Multi-stage collaborative analysis multi-modal video fire intelligent identification method and system

Abstract

The present application relates to the technical field of fire identification, in particular to a multi-stage collaborative analysis multi-modal video fire intelligent identification method and system. It comprises the following steps: S1, concurrent scheduling of video based on streaming media distribution and heterogeneous acceleration; S2, preliminary fire candidate region identification based on a target detection model; S3, video sequence consistency verification based on flame physical evolution characteristics; S4, scene fire behavior semantic discrimination based on a multi-modal large model. Compared with the prior art, it can operate stably under the condition of multi-channel video high-concurrent access, and comprehensively utilize video timing characteristics and high-level semantic understanding ability to improve detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of fire identification technology, specifically to a multimodal video fire intelligent identification method and system based on multi-stage collaborative analysis. Background Technology

[0002] With the large-scale deployment of video surveillance systems in urban security, industrial production, and public places, video-based automatic fire identification technology is gradually becoming an important supplement to traditional smoke and heat alarms. Compared to single-point sensors, video fire identification can provide more intuitive scene information and richer fire characteristics, giving it a clear advantage in early fire detection and decision support.

[0003] However, existing video fire detection technologies still face several technical bottlenecks in practical applications. Most current solutions rely on a single deep learning model to detect flames or smoke targets in a single frame of image. While this approach can achieve some detection results in controlled experimental environments, it has significant limitations in real-world, complex scenarios. Firstly, single-frame images lack temporal information, making it difficult to effectively distinguish between instantaneous light spots, light reflections, screen displays, and other non-fire-related interference factors, easily leading to frequent false alarms. Secondly, in application scenarios requiring simultaneous access to a large number of video streams, traditional solutions often struggle to balance detection accuracy and real-time processing capabilities. As the number of videos increases, system latency and frame loss become increasingly prominent.

[0004] Furthermore, existing technologies generally simplify fire identification to a binary judgment of "whether a fire has occurred," lacking the ability to comprehensively analyze the authenticity of the fire, its evolution trend, and the risk level of the scene, and are unable to distinguish between normal fire use behavior and abnormal fire behavior.

[0005] In existing intelligent fire identification and early warning technologies for video, various solutions based on image analysis, temporal modeling, or multimodal fusion have been proposed. For example, CN121438483A discloses an early fire warning system based on AI image recognition. By analyzing the monitoring video stream in real time and using a pre-trained fire identification model to extract visual features related to flames or smoke, it can provide early warning in the initial stage of a fire. This type of solution achieves the fire detection function of "video + AI," but its technical focus is mainly on fire target identification under single-frame or weak temporal conditions. Whether a fire is established basically depends on the output result of a single model, lacking a multi-dimensional verification mechanism for the evolution characteristics of flames over time. It is difficult to effectively distinguish between instantaneous light spots, reflections, or normal fire use behavior, resulting in a high risk of false alarms.

[0006] For example, CN121354267A proposes an intelligent early warning method based on multi-source fire sensor data and video analysis. It extracts spatiotemporal joint features through a multimodal fusion model and a spatiotemporal convolutional network, analyzes the fire spread trend, and generates risk scores and heat maps. While this type of technology improves fire identification by enhancing model fusion depth and spatiotemporal feature modeling capabilities, its technical approach still revolves around model probability output. The judgment of fire risk mainly relies on network learning results, without introducing explicit temporal consistency constraints based on the physical evolution of flames, nor does it differentiate the rationality of fire behavior in specific scenarios from a decision-making perspective.

[0007] Furthermore, CN119295996A discloses a multimodal fire detection system for energy storage power stations based on image and time-series analysis. This system combines short-term and long-term fire decision-making, utilizing an LSTM network to perform dynamic behavior analysis on continuous video frames to improve fire detection accuracy. While this scheme improves the stability of fire identification by introducing time-series modeling, its time-series analysis is essentially still based on neural network-based temporal feature learning. It does not perform multidimensional modeling of physical evolution characteristics such as flame flickering, area changes, and spread direction. Additionally, its application scenarios are relatively fixed, failing to consider the problem of distinguishing between normal fire use behavior and abnormal fire behavior under complex multi-scenario conditions.

[0008] In summary, existing technologies generally suffer from the following shortcomings: 1. The validity of a fire largely depends on the output of a single model or a single temporal network, lacking a multi-stage, veto-rejection collaborative judgment mechanism, making it difficult to effectively suppress false alarms in scenarios with complex lighting and frequent normal fire use; 2. The utilization of video temporal sequences mostly remains at the model level of temporal feature learning, without introducing explicit consistency verification based on the physical evolution of flames, resulting in limited reliability in judging the authenticity of fires; 3. In practical application scenarios with concurrent access to multiple video streams, existing solutions rarely design the overall system architecture for stable access, scheduling, and inference efficiency of large-scale video streams, making it difficult to balance concurrency scale and real-time requirements.

[0009] In view of the above problems, there is an urgent need for a video fire identification method that can operate stably under high-concurrency access of multiple video streams and comprehensively utilize video temporal features and high-level semantic understanding capabilities, so as to improve detection accuracy and achieve intelligent assessment of fire risk levels. Summary of the Invention

[0010] To overcome the shortcomings of existing technologies, this invention provides a video fire identification method that can operate stably under high-concurrency access of multiple video streams and comprehensively utilizes video temporal features and high-level semantic understanding capabilities to improve detection accuracy.

[0011] To achieve the above objectives, a multi-modal video fire intelligent identification method based on multi-stage collaborative analysis is designed, including the following steps: S1, Video Concurrency Scheduling Based on Streaming Media Distribution and Heterogeneous Acceleration: Acquire video streams from at least one video acquisition device. The video streams are uniformly accessed through the streaming media access module. The streaming media access module performs real-time decapsulation and format unification processing on the accessed video streams and sends the video data in the form of a stream into a pipeline-based video processing framework. The video processing framework decodes, buffers, and performs batch scheduling of the video data. S2, Preliminary fire candidate region identification based on target detection model: Input the video frame obtained in step S1 into the first fire feature detection model to detect flame targets, smoke targets and bright abnormal areas in the video frame to obtain at least one fire candidate region; S3, Video sequence consistency verification based on flame physical evolution characteristics: For the fire candidate area output in step S2, cross-frame target association and tracking are performed on the candidate area in consecutive video frames, and a multi-dimensional evolution feature sequence of the candidate target in the time dimension is constructed to verify its temporal consistency. S4, Scene fire behavior semantic discrimination based on multimodal large model: Input the local image information corresponding to the suspected fire target verified by step S3 and the global image of the scene where it is located into the multimodal semantic analysis model to perform high-level semantic understanding of the fire situation.

[0012] In step S1, the video acquisition device includes a visible light camera and / or an infrared camera.

[0013] In step S1, the streaming media access module adopts a session-management-based streaming media distribution mechanism to independently maintain and manage the status of each video stream.

[0014] In step S2, the fire candidate region includes the spatial coordinates of the candidate target in the image, the target category, and the corresponding confidence information.

[0015] In step S3, the timing consistency verification includes at least the following analysis: S31, Flame jitter feature sequence analysis: Analyze the high-frequency irregular deformation of the candidate region boundary in consecutive frames to determine whether it conforms to the random jitter characteristics of a naturally burning flame; S32, Flame Area Change Sequence Analysis: Calculate the change trend of candidate region pixel area over time to determine whether there is continuous growth, fluctuating growth or stable expansion of flame evolution behavior; S33, Flame spread direction and speed analysis: Analyze the spatial expansion direction and expansion rate of the candidate area to determine whether it conforms to the physical characteristics of flame spreading upward or outward. S34, Brightness and Color Coupling Change Analysis: Analyze the synchronous relationship between brightness changes and color channel distribution changes in candidate regions to distinguish between combustion flames and non-combustion light sources.

[0016] In step S3, when multiple time series exhibit consistent flame evolution characteristics within the same time window, the candidate target is determined to be a temporally stable suspected fire target; when the candidate target appears only in a single frame or a very short time, or its time series characteristics do not conform to the physical evolution law of flame, it is determined to be an interfering target and is eliminated.

[0017] In step S4, the multimodal semantic analysis model, based on the fusion results of visual features, scene structure features, and contextual semantic information, determines the rationality and compliance of the fire behavior in the current scene, specifically including: S41, Identify the scene type where the flame target is located; S42, combine scene semantics to determine whether the currently detected flame behavior belongs to normal fire use behavior allowed by the scene, or to abnormal fire risk behavior; S43, when the scene does not allow open flames or the flame behavior is obviously abnormal, the suspected fire target is determined as a highly reliable fire event.

[0018] Step S4 further includes: based on the completion of fire behavior semantic discrimination, the multimodal semantic analysis model further generates fire description information, and assesses the fire risk level by combining the flame size, duration and potential impact range.

[0019] It also includes step S5, fire classification early warning and alarm information output: based on the fire semantic description information and corresponding risk level obtained in step S4, generate classification early warning or alarm information, and output at least including alarm time, alarm location, fire risk level and corresponding image or video clip.

[0020] To achieve the above objectives, a system for a multi-modal video fire intelligent identification method based on multi-stage collaborative analysis is designed, including: Multi-stream video access and management module: used to access multiple monitoring video streams, perform video session management and stream status maintenance, control frame rate and time synchronization, and stably output continuous video frames; Batch processing model acceleration computation and scheduling module: used for batch aggregation of video frames across video streams, decoding and image preprocessing, parallel scheduling of heterogeneous computing resources, and high-throughput model inference output; Fire candidate area preliminary identification module: used for single-frame flame or smoke target detection, highlighting abnormal areas identification, and outputting fire candidate area and location information; The temporal consistency analysis module based on flame physical evolution is used for cross-frame target association and tracking, fire analysis of flame jitter characteristics, analysis of area change and spread trend, elimination of instantaneous interference and non-fire targets, and output of temporally stable suspected fire targets. The scene fire semantic discrimination module based on a multimodal large model integrates local flame features with global scene information to identify scene type and fire behavior attributes, determine the rationality and compliance of fire behavior, and output a semantic description of the fire situation.

[0021] Compared with the prior art, the present invention has the following advantages: 1. Maintaining stable real-time fire detection capability under ultra-large-scale concurrent video access conditions: By introducing a concurrent scheduling architecture that combines a multi-stream management module with a batch processing model to accelerate computation, the system achieves session-level management, load balancing scheduling, and cross-video stream batch inference processing for multiple video streams. This enables the system to maintain a predetermined frame rate output even when the scale of video access increases significantly, avoiding frame blocking or processing delay accumulation caused by the increase in the number of videos.

[0022] 2. Significantly reduce the false alarm rate caused by instantaneous light source changes, reflections, or flickering: By using a multi-dimensional video temporal consistency analysis method based on the physical evolution characteristics of flames, the time series of fire candidate targets in continuous video frames are jointly verified, including jitter characteristics, area changes, and spread trends. Only when multiple temporal features simultaneously meet the flame evolution law are they identified as valid fire targets, thereby effectively suppressing false judgments caused by single-frame or short-term interference.

[0023] 3. Achieve intelligent and scenario-based assessment of fire risk levels, and improve the accuracy and practicality of alarm results: By introducing a scenario fire behavior semantic discrimination mechanism based on a multimodal large model, the rationality and compliance of detected flame behavior in the current scenario are analyzed, enabling the system to distinguish between normal fire use behavior and abnormal fire behavior, and on this basis, generate fire level assessment results that are more in line with the actual risk level.

[0024] 4. Strong robustness and adaptability in daytime, nighttime and complex lighting environments: Through a multi-stage collaborative analysis mechanism, single-frame visual detection, video sequence temporal verification and high-level semantic understanding are organically combined, so that the system no longer relies on a single lighting condition or a single feature judgment, thus maintaining stable fire identification performance under backlight, low light at night or complex background conditions.

[0025] 5. Provide more decision-making value alarm information for fire emergency response and reduce the cost of manual intervention: By introducing fire risk level and scene semantic description information in the alarm output stage, the alarm result includes not only "whether a fire has occurred", but also the severity of the fire and potential risk assessment, thereby providing a clear basis for subsequent emergency response and reducing the time and manpower costs caused by manual review and misjudgment. Detailed Implementation

[0026] The present invention will now be further described.

[0027] The multi-modal video fire intelligent identification method with multi-stage collaborative analysis in this embodiment includes the following steps: S1, Video Concurrency Scheduling Based on Streaming Media Distribution and Heterogeneous Acceleration: Acquire video streams from at least one video acquisition device. The video streams are uniformly accessed through the streaming media access module. The streaming media access module performs real-time decapsulation and format unification processing on the accessed video streams and sends the video data in the form of a stream into a pipeline-based video processing framework. The video processing framework decodes, buffers, and performs batch scheduling of the video data. S2, Preliminary fire candidate region identification based on target detection model: Input the video frame obtained in step S1 into the first fire feature detection model to detect flame targets, smoke targets and bright abnormal areas in the video frame to obtain at least one fire candidate region; S3, Video sequence consistency verification based on flame physical evolution characteristics: For the fire candidate area output in step S2, cross-frame target association and tracking are performed on the candidate area in consecutive video frames, and a multi-dimensional evolution feature sequence of the candidate target in the time dimension is constructed to verify its temporal consistency. S4, Scene fire behavior semantic discrimination based on multimodal large model: Input the local image information corresponding to the suspected fire target verified by step S3 and the global image of the scene where it is located into the multimodal semantic analysis model to perform high-level semantic understanding of the fire situation.

[0028] In step S1, the video acquisition device includes a visible light camera and / or an infrared camera. The streaming media access module adopts a session-management-based streaming media distribution mechanism to independently maintain and manage the status of each video stream.

[0029] During video processing, the system, based on a heterogeneous computing architecture, assigns video decoding, image preprocessing, and fire feature detection tasks to the graphics processing unit and the general processing unit for collaborative execution. It also uses batch processing and pipeline parallel mechanisms to uniformly schedule and process video frames from different video streams in parallel.

[0030] Through the synergistic effect of the streaming media distribution mechanism, pipelined parallel processing mechanism, and heterogeneous computing scheduling mechanism, the system can simultaneously access and stably process hundreds of video streams on a single computing node or distributed computing nodes. This enables real-time capture and fire detection analysis of up to hundreds, such as 512, video streams, while maintaining a predetermined frame rate output even as the number of video streams increases, avoiding problems such as frame blocking, frame loss, or accumulated processing latency. This concurrent scheduling architecture provides a continuous, stable, and scalable video data foundation for subsequent video sequence-based fire identification, temporal consistency analysis, and multimodal semantic discrimination.

[0031] In step S2, the fire candidate region includes the spatial coordinates of the candidate target in the image, the target category, and the corresponding confidence information. Step S2 is used to quickly filter video images at the single-frame level, sending only local areas with potential fire characteristics into the subsequent multi-stage analysis process. This significantly reduces the overall computational load while ensuring detection coverage and improves the system's real-time processing capability under multi-channel video conditions.

[0032] In step S3, the timing consistency verification includes at least the following analysis: S31, Flame jitter feature sequence analysis: Analyze the high-frequency irregular deformation of the candidate region boundary in consecutive frames to determine whether it conforms to the random jitter characteristics of a naturally burning flame; S32, Flame Area Change Sequence Analysis: Calculate the change trend of candidate region pixel area over time to determine whether there is continuous growth, fluctuating growth or stable expansion of flame evolution behavior; S33, Flame spread direction and speed analysis: Analyze the spatial expansion direction and expansion rate of the candidate area to determine whether it conforms to the physical characteristics of flame spreading upward or outward. S34, Brightness and Color Coupling Change Analysis: Analyze the synchronous relationship between brightness changes and color channel distribution changes in candidate regions to distinguish between combustion flames and non-combustion light sources.

[0033] Step S3 effectively reduces false alarms caused by non-fire factors such as light reflection, instantaneous bright spots, and screen display by introducing a multi-time series consistency verification mechanism based on flame physical evolution characteristics.

[0034] In step S3, when multiple time series exhibit consistent flame evolution characteristics within the same time window, the candidate target is determined to be a temporally stable suspected fire target; when the candidate target appears only in a single frame or a very short time, or its time series characteristics do not conform to the physical evolution law of flame, it is determined to be an interfering target and is eliminated.

[0035] In step S4, the multimodal semantic analysis model, based on the fusion results of visual features, scene structural features, and contextual semantic information, determines the rationality and compliance of the fire behavior in the current scenario, specifically including: S41, identify the scene type where the flame target is located, such as industrial production scene, restaurant kitchen scene, warehouse scene or public space scene; S42, combine scene semantics to determine whether the currently detected flame behavior belongs to normal fire use behavior allowed by the scene, or to abnormal fire risk behavior; S43, when the scene does not allow open flames or the flame behavior is obviously abnormal, the suspected fire target is determined as a highly reliable fire event.

[0036] Step S4 also includes: based on the completion of fire behavior semantic discrimination, the multimodal semantic analysis model further generates fire description information, and assesses the fire risk level by combining the flame size, duration and potential impact range. The risk level includes at least one of no risk, low risk, medium risk and high risk.

[0037] It also includes step S5, fire classification early warning and alarm information output: based on the fire semantic description information and corresponding risk level obtained in step S4, generate classification early warning or alarm information, and output at least including alarm time, alarm location, fire risk level and corresponding image or video clip.

[0038] Under different risk levels, the system triggers different levels of early warning strategies to realize a hierarchical response mechanism from prompts and warnings to alarms, thereby avoiding frequent false alarms in low-risk scenarios and ensuring that high-risk fire events can be alarmed in a timely and accurate manner.

[0039] To achieve the above objectives, a system for a multi-modal video fire intelligent identification method based on multi-stage collaborative analysis is designed, including: Multi-stream video access and management module: used to access multiple monitoring video streams, perform video session management and stream status maintenance, control frame rate and time synchronization, and stably output continuous video frames; Batch processing model acceleration computation and scheduling module: used for batch aggregation of video frames across video streams, decoding and image preprocessing, parallel scheduling of heterogeneous computing resources, and high-throughput model inference output; Fire candidate area preliminary identification module: used for single-frame flame or smoke target detection, highlighting abnormal areas identification, and outputting fire candidate area and location information; The temporal consistency analysis module based on flame physical evolution is used for cross-frame target association and tracking, fire analysis of flame jitter characteristics, analysis of area change and spread trend, elimination of instantaneous interference and non-fire targets, and output of temporally stable suspected fire targets. The scene fire semantic discrimination module based on a multimodal large model integrates local flame features with global scene information to identify scene type and fire behavior attributes, determine the rationality and compliance of fire behavior, and output fire semantic description and risk assessment results.

[0040] In practical use, it also includes a fire risk classification and early warning output module, which determines the fire risk level, classifies and issues early warnings or triggers alarm strategies, and outputs alarm information and images or video clips.

[0041] This invention provides a multi-stage collaborative intelligent video fire identification and processing architecture. It divides the fire identification process into several stages: rapid fire candidate screening, temporal consistency verification based on flame physical evolution characteristics, and fire situation discrimination based on multimodal semantic understanding. A hierarchical screening and rejection mechanism is established between each stage, thereby significantly reducing the false alarm rate while ensuring real-time performance. Multi-dimensional video temporal consistency analysis is performed based on flame physical evolution characteristics. This involves collaboratively analyzing multiple time series of fire candidate targets in continuous video frames, including jitter characteristics, area changes, spatial spread direction, and brightness and color coupling changes, to distinguish real fire targets from instantaneous light spots, reflections, or other non-fire interference targets. The rationality and compliance of scene fire behavior are judged based on a multimodal large model. By fusing visual features of the flame target, scene structural features, and contextual semantic information, the rationality of detected flame behavior in the current scene is determined, and the judgment result serves as an important decision-making basis for fire establishment and risk assessment. The architecture of a concurrent fire detection system for ultra-large-scale video access includes a multi-stream management module and a batch processing model acceleration module. Through session-level management, load balancing scheduling, and cross-stream batch inference processing of multiple video streams, it achieves stable access and real-time fire detection and analysis for hundreds of video streams. Based on fire detection results and tiered early warning strategies, it makes joint decisions, classifies fire situations according to multi-stage analysis results and multi-modal semantic risk assessment results, and triggers differentiated early warning or alarm strategies at different risk levels to improve the availability and response efficiency of alarm results.

[0042] This invention does not simply improve the detection effect by improving the fire identification model or enhancing the depth of multi-source data fusion. Instead, it proposes a multi-stage collaborative intelligent fire identification and hierarchical early warning method for high-concurrency multi-channel video scenarios, starting from three levels: system architecture, fire establishment logic, and decision-making hierarchy.

[0043] This invention constructs a concurrent processing architecture, including a multi-stream management module and a batch processing model acceleration module, to achieve stable access and unified scheduling of hundreds of video streams. While ensuring overall real-time processing capabilities, it provides continuous and reliable video sequence input for subsequent fire analysis. Based on this, the invention divides the fire identification process into multiple stages: rapid fire candidate screening, temporal consistency verification based on flame physical evolution characteristics, and scene fire behavior discrimination based on multimodal semantic understanding. A collaborative decision-making mechanism of progressive confirmation or rejection is formed between each stage.

[0044] Furthermore, this invention does not merely determine whether there is a flame or smoke target in the video, but rather analyzes the physical evolution characteristics such as flame jitter, area change trend and spread direction, and combines them with scene semantic information to determine the rationality and compliance of fire behavior in the current scene, thereby distinguishing between normal fire use behavior and abnormal fire risk events, and generating graded early warning results accordingly.

[0045] Compared with existing fire early warning technologies that rely on single-model judgment, temporal network learning, or multi-source information fusion, this invention proposes a significantly different technical solution in terms of ultra-large-scale video concurrent processing capabilities, multi-stage veto-fire establishment logic, and fire behavior discrimination mechanism at the scene semantic level. Its technical ideas and protection priorities have clear innovative boundaries and significant differences.

Claims

1. A multi-modal video fire intelligent identification method based on multi-stage collaborative analysis, characterized in that: Includes the following steps: S1, Video Concurrency Scheduling Based on Streaming Media Distribution and Heterogeneous Acceleration: Acquire video streams from at least one video acquisition device. The video streams are uniformly accessed through the streaming media access module. The streaming media access module performs real-time decapsulation and format unification processing on the accessed video streams and sends the video data in the form of a stream into a pipeline-based video processing framework. The video processing framework decodes, buffers, and performs batch scheduling of the video data. S2, Preliminary fire candidate region identification based on target detection model: Input the video frame obtained in step S1 into the first fire feature detection model to detect flame targets, smoke targets and bright abnormal areas in the video frame to obtain at least one fire candidate region; S3, Video sequence consistency verification based on flame physical evolution characteristics: For the fire candidate area output in step S2, cross-frame target association and tracking are performed on the candidate area in consecutive video frames, and a multi-dimensional evolution feature sequence of the candidate target in the time dimension is constructed to verify its temporal consistency. S4, Scene fire behavior semantic discrimination based on multimodal large model: Input the local image information corresponding to the suspected fire target verified by step S3 and the global image of the scene where it is located into the multimodal semantic analysis model to perform high-level semantic understanding of the fire situation.

2. The multi-modal video fire intelligent identification method based on multi-stage collaborative analysis according to claim 1, characterized in that: In step S1, the video acquisition device includes a visible light camera and / or an infrared camera.

3. The multi-modal video fire intelligent identification method based on multi-stage collaborative analysis according to claim 1, characterized in that: In step S1, the streaming media access module adopts a session-management-based streaming media distribution mechanism to independently maintain and manage the status of each video stream.

4. The multi-modal video fire intelligent identification method based on multi-stage collaborative analysis according to claim 1, characterized in that: In step S2, the fire candidate region includes the spatial coordinates of the candidate target in the image, the target category, and the corresponding confidence information.

5. The multi-modal video fire intelligent identification method based on multi-stage collaborative analysis according to claim 1, characterized in that: In step S3, the timing consistency verification includes at least the following analysis: S31, Flame jitter feature sequence analysis: Analyze the high-frequency irregular deformation of the candidate region boundary in consecutive frames to determine whether it conforms to the random jitter characteristics of a naturally burning flame; S32, Flame Area Change Sequence Analysis: Calculate the change trend of candidate region pixel area over time to determine whether there is continuous growth, fluctuating growth or stable expansion of flame evolution behavior; S33, Flame spread direction and speed analysis: Analyze the spatial expansion direction and expansion rate of the candidate area to determine whether it conforms to the physical characteristics of flame spreading upward or outward. S34, Brightness and Color Coupling Change Analysis: Analyze the synchronous relationship between brightness changes and color channel distribution changes in candidate regions to distinguish between combustion flames and non-combustion light sources.

6. The multi-modal video fire intelligent identification method based on multi-stage collaborative analysis according to claim 1 or 5, characterized in that: In step S3, when multiple time series exhibit consistent flame evolution characteristics within the same time window, the candidate target is determined to be a temporally stable suspected fire target; when the candidate target appears only in a single frame or a very short time, or its time series characteristics do not conform to the physical evolution law of flame, it is determined to be an interfering target and is eliminated.

7. The multi-modal video fire intelligent identification method based on multi-stage collaborative analysis according to claim 1, characterized in that: In step S4, the multimodal semantic analysis model, based on the fusion results of visual features, scene structure features, and contextual semantic information, determines the rationality and compliance of the fire behavior in the current scene, specifically including: S41, Identify the scene type where the flame target is located; S42, combine scene semantics to determine whether the currently detected flame behavior belongs to normal fire use behavior allowed by the scene, or to abnormal fire risk behavior; S43, when the scene does not allow open flames or the flame behavior is obviously abnormal, the suspected fire target is determined as a highly reliable fire event.

8. The multi-modal video fire intelligent identification method based on multi-stage collaborative analysis according to claim 1 or 7, characterized in that: Step S4 further includes: based on the completion of fire behavior semantic discrimination, the multimodal semantic analysis model further generates fire description information, and assesses the fire risk level by combining the flame size, duration and potential impact range.

9. The multi-modal video fire intelligent identification method based on multi-stage collaborative analysis according to claim 1, characterized in that: It also includes step S5, fire classification early warning and alarm information output: based on the fire semantic description information and corresponding risk level obtained in step S4, generate classification early warning or alarm information, and output at least including alarm time, alarm location, fire risk level and corresponding image or video clip.

10. A multi-modal video fire intelligent recognition system based on multi-stage collaborative analysis, used to execute the multi-modal video fire intelligent recognition method based on multi-stage collaborative analysis as described in any one of claims 1-9, characterized in that: include: Multi-stream video access and management module: used to access multiple monitoring video streams, perform video session management and stream status maintenance, control frame rate and time synchronization, and stably output continuous video frames; Batch processing model acceleration computation and scheduling module: used for batch aggregation of video frames across video streams, decoding and image preprocessing, parallel scheduling of heterogeneous computing resources, and high-throughput model inference output; Fire candidate area preliminary identification module: used for single-frame flame or smoke target detection, highlighting abnormal areas identification, and outputting fire candidate area and location information; The temporal consistency analysis module based on flame physical evolution is used for cross-frame target association and tracking, fire analysis of flame jitter characteristics, analysis of area change and spread trend, elimination of instantaneous interference and non-fire targets, and output of temporally stable suspected fire targets. The scene fire semantic discrimination module based on a multimodal large model integrates local flame features with global scene information to identify scene type and fire behavior attributes, determine the rationality and compliance of fire behavior, and output a semantic description of the fire situation.