Disaster identification and disposal method based on real-time video information and knowledge graph

Abstract

The invention relates to the technical field of disaster prevention and control, in particular to a disaster identification and disposal method based on real-time video information and a knowledge graph. The method comprises the following steps: preprocessing video data of a disaster area to obtain disaster images at all moments; inputting the disaster image at the current moment into the trained first detection model to obtain a first disaster type, a personnel density and disaster duration; inputting the disaster image at the current moment into a trained second detection model, and obtaining personnel behaviors, disaster features, disaster grades and second disaster types; obtaining a current disaster type according to the first disaster type and the second disaster type; and constructing a disaster knowledge graph, and searching an emergency response scheme in the disaster knowledge graph according to the current disaster type, the personnel density, the disaster duration, the personnel behavior, the disaster characteristics and the disaster level. According to the invention, the accuracy and real-time performance of disaster monitoring can be improved, and a corresponding emergency scheme is provided for disaster disposal.

Description

Technical Field

[0001] The present invention relates to the technical field of disaster prevention and control, and in particular to a disaster identification and disposal method based on real-time video information and knowledge graphs. Background Art

[0002] When natural disasters or emergencies occur, accurate and timely understanding of the disaster situation is the key to emergency response. Traditional disaster assessment methods mainly rely on manual field surveys or analysis based on static images, which is not only time-consuming but also difficult to reflect the dynamic changes of disasters in a timely manner. At the same time, with the widespread use of equipment such as drones and surveillance cameras in recent years, video data has become an important source of information on disaster sites.

[0003] At present, video-based disaster situation analysis technology mainly relies on target detection and scene understanding algorithms in the field of computer vision. However, the existing technology has the following problems: Insufficient real-time performance: Most existing video analysis systems cannot process massive amounts of video data in real time, and the delay is high, which affects the timeliness of emergency response. Low data upload efficiency: The amount of video data at the disaster site is large, and the existing system has a bandwidth bottleneck during the upload process, resulting in untimely upload of disaster assessment information, affecting decision-making. Low accuracy of disaster situation assessment: The existing video analysis algorithm has a low recognition accuracy when facing complex disaster scenes, especially in complex environments such as lighting changes, smoke, and occlusion, the algorithm performance decreases significantly. Insufficient professionalism in emergency response: The existing video analysis algorithm system does not have the ability to generate professional and accurate emergency response plans. Summary of the invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide a disaster identification and disposal method based on real-time video information and knowledge graph. The technical solution adopted is as follows:

[0005] An embodiment of the present invention provides a disaster identification and handling method based on real-time video information and knowledge graph, the method comprising:

[0006] Obtain video data of the disaster area, and pre-process the video data to obtain disaster images at each moment;

[0007] Constructing a first detection model and training the first detection model; inputting the disaster image at the current moment into the trained first detection model to obtain the first disaster type, population density and disaster duration;

[0008] Constructing a second detection model and training the second detection model; inputting the disaster image at the current moment into the trained second detection model to obtain personnel behavior, disaster characteristics, disaster level and second disaster type;

[0009] The current disaster type is obtained based on the first disaster type and the second disaster type; a disaster knowledge graph is constructed, and emergency response plans are found in the disaster knowledge graph based on the current disaster type, population density, disaster duration, personnel behavior, disaster characteristics and disaster level.

[0010] Preferably, the video data is preprocessed to obtain disaster images at each moment, including:

[0011] The preprocessing of video data includes noise reduction, deblurring and frame difference; the preprocessed video data is cropped to obtain the disaster images at each moment.

[0012] Preferably, constructing a first detection model and training the first detection model includes:

[0013] The yolov8 model is improved by introducing a pyramid network structure and adding an attention mechanism. The improved yolov8 model is used as the first detection model. Historical disaster image data is obtained and annotated to form a first training set. The first training set is used to train the first detection model to obtain the first detection model after training.

[0014] Preferably, constructing a second detection model and training the second detection model includes:

[0015] The second detection model is a Slowfast model with a fully connected layer. The historical disaster image data is annotated to form a second training set, and the second detection model is trained using the second training set to obtain the second detection model after training.

[0016] Preferably, the first disaster type, population density and disaster duration include:

[0017] The first disaster type includes earthquakes, fires and floods; the population density is the ratio of the number of people in the disaster area to the area of ​​the disaster area; and the disaster duration is the time from the start of video data shooting to the current moment.

[0018] Preferably, after obtaining the personnel behavior, disaster characteristics, disaster level and the second disaster type, the method further includes:

[0019] The disaster levels are divided into severe, relatively severe, generally severe, minor and no disaster; for disaster levels above minor, the current moment is used as the middle moment of the video clip to obtain the key video clip, where the length of the key video clip is the preset length; the key video clip is uploaded to the command center.

[0020] Preferably, obtaining the current disaster type according to the first disaster type and the second disaster type includes:

[0021] When the first disaster type and the second disaster type are the same, the current disaster type is any one of the first disaster type and the second disaster type; when the first disaster type and the second disaster type are different, the predicted probabilities of the two models are obtained and weighted averaged to calculate the final probability of each category, and the category with the highest final probability is selected as the current disaster type.

[0022] Preferably, constructing a disaster knowledge graph includes:

[0023] Obtain historical disaster accident reports, emergency plans, emergency drill reports and emergency regulations, and record them as text data; clean, deduplicate and format the text data; extract key information related to the disaster accident based on the processed text data, perform structured processing on the extracted key information, use natural language processing technology to annotate the information, establish associations between entities and relationships, generate a disaster knowledge graph, and update the disaster knowledge graph in real time.

[0024] Preferably, an emergency response plan is searched in the disaster knowledge graph according to the current disaster type, population density, disaster duration, personnel behavior, disaster characteristics and disaster level, including:

[0025] Each historical disaster case in the disaster knowledge graph has a corresponding index, which includes disaster time, disaster location, disaster type, population density, disaster duration, personnel behavior, disaster characteristics and disaster level; the index of the historical disaster case is used as an index vector; the previous disaster type, population density, disaster duration, personnel behavior, disaster characteristics, disaster level, disaster time and disaster location at the current moment are obtained to form a disaster feature vector; the distance between the disaster feature vector and each index vector is calculated respectively; the historical disaster case corresponding to the nearest index vector is obtained, and a disaster assessment and emergency response plan is generated based on key information such as the causal relationship and processing flow of the historical disaster case.

[0026] The embodiments of the present invention have at least the following beneficial effects: the present invention collects video data of the disaster area, and pre-processes the video data to obtain a disaster image, thereby improving the data quality as well as the accuracy and efficiency of subsequent analysis; constructs a first detection model and a second detection model, and trains the first detection model and the second detection model respectively, and then inputs the disaster image respectively to obtain a first disaster type, a density of people, a duration of the disaster, human behavior, disaster characteristics, a disaster level, and a second disaster type, thereby facilitating analysis of the disaster situation, and further obtains the current disaster type based on the first disaster type and the second disaster type, and combines the analysis of the first detection model and the second detection model to make the judgment of the disaster type more accurate; at the same time, a disaster knowledge graph is constructed, and an emergency response plan is searched in the disaster knowledge graph based on the current disaster type, density of people, duration of the disaster, human behavior, disaster characteristics, and disaster level, so that when a disaster occurs, it can be dealt with in a timely manner to reduce the harmfulness of the disaster. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0028] Figure 1 A method flow chart of a disaster identification and disposal method based on real-time video information and knowledge graph provided by an embodiment of the present invention;

[0029] Figure 2 A schematic diagram of a Slowfast model of a disaster identification and disposal method based on real-time video information and knowledge graph provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the specific implementation method, structure, features and effects of a disaster identification and disposal method based on real-time video information and knowledge graph proposed by the present invention in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures, or characteristics in one or more embodiments may be combined in any suitable form.

[0031] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0032] The following is a detailed description of a specific scheme of a disaster identification and disposal method based on real-time video information and knowledge graph provided by the present invention in conjunction with the accompanying drawings.

[0033] Example:

[0034] The main application scenarios of the present invention are: after a disaster occurs, the analysis of the on-site situation plays a vital role in emergency response and rescue. The existing analysis of the disaster scene is based on intuitive analysis of on-site video. There are problems with the real-time performance of existing video analysis technology in disaster situation assessment, low data upload efficiency and insufficient assessment accuracy. Therefore, how to quickly and accurately analyze video information at the disaster scene, upload data to the command center in real time, and generate professional disposal plans has become an urgent problem to be solved.

[0035] See also Figure 1 , which shows a method flow chart of a disaster identification and disposal method based on real-time video information and knowledge graph provided by an embodiment of the present invention, the method comprising the following steps:

[0036] Step S1, obtaining video data of the disaster area, and preprocessing the video data to obtain disaster images at each moment.

[0037] The present invention relates to the fields of emergency management, computer vision and artificial intelligence, and in particular to the technology of analyzing disaster situations through video information and uploading data in real time, aiming to improve the efficiency and accuracy of information acquisition in disaster response. At the same time, the real-time processing capability of the video analysis system is improved to ensure that the disaster situation can be quickly assessed; the data upload mechanism is optimized to reduce bandwidth usage and improve the efficiency of data upload; the recognition accuracy in complex disaster scenarios is improved to ensure the accuracy of the assessment results; and professional and accurate emergency response plans are generated in combination with knowledge graphs.

[0038] First, it is necessary to determine the disaster area and further collect video data of the disaster area. It should be noted that video data of the disaster area can be collected by drones, surveillance cameras or mobile devices to ensure the diversity and coverage of video sources. It should be noted that the equipment for collecting video data needs to be determined by the implementer based on the actual situation of the disaster area.

[0039] Furthermore, after the video data is collected, it needs to be transmitted. The present invention adopts high-efficiency video coding formats such as H.265 to reduce the transmission bandwidth requirements of the video data (compared with traditional video coding formats such as H.264, more advanced coding methods are adopted, including more efficient intra-frame prediction, inter-frame prediction and entropy coding, which can reduce the bandwidth requirements by about 50% while maintaining the same video quality).

[0040] Furthermore, it is necessary to preprocess the video data, including noise reduction, deblurring, frame difference and other operations, to improve the quality of the video data and reduce the impact of environmental noise on the analysis results. Furthermore, it is necessary to crop the preprocessed video data at 1 frame per second to obtain disaster images. The purpose of such cropping is to be used for labeling. The labeling information of each disaster image includes video ID, frame number (frame), corner coordinates of the human body frame, labels, etc. At the same time, the disaster images form a continuous image set, and the images in the set are arranged in chronological order.

[0041] Step S2, constructing a first detection model and training the first detection model; inputting the disaster image at the current moment into the trained first detection model to obtain the first disaster type, population density and disaster duration.

[0042] After obtaining the video data, it is necessary to build a detection model and further improve the yolov8 model. Specifically, the feature pyramid network (FPN) structure is introduced, which can detect objects at different scales, especially the detection accuracy of small targets (such as trapped people in the distance) is significantly improved; add attention mechanism: the attention mechanism (such as SE module or CBAM module) is applied, so that the model can pay more attention to the salient areas in the disaster scene, thereby improving the detection accuracy and obtaining the first detection model.

[0043] At the same time, in order to increase the accuracy of subsequent detection, it is necessary to enhance the disaster image data, which is aimed at the illumination changes and occlusion problems (such as smoke occlusion) in the disaster scene, and can enhance the robustness of the model in complex scenes. Furthermore, historical disaster image data is obtained, annotated to form a first training set, and the first detection model is trained using the first training set to obtain the first detection model after training.

[0044] Furthermore, the disaster image at the current moment is input into the trained first detection model to obtain the first disaster type, population density and disaster duration, where the disaster types include earthquakes (house damage and collapse), fires (flames), floods, as well as population density and disaster duration; it should be noted that the population density is the ratio of the number of people in the disaster area to the area of ​​the disaster area, and the disaster duration is the time from the start of video data shooting to the current moment.

[0045] Therefore, the first detection model can perform preliminary detection of the disaster area. It should be noted that during the detection process, video data is collected continuously, and the disaster images obtained from the video data are arranged in the image set in chronological order. Therefore, the detection of the disaster area is also continuous. When the previous frame of the disaster image is input into the first detection model, the next input is the next frame of the disaster image, that is, the disaster image is input into the first detection model in chronological order to obtain the corresponding detection results.

[0046] Step S3, construct a second detection model and train the second detection model; input the disaster image at the current moment into the trained second detection model to obtain personnel behavior, disaster characteristics, disaster level and second disaster type.

[0047] In the process of detecting the disaster area, it is not possible to accurately describe the situation at the disaster site by simply identifying the objects in the disaster area. Therefore, it is also necessary to identify the actions of people in the disaster area, as well as the disaster characteristics, etc. Therefore, it is necessary to build a second detection model.

[0048] The second detection model is the Slowfast model with a fully connected layer. The Slowfast model concept was proposed by Facebook AI Research (FAIR). The core idea is to use two paths with different time scales to capture different characteristics of actions when processing videos. The target is separated from the background, the target is captured at a higher frame rate, and the background is captured at a lower frame rate. The focus is on changes in the time dimension, and three-dimensional convolution is used to decouple changes in the time dimension from changes in the spatial dimension.

[0049] The Slowfast algorithm separates the background and the target into a fast path and a slow path. For the slow path, a longer frame interval is used. For the fast path, there are more image inputs, so a simpler convolution operation is used to reduce the computational complexity, because the fast path is less dependent on the information of a single image and considers more changes in the continuous sequence. This design can align the processing speeds of the two paths. The schematic diagram is as follows Figure 2 shown.

[0050] Among them, the convolution kernel size is recorded as {T*S 2 , C}, T time series, S space, C channel, α is the speed ratio, β is the channel ratio;

[0051] The slow path feature matrix is:

[0052] {T,S 2 ,C}

[0053] The fast path feature matrix is:

[0054] {αT,S2 ,βC}

[0055] This convolution method is used to perform multiple rounds of convolution and pooling on the two paths.

[0056] Lateral connection: After each round of convolution, the fast path needs to fuse the extracted action features into the slow path. The fusion strategy adopted is to perform 3D convolution on the feature matrix of the fast path again, output the feature matrix with the same time dimension as the slow path, and then connect it. Finally, the output result is input into the fully connected layer for fusion and output.

[0057] The second detection model uses the R-CNN series of models, including Faster R-CNN, Mask R-CNN and Keypoint R-CNN. Faster R-CNN is a target detection model based on region proposal. It performs target detection through the following steps:

[0058] 1. Feature extraction: Use convolutional neural networks (such as ResNet, VGG) to extract features of the input image.

[0059] 2. Region Proposal Network (RPN): Generates a series of candidate regions (called region proposals) on the extracted feature map. These regions may contain targets.

[0060] 3. Classification and regression: Classify (identify the target category) and regress the bounding box (adjust the bounding box position) for each proposed region.

[0061] Faster R-CNN has very high detection accuracy, but is relatively slow because it involves region proposals.

[0062] Mask R-CNN is an extended version of Faster R-CNN. In addition to performing object detection, it can also perform instance segmentation. Mask R-CNN adds a branch based on Faster R-CNN to generate pixel-level segmentation masks for each detected object.

[0063] In Detectron2, Mask R-CNN is widely used for instance segmentation tasks, which is to distinguish the pixel-level regions of each object in the image. This pixel-level segmentation can be regarded as the separation of the target from the background.

[0064] Detectron2's Mask R-CNN model can generate a pixel-level segmentation mask for each detected object. This segmentation mask separates the object in the image from the background. For example, in an image containing multiple people, Mask R-CNN can not only detect the bounding box of each person, but also generate the precise outline of each person (i.e. instance segmentation), thereby distinguishing people from the background. Semantic segmentation: Detectron2 also supports semantic segmentation, which assigns a category label to each pixel in the image. Semantic segmentation can distinguish between objects and background in an image at the pixel level, such as distinguishing scene elements such as "road", "building", and "sky".

[0065] After the second detection model is constructed, the historical disaster image data is annotated to form a second training set, and the second detection model is trained using the second training set to obtain the trained second detection model. The disaster image at the current moment is input into the trained second detection model to obtain personnel behavior, disaster characteristics, disaster level and second disaster type.

[0066] Human behavior refers to the actions of people in the disaster area, such as running, walking, calling for help, etc. Disaster characteristics refer to the characteristics of the disaster, such as the flood water level, flow rate, coverage area, flame intensity, diffusion speed, smoke concentration, degree of building collapse, etc.

[0067] The disaster levels are divided into severe, relatively severe, generally severe, minor and no disaster. After obtaining the disaster level, it is necessary to intercept high-resolution video clips according to the current disaster level and upload them for further analysis by the command center. If the disaster level is above minor, the current moment is used as the middle moment of the intercepted video clip to intercept and obtain key video clips, wherein the length of the key video clip is a preset duration. Preferably, in the embodiment of the present invention, the preset duration is 1 minute, that is, the clips 30 seconds before the current moment and 30 seconds after the current moment are intercepted as key video clips. The length of the intercepted video clip can be determined by the implementer according to the actual situation.

[0068] According to the disaster level, only key video clips are uploaded to reduce the upload amount of original video data, and the transmission efficiency is further optimized through differential coding, data compression and other technologies. The present invention also adopts a graded upload mechanism, and low-resolution videos and simplified assessment reports are uploaded for disaster levels below minor, including minor.

[0069] In this way, the personnel behavior, disaster characteristics, disaster level and second disaster type at the current moment can be obtained. It should be noted that since the detection of the disaster area is monitored in real time, detection will be carried out at every moment. In this embodiment, only the current moment is used as an example for explanation.

[0070] Step S4, obtaining the current disaster type according to the first disaster type and the second disaster type; constructing a disaster knowledge graph, and searching for an emergency response plan in the disaster knowledge graph according to the current disaster type, population density, disaster duration, personnel behavior, disaster characteristics, and disaster level.

[0071] Due to model errors and video quality problems caused by the on-site environment, the first detection model and the second detection model may have errors in detecting disaster types. Therefore, it is necessary to comprehensively judge the detection results of the two models and obtain the current disaster type based on the first disaster type and the second disaster type. Specifically, when the first disaster type and the second disaster type are the same, the current disaster type is any one of the first disaster type and the second disaster type; when the first disaster type and the second disaster type are different, the predicted probabilities of the two models are obtained for weighted averaging, the final probability of each category is calculated, and the highest category is selected as the current disaster type. When performing weighted averaging, the weights of the first detection model and the second detection model are adjusted according to the performance of the two models on the validation set. In this way, the current disaster type can be obtained.

[0072] Furthermore, after obtaining various information about the disaster area, it is necessary to obtain the next emergency response plan based on this information and build a disaster knowledge graph. Specifically, a large amount of text data related to emergency management, such as disaster accident reports, emergency plans, emergency drill reports, and emergency regulations, is collected; the collected text data is cleaned, deduplicated, formatted, and other preprocessing operations are performed to ensure the validity and standardization of the data; based on the preprocessed data, key information related to the disaster accident is extracted, including but not limited to the type of accident, accident level, accident responsibility division, accident cause, accident impact, emergency response measures, etc.; the extracted key information is structured, and natural language processing technology is used to annotate the information to establish the association between entities (such as accident type, accident location, accident time) and relationships (such as accident cause, accident consequences); a knowledge graph in the emergency field is generated to comprehensively and systematically display the causal relationship, processing flow, and other key information of various disaster accidents. Finally, the knowledge graph is used to ensure the accuracy of the content of the question-answering model, and accurate data is used to answer questions to generate emergency response plans. At the same time, it should be noted that the knowledge graph is updated in real time. After the knowledge graph is initially constructed, it needs to be updated in real time over time.

[0073] In the disaster knowledge graph, each historical disaster case has a corresponding index, which includes disaster time, disaster location, disaster type, population density, disaster duration, personnel behavior, disaster characteristics and disaster level. The index of historical disaster cases can be used as an index vector; obtain the current disaster type, population density, disaster duration, personnel behavior, disaster characteristics, disaster level, disaster time and disaster location to form a disaster feature vector; calculate the distance between the disaster feature vector and each index vector respectively, and obtain the historical disaster case corresponding to the nearest index vector. According to the causal relationship and processing flow of the historical case, the disaster assessment (casualties, economic losses, accident causes, accident responsible parties, etc.) and emergency response plan (resource scheduling, personnel scheduling, etc.) are generated. At this point, the situation at the disaster site can be obtained, and the corresponding emergency response plan can be obtained to facilitate the handling of the disaster and reduce the harm of the disaster.

[0074] The present invention can integrate multi-source video data, comprehensively grasp the multi-angle information of the disaster scene, and provide more comprehensive disaster assessment results. By building a knowledge graph based on historical reports, drill data and other information in the emergency field, the model can accurately grasp the professional knowledge in the field of emergency management, provide more targeted answers, and provide guidance for disaster disposal.

[0075] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The above is a description of a specific embodiment of this specification. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0076] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0077] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A disaster identification and disposal method based on real-time video information and knowledge graph, characterized in that: The method includes: Obtain video data of the disaster area, and pre-process the video data to obtain disaster images at each moment; Constructing a first detection model and training the first detection model; inputting the disaster image at the current moment into the trained first detection model to obtain the first disaster type, population density and disaster duration; Constructing a second detection model and training the second detection model; inputting the disaster image at the current moment into the trained second detection model to obtain personnel behavior, disaster characteristics, disaster level and second disaster type; The current disaster type is obtained based on the first disaster type and the second disaster type; a disaster knowledge graph is constructed, and emergency response plans are found in the disaster knowledge graph based on the current disaster type, population density, disaster duration, personnel behavior, disaster characteristics and disaster level.

2. A disaster identification and handling method based on real-time video information and knowledge graph according to claim 1, characterized in that: The preprocessing of the video data to obtain disaster images at each moment includes: The preprocessing of video data includes noise reduction, deblurring and frame difference; the preprocessed video data is cropped to obtain the disaster images at each moment.

3. A disaster identification and handling method based on real-time video information and knowledge graph according to claim 1, characterized in that: The step of constructing a first detection model and training the first detection model includes: The yolov8 model is improved by introducing a pyramid network structure and adding an attention mechanism. The improved yolov8 model is used as the first detection model. Historical disaster image data is obtained and annotated to form a first training set. The first training set is used to train the first detection model to obtain the first detection model after training.

4. A disaster identification and handling method based on real-time video information and knowledge graph according to claim 1, characterized in that: The step of constructing a second detection model and training the second detection model includes: The second detection model is a Slowfast model with an added fully connected layer. The historical disaster image data is annotated to form a second training set, and the second detection model is trained using the second training set to obtain a trained second detection model.

5. A disaster identification and handling method based on real-time video information and knowledge graph according to claim 1, characterized in that: The first disaster type, population density and disaster duration include: The first disaster type includes earthquakes, fires and floods; the population density is the ratio of the number of people in the disaster area to the area of ​​the disaster area; and the disaster duration is the time from the start of video data shooting to the current moment.

6. A disaster identification and handling method based on real-time video information and knowledge graph according to claim 1, characterized in that: After obtaining the behavior of personnel, disaster characteristics, disaster level and second disaster type, it also includes: The disaster levels are divided into severe, relatively severe, generally severe, minor and no disaster; for disaster levels above minor, the current moment is used as the middle moment of the video clip to obtain the key video clip, where the length of the key video clip is the preset length; the key video clip is uploaded to the command center.

7. The method for disaster identification and handling based on real-time video information and knowledge graph according to claim 1, characterized in that: The obtaining of the current disaster type according to the first disaster type and the second disaster type comprises: When the first disaster type and the second disaster type are the same, the current disaster type is any one of the first disaster type and the second disaster type; when the first disaster type and the second disaster type are different, the predicted probabilities of the two models are obtained and weighted averaged to calculate the final probability of each category, and the category with the highest final probability is selected as the current disaster type.

8. The method for disaster identification and handling based on real-time video information and knowledge graph according to claim 1 is characterized in that: The construction of the disaster knowledge graph includes: Obtain historical disaster accident reports, emergency plans, emergency drill reports and emergency regulations, and record them as text data; clean, deduplicate and format the text data; extract key information related to the disaster accident based on the processed text data, perform structured processing on the extracted key information, use natural language processing technology to annotate the information, establish associations between entities and relationships, generate a disaster knowledge graph, and update the disaster knowledge graph in real time.

9. The method for disaster identification and handling based on real-time video information and knowledge graph according to claim 1, characterized in that: The search for emergency response plans in the disaster knowledge graph based on the current disaster type, population density, disaster duration, personnel behavior, disaster characteristics and disaster level includes: Each historical disaster case in the disaster knowledge graph has a corresponding index, which includes disaster time, disaster location, disaster type, population density, disaster duration, personnel behavior, disaster characteristics and disaster level; the index of the historical disaster case is used as an index vector; the previous disaster type, population density, disaster duration, personnel behavior, disaster characteristics, disaster level, disaster time and disaster location at the current moment are obtained to form a disaster feature vector; the distance between the disaster feature vector and each index vector is calculated respectively; the historical disaster case corresponding to the nearest index vector is obtained, and a disaster assessment and emergency response plan is generated based on key information such as the causal relationship and processing flow of the historical disaster case.

Images