Fire-fighting hidden danger rapid identification method and system using AI visual technology

Abstract

The invention relates to the technical field of intelligent recognition of fire-fighting hidden dangers, and discloses a method and a system for quickly recognizing fire-fighting hidden dangers by using an AI vision technology. The method comprises the following steps: continuously acquiring a scene video stream through deployed image acquisition equipment, and performing frame extraction and denoising to obtain a sharp image sequence; inputting the image sequence into a hidden danger identification network to obtain a preliminary identification result; performing region growth processing on the preliminary result to obtain accurate hidden danger region contour information; constructing a three-dimensional point cloud model of the hidden danger area according to the contour information; dense scanning path clusters are generated on the surface of the model, and a final hidden danger scanning track is formed through optimized sorting and injection of a control instruction. According to the method, the recognition precision of hidden danger contours is improved through region growth processing, automatic precise scanning is realized based on three-dimensional point cloud path planning, the problems of rough boundary recognition and lack of active probing ability in the prior art are solved, and the automation and precision level of hidden danger investigation is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent fire hazard identification technology, specifically to a method and system for rapid fire hazard identification using AI vision technology. Background Technology

[0002] Currently, AI-based fire hazard identification technology primarily relies on deep neural network analysis of surveillance video streams or images. These methods typically employ an end-to-end approach, directly inputting the acquired images into a trained model, which outputs bounding boxes or coarse segmentation masks to identify hazard locations. However, when dealing with complex backgrounds, lighting variations, or object edge details, the output region boundaries of neural network models often exhibit uncertainty, easily resulting in rough edges, internal voids, or poor alignment with the actual object contours. This inherent limitation leads to insufficient contour accuracy in the identification results, making it difficult to meet the application requirements for precise measurement and refined operations of hazard areas.

[0003] Existing technical solutions, after identifying potential hazards, mostly stop at issuing alarms or generating reports, lacking the ability to proactively and thoroughly automate the exploration of hazard-prone areas. Hazard confirmation and assessment still heavily rely on manual on-site verification, especially for large or hard-to-reach areas, where manual inspection suffers from low efficiency, high risk, and strong subjectivity. Automated inspection equipment, when performing detailed scanning tasks, lacks intelligent path guidance based on 3D spatial information of hazards, and cannot autonomously plan efficient and comprehensive scanning trajectories, limiting the practical application of automated inspection technology. A technical solution is needed that can extend from 2D image recognition to proactive 3D spatial perception and achieve precise automated operation. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for rapid identification of fire hazards using AI vision technology, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for rapid identification of fire hazards using AI vision technology, the method comprising: Continuously acquire scene video streams through image acquisition devices deployed in fire-fighting scenarios; Frame extraction and denoising are performed on the scene video stream to obtain a sharpened image sequence; The clarified image sequence is input into the hazard identification network, and the network outputs preliminary identification results containing potential hazard areas; The preliminary identification results are processed by region growing to obtain accurate contour information of the potential hazard area; A three-dimensional point cloud model of the potential hazard area is constructed based on the precise contour information of the hazard area. A dense cluster of scanning paths is generated on the surface of the 3D point cloud model; The scanning path clusters are optimized and sorted, and control commands are injected to form the final hazard scanning trajectory.

[0006] Preferably, the operation of frame extraction and denoising of the scene video stream specifically includes: Set an extraction time interval and extract keyframe images from the scene video stream; Brightness equalization correction is performed on the extracted keyframe images; A nonlocal means denoising algorithm is used to eliminate noise interference in keyframe images; The denoised keyframe images are arranged in chronological order to form the sharpened image sequence.

[0007] Preferably, the operation of inputting the sharpened image sequence into the hazard identification network specifically includes: The hazard identification network adopts an encoder-decoder architecture, with the encoder part using a deep residual network to extract multi-scale image features. The decoder part gradually restores the feature map resolution through upsampling and skip connections; At the network output, the Softmax function is used to calculate the probability that each pixel belongs to a potential danger zone. The set of pixels with a probability greater than a set threshold is taken as the preliminary identification result.

[0008] Preferably, the operation of performing region growing on the preliminary identification results specifically includes: The high-probability pixels in the preliminary identification results are used as seed points; Calculate the feature similarity of pixels surrounding the seed point, whereby the feature similarity includes color, texture, and gradient information; Pixels whose feature similarity meets the growth criteria are merged into the current potential hazard area; The growth process is iteratively executed until no new pixels can be merged, thereby obtaining the precise contour information of the potential hazard area.

[0009] Preferably, the operation of constructing the three-dimensional point cloud model of the potential hazard area specifically includes: Use a binocular vision camera to acquire multi-view image pairs of the potential hazard area; Stereo matching is performed on the multi-view image pairs to calculate the disparity map; The three-dimensional coordinates of the surface points in the potential hazard area are obtained by converting the camera intrinsic parameters and the parallax map. All three-dimensional coordinate points are registered and fused to construct the complete three-dimensional point cloud model.

[0010] Preferably, the operation of generating a dense cluster of scan paths on the surface of the 3D point cloud model specifically includes: The three-dimensional point cloud model is triangulated to obtain the model surface mesh; Calculate the normal field of the surface mesh, and plan the initial scan line direction based on the direction of the normal field; Multiple scan lines are generated in parallel along the initial scan line direction with a fixed line spacing to form the dense scan path cluster.

[0011] Preferably, the operation of optimizing and sorting the scan path cluster specifically includes: Calculate the length and curvature of each path in the scan path cluster; Establish a cost function for path switching, which takes into account changes in distance and angle between paths; A heuristic search algorithm is used to find the path access sequence with the lowest total cost, and the order of scanning path clusters is rearranged according to the path access sequence.

[0012] Preferably, the operation of the injection control command specifically includes: Control points are set at equal intervals on the optimized and sorted scanning path, and each control point is assigned a velocity value, an acceleration value, and a sensor trigger command. All control points and their corresponding instructions are sequentially encapsulated into a data packet, which is the final hazard scanning trajectory and can be directly sent to the execution agency.

[0013] Preferably, the training process of the hazard identification network adopts an incremental learning strategy, specifically including: Initial training uses a large-scale public dataset for network parameter pre-training; When the system is deployed to a new scenario, it collects new labeled data for that scenario; The pre-trained network is fine-tuned periodically using new labeled data, and the fine-tuned network parameters are saved to update the hazard identification network.

[0014] Preferably, the present invention also includes a fire hazard rapid identification system utilizing AI vision technology. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the fire hazard rapid identification method utilizing AI vision technology as described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are: After the AI ​​visual model outputs preliminary recognition results, a region growing algorithm is introduced for post-processing. This technique uses potential hazard areas output by the model as seed points and iteratively grows pixels based on the similarity of their colors, textures, or gradients in feature spaces. This effectively bridges boundary breaks in the preliminary recognition results, fills in voids caused by noise or occlusion within the region, and improves the fit between the hazard area contour and the edge of the real object. This process combines the rapid recognition capabilities of AI with the advantages of precise local segmentation in traditional image processing algorithms. The generated precise contour information provides high-quality input data for subsequent geometric measurements and 3D reconstruction, avoiding model distortion and measurement errors caused by blurred boundaries.

[0016] A 3D point cloud model of the hazard area is constructed based on precise contour information, and a dense cluster of scanning paths is generated on the model surface. These paths are then optimized and ordered, for example, through global planning based on the shortest path principle or coverage priority principle, ultimately forming an efficient and comprehensive scanning trajectory. This technology transforms hazard identification from passive visual observation to active spatial perception and control, enabling automated inspection equipment to perform systematic close-range scanning or data collection of hazard targets based on this trajectory command. This not only improves the automation and coverage completeness of hazard investigation but also reduces reliance on human experience, enabling targeted and refined review and monitoring of specific hazard areas. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the working principle of the fire hazard rapid identification method utilizing AI vision technology described in this invention. Figure 2 This is a flowchart of frame extraction and noise reduction processing. Figure 3 A flowchart for network processing of hazard identification; Figure 4 A comprehensive evaluation chart of processing time and accuracy at each stage of the fire hazard identification process; Figure 5 Plan the characteristic diagram for the S-shaped velocity curve. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1This invention provides a method for rapid identification of fire hazards using AI vision technology. The method includes: deploying an image acquisition device in a fire scene to continuously acquire scene video streams; the image acquisition device can be a fixedly installed high-definition camera or a vision sensor mounted on a mobile robot. Frame extraction and denoising processing are performed on the scene video stream to obtain a sharpened image sequence. This step aims to extract high-quality static images from dynamic video to reduce data volume and improve the accuracy of subsequent processing. The sharpened image sequence is input into a hazard identification network, which is designed based on a deep learning architecture and outputs preliminary identification results containing potential hazard areas. The preliminary identification results are presented in the form of pixel-level probability maps. Region growing processing is performed on the preliminary identification results. By iteratively merging similar pixels, accurate hazard area contour information is obtained, thereby eliminating isolated noise points and improving area boundaries. A three-dimensional point cloud model of the hazard area is constructed based on the accurate hazard area contour information. This model is implemented through multi-view geometric calculations and can accurately reflect the spatial morphology of the hazard object. Dense scanning path clusters are generated on the surface of the three-dimensional point cloud model, covering the model surface to plan the detection path. The scanning path clusters are optimized and sorted, and control commands are injected to form the final hazard scanning trajectory. This trajectory can be directly sent to the actuator, such as a scanner or robot, for automated hazard detection.

[0020] Example 1: See Figure 2 In practice, frame extraction and denoising are performed by extracting keyframe images from a continuously input scene video stream by setting a fixed extraction time interval. The value of the extraction time interval needs to be configured according to the dynamic characteristics of the actual application scenario. For example, a shorter time interval is set in areas with frequent movement of people or vehicles, while a longer time interval can be set in relatively static environments to effectively control the amount of data. Each extracted keyframe image needs to undergo brightness equalization correction. Brightness equalization correction uses global histogram normalization or adaptive histogram equalization algorithms based on local regions. The purpose is to eliminate the brightness differences between frames caused by changes in lighting conditions or automatic exposure adjustment of the camera, so that the entire sharpened image sequence has a consistent visual performance.

[0021] In practical implementation, a nonlocal means denoising algorithm is used to eliminate noise interference in keyframe images. The core idea of ​​the nonlocal means denoising algorithm is to utilize the redundant information existing within the image and perform weighted average filtering on the central pixel by calculating the similarity weight between two pixel blocks in the image. The size of the pixel block and the range of the search window are key parameters affecting the effect of the nonlocal means denoising algorithm and need to be finely adjusted according to the image resolution and noise level. All denoised keyframe images are arranged in their original timestamp order to form the final sharpened image sequence used as input. The sharpened image sequence is stored in memory as an ordered array or linked list data structure and indexed so that the hazard identification network can read it sequentially. In some embodiments, the parameter settings of the image acquisition device have a direct impact on the effect of frame extraction and denoising. The shutter speed, ISO sensitivity, and white balance mode of the image acquisition device need to be pre-calibrated during the system initialization stage to ensure that the acquired scene video stream has a high signal-to-noise ratio and color fidelity. The frame extraction module can be implemented on the embedded video processing unit. This module monitors the video stream in real time and triggers frame capture operations when a preset extraction time interval is reached, based on the system clock or video frame counter. The captured keyframe images are temporarily stored in a first-in-first-out image buffer. Brightness equalization correction can be performed immediately after frame capture. The correction process typically includes calculating the image's grayscale histogram and performing histogram matching or specification processing to ensure that the histogram distribution of the corrected image conforms to a predefined ideal distribution model, thereby improving image contrast and compressing dynamic range.

[0022] It is understandable that nonlocal means denoising algorithms have relatively high computational complexity. To meet real-time processing requirements, a parallel computing implementation based on GPU acceleration can be adopted. In parallel implementation, the input image is segmented into multiple patches, and each patch is assigned to an independent processing unit for similarity weight calculation and filtering operations. Finally, the processing results are merged to generate the denoised image. The similarity weight function in nonlocal means denoising algorithms typically uses Gaussian weighted Euclidean distance to measure the difference between two pixel blocks. The attenuation parameter of the weight function controls the smoothness of the filtering and needs to be adaptively set according to the noise variance. The arrangement and management of the sharpened image sequence is handled by a dedicated sequence management module. This module adds a timestamp and sequence number metadata to each image and monitors the integrity and continuity of the sequence to prevent data inconsistencies caused by frame loss or processing delays.

[0023] In practical implementation, the frame extraction time interval needs to be set in a balance between processing efficiency and information integrity. Too short a time interval leads to data redundancy and wasted computational resources, while too long a time interval may miss important scene state changes. A feasible strategy is to combine motion detection algorithms to automatically extend the extraction time interval during static scenes and switch to a denser time interval when significant motion is detected. For brightness equalization correction, in addition to global correction methods, local correction methods based on Retinex theory can also be used. These methods can better handle complex lighting scenes with both strong light and shadow areas, restoring image details by estimating and compensating for the illumination components. In the implementation of nonlocal mean denoising algorithms, accurate estimation of the noise level is crucial. The gray-level variance can be calculated in flat areas of the image as an estimate of the noise variance, and the filtering parameters can be dynamically adjusted accordingly.

[0024] Optionally, frame extraction can be based not only on fixed time intervals but also linked to scene transition detection or I-frame occurrence events in the video stream. When a significant change in scene content is detected, a frame is extracted immediately, thereby further improving extraction efficiency while ensuring no information loss. The brightness equalization correction process can introduce color constancy constraints, maintaining the image's chromaticity information without distortion while adjusting the luminance component, avoiding color distortion due to over-correction. The non-local means denoising algorithm can be combined with a human visual characteristic model, introducing a visual masking effect in the weight calculation. Weaker smoothing is applied to textured areas to preserve details, while stronger denoising is applied to flat areas.

[0025] In some embodiments, the storage format for the sharpened image sequence can be a lossless compression format such as PNG or a lossy compression format such as JPEG, depending on the trade-off between storage space and image quality. If the subsequent hazard identification network has extremely high requirements for image accuracy, lossless compression or direct storage of the original bitmap data is recommended. The entire frame extraction and denoising process can be encapsulated as an independent software service, receiving video stream input and outputting a sharpened image sequence through a well-defined application programming interface, facilitating system integration and decoupling of functional modules. Various parameters in the processing, such as extraction time interval, equalization intensity, and denoising degree, can be externally set through configuration files, enabling the system to flexibly adapt to different deployment environments.

[0026] It is understandable that the performance of frame extraction and denoising processing needs thorough testing and optimization, especially with high-resolution video stream input. The end-to-end processing latency from frame extraction to generating a sharpened image sequence needs to be evaluated to ensure it meets the overall system's real-time requirements. For scenarios with extremely poor lighting conditions, such as strong reflections or severe underlighting, it may be necessary to add optical filters or supplementary lighting devices to the front-end image acquisition equipment to physically improve the quality of the input video stream and create better conditions for subsequent digital image processing. Non-local means denoising algorithms have high requirements for computational resources and memory bandwidth when processing high-resolution images. In resource-constrained embedded deployment environments, wavelet transform-based denoising methods or variants of the fast non-local means algorithm can be considered to achieve acceptable denoising results with lower computational costs.

[0027] Optionally, before feeding the sharpened image sequence into the hazard identification network, an image enhancement post-processing step can be added. For example, an adaptive contrast-limited histogram equalization algorithm can be used to further sharpen image edges and enhance texture features. This step can be executed sequentially in the same processing pipeline as the denoising process. The entire frame extraction and denoising process should have an exception handling mechanism. For example, when the video stream is interrupted, the image format is abnormal, or the processing algorithm fails, the system should be able to record error logs and attempt to recover, or degrade to using the unprocessed original frames for subsequent analysis, ensuring the robustness of the system. The final sharpened image sequence can be stored in memory or selectively persisted to a disk or network storage system for post-event review, algorithm debugging, or accumulation of model training data.

[0028] Example 2: See Figure 3In its implementation, the hazard identification network adopts an encoder-decoder architecture. The encoder uses a deep residual network to extract multi-scale image features. The deep residual network consists of multiple stacked residual blocks, each containing two sets of convolutional layers, batch normalization layers, and ReLU activation functions. The input and convolutional output are added together via skip connections. The encoder output consists of feature maps at different scales, capturing hierarchical information from low-level edges to high-level semantics. The decoder gradually restores the feature map resolution through upsampling and skip connections. Upsampling uses bilinear interpolation or transposed convolution to enlarge the feature map size. Skip connections concatenate the feature map at the corresponding resolution from the encoder stage with the current feature map from the decoder along the channel dimension, thus fusing spatial details and contextual information. At the network output, the Softmax function calculates the probability that each pixel belongs to a hazard region. The Softmax function converts the score vectors of multiple categories for each pixel output from the last layer of the network into a probability distribution vector; the category with the highest probability value is the predicted category for that pixel. The set of pixels with a probability greater than a set threshold is used as the initial identification result. The set threshold is a configurable parameter between 0 and 1, used to balance the detection sensitivity and false alarm rate.

[0029] In practice, the region growing process on the preliminary identification results uses high-probability pixels as seed points. The selection criteria for seed points are that the pixel's probability value not only exceeds a set threshold but also exhibits a local maximum within its 8-neighborhood, ensuring that the seed point is located in the center of the potential hazard area rather than at the edge. Feature similarity is calculated for pixels surrounding the seed point, including color, texture, and gradient information. Color similarity is calculated based on the Euclidean distance between pixels using the CIELAB color space. Texture similarity is measured by calculating the Bach distance of the local binary pattern histogram within a local window surrounding the pixel. Gradient similarity is calculated by comparing the gradient magnitude and orientation angle calculated by the Sobel operator at the pixel. Pixels whose feature similarity meets the growth criteria are merged into the current potential hazard area. The growth criteria are set as a dynamic similarity threshold, which is adaptively adjusted during the region growing process based on the average feature value of the merged areas. The growing process is iteratively executed until no new pixels can be merged, thus obtaining accurate contour information of the potential hazard area. The iterative process uses a queue data structure for breadth-first search to ensure the uniformity of region growth.

[0030] In some embodiments, the deep residual network can be configured using ResNet-50 or ResNet-101 as the encoder backbone network. These networks are pre-trained on the ImageNet dataset, and their weights are used as initialization parameters for the hazard identification network. The upsampling factor of the decoder corresponds to the downsampling factor of the encoder. For example, when the encoder downsamples the input image by 32 times, the decoder needs to restore the original resolution through 5 upsampling operations of 2x. The implementation of skip connections needs to ensure that the sizes of the corresponding feature maps of the encoder and decoder are perfectly matched. When the sizes do not match, adjustments need to be made through cropping or interpolation. The calculation of the Softmax function can be expressed as: ; in: This represents the probability that pixel i belongs to category c. It is the score of pixel i for category c, which is the output of the last layer of the network. This represents the total number of categories, and the summation operation covers all categories. Typical threshold values ​​range from 0.6 to 0.8; the specific value needs to be determined using performance curves on the validation set.

[0031] It's understandable that calculating feature similarity in region growing is a computationally intensive operation. To improve processing speed, integral image techniques can be used to accelerate the calculation of local binary pattern histograms, or the calculation of color and gradient features can be parallelized across multiple processor cores. The dynamic similarity threshold adjustment strategy can be based on the statistical characteristics of region growing; for example, a more lenient threshold can be used in the early stages of growth to rapidly expand the region, while a stricter threshold can be used in the later stages to refine the boundaries. Besides the absence of new pixels to merge, the termination condition for the region growing process can also include setting a maximum number of iterations or a maximum region area limit to prevent the algorithm from getting stuck in infinite loops or oversegmenting in complex texture regions.

[0032] Optionally, the training process of the hazard identification network can employ the Adam optimizer with weight decay, using a weighted sum of cross-entropy loss and Dice loss as the loss function to simultaneously optimize pixel-level classification accuracy and region-level overlap. For region growing, a multi-scale strategy can be introduced: first, rapid region growing is performed on the downsampled preliminary identification results to determine the approximate region range; then, fine-grained growing is performed on the original resolution to obtain accurate boundaries. Spatial distance weights can be introduced into the calculation of feature similarity, giving higher weights to pixels closer to the seed point in the similarity evaluation, thus prioritizing the growing of neighboring pixels and ensuring region connectivity.

[0033] In some embodiments, the sharpened image sequence can be subjected to online data augmentation, such as random rotation, flipping, or color jitter, before being input into the hazard identification network to improve the network's generalization ability. The preliminary identification results can undergo morphological post-processing before region growing, such as using opening operations to remove small noise points or using closing operations to fill holes within the region to improve the quality of the seed points. The accurate hazard region contour information obtained from region growing can be stored as a binary mask image or a sequence of polygon vertices for subsequent use by the 3D point cloud model construction module.

[0034] It is understandable that the inference performance of the hazard identification network is crucial to the real-time performance of the system. Network pruning or quantization techniques can be used to reduce the computational load and memory consumption of the model, thereby achieving efficient inference on embedded devices. The parameters of the region growing algorithm, such as the initial similarity threshold, neighborhood size, and maximum number of iterations, need to be individually optimized for different types of fire hazards to achieve the best segmentation results. The entire processing chain from network inference to region growing should be designed as a pipeline, allowing the processing of one frame to overlap with the reading of the next, maximizing the utilization of computational resources and reducing end-to-end latency.

[0035] Optionally, for specific types of hazards, a dedicated task head can be added to the output of the hazard identification network, such as for predicting the hazard category label or hazard level score, to achieve multi-task learning. Region growing can be extended to multiple sub-point competitive growing modes. When multiple disconnected hazard regions exist in the image, independent seed points are assigned to each region and growing is performed synchronously. Finally, over-segmentation is eliminated through region merging criteria. Accurate hazard region contour information can be further used to calculate the region's geometric features, such as area, perimeter, and density. These features can be used as high-level semantic information input into subsequent decision-making modules.

[0036] Example 3: In specific implementation, a binocular vision camera is used to acquire multi-view image pairs of the hazard area. The binocular vision camera consists of two image sensors fixed in the horizontal direction. The baseline distance between the two image sensors is determined by precision machining at the factory and recorded in the configuration file. When acquiring multi-view image pairs, it is necessary to control the binocular vision camera to move around the hazard area or to trigger it synchronously through multiple fixed-position cameras to ensure that the image pairs captured from different perspectives can cover all surface features of the hazard area. The overlap between adjacent perspectives is generally required to be no less than 60% to ensure the reliability of subsequent registration. Stereo matching is performed on the multi-view image pairs, and disparity maps are calculated. The stereo matching process adopts a semi-global matching algorithm. The semi-global matching algorithm aggregates the matching cost along multiple one-dimensional paths, thereby approximating two-dimensional smooth constraints and effectively reducing mismatches in occluded areas and weak texture areas. The disparity map is represented in grayscale image form, where the grayscale value of each pixel represents the difference in horizontal displacement of that point in the left and right images. The three-dimensional coordinates of the surface points in the hazard area are obtained by converting the camera intrinsic parameters and the disparity map. The camera intrinsic parameters include focal length, principal point coordinates and radial distortion coefficient. These parameters are obtained in advance by Zhang Zhengyou calibration method and stored in the calibration file. The conversion process is based on the principle of perspective geometry, which maps the two-dimensional point coordinates in the image coordinate system to the three-dimensional coordinate system with the optical center of the left camera as the origin, combined with their disparity values.

[0037] In the specific implementation, all 3D coordinate points are registered and fused to construct a complete 3D point cloud model. The registration process uses the Iterative Nearest Point Algorithm (TNP) to continuously find the nearest neighbor correspondence between two point cloud sets and calculate the optimal rigid body transformation matrix to minimize the average distance between corresponding points. The fusion process uses a voxel mesh downsampling method, dividing the 3D space into a uniform cubic mesh. For all points falling within each voxel, the geometric center is taken as the representative point, and the average color value is calculated, thus significantly reducing the amount of data while maintaining the model shape. The acquisition of multi-view image pairs requires strict time synchronization and position recording. When using a moving camera for acquisition, an inertial measurement unit and encoder need to be installed on the camera platform to record the six-degree-of-freedom pose information of each shooting position. This information serves as the initial transformation matrix for the TNP, greatly accelerating the registration convergence speed. The cost calculation for stereo matching uses the Census transform. The Census transform generates a bit string by comparing the grayscale relationship between the center pixel and surrounding pixels in the pixel's neighborhood. The Hamming distance between the bit strings is calculated as the matching cost. This transform has good robustness to changes in illumination. Post-processing of disparity maps includes left-right consistency checking, subpixel refinement, and small connected component filtering. Left-right consistency checking detects occluded regions by comparing the left and right disparity maps. Subpixel refinement improves disparity accuracy through quadratic curve fitting. Small connected component filtering removes noise regions with areas smaller than a threshold.

[0038] In some embodiments, the mathematical expression of three-dimensional coordinate transformation can be achieved by the following formula: ; in: This represents the depth value of a point in the camera coordinate system. This represents the horizontal coordinates of a point in the point cloud in the camera coordinate system. This represents the vertical coordinates of a point in the point cloud in the camera coordinate system. Indicates the camera's focal length. Indicates the baseline length of the stereo camera. Indicates the disparity value. and Represents the column and row coordinates of a pixel in an image. and This represents the coordinates of the principal point in the image. The stopping condition for the iterative nearest-neighbor algorithm is typically set to the change in the transformation matrix being less than a threshold or reaching the maximum number of iterations. To avoid getting trapped in local optima, a random sampling consensus algorithm can be used to filter high-quality corresponding point pairs. The size of the voxel grid needs to be determined based on the density of the point cloud and application requirements. Smaller voxel sizes can be used for regions that need to retain detail, while larger voxel sizes can be used for flat regions.

[0039] It is understandable that the accuracy of 3D point cloud models is significantly affected by the calibration errors of binocular vision cameras, especially the calibration errors of baseline distance B and focal length f, which are directly and linearly transmitted to the depth measurement results. Therefore, it is necessary to periodically recalibrate the camera parameters using a high-precision calibration board, especially after the camera has undergone mechanical shock or drastic temperature changes. Stereo matching is prone to matching ambiguity in weakly textured regions. This can be improved by actively adding texture information by projecting random speckle patterns in front of the camera, thus enhancing the reliability of the matching. The iterative nearest-point algorithm is sensitive to the initial pose. When the initial relative pose deviation between two point clouds is too large, it is prone to converge to an erroneous local extremum. In this case, it is necessary to rely on the initial pose provided by an external sensor or to use a coarse registration method based on feature descriptors.

[0040] Optionally, each individual point cloud can be preprocessed before point cloud registration, including statistical outlier removal and radius filtering. Statistical outlier removal is based on the average distance distribution from each point in the point cloud to its neighbors, removing points whose mean distance exceeds three standard deviations. Radius filtering removes points whose number of neighbors within a given radius is less than a threshold. For scenes rich in color information, color consistency constraints can be introduced into the iterative nearest-point algorithm, considering both geometric distance and color difference when finding corresponding points. After point cloud fusion, surface smoothing can be performed, using moving least squares to resample and smooth the point cloud, improving the model's visual effect while providing more continuous surface geometric information for subsequent path planning.

[0041] In some embodiments, the complete 3D point cloud model can be imported into professional point cloud processing software for visualization inspection and quality assessment. The reconstruction accuracy can be verified by measuring the length of objects of known dimensions in the model. For the reconstruction of large potential hazard areas, a block reconstruction strategy can be adopted. The scene is first divided into multiple sub-regions for reconstruction, and then the sub-models are stitched together into a complete model using marker points. The data structure of the 3D point cloud model is usually organized using a kd-tree. The kd-tree structure supports fast nearest neighbor search, which greatly improves the efficiency of subsequent registration and surface analysis.

[0042] It's understandable that the computational complexity of 3D reconstruction is positively correlated with image resolution and the number of point clouds. In practical deployments, a trade-off between accuracy and efficiency is necessary. For applications with high real-time requirements, image resolution or the number of acquisition viewpoints can be reduced, while multi-threaded parallel computing technology can be used to accelerate stereo matching and point cloud registration. The choice of point cloud data storage format is also crucial. The PLY format can simultaneously store point coordinates, color, and normal vector information, while the PCD format is optimized for point cloud processing libraries, supporting more efficient file read and write operations. The entire 3D reconstruction process should have an integrity check mechanism. When a pair of images fails to match or the registration error exceeds a threshold, the system should automatically mark the problematic data and attempt to re-acquire it or prompt for manual intervention.

[0043] Optionally, the completed 3D point cloud model can be further semantically segmented. A trained deep learning model can then identify different object categories within the point cloud, such as smoke detectors, fire extinguishers, and evacuation routes, providing richer contextual information for hazard analysis. For hazard objects with transparent or highly reflective surfaces, polarization imaging technology or multi-exposure fusion methods can be combined to improve the 3D reconstruction quality of these challenging surfaces. The 3D point cloud model can also be fused with thermal imaging camera data, overlaying temperature distribution information onto the geometric model to achieve hotspot detection and temperature anomaly analysis for fire hazards.

[0044] Example 4: In specific implementation, the 3D point cloud model is triangulated to obtain the model surface mesh. The triangulation process uses a greedy projection-based triangulation algorithm, which projects the 3D point cloud onto a 2D parametric plane, performs Delaunay triangulation on the 2D plane, and then maps the triangular mesh back to 3D space. The vertices of the surface mesh are the points in the original point cloud, while the edges and faces of the mesh describe the topological connections between the points, forming a continuous surface model. The normal vector field of the surface mesh is calculated, and the initial scan line direction is planned based on the direction of the normal vector field. The normal vector field is obtained by calculating the unit normal vector of each triangular facet and weighted averaging it to each vertex. The initial scan line direction is usually chosen to be aligned with the principal direction of the model surface on the parametric plane, or a global direction vector can be specified by the user according to the detection requirements. Multiple scan lines are generated parallel to each other along the initial scan line direction with a fixed row spacing, forming a dense cluster of scan paths. The value of the fixed row spacing is determined according to the scanning accuracy requirements and the performance of the actuator. The scan line consists of a series of ordered 3D points, which are the intersections of the scan path and the surface mesh.

[0045] In practical implementation, the operation of optimizing and sorting the scan path cluster calculates the length and curvature of each path in the scan path cluster. The path length is obtained by accumulating the Euclidean distances between adjacent points on the path, and the path curvature is estimated by calculating the first and second derivatives of the path points to determine the degree of local bending. A cost function for path switching is established, which considers the distance and angle changes between paths. The distance cost is defined as the straight-line distance from the end point of the current path to the start point of the next path, and the angle cost is defined as a function of the angle between the tangent vector at the end of the current path and the tangent vector at the beginning of the next path. A heuristic search algorithm is used to find the path access sequence with the lowest total cost. Heuristic search algorithms, such as genetic algorithms or simulated annealing algorithms, randomly walk in the permutation space of the path sequence and accept transformations that reduce the total cost, eventually converging to an approximately optimal solution. The order of the scan path cluster is rearranged according to the path access sequence, making adjacent paths spatially continuous and with gradual directional changes, reducing idle travel time and wear caused by sudden stops and starts of the mechanism.

[0046] In some embodiments, the 3D point cloud needs to be preprocessed before triangulation, including removing outliers through statistical filtering and smoothing the point cloud using a radius-based filtering method to improve mesh quality. When calculating the normal field, the neighborhood search radius is a key parameter; a radius that is too small will make the normal vector sensitive to noise, while a radius that is too large will smooth out the model's detailed features. The initial scan line direction can be determined using principal component analysis to perform eigenvalue decomposition on the local point cloud covariance matrix. The direction of the eigenvector with the smallest eigenvalue is the normal vector direction, while the direction of the eigenvector with the largest eigenvalue can be used as the initial scan line direction. The scan path is generated using an isoparametric line method, generating a set of parallel lines on the 2D parametric domain of the model surface, and then back-calculating to the 3D model surface through parametric mapping relationships.

[0047] It is understandable that the path optimization and sorting problem is a typical variant of the Traveling Salesman Problem, belonging to the NP-hard category, making it impossible to find an exact optimal solution in polynomial time. Heuristic search algorithms sacrifice accuracy for computational efficiency, achieving an engineering-acceptable approximate optimal solution within a reasonable timeframe. The design of the path switching cost function directly affects the optimization result; the relative magnitudes of the distance cost weight and the angle cost weight determine whether the optimization algorithm tends to choose spatially close paths or paths with consistent directions. The reordering of scan path clusters not only considers the switching cost between individual paths but also the continuity of the entire scanning process, avoiding long-distance empty return movements during the scan.

[0048] Optionally, referring to Table 1, an adaptive row spacing strategy can be adopted when generating scan path clusters. This automatically densifies scan lines in areas of high model curvature and uses sparser row spacing in flat areas, improving efficiency while maintaining scanning accuracy. Smoothing can be introduced into the path curvature calculation, using a moving average filter to eliminate the influence of random noise at path point positions on curvature estimation. A time cost term can be added to the cost function, considering the maximum acceleration limits of the actuator in different directions, thus incorporating dynamic constraints into the optimization considerations. The parameter settings of heuristic search algorithms, such as the population size and mutation probability of genetic algorithms, need to be calibrated through extensive experiments to achieve a balance between solution quality and computation time.

[0049] Table 1: Scan Path Optimization Parameters Parameter name Parameter symbol Typical values unit illustrate Distance cost weight ω_d 0.7 none The weighting coefficient of the distance term in the total cost function Angle cost weight ω_θ 0.3 none Weighting coefficients of the angle term in the total cost function Maximum line spacing D_max 10.0 millimeters Maximum allowable distance between adjacent scan lines Minimum radius of curvature R_min 5.0 millimeters Minimum bending radius allowed by path planning Population size N_pop 100 indivual Number of individuals per generation in a genetic algorithm Maximum number of iterations K_max 500 Second-rate Maximum number of iterations for optimization algorithm

[0050] The total cost of path switching can be calculated using the following formula: ; in: This represents the total cost of switching from path i to path j. and These are the distance cost weight and the angle cost weight, respectively. This represents the straight-line distance from the end point of path i to the start point of path j. The maximum distance between all possible paths is used for normalization. and These are the tangent vectors at the end of path i and the beginning of path j, respectively. Representing vectors The magnitude is the length of the vector. Representing vectors The modulus is the length of the vector, and the dot product operation... Calculate the cosine similarity between two vectors.

[0051] In some embodiments, the optimized and sorted scan path clusters can be exported as standard G-code or a custom trajectory file format for parsing and use by different actuators. For model surfaces with complex topologies, it may be necessary to divide the model into multiple regions to generate scan paths separately, and then connect them into a complete scan trajectory through transition paths between regions. During scan path generation, the physical constraints of the actuators, such as maximum velocity, acceleration limits, and workspace boundaries, need to be considered to avoid generating unexecutable paths.

[0052] It is understandable that the quality of the triangular mesh directly affects the smoothness and integrity of the scan path. The presence of non-manifold edges or self-intersecting triangles in the mesh can lead to abnormal path generation, requiring inspection and repair after meshing. The consistency of the normal vector field is crucial; all normal vectors should point in a unified direction, either outside or inside the model. Otherwise, the scan path will jump between sides of the surface. Normal vector consistency can be corrected through propagation using the minimum spanning tree method. Dense clusters of scan paths are stored in memory using a linked list structure. Each path object contains attributes such as path point sequence, length, and curvature, facilitating quick access and modification by the optimization algorithm.

[0053] Optionally, for particularly complex models, a multi-resolution strategy can be used to generate scan paths. First, a coarse path is generated and optimized on a simplified low-resolution grid. Then, the path is finely adjusted on the original resolution grid. The path optimization and sorting algorithm can display the optimization process in real time, dynamically showing the decrease curve of total cost with the number of iterations, facilitating user monitoring of optimization progress and parameter adjustment. The optimized path sequence can be visualized, using different colors to distinguish different paths and arrows to indicate the scan direction, providing an intuitive means of inspection. Scan path data can be serialized and saved to a file, supporting the function of resuming scans after interruption. If the scan process is interrupted for any reason, execution can continue from the breakpoint.

[0054] See Figure 4In the performance evaluation of the complete fire hazard identification process, the processing time (in seconds) and accuracy score (in percentage) of each processing stage showed differentiated correlation characteristics. Specifically, the processing time of each stage was quantified by a bar chart, and the accuracy score was represented by a line chart: the video frame processing stage had a short processing time (approximately 2.5 seconds) and the lowest accuracy score (90.5%), reflecting the lightweight nature of the initial preprocessing stage and the limitations of basic feature extraction; the hazard identification network stage improved the processing time to 8.2 seconds, and the accuracy score rapidly climbed to 95.2%, reflecting the efficiency and accuracy of the encoder-decoder architecture in feature extraction and preliminary hazard area identification; the region growing optimization stage reduced the processing time to 4.3 seconds, and the accuracy score reached a peak of 98.5%, indicating that the region growing algorithm achieved precise refinement of the hazard area contour through seed point expansion and feature similarity matching; the 3D point cloud construction, 3D... The processing time for stages such as corner meshing increased significantly (approximately 12.8 seconds and 9.6 seconds respectively), and the accuracy score rebounded after a period of decline. This is related to the increased computational complexity of 3D reconstruction and mesh generation, and also reflects the support of multi-view matching and Delaunay triangulation techniques for the accuracy of spatial modeling of hazard areas. The path optimization and sorting stage took 15 seconds (the longest in the entire process), with the corresponding accuracy score rebounding to 97.8%, indicating that the heuristic search algorithm for path optimization consumes high computational resources while ensuring the rationality of the scanning path. The processing time for the control command injection stage dropped to 3.8 seconds, and the accuracy score finally reached 99.2%, reflecting the lightweight nature of the command encapsulation stage and the final convergence of accuracy throughout the entire process. Overall, the dynamic trade-off between "processing time and accuracy" in the fire hazard identification process is clearly presented. The resource consumption and performance output of each stage match the functional positioning of its technical links: the preprocessing and command stages are lightweight and efficient, the core identification and 3D modeling stages embody the characteristic of "trading time for accuracy", and the path optimization stage is the balance point between computational complexity and accuracy assurance.

[0055] Example 5: In specific implementation, the operation of injecting control commands involves setting control points at equal intervals along the optimized and sorted scanning path. The spacing between control points is determined based on the positioning accuracy and motion stability requirements of the actuator, typically set between 1 / 2 and 1 / 5 of the scanning line spacing. Each control point is assigned a speed value, acceleration value, and sensor trigger command. The speed value is dynamically adjusted according to the path curvature, using higher speeds on straight sections and lower speeds on curved sections to ensure scanning quality. The acceleration value is limited based on the inertial characteristics of the actuator to avoid sudden starts and stops. The sensor trigger command precisely controls the laser scanner or infrared thermal imager to begin data acquisition at a specific spatial location. All control points and their corresponding commands are sequentially encapsulated into a data packet. The data packet uses a layered structure, containing a header, a control point sequence, and a footer. The header records the total number of points on the path, a timestamp, and coordinate system information. Each control point contains three-dimensional coordinates, speed, acceleration, and command code. The footer contains a checksum for verifying data transmission integrity. This data packet constitutes the final hazard scanning trajectory and can be directly sent to the actuator via industrial Ethernet or real-time bus protocols.

[0056] In practical implementation, the hazard identification network training process adopts an incremental learning strategy. Initial training uses a large-scale public dataset for network parameter pre-training. This dataset contains millions of labeled fire scene images, covering various hazard types such as fire, smoke, and obstruction of fire-fighting facilities. The pre-training process employs a stochastic gradient descent algorithm, optimizing network weights by minimizing the cross-entropy loss function. When the system is deployed to a new scene, new labeled data for that scene is collected. This new labeled data is obtained through a manual annotation platform, where annotators use rectangular or polygonal tools to accurately mark hazard areas in the images, while simultaneously recording the hazard category and confidence level. The pre-trained network is periodically fine-tuned using the new labeled data. Fine-tuning employs a small learning rate, updating only the parameters of the last few layers of the network to avoid catastrophic forgetting of pre-trained knowledge. The fine-tuned network parameters are saved, and the hazard identification network is updated. The update operation uses hot-swapping technology, loading the new parameter file into the inference engine without interrupting system operation.

[0057] In some embodiments, the control point velocity planning employs an S-shaped velocity curve algorithm. This algorithm achieves continuous acceleration variation through a seven-segment acceleration / deceleration planning process, ensuring smooth motion. Acceleration value allocation needs to consider the torque limitations and load characteristics of the actuator. The maximum allowable acceleration is calculated using the following formula: ; in: Indicates the maximum permissible acceleration. This indicates the maximum output torque of the motor. This indicates the frictional force of the guide rail. Indicates the radius of the transmission mechanism. This indicates the mass of the moving parts. The sensor trigger command is implemented using a hardware interrupt method. When the motion controller reaches the designated control point, it sends a pulse signal through the digital output module to trigger the external sensor to collect data.

[0058] It is understandable that incremental learning strategies need to address the catastrophic forgetting problem. The elastic weight consolidation method can protect network parameters important for older tasks during the learning process for new tasks. This method quantifies parameter importance by calculating the Fisher information matrix of the parameters on older tasks. The quality of newly labeled data directly affects model performance improvement, requiring a multi-level review mechanism, including annotation by junior annotators, verification by senior annotators, and expert sampling. The frequency of network fine-tuning needs to be adjusted based on the degree of difference in the distribution of data in the new scene. The fine-tuning frequency should be increased when the new scene differs significantly from the original training scene, and decreased when the difference is small to save computational resources. Optionally, control point data packets can use differential coding compression technology, storing only the coordinate differences between adjacent control points instead of absolute coordinates, significantly reducing data volume. Sensor trigger commands can integrate multiple working modes, supporting the simultaneous acquisition of data from multiple sensor devices at the same control point, achieving multimodal data acquisition. The incremental learning process can be combined with an active learning strategy, allowing the model to automatically select the unlabeled samples with the most information for priority manual annotation, improving data utilization efficiency. Before updating the network, a comprehensive test must be conducted on the validation set of the new scenario. Only models that meet the thresholds for both accuracy and recall will be deployed to the production environment.

[0059] In some embodiments, the data packet format of the final hazard scanning trajectory is compatible with the ISO14649 standard, supporting direct parsing and execution by various CNC systems. The interpolation algorithm between control points employs non-uniform rational B-spline curve interpolation to achieve a smooth path with high-order continuity. Incremental learning of the hazard identification network can integrate knowledge distillation techniques, allowing the new model to simultaneously learn the output distribution of both newly labeled data and the old model, further improving model stability. It is understood that the synchronization accuracy of control point commands depends on the clock synchronization mechanism between the motion controller and the sensor system; using the IEEE 1588 precise time protocol can achieve microsecond-level time synchronization accuracy. During incremental learning, it is necessary to monitor the performance balance of the model on new and old data. When the model's performance improves on new data but significantly deteriorates on old data, the regularization strength of the elastic weight consolidation method needs to be adjusted. During hot switching of network parameters, it is necessary to ensure the atomicity and transactionality of memory operations to prevent inconsistent inference results during parameter updates.

[0060] Optionally, the control point data packet can include redundancy check information, with each control point appended with a cyclic redundancy check code to prevent errors during data transmission. Sensor triggering commands can support both soft and hard triggering modes. Soft triggering is implemented through software commands, while hard triggering uses dedicated hardware signals for higher-precision synchronization. The incremental learning strategy can be designed with a model rollback mechanism, automatically reverting to the previous version when the performance of the new model on the validation set is lower than that of the old model. The network training process can display the loss function curve and evaluation metric trends in real time, providing visual feedback for hyperparameter adjustment.

[0061] See Figure 5 In the S-shaped velocity curve algorithm for control point velocity planning, the smoothness of the motion process is achieved through a seven-segment acceleration / deceleration planning process that continuously changes the acceleration. Specifically, the velocity curve (blue curve) exhibits a smooth "S"-shaped transition, corresponding to the continuous changes in the acceleration curve (red curve) at different stages. In the initial stage, the acceleration gradually increases (the slope of the velocity curve increases), then enters a stable acceleration stage (the velocity curve rises linearly), and then transitions to a stage where the acceleration gradually decreases (the slope of the velocity curve slows down). Afterward, the velocity remains stable, and the deceleration stage replicates this process in reverse (the acceleration first increases negatively, then stabilizes, and finally decreases). The core advantage of this curve is that by continuously changing the acceleration, it avoids abrupt starts and stops during motion, can match the torque limitations and load characteristics of the actuator, and ensures motion smoothness and scanning quality.

[0062] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0063] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for rapid identification of fire hazards using AI vision technology, characterized in that, The method includes the following operations: Continuously acquire scene video streams through image acquisition devices deployed in fire-fighting scenarios; Frame extraction and denoising are performed on the scene video stream to obtain a sharpened image sequence; The clarified image sequence is input into the hazard identification network, and the network outputs preliminary identification results containing potential hazard areas; The preliminary identification results are processed by region growing to obtain accurate contour information of the potential hazard area; A three-dimensional point cloud model of the potential hazard area is constructed based on the precise contour information of the hazard area. A dense cluster of scanning paths is generated on the surface of the 3D point cloud model; The scanning path clusters are optimized and sorted, and control commands are injected to form the final hazard scanning trajectory.

2. The method for rapid identification of fire hazards using AI vision technology as described in claim 1, characterized in that, The specific operations of frame extraction and denoising processing of the scene video stream include: Set an extraction time interval and extract keyframe images from the scene video stream; Brightness equalization correction is performed on the extracted keyframe images; A nonlocal means denoising algorithm is used to eliminate noise interference in keyframe images; The denoised keyframe images are arranged in chronological order to form the sharpened image sequence.

3. The method for rapid identification of fire hazards using AI vision technology as described in claim 1, characterized in that, The operation of inputting the clarified image sequence into the hazard identification network specifically includes: The hazard identification network adopts an encoder-decoder architecture, with the encoder part using a deep residual network to extract multi-scale image features. The decoder part gradually restores the feature map resolution through upsampling and skip connections; At the network output, the Softmax function is used to calculate the probability that each pixel belongs to a potential danger zone. The set of pixels with a probability greater than a set threshold is taken as the preliminary identification result.

4. The method for rapid identification of fire hazards using AI vision technology as described in claim 1, characterized in that, The operation of performing region growing on the preliminary identification results specifically includes: The high-probability pixels in the preliminary identification results are used as seed points; Calculate the feature similarity of pixels surrounding the seed point, whereby the feature similarity includes color, texture, and gradient information; Pixels whose feature similarity meets the growth criteria are merged into the current potential hazard area; The growth process is iteratively executed until no new pixels can be merged, thereby obtaining the precise contour information of the potential hazard area.

5. The method for rapid identification of fire hazards using AI vision technology as described in claim 1, characterized in that, The specific operations for constructing the 3D point cloud model of the potential hazard area include: Use a binocular vision camera to acquire multi-view image pairs of the potential hazard area; Stereo matching is performed on the multi-view image pairs to calculate the disparity map; The three-dimensional coordinates of the surface points in the potential hazard area are obtained by converting the camera intrinsic parameters and the parallax map. All three-dimensional coordinate points are registered and fused to construct the complete three-dimensional point cloud model.

6. The method for rapid identification of fire hazards using AI vision technology as described in claim 1, characterized in that, The operation of generating a dense cluster of scan paths on the surface of the 3D point cloud model specifically includes: The three-dimensional point cloud model is triangulated to obtain the model surface mesh; Calculate the normal field of the surface mesh, and plan the initial scan line direction based on the direction of the normal field; Multiple scan lines are generated in parallel along the initial scan line direction with a fixed line spacing to form the dense scan path cluster.

7. The method for rapid identification of fire hazards using AI vision technology as described in claim 1, characterized in that, The operation of optimizing and sorting the scan path cluster specifically includes: Calculate the length and curvature of each path in the scan path cluster; Establish a cost function for path switching, which takes into account changes in distance and angle between paths; A heuristic search algorithm is used to find the path access sequence with the lowest total cost, and the order of scanning path clusters is rearranged according to the path access sequence.

8. The method for rapid identification of fire hazards using AI vision technology as described in claim 1, characterized in that, The specific operations of the injection control command include: Control points are set at equal intervals on the optimized and sorted scanning path, and each control point is assigned a velocity value, an acceleration value, and a sensor trigger command. All control points and their corresponding instructions are sequentially encapsulated into a data packet, which is the final hazard scanning trajectory and can be directly sent to the execution agency.

9. The method for rapid identification of fire hazards using AI vision technology as described in claim 1, characterized in that, The training process of the hazard identification network adopts an incremental learning strategy, specifically including: Initial training uses a large-scale public dataset for network parameter pre-training; When the system is deployed to a new scenario, it collects new labeled data for that scenario; The pre-trained network is fine-tuned periodically using new labeled data, and the fine-tuned network parameters are saved to update the hazard identification network.

10. A rapid fire hazard identification system utilizing AI vision technology, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for rapid identification of fire hazards using AI vision technology as described in any one of claims 1 to 9.

Images