline is established. Each frame corresponds to a time point, which is used to record the multimodal feature vector at that time point.

[0109] Sequence construction: The multimodal feature vectors corresponding to each frame are arranged in chronological order to form a multimodal feature sequence. This sequence contains not only the static feature information of the fire, but also the dynamic feature information of the fire changing over time.

[0110] Multimodal feature sequences can comprehensively reflect the dynamic change process of a fire from its inception to the current time point, providing richer information for the accurate identification of fire types. By considering the spatiotemporal variation characteristics of fires, fire types can be more accurately identified, especially when fire types are complex and changeable.

[0111] S5: Compare the multimodal feature sequence with a pre-built fire type feature library to obtain the fire type of the current fire, wherein the fire type feature library includes multiple fire events of known fire types and corresponding known multimodal feature sequences.

[0112] Specifically, step S5 includes the following steps:

[0113] S51: Calculate the similarity between the multimodal feature sequence and the known multimodal feature sequence of each known fire type in the fire type feature library, wherein the similarity is represented by Euclidean distance, and the calculation formula of Euclidean distance is:

[0114]

[0115] Where a and b represent the multimodal feature sequence of the current fire and the known multimodal feature sequence, respectively, T represents the total number of frames in the video stream of the current fire, d(a, b) represents the Euclidean distance, and a t represents the multimodal features of the current fire at frame t, b t The known multimodal features representing the known fire type at the t-th frame.

[0116] In this embodiment, the multimodal feature sequence of the current fire has been obtained in step S4, and the pre-constructed fire type feature library includes multiple fire events of known fire types and corresponding known multimodal feature sequences.

[0117] The Euclidean distance is used as the similarity measure to calculate the distance between the multimodal feature sequence of the current fire and the known multimodal feature sequence of each known fire type in the fire type feature library.

[0118] S52: Selecting a multimodal feature sequence of a known fire type that has the smallest Euclidean distance with the multimodal feature sequence, and using the fire type corresponding to the multimodal feature sequence as the fire type of the current fire.

[0119] In this embodiment, according to the Euclidean distance calculated in step S51, the multimodal feature sequence of the known fire type with the smallest distance is selected, and the fire type corresponding to the multimodal feature sequence of the known fire type that is most similar to the multimodal feature sequence of the current fire is used as the fire type of the current fire.

[0120] By selecting the multimodal feature sequence of the known fire type that is most similar to the multimodal feature sequence of the current fire, the type of the current fire can be accurately identified. Accurate fire type identification helps to formulate effective fire extinguishing strategies and take appropriate countermeasures.

[0121] The present invention provides a method for analyzing the cause of fire based on visual features, the main purpose of which is to automatically identify the type of fire and its possible cause by analyzing the current fire video stream. The working principle is as follows:

[0122] First, a video stream from the fire scene is received and preprocessed, including but not limited to brightness and contrast adjustment, dehazing processing, etc., to improve image quality.

[0123] From the preprocessed video stream, image segmentation techniques are used to separate the fire area from the background.

[0124] In the separated fire area, brightness features, color features, texture features, shape features and motion features are extracted respectively.

[0125] Secondly, the constructed machine learning model is trained using the collected training data, which contains video stream data of multiple known fire events and the corresponding labeled fire development stages.

[0126] The extracted multiple fire features are input into the trained machine learning model, and the machine learning model outputs the fire development stage of the current fire. The development stages include the initial stage, rising stage and stable stage.

[0127] Next, a weight assignment formula is established to assign weights to various fire characteristics according to the fire development stage.

[0128] Dynamically adjust impact factors to reflect the different impacts of different development stages on feature importance.

[0129] The weights of various fire characteristics are dynamically calculated using the adjusted impact factors.

[0130] Next, multiple fire features are weightedly fused based on the weights to obtain a multimodal feature vector.

[0131] Taking the video frame number of the current fire video stream as the time axis, a multimodal feature sequence of the multimodal feature vector is established.

[0132] Finally, the similarity between the multimodal feature sequence and the known multimodal feature sequences of each known fire type in the fire type feature library is calculated, and the Euclidean distance is used as the similarity measure.

[0133] The multimodal feature sequence of the known fire type with the smallest Euclidean distance to the multimodal feature sequence is selected, and the fire type corresponding to the sequence is used as the fire type of the current fire.

[0134] In summary, the present invention extracts multimodal features from fire video streams, uses machine learning technology to identify fire development stages and assign feature weights, and finally identifies the type of current fire and its possible cause by comparing the similarity between the multimodal feature sequence and the multimodal feature sequence of known fire types. This method can effectively improve the accuracy and reliability of fire identification.

[0135] Embodiment 2

[0136] The present invention also provides a fire cause analysis system based on visual features, which is used to perform a fire cause analysis method based on visual features. Figure 2 As shown, the analysis system includes:

[0137] The feature extraction module 100 is used to obtain a video stream of the current fire and extract multiple fire features from the video stream of the current fire;

[0138] A fire development stage identification module 200 is used to identify the fire development stage based on multiple fire characteristics using a machine learning model;

[0139] A weight allocation module 300, for allocating weights to various fire characteristics according to the fire development stage;

[0140] The feature sequence construction module 400 is used to perform weighted fusion on multiple fire features based on weights to obtain a multimodal feature vector, and to establish a multimodal feature sequence of the multimodal feature vector using the video frame number of the current fire video stream as the time axis;

[0141] The fire type determination module 500 is used to compare the multimodal feature sequence with a pre-built fire type feature library to obtain the fire type of the current fire, wherein the fire type feature library includes fire events of multiple known fire types and corresponding known multimodal feature sequences.

[0142] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0143] A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable rewritable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0144] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

Claims

1. A method for analyzing the cause of fire based on visual features, characterized in that: The analytical method comprises the following steps: S1: Obtain a video stream of a current fire, and extract multiple fire features from the video stream of the current fire; S2: Based on the multiple fire characteristics, identifying the fire development stage using a machine learning model; S3: assigning weights to the various fire characteristics according to the fire development stage; S4: performing weighted fusion on the multiple fire features based on the weights to obtain a multimodal feature vector, and establishing a multimodal feature sequence of the multimodal feature vector using the video frame number of the video stream of the current fire as the time axis; S5: Compare the multimodal feature sequence with a pre-constructed fire type feature library to obtain the fire type of the current fire, wherein the fire type feature library includes fire events of multiple known fire types and corresponding known multimodal feature sequences.

2. A method for analyzing fire causes based on visual features according to claim 1, characterized in that: The plurality of fire features include brightness features, color features, texture features, shape features and motion features.

3. The method for analyzing the cause of fire based on visual features according to claim 2, characterized in that: Step S1 includes the following steps: S11: receiving a video stream from the fire scene and preprocessing the received video stream; S12: Separate the fire area from the background using image segmentation technology from the preprocessed video stream; S13: In the separated fire area, brightness features, color features, texture features, shape features and motion features are extracted respectively.

4. The method for analyzing the cause of fire based on visual features according to claim 3, characterized in that: Step S2 includes the following steps: S21: training the constructed machine learning model using the collected training data, wherein the training data includes video stream data of a plurality of known fire events and corresponding annotated fire development stages; S22: Input the extracted multiple fire features into a trained machine learning model, and the machine learning model outputs the fire development stage of the current fire, wherein the fire development stage includes an initial stage, a rising stage, and a stable stage.

5. The method for analyzing the cause of fire based on visual features according to claim 4, characterized in that: Step S3 includes the following steps: S31: Establish a weight distribution formula, wherein the weight distribution formula is: oh i =a i ·f(D) Among them, ω i represents the weight of the i-th fire feature, α i represents the basic weight of the ith fire characteristic, f(D) represents the influencing factor of the fire development stage D on the weight of the fire characteristic; S32: Dynamically adjust the impact factor according to the fire development stage of the current fire; S33: Based on the adjusted influencing factors, the weights of the various fire characteristics are dynamically calculated using the weight allocation formula.

6. The method for analyzing the cause of fire based on visual features according to claim 5, characterized in that: Step S4 includes the following steps: S41: performing weighted fusion on the multiple fire features based on the weights to obtain a multimodal feature vector, wherein the calculation formula of the multimodal feature vector is: Among them, V t represents the multimodal feature vector of the tth frame, ω i (t) represents the weight of the i-th fire feature in the t-th frame, v i (t) represents the feature vector of the i-th fire feature in the t-th frame; S42: Using the video frame number of the video stream of the current fire as the time axis, establishing a multimodal feature sequence of the multimodal feature vector, wherein the multimodal feature sequence is expressed as: {V1,V2,……,V T } Where T represents the total number of frames in the video stream of the current fire.

7. The method for analyzing the cause of fire based on visual features according to claim 6, characterized in that: Step S5 includes the following steps: S51: Calculate the similarity between the multimodal feature sequence and the known multimodal feature sequence of each known fire type in the fire type feature library, wherein the similarity is represented by Euclidean distance, and the calculation formula of the Euclidean distance is: Where a and b represent the multimodal feature sequence of the current fire and the known multimodal feature sequence, respectively, T represents the total number of frames in the video stream of the current fire, d(a, b) represents the Euclidean distance, and a t represents the multimodal features of the current fire at frame t, b t The known multimodal features representing the known fire type at the t-th frame; S52: Selecting a multimodal feature sequence of a known fire type that has the smallest Euclidean distance to the multimodal feature sequence, and using the fire type corresponding to the multimodal feature sequence as the fire type of the current fire.

8. A fire cause analysis system based on visual features, used to execute a fire cause analysis method based on visual features as claimed in any one of claims 1 to 7, characterized in that: The analysis system comprises: A feature extraction module, used for acquiring a video stream of a current fire and extracting a plurality of fire features from the video stream of the current fire; A fire development stage identification module, configured to identify the fire development stage using a machine learning model based on the plurality of fire characteristics; A weight allocation module, used for allocating weights to various fire characteristics according to the fire development stage; A feature sequence construction module, used for weighted fusion of the plurality of fire features based on the weights to obtain a multimodal feature vector, and establishing a multimodal feature sequence of the multimodal feature vector using the video frame number of the video stream of the current fire as the time axis; The fire type determination module is used to compare the multimodal feature sequence with a pre-built fire type feature library to obtain the fire type of the current fire, wherein the fire type feature library includes fire events of multiple known fire types and corresponding known multimodal feature sequences.

Images