Non-contact visual recognition method and system for heat release rate in tunnel fires

Abstract

This invention relates to the fields of computer vision and fire safety technology, specifically a non-contact visual recognition method and system for the heat release rate of tunnel fires. The method includes conducting standard combustion experiments, constructing a fire perception and location model, a quantitative mapping model between the true value of the flame's three-dimensional morphology and the true value of its heat release rate, and a sparse view three-dimensional reconstruction model; acquiring tunnel monitoring video, locating the flame area using the fire perception and location model, and, based on the horizontal distance between the flame and the camera, rotating a camera downstream of the flame to capture images; acquiring flame images from different perspectives, performing standardization processing, obtaining a predicted three-dimensional voxel mesh of the flame using the sparse view three-dimensional reconstruction model, extracting the true value of the flame's three-dimensional morphology, and obtaining an estimated value of the heat release rate using the quantitative mapping model. This solution monitors and locates fires, performs non-contact, real-time, and accurate prediction of the heat release rate of combustion to quantify the fire scale, and is low-cost and highly practical.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and fire safety technology, specifically to a non-contact visual recognition method and system for the heat release rate of tunnel fires. Background Technology

[0002] Highway tunnels are enclosed structures, and once a fire breaks out, it can spread rapidly with extremely serious consequences. Heat release rate is a core physical parameter for measuring fire intensity and determining emergency response levels. Current tunnel fire monitoring mainly relies on traditional fire detectors, which generally suffer from problems such as response delays, inaccurate location, and an inability to quantify the scale of the fire.

[0003] In recent years, fire detection technology based on video images has been applied, but its function is mostly limited to qualitative judgment of "whether there is a fire". Heat release rate cannot be obtained through fixed monocular surveillance cameras. Although some studies have attempted to establish a correlation between two-dimensional image features and heat release rate, its accuracy and generalization ability are greatly affected by the viewing angle, making it difficult to reliably apply in real-world complex tunnel environments.

[0004] Existing research has confirmed the correlation between combustion heat release rate and three-dimensional volume, pointing the way for visual measurements, but applying it to tunnel scenarios still faces significant challenges. An idealized multi-view reconstruction scheme would require deploying a large number of synchronized cameras, which is costly and impractical.

[0005] Therefore, there is an urgent need for a non-contact visual recognition method and system for the heat release rate of tunnel fires, which can monitor and locate fires, and make accurate predictions of the heat release rate of combustion in real time to quantify the scale of fires. This method requires no additional specialized equipment, is low in cost, and is highly practical. Summary of the Invention

[0006] One of the objectives of this invention is to provide a non-contact visual recognition method for the heat release rate of tunnel fires, which can monitor and locate fires, and make accurate predictions of the heat release rate of combustion in real time to quantify the scale of the fire. This method requires no additional specialized equipment, is low in cost, and is highly practical.

[0007] The basic solution provided by this invention is a non-contact visual recognition method for the heat release rate of tunnel fires, which includes the following: Offline modeling: Conduct standard combustion experiments to construct a fire perception and location model, a quantitative mapping model between the true values ​​of flame 3D morphology and heat release rate, and a sparse view model. Figure 3 The three-dimensional reconstruction model is used to locate the flame region based on the image, obtain the heat release rate estimate based on the true value of the three-dimensional flame morphology, and generate a three-dimensional voxel mesh based on the two-dimensional flame image. Online identification: Tunnel monitoring videos captured by cameras installed within the tunnel are acquired. Using a fire perception and positioning model, the flame area is located. Based on the positioning results, the horizontal distance between the flame and the camera is calculated. Combined with the positioning results, a camera downstream of the flame is rotated to capture images of the flame. The cameras acquire flame images from different perspectives. After standardizing the flame images, sparse visual data is used to identify the flames. Figure 3 The three-dimensional reconstruction model is used to obtain the predicted three-dimensional voxel mesh of the flame. The true value of the three-dimensional morphology of the flame is extracted from the predicted three-dimensional voxel mesh. The heat release rate is estimated by using a quantitative mapping model.

[0008] The second objective of this invention is to provide a non-contact visual recognition system for the heat release rate of tunnel fires, which can monitor and locate fires, and make accurate predictions of the heat release rate of combustion in real time to quantify the scale of the fire. It does not require additional specialized equipment, is low in cost, and is highly practical.

[0009] This invention provides a second basic solution: a non-contact visual recognition system for the heat release rate of tunnel fires, used to perform the aforementioned non-contact visual recognition method for the heat release rate of tunnel fires, comprising: a server; The server is equipped with a prior knowledge base to store fire perception and location models, quantitative mapping models of flame 3D morphology truth values ​​and heat release rate truth values, and sparse vision models constructed through standard combustion experiments. Figure 3 The three-dimensional reconstruction model is used to locate the flame region based on the image, obtain the heat release rate estimate based on the true value of the three-dimensional flame morphology, and generate a three-dimensional voxel mesh based on the two-dimensional flame image. The server is also used to acquire tunnel monitoring videos captured by cameras installed inside the tunnel, locate the flame area using a fire perception and positioning model, and calculate the horizontal distance between the flame and the camera based on the positioning results. In conjunction with the positioning results, the camera downstream of the flame was rotated to take pictures of the flame; It is also used to standardize flame images obtained from different camera viewpoints, and then use sparse views... Figure 3 The three-dimensional reconstruction model is used to obtain the predicted three-dimensional voxel mesh of the flame. The true value of the three-dimensional morphology of the flame is extracted from the predicted three-dimensional voxel mesh. The heat release rate is estimated by using a quantitative mapping model.

[0010] Beneficial effects: Firstly, this scheme performs offline modeling, constructing a fire perception and location model, a quantitative mapping model between the true value of flame 3D morphology and the true value of heat release rate, and a sparse view model through standard combustion experiments. Figure 3 The three-dimensional reconstruction model is used to locate the flame region based on the image, obtain the heat release rate estimate based on the true value of the three-dimensional flame morphology, and generate a three-dimensional voxel mesh based on the two-dimensional flame image. Then, online identification is performed to obtain tunnel monitoring videos captured by cameras installed inside the tunnel. Using a fire perception and positioning model, the flame area is located, and based on the positioning results, the horizontal distance between the flame and the camera is calculated. In conjunction with the positioning results, a camera downstream of the flame is rotated to capture images of the flame; the camera acquires flame images from different perspectives, and after standardizing the flame images, they are analyzed using sparse perspective. Figure 3 The model is used to reconstruct the flame in three dimensions. The predicted three-dimensional voxel grid is obtained, which maps the captured non-standard viewpoint images to a standardized space consistent with the viewpoint distribution of the offline training data. The pre-trained model is then used to complete the three-dimensional reconstruction. The true value of the flame's three-dimensional morphology is extracted from the predicted three-dimensional voxel grid. The estimated value of the heat release rate is obtained through a quantitative mapping model. This is the core fire quantification indicator, which reflects the size of the fire.

[0011] In tunnel scenarios, online identification utilizes only ordinary side-wall monitoring cameras. Through model building and analysis, non-contact, real-time, and quantitative measurement of the key fire parameter, heat release rate, provides data for emergency decision-making. Both offline and online phases employ voxel meshes as a unified representation of 3D information. Two advanced voxelization techniques—visual shell (offline ground truth) and deep learning networks (online prediction)—ensure data consistency and model accuracy throughout the process. In the offline phase, two precise mapping models are constructed between 2D images, 3D voxel models, and heat release rate. In the online phase, only two sparse viewpoints of flame images need to be processed and standardized to a standard viewpoint consistent with the offline phase before being input into the sparse viewpoints. Figure 3 The 3D reconstruction model directly infers a 3D voxel mesh, effectively ensuring the accuracy of subsequent data prediction.

[0012] In summary, this solution can monitor and locate fires, and make accurate, non-contact, real-time predictions of the heat release rate of combustion to quantify the scale of a fire. It requires no additional specialized equipment, is low in cost, and highly practical. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating an embodiment of a non-contact visual recognition method for the heat release rate of a tunnel fire according to the present invention. Figure 2 This is a schematic diagram of the fire perception and positioning model structure in an embodiment of a non-contact visual recognition method for the heat release rate of a tunnel fire according to the present invention. Figure 3 This is a schematic diagram of the standard combustion experiment layout in an embodiment of a non-contact visual recognition method for the heat release rate of a tunnel fire according to the present invention. Figure 4This is a schematic diagram of monocular visual localization and real-time reasoning during the online identification stage in an embodiment of a non-contact visual recognition method for the heat release rate of a tunnel fire according to the present invention. Detailed Implementation

[0014] The following detailed description illustrates the specific implementation method: Example 1 This embodiment provides a non-contact visual recognition method for the heat release rate of tunnel fires, as shown in the attached figure. Figure 1 As shown, it includes the following: Offline modeling phase: Conduct standard combustion experiments to construct a fire perception and location model, a quantitative mapping model between the true values ​​of flame 3D morphology and heat release rate, and a sparse view model. Figure 3 The three-dimensional reconstruction model is used to locate the flame region based on the image, obtain the heat release rate estimate based on the true value of the three-dimensional flame morphology, and generate a three-dimensional voxel mesh based on the two-dimensional flame image. The specific process includes: S1. Standard Combustion Experiment and Synchronous Data Acquisition: Conduct a standard combustion experiment, construct a standard viewpoint space, and synchronously acquire multi-view image sequences of the flame and fuel mass. Calculate the true value of the heat release rate based on changes in fuel mass. The standard viewpoint space is constructed with the flame source centroid as the origin, and a synchronously triggered acquisition array consisting of multiple high-definition industrial cameras is deployed to acquire multi-view image sequences of the flame, such as... Figure 3 As shown; a precision electronic balance is placed below the fuel pan to record fuel quality in real time.

[0015] Specifically, in a controlled experimental tunnel environment, a series of standard combustion experiments were conducted to build a three-dimensional prior knowledge base. Various typical fuels, such as gasoline, diesel, and wood chips, were used, and the fuel load and oil pan size were systematically changed to cover the dynamic range of the complete heat release rate from the early ignition source to the fully developed fire.

[0016] Construct a standard view space with precisely known geometric relationships, and set up a synchronously triggered acquisition array consisting of multiple high-definition industrial cameras with the centroid of the fire source as the origin of the spatial coordinates.

[0017] The poses and internal parameters of all cameras in space were obtained before the experiment using high-precision calibration methods, such as PnP calculation and Zhang Zhengyou calibration, thereby establishing a unified world coordinate system. These precisely calibrated cameras themselves constitute a standard viewpoint observation system, providing a series of realistic and usable standard observation viewpoints. This system provides training samples with clear viewpoints and defined geometric relationships for subsequent deep learning models, and serves as a geometric benchmark for unified viewpoint correction during the online recognition stage.

[0018] The construction parameters of the standard viewpoint space include: number of cameras, spatial layout and angular range, observation baseline length and field of view; Number of cameras: Theoretically, the more cameras involved in the reconstruction, the more accurate the generated visual shell ground truth. However, considering system cost, synchronization triggering complexity, and data processing efficiency, it is recommended to deploy 6 to 18 synchronized high-definition industrial cameras.

[0019] Spatial layout and angular range: The cameras should be distributed in three dimensions in the horizontal and vertical directions around the center of mass of the fire source (i.e., the origin of the spatial coordinates) to form an enclosed observation network; Horizontal angular coverage: Covering the complete horizontal azimuth from 0° to 360°. In actual deployment, given the limitations of the tunnel testing environment, at least an effective horizontal viewing angle range of 180° to 270° should be guaranteed. Vertical Angle Coverage: With the fire source's center of mass horizontal plane as 0°, the pitch angle covers a range of -30° (view from above) to +45° (view from below). To accurately capture the vertical development profile of the flame, cameras should not all be located at the same height. It is recommended to distribute cameras at two or more different height levels, so that their pitch angles cover a range of -30° (view from above) to +45° (view from below) (with the fire source's center of mass horizontal plane as 0°). This layout effectively improves the vertical reconstruction accuracy of the visual shell, especially for the representation of tall flames.

[0020] Observation baseline length and field of view design: The observation baseline length is the distance from the optical center of the camera to the centroid of the fire source. It is set according to actual needs. The principle for determining it is to ensure that the expected maximum flame size can be completely within the field of view of all cameras throughout the entire combustion process and maintain sufficient image resolution. At the same time, it is necessary to ensure that the camera equipment is in a safe working environment to avoid damage or serious degradation of image quality due to high temperature, flame or smoke.

[0021] Throughout the combustion process, the acquisition array synchronously acquires multi-view image sequences of the flame at a fixed frame rate; A precision electronic balance was placed beneath the fuel pan in the experiment to obtain the fuel mass. Real-time recording of fuel mass changes allows for the acquisition of the true value of the heat release rate. Specifically, the true value of the heat release rate is calculated using the weightlessness method: ; in, For the rate of mass loss, The effective heat of combustion of fuel.

[0022] S2. Construct a fire perception and localization model to perceive and localize the flame region of the input image (tunnel fire image) and output a flame detection box; The fire perception and localization model aims to robustly perceive fires and determine their spatial location from any monitoring perspective. In this embodiment, the fire perception and localization model employs a deep learning object detection network (such as YOLO, Faster R-CNN, or variants thereof). Figure 2 The method is trained on a large number of labeled images of tunnel fires to enable it to identify and locate flame areas in images and output flame detection boxes. This facilitates the subsequent ability to quickly and accurately locate flame image areas from complex tunnel monitoring screens and output flame detection boxes. The midpoint of the lower edge of the flame detection box is defined as a representative point and used as the input for subsequent monocular vision geometric localization calculations. By combining camera calibration parameters with the dedicated geometric localization model constructed in this invention, the two-dimensional pixel is accurately mapped to its position coordinates in the three-dimensional space of the tunnel, providing a spatial reference for subsequent viewpoint standardization.

[0023] S3. Three-dimensional reconstruction based on visual shell and establishment of the first mapping relationship model: Based on the multi-view image sequence, generate the true value of the three-dimensional shape of the flame, and establish a quantitative mapping model between it and the true value of the heat release rate, as the first mapping relationship model; Using synchronized image data from a multi-view image sequence acquired in the standard view space constructed by S1, the true three-dimensional morphology of the flame is generated, and a quantitative mapping model between the flame and the heat release rate (HRR) is established. The specific steps are as follows: S301, Image Segmentation and Calibration (Camera Calibration and Image Preprocessing): In a multi-view image sequence acquired synchronously by an array under standard viewpoint space, each group of images at each timestamp... The following images are all processed using image segmentation algorithms (such as adaptive thresholding or U-Net-based semantic segmentation models). Extracting the flame segmentation mask , which is the binary mask of the flame.

[0024] S302, 3D Reconstruction (Spatial Sculpture and Visual Shell Generation): The 3D space of interest (i.e., the flame region) is discretized into a dense 3D voxel mesh; all voxels are traversed (a 3D voxel mesh refers to the entire 3D data structure, composed of regularly arranged voxels, where a voxel is the smallest unit of volume in the mesh), for each voxel... The three-dimensional coordinates are transformed onto the plane of the images captured by each camera using the camera projection model to obtain the projection points. If the projection points of the voxel are all located in the corresponding flame segmentation mask under all (or a set threshold, such as 85%) camera views. If the voxel is within the specified range, it is determined that the voxel belongs to the flame entity; after traversal, all retained voxels constitute the time frame. Flame Visual Shell Using three-dimensional voxel mesh Formal storage; Flame visual shell It is the ground truth of the three-dimensional flame shape obtained through multi-view images and spatial sculpting in the offline stage, that is, the three-dimensional morphological truth. The storage form of this truth in the computer is a voxel grid.

[0025] S303, Extraction of 3D Morphological Parameters: From the flame visual shell The following morphological parameters are extracted from them, including volume, surface area, and 3D bounding box size, to form a 3D morphological feature vector; Volume Count the total number of occupied voxels and multiply by the physical volume of a single voxel: The number of voxels occupied. , , The physical dimensions of a voxel in three directions.

[0026] Surface area The total area is calculated after reconstructing the surface mesh using the moving cube algorithm.

[0027] 3D bounding box dimensions, calculate the height of the minimum circumscribed cuboid. ,width ,depth The three-dimensional bounding box is the smallest circumscribed cuboid of the flame visual shell. Ultimately, a three-dimensional morphological feature vector representing the true value of the three-dimensional morphology is formed. .

[0028] S304. Establishment of the First Mapping Relationship Model Building a dataset The regression model is trained using algorithms such as support vector regression or neural networks to establish a mapping relationship from three-dimensional morphological feature vectors to heat release rate. By minimizing the predicted value Compared with measured values Mean squared error, optimize model parameters and The first mapping relationship model was established. This model realizes the quantitative inversion from the three-dimensional shape measured by non-contact methods to the key fire parameter HRR.

[0029] S4, Sparse Vision Figure 3 3D Reconstruction Model Construction and Training: Constructing Sparse Views Figure 3 A 3D reconstruction model, serving as the second mapping model, is used to generate a 3D voxel mesh based on a 2D flame image; and images from a multi-view image sequence acquired in a standard combustion experiment are used as training samples. The three-dimensional voxel grid corresponding to the image As 3D labels, construct the training dataset , for sparse views Figure 3 Training the 3D reconstruction model; A deep learning model is trained to directly infer a complete 3D voxel model from sparse 2D flame images. This model will serve as the core of the online recognition stage. The specific steps are as follows: S401. Data Preparation: Use all flame images acquired by all cameras from all perspectives in the offline experiment as training samples. Each image Each corresponds to a three-dimensional voxel mesh of the visual shell generated at the same time through step S3. This serves as its actual 3D label, thus forming a large-scale training dataset. S402, Model Building: Constructing a Sparse View Figure 3 The 3D reconstruction model uses a deep learning network architecture based on an encoder-decoder structure; the specific structure is as follows: Encoder: Uses a pre-trained convolutional neural network (such as ResNet-34 or ResNet-50) to process the input image. Encode into a compact feature vector ; Decoder: Consists of a series of 3D deconvolution layers, which convert feature vectors Upsampled and decoded into a three-dimensional probabilistic voxel grid. Each value represents the probability that the corresponding voxel is occupied by flame.

[0030] S403, Loss Function and Training Process: Using the training dataset to construct a sparse view Figure 3 The 3D reconstruction model is trained with the goal of minimizing the predicted 3D voxel mesh. Compared with real 3D voxel mesh The differences between them; the loss function uses binary cross-entropy loss: in, It is the number of training samples; Optimize network parameters using the backpropagation algorithm. Ultimately, a high-performance sparse view is obtained. Figure 3 3D reconstruction model .

[0031] Online identification phase: Acquire tunnel monitoring video from cameras installed inside the tunnel; locate the flame area using a fire perception and positioning model; and calculate the horizontal distance between the flame and the camera based on the positioning results. In conjunction with the positioning results, a camera downstream of the flame is rotated to capture images of the flame; the camera acquires flame images from different perspectives, and after standardizing the flame images, they are analyzed using sparse perspective. Figure 3 The three-dimensional reconstruction model is used to obtain the predicted three-dimensional voxel mesh of the flame. The true value of the three-dimensional morphology of the flame is extracted from the predicted three-dimensional voxel mesh. The estimated value of the heat release rate is obtained through the quantitative mapping model. If a flame area is detected in the surveillance video, an alert will be issued. If the detected flame area indicates a vehicle fire, the alert will be based on the real-time estimated heat release rate. Typical peak heat release rate value of the vehicle model that caught fire Based on the relationship, conduct fire development stage assessment and early warning.

[0032] The specific process includes: S5. Fire Sensing Perception and Monocular Visual Localization: Acquire tunnel monitoring video, extract frame images from the video, locate the flame area using a fire sensing perception and localization model, and calculate the horizontal distance between the flame and the camera based on the localization results. In conjunction with the positioning results, the camera downstream of the flame was rotated to take pictures of the flame; Specifically, for real-time analysis of tunnel monitoring video, the fire perception and localization model trained in the offline phase is first invoked to process the input video frames, achieving real-time detection and preliminary localization of the flames, such as... Figure 4 As shown, the specific steps are as follows: S501, Flame Region Detection: Acquire tunnel monitoring video captured by a fixed camera, extract frame images from the tunnel monitoring video, and obtain pixel-level bounding boxes of the flame region using a fire perception and localization model; using the midpoint of the lower edge of this bounding box... This point, representing the two-dimensional coordinates of the flame base in the image, will serve as the calculation reference for subsequent geometric positioning. The fixed camera setup used in this embodiment is suitable for fire perception and positioning, as its pose remains constant after calibration, ensuring accurate subsequent positioning calculations.

[0033] S502. Monocular Vision Positioning Calculation: Based on the two-dimensional coordinates of the flame base in the image, estimate the horizontal distance from the flame base to the camera using the principle of monocular vision positioning. ; Specifically, assuming the camera is calibrated, the intrinsic parameter matrix is... Installation height is The optical axis pitch angle is The deflection angle is The pixel coordinates of the flame base in the image are ; in, The coordinates of the principal point in the image. This refers to the camera's focal length parameter.

[0034] S503, Multi-view Collaborative Trigger: After the fixed camera completes the initial fire location (i.e., monocular visual location), it automatically dispatches the PTZ camera deployed downstream of the fire source, based on the midpoint... and horizontal distance The camera controls its horizontal rotation and aligns it with the fire source. Through this collaborative mechanism, a pair of flame views with significant visual differences are captured from two directions: upstream (fixed camera) and downstream (PTZ camera) of the fire source, providing the necessary observation baseline for subsequent 3D reconstruction.

[0035] S6. Viewpoint Unification Correction and 3D Reconstruction: The camera acquires flame images from different viewpoints. After standardizing the flame images, sparse viewpoints are used for 3D reconstruction. Figure 3 The model is reconstructed to obtain a predicted three-dimensional voxel mesh of the flame; This step is the core of the online recognition phase. Its purpose is to map the captured non-standard viewpoint images to a standardized space consistent with the viewpoint distribution of the offline training data, and then use a pre-trained model to complete 3D reconstruction. The specific process is as follows: S601, Viewpoint Correction Transformation: To overcome the difference between online camera pose and the standard view space in offline modeling, the following view uniformity correction process is performed: S60101, Virtual Viewpoint Definition: Based on the standard observation coordinate system defined in the offline phase, generate corresponding virtual camera viewpoint parameters; the pose and optical parameters of these virtual viewpoints are consistent with the offline training data.

[0036] S60102. Homography Transformation Calculation: For each actual camera (upstream fixed camera and downstream PTZ camera), calculate the homography transformation matrix between them based on their actual pose and the pose parameters of the corresponding virtual standard viewpoint. .

[0037] S60103, Image Remapping: Based on the calculated homography transformation matrix Geometric correction and resampling are performed on the original flame images captured by each camera: This process generates a set of corrected images with normalized appearance and eliminated perspective distortion.

[0038] S602, 3D Reconstruction: S60201, Multi-view fusion input: Combines the calibrated, standardized views from both the fixed camera and the PTZ camera as a sparse multi-view input set; S60202. Based on the sparse multi-view input set, through the second mapping relationship model... By fusing information from multiple aligned viewpoints, a three-dimensional voxel mesh of the predicted flame is obtained. .

[0039] S7. Real-time estimation of heat release rate: from the predicted 3D voxel mesh. In the process, a three-dimensional morphological parameter vector representing the true value of the three-dimensional morphology of the flame is extracted. As input, through the first mapping relationship model Obtain the estimated heat release rate. : This value is the core fire quantification index that the system ultimately outputs.

[0040] S8. Multi-source information fusion and decision-making: If a flame area is detected in the monitoring video, an early warning is issued. If the detected flame area is a vehicle fire, the warning is based on the real-time estimated heat release rate. Typical peak heat release rate value of the vehicle model that caught fire Based on the relationship, conduct fire development stage assessment and early warning; To improve system reliability and enable tiered and precise fire responses, the system executes the following information fusion and decision-making processes in parallel: S701, Vehicle Model Recognition and Peak Heat Release Rate Query: Based on tunnel surveillance video, the model of the burning vehicle was identified using a deep neural network for license plate recognition and vehicle type classification. And from the preset vehicle model-peak heat release rate database, obtain the typical peak heat release rate value corresponding to the vehicle model. ; Using the same surveillance video, a deep neural network for license plate recognition and vehicle type classification was run simultaneously to determine the specific model of the vehicle on fire. The system comes pre-loaded with a vehicle model-peak heat release rate database. This database integrates NFPA 502 standards, large-scale combustion test data, and research findings on electric vehicle battery fires, and includes typical peak heat release rate values ​​for various vehicle models. The core content of the database is shown in Table 1 below (example): Table 1: Example of Vehicle Model-Peak Heat Release Rate Database S702, Fusion Decision and Alarm: Employs a three-stage progressive decision logic to integrate the estimated visual heat release rate. Peak heat release rate of vehicle model This integration enables comprehensive decision support, ranging from immediate response to strategic foresight, as detailed below: Level 1 Alarm: Fire Confirmation and Location Triggering condition: A flame area is detected in any surveillance video. In this embodiment, the flame is continuously identified and confirmed in any camera frame. Response action: Immediately trigger the highest priority alarm, provide audible and visual warnings at the control center, accurately locate the fire source based on monocular visual positioning technology, and automatically activate fire-fighting equipment near the fire source. This stage aims to achieve a "second-level" response.

[0041] Level 2 Alarm: Fire Development Stage Assessment Triggering condition: Continuous and stable output ; Response action: based on and The ratio of [value] to dynamically assess the stage of fire development: Initial stage: < 0.3 * ; Development stage: 0.3 * ≤ < 0.7 * ; Intense Phase: ≥ 0.7 * ; Level 3 Alarm: Disaster Peak Prediction and Resource Allocation Triggering condition: If the detected flame area indicates a vehicle fire, obtain the vehicle model. ; Response Action: The command interface will prominently display "Expected Maximum Fire Size: "MW"; based on this forecast value, strategic-level decision support is provided to commanders: Firefighting resource pre-positioning: If identified as a dangerous goods tanker truck ( If the fire is larger than 200 MW, it is recommended to immediately dispatch special fire trucks and remote fire extinguishing systems to deal with the super-large fire. Structural safety early warning: Calculate the theoretical withstand time of the tunnel structure under expected thermal radiation to provide a safety window reference for personnel evacuation and internal fire fighting; Traffic control scope: Based on the predicted scale, suggest a reasonable scope for traffic control upstream and downstream.

[0042] Compared with existing technologies, this solution has the following significant advantages: 1. Achieving a leap from detection to quantification: This is the first exploration of using ordinary sidewall monitoring cameras in a tunnel scenario to achieve non-contact, real-time, and quantitative measurement of the key fire parameter of heat release rate, providing unprecedented data accuracy for emergency decision-making; 2. Fundamentally solves the perspective problem: Through monocular vision positioning and perspective correction technology, the side-view perspective image is accurately transformed into a virtual front view consistent with the offline training perspective, which fundamentally overcomes the serious impact of perspective distortion of the side-view camera on the measurement and ensures the accuracy of 3D reconstruction and heat release rate estimation. 3. Advanced and consistent technology: Voxels are used as a unified representation of 3D information in both offline and online stages. Through two advanced voxelization technologies, namely visual shell (offline ground truth) and deep learning network (online prediction), the consistency of data and the accuracy of the model are ensured throughout the process. 4. Extremely cost-effective: It makes full use of the existing monitoring camera hardware in the tunnel, eliminating the need for additional expensive special measuring equipment (such as infrared thermal imagers or laser scanners), which greatly reduces the threshold for system deployment and transformation.

[0043] This solution adopts a two-stage framework of "offline knowledge base construction and online sparse perspective reasoning," including: Prior knowledge base: In the offline stage, through multi-view reconstruction and deep learning, two precise mapping relationship models between "two-dimensional image - three-dimensional voxel mesh model - heat release rate" were constructed.

[0044] Sparse Viewpoint Reconstruction and Viewpoint Standardization: In the online phase, only flame images from two sparse viewpoints (upstream fixed camera + downstream PTZ camera) need to be scheduled. These images are then unified to a "standard viewpoint" consistent with the knowledge base through geometric correction (homography transformation), and then input into a deep learning model to directly infer a 3D voxel mesh model.

[0045] Parameter inversion and fusion decision-making: Morphological parameters are extracted from the predicted 3D model, and the heat release rate is calculated through a mapping model. Vehicle model identification information is then fused for cross-validation and tiered alarm processing.

[0046] This embodiment also provides a non-contact visual recognition system for the heat release rate of a tunnel fire, used to perform the above-described non-contact visual recognition method for the heat release rate of a tunnel fire, including: a server; The server is equipped with a prior knowledge base to store fire perception and location models, quantitative mapping models of flame 3D morphology truth values ​​and heat release rate truth values, and sparse vision models constructed through standard combustion experiments. Figure 3The three-dimensional reconstruction model is used to locate the flame region based on the image, obtain the heat release rate estimate based on the true value of the three-dimensional flame morphology, and generate a three-dimensional voxel mesh based on the two-dimensional flame image. The server is also used to acquire tunnel monitoring videos captured by cameras installed inside the tunnel, locate the flame area using a fire perception and positioning model, and calculate the horizontal distance between the flame and the camera based on the positioning results. In conjunction with the positioning results, the camera downstream of the flame was rotated to take pictures of the flame; It is also used to standardize flame images obtained from different camera viewpoints, and then use sparse views... Figure 3 The three-dimensional reconstruction model is used to obtain the predicted three-dimensional voxel mesh of the flame. The true value of the three-dimensional morphology of the flame is extracted from the predicted three-dimensional voxel mesh. The heat release rate is estimated by using a quantitative mapping model.

[0047] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A non-contact visual recognition method for the heat release rate of a tunnel fire, characterized in that, Includes the following: Offline modeling: Standard combustion experiments were conducted to construct a fire perception and localization model, a quantitative mapping model of the true value of flame 3D morphology and the true value of heat release rate, and a sparse view 3D reconstruction model. These models were used to locate the flame region based on the image, obtain the estimated value of heat release rate based on the true value of flame 3D morphology, and generate a 3D voxel mesh based on the 2D flame image. Online identification: The system acquires tunnel monitoring videos from cameras installed within the tunnel. Using a fire perception and positioning model, it locates the flame area. Based on the positioning results, it calculates the horizontal distance between the flame and the camera. Combining the positioning results, it retrieves images of the flame from a camera downstream of the flame. The camera acquires flame images from different perspectives. After standardizing the flame images, a sparse view 3D reconstruction model is used to obtain a predicted 3D voxel grid of the flame. The true value of the flame's 3D morphology is extracted from the predicted 3D voxel grid. Using a quantitative mapping model, an estimated value of the heat release rate is obtained.

2. The non-contact visual recognition method for tunnel fire heat release rate according to claim 1, characterized in that, The offline modeling involves standard combustion experiments, including: conducting standard combustion experiments, constructing a standard view space, simultaneously acquiring multi-view image sequences of the flame and fuel mass, and calculating the true value of the heat release rate based on changes in fuel mass. The standard view space is constructed, with the centroid of the fire source as the origin of the spatial coordinates. A synchronously triggered acquisition array composed of multiple cameras is arranged to acquire multi-view image sequences of the flame.

3. The non-contact visual recognition method for tunnel fire heat release rate according to claim 1, characterized in that, The offline modeling process constructs a fire perception and localization model, which is used to perceive and locate the flame region in the input image, outputs a flame detection box, and defines the midpoint of the lower edge of the flame detection box as the representative point.

4. The non-contact visual recognition method for tunnel fire heat release rate according to claim 2, characterized in that, The offline modeling process involves constructing a quantitative mapping model, including: In a multi-view image sequence acquired synchronously by an array under standard viewpoint space, each group of images at each timestamp... The following images are all processed using image segmentation algorithms. Extracting the flame segmentation mask ; Discretize the flame region into a dense voxel grid; iterate through all voxels, for each voxel... The three-dimensional coordinates are transformed onto the plane of the images captured by each camera using the camera projection model to obtain the projection points. If the projection point of the voxel is located in the binary mask of the corresponding flame in all camera views. If the voxel is within the specified range, it is determined that the voxel belongs to the flame entity; after traversal, all retained voxels constitute the time frame. Flame Visual Shell Using three-dimensional voxel mesh Formal storage; From the flame visual shell Morphological parameters extracted as the true values ​​of the three-dimensional morphology of the flame include: volume. Surface area 3D bounding box dimensions This forms a three-dimensional morphological feature vector that represents the true value of the three-dimensional morphology. ; Building a dataset The regression model is trained using algorithms such as support vector regression or neural networks to establish a mapping relationship from three-dimensional morphological feature vectors to heat release rate. By minimizing the predicted heat release rate value Compared with the measured heat release rate value Mean squared error, optimize model parameters and .

5. The non-contact visual recognition method for tunnel fire heat release rate according to claim 2, characterized in that, The offline modeling process involves constructing a sparse view 3D reconstruction model, including: Using multi-view image sequences from standard combustion experiments as training samples, each image... Each corresponds to a three-dimensional voxel mesh generated at the same time. As its true 3D labels, it constitutes the training dataset. ; A sparse view 3D reconstruction model is constructed, with a deep learning network architecture based on an encoder-decoder architecture. The encoder uses a pre-trained convolutional neural network to process the input image. Encode as feature vector The decoder, consisting of a series of 3D deconvolution layers, converts the feature vectors... Upsampled and decoded into a three-dimensional probabilistic voxel grid. Each value represents the probability that the corresponding voxel is occupied by flame; The sparse view 3D reconstruction model is trained using a training dataset, with the training objective being to minimize the predicted 3D voxel mesh. Compared with real 3D voxel mesh The differences between them; the loss function uses binary cross-entropy loss: in, It is the number of training samples; Optimize network parameters using the backpropagation algorithm. Finally, a sparse view 3D reconstruction model is obtained. .

6. The non-contact visual recognition method for tunnel fire heat release rate according to claim 1, characterized in that, The online identification process involves acquiring tunnel monitoring video, extracting frame images from the video, locating the flame area using a fire perception and positioning model, calculating the horizontal distance between the flame and the camera based on the positioning results, and then, in conjunction with the positioning results, rotating a camera downstream of the flame to capture images of it. This includes: Acquire tunnel monitoring video captured by cameras, extract frame images from the video, and obtain pixel-level bounding boxes of the flame area using a fire perception and localization model; then, use the midpoint of the lower edge of this bounding box as the reference point. The two-dimensional coordinates of the flame substrate in the image; Based on the two-dimensional coordinates of the flame base in the image The horizontal distance from the flame base to the camera is estimated based on the principle of monocular vision localization. ; After the fixed camera completes the initial fire location, the system automatically dispatches cameras deployed downstream of the fire source, based on the midpoint. and horizontal distance Control its rotation and aim it at the fire source to obtain flame images from at least two different perspectives.

7. The non-contact visual recognition method for tunnel fire heat release rate according to claim 1, characterized in that, The online identification process acquires flame images from at least two different perspectives. After standardizing the flame images, a predicted three-dimensional voxel mesh of the flame is obtained through a sparse view 3D reconstruction model, including: Viewpoint correction transformation: Based on the standard observation coordinate system defined in the offline phase, generate the corresponding virtual camera view parameters; For each actual camera, calculate the homography transformation matrix between its actual pose and the pose parameters of the corresponding virtual standard viewpoint. ; Based on the homography transformation matrix Geometric correction and resampling are performed on the original flame images captured by each camera: ； 3D Reconstruction: The corrected and standardized views are used together as a sparse multi-view input set; Based on a sparse multi-view input set, a 3D model is reconstructed using sparse views. By fusing information from multiple aligned viewpoints, a three-dimensional voxel mesh of the predicted flame is obtained. .

8. The non-contact visual recognition method for the heat release rate of a tunnel fire according to claim 1, characterized in that, In the online identification process, the true value of the three-dimensional flame morphology is extracted from the predicted three-dimensional voxel grid, and the estimated value of the heat release rate is obtained through a quantitative mapping model, including: From the predicted 3D voxel mesh In the process, extract the three-dimensional morphological parameter vector. As input, through a quantitative mapping model Obtain the estimated heat release rate. : 。 9. The non-contact visual recognition method for tunnel fire heat release rate according to claim 1, characterized in that, The online identification also includes: issuing an early warning if a flame area is detected in the monitoring video; and if the detected flame area is a vehicle fire, estimating the value based on the real-time estimated heat release rate. Typical peak heat release rate value of the vehicle model that caught fire Based on the relationship, conduct fire development stage assessment and early warning.

10. A non-contact visual recognition system for the heat release rate in tunnel fires, characterized in that, A non-contact visual recognition method for performing the tunnel fire heat release rate as described in any one of claims 1-9, comprising: a server; The server is equipped with a prior knowledge base, which stores the fire perception and localization model, the quantitative mapping model of the true value of the three-dimensional shape of the flame and the true value of the heat release rate, and the sparse view three-dimensional reconstruction model constructed by conducting standard combustion experiments. These models are used to locate the flame area based on the image, obtain the estimated value of the heat release rate based on the true value of the three-dimensional shape of the flame, and generate a three-dimensional voxel mesh based on the two-dimensional flame image. The server is also used to acquire tunnel monitoring videos captured by cameras installed inside the tunnel, locate the flame area using a fire perception and positioning model, and calculate the horizontal distance between the flame and the camera based on the positioning results. In conjunction with the positioning results, the camera downstream of the flame was rotated to take pictures of the flame; It is also used to standardize flame images from different perspectives obtained by the camera, obtain a predicted three-dimensional voxel grid of the flame through a sparse view three-dimensional reconstruction model, extract the true value of the three-dimensional morphology of the flame from the predicted three-dimensional voxel grid, and obtain an estimated value of the heat release rate through a quantitative mapping model.

Images