General-purpose emergency panoramic command and intelligent decision system

Abstract

The application discloses a general-purpose emergency panoramic command and intelligent decision system, relates to the technical field of emergency command, and comprises a space-air-ground collaborative sensing network unit, a multi-source data fusion and intelligent analysis unit and a dynamic sand table and collaborative decision unit. The application realizes cm-level-precision panoramic sensing and real-time environment modeling at a disaster site by constructing a space-air-ground collaborative sensing network, can quickly identify a disaster target and generate an interactive semantic map based on multi-source data fusion and intelligent analysis, realizes multi-terminal situation synchronization and intelligent preplan generation through a three-dimensional intelligent sand table and a collaborative decision mechanism, and thus the speed, precision and collaborative efficiency of emergency response are improved.

Description

Technical Field

[0001] This invention relates to the field of emergency command technology, and more specifically, to a general-purpose emergency panoramic command and intelligent decision-making system. Background Technology

[0002] With the accelerated pace of modern urban development, emergency response to natural disasters, accidents, and other emergencies faces increasingly complex challenges. At disaster relief sites, comprehensive, multi-dimensional, and real-time situational awareness and intelligent decision-making have become core requirements for emergency management. Unmanned aerial vehicle (UAV) swarms, with their flexible and mobile characteristics, can achieve rapid, large-scale patrols and high-precision data collection; vehicle-mounted mobile platforms equipped with various sensors can achieve refined perception of localized areas; and cloud-edge collaborative intelligent analysis technology provides a new technological path for multi-source data fusion and rapid decision-making. The integrated application of these technologies lays a solid foundation for building a new emergency command system that integrates panoramic perception, intelligent analysis, and collaborative decision-making, driving the transformation and upgrading of emergency management towards digitalization and intelligence.

[0003] However, existing technologies suffer from limitations such as a single perception dimension and insufficient data fusion capabilities, making it difficult to achieve centimeter-level accuracy in panoramic environmental perception. The lack of effective spatiotemporal alignment and cross-modal fusion mechanisms for multi-source heterogeneous data leads to low efficiency in disaster target identification. Furthermore, the two-dimensional planar command mode cannot support three-dimensional dynamic simulations and real-time multi-terminal collaboration, thus hindering the accuracy and speed of emergency decision-making.

[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a general-purpose emergency panoramic command and intelligent decision-making system. This system solves the problems mentioned in the background section regarding the single dimension of presence perception and insufficient data fusion capabilities, making it difficult to achieve centimeter-level precision in panoramic environmental perception. Furthermore, the lack of effective spatiotemporal alignment and cross-modal fusion mechanisms for multi-source heterogeneous data leads to low efficiency in disaster target identification. Simultaneously, the two-dimensional planar command mode cannot support three-dimensional dynamic simulation and real-time multi-terminal collaboration, thus limiting the accuracy and response speed of emergency decision-making.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A general-purpose emergency panoramic command and intelligent decision-making system, including:

[0008] The air-space-ground collaborative sensing network unit is used to acquire multi-source data of the entire disaster site, generate panoramic digital surface models and vehicle surrounding environment models based on multi-source data, and use dynamic path planning algorithms to achieve automatic obstacle avoidance and route optimization.

[0009] The multi-source data fusion and intelligent analysis unit is used to perform spatiotemporal alignment and cross-modal fusion of multi-source data using spatiotemporal alignment algorithms, and to complete disaster target identification and scene semantic segmentation through anomaly target detection model and panoramic semantic segmentation model, generating interactive semantic maps and multi-dimensional feature vectors.

[0010] The dynamic sand table and collaborative decision-making unit are used to construct a three-dimensional intelligent sand table based on semantic maps and multi-dimensional feature vectors, and overlay a dynamic data layer. Through contingency plan generation and resource scheduling, decision-making schemes are formed to achieve multi-terminal sand table status synchronization, collaborative annotation, and automatic triggering and push of alarm work orders.

[0011] Furthermore, the air-space-ground collaborative sensing network unit includes:

[0012] The UAV swarm full-domain scanning module is used to scan the entire disaster site using five-lens oblique photography, infrared thermal imaging and multispectral sensors carried by the UAV, in order to generate a panoramic digital surface model and identify thermal anomaly targets and chemical pollutant outlines.

[0013] The vehicle-mounted command vehicle's environmental perception module integrates lidar, surround-view cameras, and weather sensors to build a real-time model of the vehicle's surrounding environment and run dynamic path planning algorithms to achieve automatic obstacle avoidance and route optimization based on changes in the on-site environment.

[0014] Among them, the air-space-ground collaborative perception network unit is connected to the multi-source data fusion and intelligent analysis unit and the dynamic sand table and collaborative decision-making unit, and is connected to the UAV cluster full-domain scanning module and the vehicle-mounted command vehicle environmental perception module.

[0015] Furthermore, integrating LiDAR, surround-view cameras, and weather sensors, a real-time model of the vehicle's surrounding environment is constructed, and a dynamic path planning algorithm is run to achieve automatic obstacle avoidance and route optimization based on changes in the on-site environment, including:

[0016] Data is acquired from LiDAR, surround-view cameras, and weather sensors. A real-time model of the vehicle's surrounding environment is constructed based on this data, and the parameters of the dynamic path planning algorithm are initialized, with a maximum number of iterations set. Random expansion exploration is performed using a preset adaptive step size, while simultaneously detecting collisions between the expanded path and obstacles in the vehicle's surrounding environment model. An adaptive gridded model is applied to the obstacle-free area, and a search algorithm is used to quickly search for an effective path connecting the starting point and the target point within the feasible area. Node redundancy removal and segmentation are performed on the effective path, and a high-order Bézier curve equation is used to smooth the path and improve vehicle tracking performance. Combining real-time data from the vehicle's surrounding environment model, a preset evaluation function is used to perform multi-dimensional evaluation and optimization of the effective path to ensure its safety and feasibility. If the maximum number of iterations is reached, the final optimized path is executed, and automatic obstacle avoidance and route optimization are implemented based on changes in the on-site environment; otherwise, iterative optimization continues.

[0017] Furthermore, adaptive gridded modeling is performed on the barrier-free area, and a search algorithm is used to quickly search for effective paths connecting the starting point and the target point within the feasible area, including:

[0018] Adaptive raster modeling is performed on the accessible area. A candidate path set is initialized, and multiple initial path schemes are randomly generated within the feasible area, with their path lengths evaluated. The parameters and maximum number of iterations of the search algorithm are initialized. An existing path is selected from the candidate path set, and its node sequence is locally optimized and adjusted according to a preset adjustment probability. The feasibility of the locally optimized existing path is verified to ensure that the path is completely within the accessible area and meets the minimum turning radius, thus obtaining a new path. The quality of the new path is compared with the original paths in the candidate path set. If the new path is optimal, the worst path in the candidate path set is replaced, and the candidate path set is updated. If the maximum number of iterations is reached, the optimal path in the candidate path set is output as the valid path connecting the starting point and the target point; otherwise, the path optimization and update process continues.

[0019] Furthermore, the formula for locally optimizing the node sequence of the existing path according to the preset adjustment probability is as follows:

[0020] ;

[0021] In the formula, P a This represents the node sequence of the existing path. a The coordinates of each node; This represents the node sequence of the existing path after local optimization and adjustment. a The coordinates of each node; e Represents extremely small positive numbers; or Indicates the random disturbance intensity coefficient; l Indicates the curvature optimization strength coefficient; d Indicates a Gaussian distribution; s 2 Indicates variance; r Indicates the preset adjustment probability; r 1. r 2 represents a random number.

[0022] Furthermore, the multi-source data fusion and intelligent analysis unit includes:

[0023] The multimodal data fusion engine module is used to unify UAV, vehicle sensor and geographic information data into the same spatiotemporal coordinate system through spatiotemporal alignment algorithm, and to use graph neural network to extract cross-modal features to construct multi-dimensional feature vectors of disaster scenarios;

[0024] The panoramic image anomaly recognition module is used to identify disaster targets based on the fused multidimensional feature vectors using the YOLOv8 model, and to perform semantic segmentation of disaster scene elements using a panoramic semantic segmentation model, ultimately generating an interactive semantic map.

[0025] The multimodal data fusion engine module and the panoramic image anomaly recognition module are connected.

[0026] Furthermore, a spatiotemporal alignment algorithm is used to unify UAV, vehicle-mounted sensor, and geographic information data into the same spatiotemporal coordinate system, and a graph neural network is used for cross-modal feature extraction to construct a multi-dimensional feature vector for disaster scenarios, including:

[0027] A unified spatiotemporal coordinate system was established to perform spatiotemporal registration of UAV imagery, vehicle-mounted sensor data cloud, and geographic information data. Spatial correspondences between different modal data were established, and visible light image pixels, infrared hotspot regions, and lidar point clouds were associated and mapped. The registered UAV imagery, vehicle-mounted sensor data cloud, and geographic information data were used to construct a graph neural network input structure, with spatial location as nodes and inter-modal feature associations as edges, forming a topological graph representation of the disaster scene. The message passing mechanism of the graph neural network was used to aggregate multimodal feature information within the neighborhood of nodes, and a deep fusion representation of visible light texture, infrared thermal radiation, and lidar geometric features was learned. The node feature representation was iteratively optimized through multi-layer graph convolution operations to eliminate feature conflicts and enhance complementary information, ultimately constructing a multi-dimensional feature vector of the disaster scene.

[0028] Furthermore, based on the fused multidimensional feature vectors, the YOLOv8 model is used to identify disaster targets, and a panoramic semantic segmentation model is employed to perform semantic segmentation on disaster scene elements, ultimately generating an interactive semantic map including:

[0029] The YOLOv8 model is iteratively optimized using a few-shot incremental learning method to obtain an optimized YOLOv8 model. The fused multi-dimensional feature vectors are then input into the optimized YOLOv8 model and the panoramic semantic segmentation model, respectively. Multi-scale features are extracted through the feature pyramid network of the optimized YOLOv8 model to complete the identification of preset disaster targets. The hierarchical attention mechanism of the panoramic semantic segmentation model is used to parse the global context relationship and obtain pixel-level scene understanding. The feature map spatial resolution is restored through the segmentation decoder, and the segmentation mask of disaster scene elements is output. The target recognition and semantic segmentation results are fused, and geographic information is overlaid to generate an interactive semantic map.

[0030] Furthermore, the YOLOv8 model is iteratively optimized using the few-shot incremental learning method, resulting in the optimized YOLOv8 model, which includes:

[0031] Initialize the parameters of the few-shot incremental learning algorithm and set the maximum number of iterations. Perform a preliminary evaluation of the YOLOv8 model based on the initial few-shot dataset, and select a set of candidate model parameters to be optimized based on the detection accuracy. If the current YOLOv8 model accuracy meets the preset threshold, output the optimization result directly; otherwise, proceed to the incremental learning process. Construct a training dataset using newly added few-shot data and label the samples as positive and negative examples to form the labeled data required for incremental learning. If it is the first incremental learning iteration, construct a base classifier; otherwise, use the newly added labeled data to incrementally update the parameters of the existing classifier. Use the updated classifier to filter the YOLOv8 model parameter space, calculate the correlation of YOLOv8 model parameters, and select the optimal parameter combination to update the YOLOv8 model. Repeat the incremental learning process until the maximum number of iterations is reached, and finally output the optimized YOLOv8 model.

[0032] Furthermore, the dynamic sand table and collaborative decision-making unit include:

[0033] The 3D intelligent sand table construction module is used to develop real-time rendering technology based on game engines. It integrates semantic maps and multi-dimensional feature vectors to construct a 3D sand table, and overlays dynamic data layers of disaster situation, resource distribution and risk warning to achieve panoramic roaming and multi-scale detailed display.

[0034] The intelligent decision engine module is used to automatically match historical case databases based on disaster type and level to generate graded response plans, and optimize rescue routes and material allocation through operations research models to form a decision plan that takes into account time cost, risk coefficient and resource utilization.

[0035] The cross-platform collaboration module is used to realize real-time synchronization of the sandbox status across multiple terminals, support collaborative annotation operations by users, and automatically trigger alarm work order push and processing progress tracking when abnormal events occur.

[0036] The 3D intelligent sand table construction module is connected through the intelligent decision engine module and the cross-platform collaboration module.

[0037] The beneficial effects of this invention are as follows:

[0038] 1. This invention achieves centimeter-level precision panoramic perception and real-time environmental modeling at disaster sites by constructing a space-air-ground collaborative sensing network. Based on multi-source data fusion and intelligent analysis, the system can quickly identify disaster targets and generate interactive semantic maps. Through a 3D intelligent sand table and collaborative decision-making mechanism, it realizes multi-terminal situational synchronization and intelligent contingency plan generation, thereby improving the speed, accuracy, and collaborative efficiency of emergency response.

[0039] 2. This invention achieves high-precision real-time perception of the vehicle's surrounding environment through multi-sensor fusion and adaptive gridded modeling. Employing a path planning algorithm, it improves path quality and vehicle tracking performance through local optimization adjustments and Bézier curve smoothing. Combining a dynamic evaluation function and iterative optimization mechanism ensures the rapid generation of safe and reliable optimal paths in complex disaster environments, thereby effectively enhancing the system's autonomous obstacle avoidance and dynamic response capabilities.

[0040] 3. This invention utilizes a multimodal data fusion engine to achieve spatiotemporal alignment and cross-modal feature extraction of multi-source data, constructing accurate multidimensional feature vectors for disaster scenarios. Based on the YOLOv8 model and panoramic semantic segmentation technology, it achieves accurate identification of disaster targets and fine segmentation of scene elements. Combined with a few-shot incremental learning method, it enhances the model's adaptability to new scenarios, ultimately generating a high-precision, interactive semantic map, providing comprehensive and reliable situational awareness support for emergency decision-making.

[0041] 4. This invention enables multi-dimensional visualization of disaster scenarios through a three-dimensional intelligent sand table, automatically generates the optimal rescue plan by combining an intelligent decision engine, and ensures real-time synchronization and efficient collaboration among multiple terminals by utilizing a cross-platform collaboration module, thereby improving the decision-making efficiency and execution accuracy of emergency command. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a schematic diagram of a general-purpose emergency panoramic command and intelligent decision-making system according to an embodiment of the present invention;

[0044] Figure 2This is a flowchart illustrating the application of the vehicle-mounted command vehicle environmental perception module in a general-purpose emergency panoramic command and intelligent decision-making system according to an embodiment of the present invention.

[0045] In the picture:

[0046] 1. Space-Air-Ground Collaborative Sensing Network Unit; 2. Multi-Source Data Fusion and Intelligent Analysis Unit; 3. Dynamic Sand Table and Collaborative Decision-Making Unit. Detailed Implementation

[0047] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0048] In the description of this invention, unless otherwise stated, "a plurality of" means two or more. Furthermore, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0049] According to embodiments of the present invention, a general-purpose emergency panoramic command and intelligent decision-making system is provided.

[0050] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1-Figure 2 As shown, the general-purpose emergency panoramic command and intelligent decision-making system according to an embodiment of the present invention includes:

[0051] The air-space-ground collaborative sensing network unit 1 is used to acquire multi-source data of the entire disaster site, generate a panoramic digital land surface model and a vehicle surrounding environment model based on the multi-source data, and use dynamic path planning algorithms to achieve automatic obstacle avoidance and route optimization.

[0052] Specifically, the multi-rotor drone is equipped with a five-lens oblique photography system, flies at an altitude of 200 meters, and has a single-unit coverage radius of 3 kilometers. It generates a panoramic digital terrain model (2cm resolution) through real-time image stitching, supporting terrain change detection (e.g., earthquake cracks, landslide displacement, etc.). An infrared thermal imaging channel (sensitivity 0.05℃) identifies thermal anomalies (e.g., smoldering fire spots, thermal signals from personnel), and combined with multispectral imaging (400-1000nm) detects the diffusion profile of chemical pollutants. It integrates a 128-line lidar, a 360° surround-view camera, and meteorological sensors (temperature, humidity, wind speed) to construct a high-precision environmental model within a 200-meter radius around the vehicle in real time (update frequency 10Hz). Utilizing a dynamic path planning algorithm, it automatically optimizes its route based on real-time changes at the disaster site (e.g., road collapse, toxic gas diffusion), with an obstacle avoidance response time of <0.5 seconds.

[0053] The multi-source data fusion and intelligent analysis unit 2 is used to perform spatiotemporal alignment and cross-modal fusion of multi-source data using a spatiotemporal alignment algorithm, and to complete disaster target identification and scene semantic segmentation through an anomaly target detection model and a panoramic semantic segmentation model, generating an interactive semantic map and multi-dimensional feature vectors.

[0054] Specifically, based on the improved YOLOv8 architecture (anomaly target detection model), it supports the identification of six typical disaster targets: building damage (cracks > 5cm, structural tilt > 3°); trapped personnel (attitude recognition plus vital sign thermal signals); fire smoke (RGB plus infrared feature fusion); chemical leaks (multispectral anomalous reflective areas); traffic disruptions (e.g., road cracks, traffic congestion); infrastructure failures (e.g., collapsed power towers, ruptured pipelines). Utilizing a few-shot incremental learning algorithm, the model iteration cycle for new scenarios is shortened from 7 days to 12 hours. The Swin Transformer model (panoramic semantic segmentation model) is used to segment disaster scene elements (e.g., damaged areas, safety passages, hazard sources), generating an interactive semantic map (IoU > 85%). A spatiotemporal alignment algorithm is used to unify UAV imagery, vehicle-mounted sensor data, and geographic information (DEM, building outlines) into the same spatiotemporal coordinate system (error < 0.3m, time synchronization < 100ms). A graph neural network (GNN) is used to connect visible light, infrared, and lidar data to construct a multidimensional feature vector for the disaster scene.

[0055] The dynamic sand table and collaborative decision-making unit 3 are used to construct a three-dimensional intelligent sand table based on semantic maps and multi-dimensional feature vectors and overlay a dynamic data layer. Through pre-plan generation and resource scheduling, a decision-making scheme is formed to realize the synchronization of sand table status across multiple terminals, collaborative annotation, and automatic triggering and push of alarm work orders.

[0056] Specifically, a real-time rendering system based on a game engine is developed, supporting the overlay of hundreds of millions of triangular facet models and dynamic data layers: a disaster situation layer, displaying the degree of damage and pollution spread predictions via heat maps; a resource distribution layer, marking rescue teams and material distribution points in real time; and a risk warning layer, using probabilistic models to predict secondary disasters (e.g., landslides induced by aftershocks). A panoramic roaming function is developed, supporting free switching between first-person and top-down perspectives, with key areas zoomed in to centimeter-level detail. Inputting disaster type and level automatically matches historical case libraries (over 100,000 events) to generate tiered response plans, supporting adjustments via natural language commands (e.g., prioritizing the evacuation of residents in the western area). Rescue routes and material allocation are planned based on an operations research model, with the objective function considering time cost, risk coefficient, and resource utilization. The sand table status is synchronized in real time across multiple terminals (command vehicles, mobile devices, cloud), supporting collaborative annotation by 50+ users (e.g., delineating lockdown zones, marking temporary shelters), with operation latency <200ms. Abnormal events automatically trigger alarm work orders, pushing them to responsible units and tracking the progress of handling.

[0057] In this optional embodiment, the air-space-ground collaborative sensing network unit 1 includes:

[0058] The UAV swarm full-domain scanning module is used to scan the entire disaster site using five-lens oblique photography, infrared thermal imaging and multispectral sensors carried by the UAV, in order to generate a panoramic digital surface model and identify thermal anomaly targets and chemical pollutant outlines.

[0059] The vehicle-mounted command vehicle's environmental perception module integrates lidar, surround-view cameras, and weather sensors to build a real-time model of the vehicle's surrounding environment and run dynamic path planning algorithms to achieve automatic obstacle avoidance and route optimization based on changes in the on-site environment.

[0060] Among them, the air-space-ground collaborative perception network unit 1 is connected to the multi-source data fusion and intelligent analysis unit 2 and the dynamic sand table and collaborative decision-making unit 3, and is connected to the UAV cluster full-domain scanning module and the vehicle-mounted command vehicle environmental perception module.

[0061] In this optional embodiment, integrating LiDAR, surround-view cameras, and weather sensors, a real-time model of the vehicle's surrounding environment is constructed, and a dynamic path planning algorithm is run to achieve automatic obstacle avoidance and route optimization based on changes in the on-site environment, including:

[0062] Data is acquired from LiDAR, surround-view cameras, and weather sensors. A real-time model of the vehicle's surrounding environment is constructed based on this data, and the parameters of the dynamic path planning algorithm are initialized, with a maximum number of iterations set. Random expansion exploration is performed using a preset adaptive step size, while simultaneously detecting collisions between the expanded path and obstacles in the vehicle's surrounding environment model. An adaptive gridded model is applied to the obstacle-free area, and a search algorithm is used to quickly search for an effective path connecting the starting point and the target point within the feasible area. Node redundancy removal and segmentation are performed on the effective path, and a high-order Bézier curve equation is used to smooth the path and improve vehicle tracking performance. Combining real-time data from the vehicle's surrounding environment model, a preset evaluation function is used to perform multi-dimensional evaluation and optimization of the effective path to ensure its safety and feasibility. If the maximum number of iterations is reached, the final optimized path is executed, and automatic obstacle avoidance and route optimization are implemented based on changes in the on-site environment; otherwise, iterative optimization continues.

[0063] Specifically, the system scans the terrain within a 200-meter radius using LiDAR, captures visual information using surround-view cameras, and monitors wind speed and humidity using meteorological sensors, fusing these elements to construct a 3D environmental model that includes obstacles such as building debris and tilted utility poles. A dynamic path planning algorithm is initialized, employing an adaptive step-size strategy for path exploration: initially expanding the search range rapidly with a larger step size, then automatically reducing the step size for fine-tuning as obstacles approach. The algorithm continuously detects the collision probability between each expanded path and obstacles in the environmental model, automatically marking detected collapsed areas as impassable zones. The system performs raster modeling of safe areas and, driven by a harmony search algorithm, selects and optimizes basic paths from a candidate path library: adjusting path node positions with preset probabilities to avoid newly discovered obstacles; and smoothing the adjusted paths using high-order Bézier curves to eliminate sharp turns and ensure smooth vehicle passage. During the planning process, the system continuously receives updated obstacle information from the environmental model and uses multi-dimensional evaluation functions (e.g., comprehensively considering path length, safety factor, and passage difficulty) to dynamically evaluate and optimize the paths. When a sudden road collapse is detected ahead, the system replans an alternative route within 0.5 seconds and iteratively optimizes the new path to ensure its safety and efficiency. This achieves centimeter-level accuracy in real-time obstacle avoidance in complex disaster environments, reducing path planning response time to within 1 second and improving the efficiency and safety of emergency vehicles.

[0064] Specifically, the dynamic path planning algorithm is a fusion of D... The algorithm (an incremental path planning algorithm) and the Rapid Exploration Random Tree (RRT) algorithm are both path planning algorithms. RRT performs rapid random sampling and exploration in an unknown environment to generate an initial feasible path; when the environment changes, it uses D... The algorithm performs local incremental replanning on affected path segments to avoid global recalculation. In this invention, the algorithm expands the search tree in a rasterized environment through adaptive step-size replanning (RRT), while utilizing D... The algorithm dynamically adjusts the path based on the real-time environment model. This is achieved by combining RRT's global exploration capabilities with D... The algorithm's dynamic optimization capabilities enable real-time obstacle avoidance and path optimization in complex disaster environments.

[0065] In this optional embodiment, adaptive gridding modeling is performed on the barrier-free area, and a search algorithm is used to quickly search for an effective path connecting the starting point and the target point within the feasible area, including:

[0066] Adaptive raster modeling is performed on the accessible area. A candidate path set is initialized, and multiple initial path schemes are randomly generated within the feasible area, with their path lengths evaluated. The parameters and maximum number of iterations of the search algorithm are initialized. An existing path is selected from the candidate path set, and its node sequence is locally optimized and adjusted according to a preset adjustment probability. The feasibility of the locally optimized existing path is verified to ensure that the path is completely within the accessible area and meets the minimum turning radius, thus obtaining a new path. The quality of the new path is compared with the original paths in the candidate path set. If the new path is optimal, the worst path in the candidate path set is replaced, and the candidate path set is updated. If the maximum number of iterations is reached, the optimal path in the candidate path set is output as the valid path connecting the starting point and the target point; otherwise, the path optimization and update process continues.

[0067] Specifically, the system adaptively rasterizes the scanned barrier-free area, dividing the feasible region into regular grid cells. It then initializes a candidate set containing multiple random paths, each composed of connected grid nodes, and records its path length as an initial evaluation metric. During path optimization, the system employs a harmony search strategy: each time, an existing path is selected from the candidate set, and its node sequence is intelligently adjusted with a preset probability. This includes displacement optimization of key nodes in the path, maintaining path continuity while avoiding obstacle edges. The adjusted path undergoes rigorous verification to ensure all nodes are within the barrier-free grid and that the turning angle meets the vehicle's minimum turning radius requirement. The verified new path is then compared with existing paths in the candidate set in a multi-dimensional quality assessment. The system establishes a comprehensive evaluation system including path length, smoothness, and safety factor. When the quality of the new path is better than the worst path in the set, it is replaced and updated. The entire process is iterative, with the candidate path set continuously evolving and gradually improving its overall quality. After a preset number of iterations, the system finally outputs the path with the highest comprehensive score in the candidate set as the optimal solution. This route ensures both traffic safety and maximizes travel efficiency.

[0068] Specifically, the search algorithm is a harmony search algorithm, a metaheuristic optimization algorithm that simulates the process by which musicians fine-tune the pitch of their instruments to achieve perfect harmony in music composition. It maintains a harmony memory bank (a set of candidate solutions) and iteratively optimizes the quality of solutions by combining memory considerations, pitch fine-tuning, and a random innovation mechanism. In this invention, the algorithm treats each path scheme as a harmony. New paths are generated by randomly selecting existing paths from the memory bank and locally optimizing them with a preset probability (e.g., node sequence mutation). The quality of the new paths is then evaluated; if a path is better, the worst path in the memory bank is replaced. This process is iteratively updated until the optimal effective path that meets the conditions is obtained.

[0069] In this optional embodiment, the formula for locally optimizing the node sequence of the existing path according to a preset adjustment probability is as follows:

[0070] ;

[0071] In the formula, P a This represents the node sequence of the existing path. a The coordinates of each node; This represents the node sequence of the existing path after local optimization and adjustment. a The coordinates of each node; e To represent extremely small positive numbers, preventing the denominator from being zero; or This represents the random disturbance intensity coefficient, which controls the overall magnitude of the exploratory adjustment; l This represents the curvature optimization intensity coefficient, which controls the magnitude of path smoothness adjustment; d Indicates a Gaussian distribution; s 2 Indicates variance; r This indicates the preset adjustment probability, which controls whether each optimization component is activated. r 1. r 2 represents random numbers, that is, two uniformly distributed random numbers generated independently in the interval [0, 1].

[0072] It should be clarified that the node sequence of the existing path in this invention is not ordered according to the time dimension, but rather arranged sequentially according to the spatial travel path of the vehicle from the starting point to the target point. The construction and application of this node sequence is primarily aimed at optimizing the physical travel path of the vehicle-mounted command vehicle at the disaster site, serving the spatial topology construction needs of dynamic path planning, rather than recording and representing the vehicle's position information at different time points. When implementing local optimization adjustments to the existing path node sequence using local optimization adjustment formulas, a curvature optimization intensity coefficient needs to be introduced. l In the calculation, curvature, as the core geometric feature parameter of the spatial path, is included in the formula. It means that by precisely iteratively adjusting the node coordinates, the optimized node sequence is fitted to form a smooth and continuous spatial path, thereby strictly matching the spatial geometric constraints of the minimum turning radius of the vehicle, ensuring the feasibility of the planned path in physical space and the stability of the vehicle's tracking process.

[0073] In practical applications r 1. r 2 represents pseudo-random numbers based on prior data of disaster site environmental characteristics. The seed value for this generation is dynamically determined by real-time environmental data (such as obstacle density, road curvature, and weather conditions) collected by the environmental perception module of the vehicle-mounted command vehicle, and is not generated randomly without any rules. During the iteration process, r 1. rThe value range of 2 is dynamically adjusted according to the complexity of the environment: when the disaster site environment is simple (unobstructed / low curvature), the range is reduced. r 1. r 2. Value range: Reduce the magnitude of optimization adjustments; when the environment is complex (multiple obstacles / high curvature), expand the value range to enhance exploratory adjustment capabilities; the activation of optimization components is not solely determined by... r 1. r The decision is not made by 2, but by a preset adjustment probability. r and r 1. r The value of 2 is controlled by comparison results (the preset adjustment probability is a fixed threshold calibrated based on engineering practice, ranging from 0.6 to 0.8), and only when... r > r 1 / r > r Only when the time is 2 is the corresponding optimization component activated, fundamentally ensuring the controllability of the optimization component activation; during model iteration, a multi-dimensional evaluation function is set to evaluate the path after each optimization in real time, even if r 1. r 2. There are slight fluctuations, and the evaluation function will also filter the iteration results, retaining only the optimization results that meet the criteria of "path safety, smoothness, and efficiency" to ensure the stability of the model iteration effect.

[0074] In the aforementioned local optimization adjustment of the path node sequence, the core comprises two independent optimization components, which together constitute the adjustment factor for the node coordinates, specifically:

[0075] Random disturbance component: derived from random disturbance intensity coefficient or Gaussian distribution d ,variance s 2 Together, they serve to perform small-scale random exploration of node coordinates, preventing path planning from getting trapped in local optima.

[0076] Curvature optimization component: derived from curvature optimization intensity coefficient l Its function is to smoothly adjust the node coordinates, ensuring that the curvature of the path meets the minimum turning radius requirement of the vehicle-mounted command vehicle, and improving the tracking performance.

[0077] r 1. r 2 represents an independent control factor for two optimization components, corresponding to the random perturbation component and the curvature optimization component, respectively. Combined with a preset adjustment probability, this factor enables precise control of the optimization variables. The specific logic is as follows:

[0078] r 1. Control of random disturbance components: r The value of 1 is positively correlated with the amplitude of the random disturbance component. r >r At time 1, the random perturbation component is activated. r The smaller the value of 1, the greater the amplitude of the random perturbation (the stronger the exploratory nature); when r ≤ r At time 1, random perturbation components are not activated, and node coordinates are not randomly explored and adjusted;

[0079] r 2. Control of curvature optimization components: r The value of 2 is negatively correlated with the intensity of the curvature optimization component. r > r At time 2, the curvature optimization component is activated. r The smaller the value of 2, the stronger the curvature optimization (the smoother the path); when r ≤ r At time 2, the curvature optimization component is not activated, and the node coordinates maintain the original curvature;

[0080] In the above control logic, r These are fixed thresholds calibrated in engineering practice based on the driving performance of emergency vehicles and the route planning requirements at disaster sites. Technical personnel in the relevant technical field can fine-tune them according to actual application scenarios, making them clearly feasible.

[0081] In this optional embodiment, the multi-source data fusion and intelligent analysis unit 2 includes:

[0082] The multimodal data fusion engine module is used to unify UAV, vehicle sensor and geographic information data into the same spatiotemporal coordinate system through spatiotemporal alignment algorithm, and to use graph neural network to extract cross-modal features to construct multi-dimensional feature vectors of disaster scenarios;

[0083] The panoramic image anomaly recognition module is used to identify disaster targets based on the fused multidimensional feature vectors using the YOLOv8 model, and to perform semantic segmentation of disaster scene elements using a panoramic semantic segmentation model, ultimately generating an interactive semantic map.

[0084] The multimodal data fusion engine module and the panoramic image anomaly recognition module are connected.

[0085] In this optional embodiment, the UAV, vehicle-mounted sensor, and geographic information data are unified to the same spatiotemporal coordinate system through a spatiotemporal alignment algorithm, and cross-modal feature extraction is performed using a graph neural network to construct a multidimensional feature vector for the disaster scenario, including:

[0086] A unified spatiotemporal coordinate system was established to perform spatiotemporal registration of UAV imagery, vehicle-mounted sensor data cloud, and geographic information data. Spatial correspondences between different modal data were established, and visible light image pixels, infrared hotspot regions, and lidar point clouds were associated and mapped. The registered UAV imagery, vehicle-mounted sensor data cloud, and geographic information data were used to construct a graph neural network input structure, with spatial location as nodes and inter-modal feature associations as edges, forming a topological graph representation of the disaster scene. The message passing mechanism of the graph neural network was used to aggregate multimodal feature information within the neighborhood of nodes, and a deep fusion representation of visible light texture, infrared thermal radiation, and lidar geometric features was learned. The node feature representation was iteratively optimized through multi-layer graph convolution operations to eliminate feature conflicts and enhance complementary information, ultimately constructing a multi-dimensional feature vector of the disaster scene.

[0087] Specifically, the system establishes a unified spatiotemporal coordinate system, precisely registering oblique images captured by UAVs, point cloud data scanned by vehicle-mounted LiDAR, and basic geographic information data. Through coordinate transformation and temporal alignment, it ensures that all data sources remain synchronized in spatial location and timestamp, laying the foundation for subsequent fusion processing. The system then establishes spatial correspondences between multimodal data: precisely associating building outline pixels in visible light images, temperature anomaly areas detected by infrared thermal imaging, and 3D point clouds acquired by LiDAR. This cross-modal mapping enables different attribute features of the same geographical location to complement each other. Based on the registered multi-source data, the system constructs a graph neural network input structure: each spatial sampling point is a node, and the node features include the image texture, thermal radiation intensity, and 3D coordinate information of that location; adjacent spatial relationships and cross-modal feature associations are used as edges to form a complete disaster scene topology map. In the graph neural network processing stage, multimodal features of the surrounding neighborhood of each node are aggregated through a message passing mechanism. Nodes iteratively fuse visible light texture features, infrared thermal features, and point cloud geometric features through multi-layer graph convolution operations, achieving deep interaction and complementarity of cross-modal information at the feature level. After multiple rounds of graph convolution optimization, the system finally outputs a disaster scene representation vector containing 128-dimensional features. This vector comprehensively reflects the visual, thermodynamic, and three-dimensional structural features of the scene.

[0088] In this optional embodiment, based on the fused multidimensional feature vectors, the YOLOv8 model is used to identify disaster targets, and a panoramic semantic segmentation model is used to perform semantic segmentation on disaster scene elements, ultimately generating an interactive semantic map, including:

[0089] The YOLOv8 model is iteratively optimized using a few-shot incremental learning method to obtain an optimized YOLOv8 model. The fused multi-dimensional feature vectors are then input into the optimized YOLOv8 model and the panoramic semantic segmentation model, respectively. Multi-scale features are extracted through the feature pyramid network of the optimized YOLOv8 model to complete the identification of preset disaster targets. The hierarchical attention mechanism of the panoramic semantic segmentation model is used to parse the global context relationship and obtain pixel-level scene understanding. The feature map spatial resolution is restored through the segmentation decoder, and the segmentation mask of disaster scene elements is output. The target recognition and semantic segmentation results are fused, and geographic information is overlaid to generate an interactive semantic map.

[0090] Specifically, the system incrementally learns and optimizes the YOLOv8 model using small sample data collected on-site, adapting it to the unique environmental characteristics of the disaster area. Then, the 128-dimensional feature vector obtained from the fusion of multi-source data is input into both the optimized YOLOv8 model and the panoramic semantic segmentation model. The YOLOv8 model, through its feature pyramid network, analyzes the feature vectors at different scales, accurately identifying six types of disaster targets, including building damage and trapped personnel, and annotating their location information. Simultaneously, the semantic segmentation model employs a hierarchical attention mechanism to comprehensively analyze the global context of the scene, achieving fine-grained classification of each pixel. The segmentation decoder progressively upsamples the processed feature map to restore it to the original image resolution, outputting a segmentation mask that accurately identifies damaged areas, safety passages, and other elements. Finally, the system fuses the target recognition results with the semantic segmentation mask, overlays geographic coordinate information, and generates an interactive semantic map containing the locations of disaster targets and the distribution of scene elements.

[0091] In this optional embodiment, the YOLOv8 model is iteratively optimized using a few-shot incremental learning method to obtain the optimized YOLOv8 model. This includes: initializing the parameters of the few-shot incremental learning algorithm and setting the maximum number of iterations; performing a preliminary evaluation of the YOLOv8 model based on the initial few-shot dataset; selecting a set of candidate model parameters to be optimized based on the detection accuracy; if the current YOLOv8 model accuracy meets a preset threshold, the optimization result is directly output; otherwise, the incremental learning process begins; a training dataset is constructed using newly added few-shot data, and positive and negative examples are labeled in the dataset to form the labeled data required for incremental learning; if it is the first incremental learning iteration, a basic classifier is constructed; otherwise, the existing classifier is incrementally updated using the newly added labeled data; the updated classifier is used to filter the YOLOv8 model parameter space, the correlation of the YOLOv8 model parameters is calculated, and the optimal parameter combination is selected to update the YOLOv8 model; the incremental learning process is repeated until the maximum number of iterations is reached, and finally, the optimized YOLOv8 model is output.

[0092] Specifically, the system initializes learning parameters and evaluates the existing model based on over a hundred samples collected on-site, selecting network parameters with poor recognition performance in the current scene as the set to be optimized. When the model's recognition rate of collapsed buildings fails to meet the preset standard, the system enters the incremental learning phase. New samples taken by rescue personnel on-site are quickly labeled, distinguishing positive examples such as structural cracks and broken load-bearing columns from negative examples such as normal building debris, thus constructing a targeted training dataset. The system establishes a basic classifier during its initial run, and subsequently uses newly labeled data to incrementally adjust the classifier parameters. The updated classifier performs correlation analysis on key parameters such as the YOLOv8 backbone network and feature pyramid, selecting the parameter combinations most sensitive to the current disaster scenario for targeted optimization. Through multiple rounds of iterative learning, the model gradually adapts to the special environment of the disaster area, improving the detection sensitivity of specific disaster targets in the earthquake zone while maintaining its original recognition capabilities. After a complete optimization cycle, the system outputs an enhanced YOLOv8 model specifically designed for the current disaster scenario.

[0093] Specifically, the few-shot incremental learning algorithm is a biased learning algorithm, a machine learning method designed for data-scarce scenarios. Its core is to construct a biased classifier, which, while retaining existing knowledge, utilizes a small number of new samples to achieve efficient targeted optimization of the model. This algorithm prioritizes updating model components most sensitive to the current task through feature space selection and parameter correlation analysis. In this invention, the algorithm dynamically constructs a biased classifier by continuously inputting small-sample data from disaster sites to select key YOLOv8 parameters. Based on parameter correlation analysis, the optimal parameter combination is selected for targeted updates, enabling the model to quickly adapt to the characteristics of new disaster scenarios while retaining its original recognition capabilities, shortening the model iteration cycle from 7 days to 12 hours.

[0094] In this optional embodiment, the dynamic sandbox and collaborative decision-making unit 3 includes:

[0095] The 3D intelligent sand table construction module is used to develop real-time rendering technology based on game engines. It integrates semantic maps and multi-dimensional feature vectors to construct a 3D sand table, and overlays dynamic data layers of disaster situation, resource distribution and risk warning to achieve panoramic roaming and multi-scale detailed display.

[0096] The intelligent decision engine module is used to automatically match historical case databases based on disaster type and level to generate graded response plans, and optimize rescue routes and material allocation through operations research models to form a decision plan that takes into account time cost, risk coefficient and resource utilization.

[0097] The cross-platform collaboration module is used to realize real-time synchronization of the sandbox status across multiple terminals, support collaborative annotation operations by users, and automatically trigger alarm work order push and processing progress tracking when abnormal events occur.

[0098] The 3D intelligent sand table construction module is connected through the intelligent decision engine module and the cross-platform collaboration module.

[0099] Specifically, in the earthquake disaster emergency command center, the 3D intelligent sand table construction module, based on real-time acquired semantic maps and multi-dimensional feature vectors, constructs a high-precision 3D scene of the disaster area using game engine technology. The sand table dynamically overlays three data layers: a disaster situation layer presenting the degree of building damage and dangerous areas through heat maps; a resource distribution layer displaying the real-time location of rescue teams and material reserves; and a risk warning layer marking high-risk areas for secondary disasters caused by aftershocks, supporting commanders in reviewing the disaster situation from multiple angles through panoramic roaming. After receiving earthquake disaster information, the intelligent decision-making engine module automatically matches historical case libraries to generate a three-level emergency response plan. The system uses operations research models to comprehensively consider road travel time, the difficulty of debris clearing, and the distribution of rescue forces, calculating the optimal rescue route and material dispatch plan, achieving a balance between time cost and safety. The cross-platform collaboration module ensures that the command center's large screen, the on-site command vehicle's tablet, and rescue personnel's mobile phones are synchronized with the sand table status in real time. Commanders can delineate dangerous cordon areas and mark temporary medical points on the sand table, with all operations synchronized to each terminal in real time. When the system detects a new gas leak, it automatically generates an alarm work order and pushes it to the emergency response department, tracking the progress of the response throughout the process.

[0100] To facilitate understanding of the above technical solutions of the present invention, the following provides a detailed description of the universal emergency panoramic command and intelligent decision-making system of the present invention in actual earthquake and fire situations.

[0101] I. Taking an earthquake disaster scenario as an example.

[0102] 1. Implementation process of the air-space-ground collaborative sensing network.

[0103] 1) For example, after an earthquake, the system immediately activates three drone clusters (each cluster containing six drones) to scan the affected area. Specific parameter configurations are shown in Table 1.

[0104] Table 1 Body Parameter Configuration

[0105]

[0106] 2) The drone swarm completed a panoramic scan of the 10km² core disaster area within 15 minutes through real-time image stitching, identifying: 127 damaged buildings (including 42 dangerous buildings with a structural tilt >3°); 89 abnormal thermal signals (confirming 53 trapped individuals); and 7 gas leak points. Simultaneously, three vehicle-mounted command vehicles were deployed at the east, west, and north entrances of the disaster area, using 128-line lidar to construct a 200-meter radius model of the surrounding environment (updated at 10Hz), planning six safe routes for the rescue convoy.

[0107] 2. Multi-source data fusion and intelligent analysis process.

[0108] 1) The data fusion engine employs a spatiotemporal alignment algorithm to accurately register UAV imagery, vehicle-mounted LiDAR point clouds, and DEM and building outline data from the geographic information database. Spatial errors are controlled within 0.3 meters, and temporal synchronization errors are less than 100 milliseconds. Specific processing includes: establishing a unified UTM coordinate system using the geographic coordinates of the disaster area center as the origin, and transforming all perceived data to this benchmark; using a graph neural network (GNN), associating and fusing texture features from visible light images (such as crack shape), temperature features from infrared thermal imaging (such as thermal signals from people), and geometric features from LiDAR (such as the volume of collapsed buildings), generating a multi-dimensional feature vector of the disaster scene containing 128 features; the fused feature vector is input into an optimized YOLOv8 model to accurately identify six types of disaster targets; simultaneously, this vector is input into a SwinTransformer model, outputting pixel-level semantic segmentation results (IoU > 85%), accurately delineating damaged areas, safe passages, and hazard sources. Target recognition results are shown in Table 2.

[0109] Table 2 Target Recognition Results

[0110]

[0111] 2) By using an improved YOLOv8 model, incremental learning of new scenes can be completed within 12 hours, shortening the model iteration cycle from the traditional 7 days to 12 hours. Graph neural networks fuse visible light, infrared, and LiDAR data to generate feature vectors containing 128 dimensions, providing input for semantic segmentation.

[0112] 3. Dynamic sand table and collaborative decision-making process.

[0113] 1) A 3D intelligent sandbox built on a game engine renders 120 million triangular faces in real time, overlaying three dynamic data layers to form command and decision-making mechanisms. This includes: dynamically rendering building damage levels (red: severely damaged; yellow: moderately damaged) using heatmaps based on semantic segmentation results, and integrating meteorological data to predict the spread of toxic gas pollution, providing a basis for evacuation directions; real-time marking of the location and status of all rescue forces, including the locations of 15 rescue teams, 8 pieces of engineering machinery, and the inventory of supplies at 3 temporary material distribution points; and accessing seismic network data to predict the risk of secondary disasters (such as landslides induced by aftershocks) using probabilistic models, marking high-probability landslide areas on the sandbox with flashing warning zones. The resource scheduling optimization scheme is shown in Table 3.

[0114] Table 3 Resource Scheduling Optimization Scheme

[0115]

[0116] 2) Based on the panoramic situation presented in the sand table, the intelligent decision-making engine generates a tiered response plan within 3 seconds of receiving a magnitude 7 earthquake input, matching data from a database of over 100,000 cases. Through operations research models, using time cost plus risk coefficient as the objective function, the optimal rescue path is derived, optimizing the average rescue time from 52 minutes to 37 minutes.

[0117] II. Taking a fire disaster scenario as an example.

[0118] 1. Implementation process of the air-space-ground collaborative sensing network.

[0119] 1) Upon the occurrence of a high-rise building fire, the system immediately activates four drone clusters (each cluster containing eight drones) to conduct a full-area scan of the fire's core area and a surrounding 15 km² region. Specific parameter configurations are shown in Table 4.

[0120] Table 4 Parameter Configuration

[0121]

[0122] 2) The drone swarm adopted a zoned collaborative scanning strategy, using real-time image stitching technology to complete a panoramic scan of the 15km² core fire area within 12 minutes, collecting a total of 12,000 frames of image data. Preliminary identification revealed 156 damage points with cracks ≥5cm in the exterior walls of buildings, including 27 high-risk buildings with a structural tilt >3°; the infrared thermal imaging channel captured 121 thermal anomaly signals, of which 67 were confirmed as thermal signals of trapped personnel; multispectral imaging identified a toxic gas diffusion range of 0.8km², with the diffusion direction consistent with the prevailing wind direction. Simultaneously, five vehicle-mounted command vehicles were deployed at the five key entrances to the fire site (east, south, west, north, and northeast). Through the fusion of lidar and 360° surround-view cameras, a high-precision environmental model (updating at 10Hz) was constructed within a 200-meter radius, monitoring road congestion and obstacle distribution in real time. This enabled the planning of eight safe routes for the fire rescue convoy, with obstacle avoidance response times all controlled within 0.3 seconds, ensuring the rapid arrival of rescue vehicles at the fire scene.

[0123] 2. Multi-source data fusion and intelligent analysis process.

[0124] 1) The data fusion engine employs a spatiotemporal alignment algorithm to precisely register UAV imagery, vehicle-mounted LiDAR point clouds, meteorological data, and 3D building models and urban road network data from a geographic information database. Spatial errors are controlled within 0.25 meters, and temporal synchronization errors are less than 80 milliseconds. The specific processing flow is as follows: A unified UTM coordinate system is established with the fire command center at the fire scene as the origin, and all perceived data is transformed to this reference coordinate system. Through a graph neural network (GNN), the building damage texture features of visible light imagery, the temperature gradient features of infrared thermal imaging, the 3D geometric features of the building structure from LiDAR, and the gas composition features from multispectral sources are deeply correlated and fused to generate a multidimensional feature vector of the fire scene containing 160 dimensions.

[0125] The fused feature vectors are simultaneously input into the improved YOLOv8 anomaly target detection model and the SwinTransformer panoramic semantic segmentation model. The improved YOLOv8 model optimizes the target detection anchor boxes for fire scenarios, achieving accurate identification of six typical fire disaster targets; the SwinTransformer model performs pixel-level segmentation of fire scene elements, generating an interactive semantic map with an IoU value of 87.2%, clearly outlining the fire combustion zone, high-temperature hazard zone, trapped personnel zone, safe evacuation route, and toxic gas diffusion zone. The target recognition results are shown in Table 5.

[0126] Table 5 Target Recognition Results

[0127]

[0128] 2) For new types of hazardous targets appearing at fire scenes (such as special smoke produced by the combustion of external wall insulation materials), the system adopts a small-sample incremental learning algorithm, collecting only 50 sets of sample data and completing model iteration updates within 12 hours, which is 14 times more efficient than the traditional 7-day iteration cycle. The updated model achieves a 90.1% accuracy rate in identifying new smoke targets, meeting the needs of on-site emergency command.

[0129] 3. Dynamic sand table and collaborative decision-making process.

[0130] 1) A real-time rendering system based on a game engine constructs a 3D intelligent sandbox of fire scenes, supporting smooth rendering of 150 million triangular facet models and overlaying three dynamic data layers to provide panoramic situational support for command and decision-making:

[0131] Disaster Situation Layer: Based on semantic segmentation results, a heat map is used to dynamically render the fire combustion intensity (red represents the core combustion zone, temperature > 800℃; orange represents the high temperature zone, temperature 300-800℃; yellow represents the residual heat zone, temperature 80-300℃). Combined with wind speed and direction data collected by meteorological sensors, a diffusion model is used to predict the diffusion range of toxic gases in the next hour, providing a scientific basis for the direction of personnel evacuation.

[0132] Resource distribution layer: Real-time tracking of the location and status of all rescue forces and materials on site, including the real-time locations of 22 fire and rescue teams, 15 aerial ladder trucks, and 8 high-pressure water cannon trucks, as well as the inventory of fire extinguishing agents and protective equipment at 5 temporary material distribution points.

[0133] Risk warning layer: Access building structure monitoring data and use probability models to predict the probability of secondary disasters such as building collapse and falling objects from heights. High-risk areas are marked with flashing red areas on the sand table. The warning level is divided into three levels (Level I: extremely high risk, immediate evacuation is required; Level II: high risk, proceed with caution; Level III: medium risk, continuous monitoring).

[0134] The sand table supports panoramic roaming, allowing commanders to freely switch between first-person and top-down perspectives. The magnification accuracy for key areas such as cracks in building exterior walls and trapped people in windows can reach the centimeter level, enabling clear viewing of target details.

[0135] 2) Upon receiving disaster information about a high-rise building fire with a burning area of ​​500m² and approximately 60 people trapped, the intelligent decision-making engine matches 12 similar cases from a database of over 100,000 historical cases within 3 seconds and generates a three-level response plan. By inputting natural language commands to prioritize rescuing those trapped on floors 10-15 of the high-rise building, the system automatically adjusts decision parameters. Based on an operations research model, an objective function is constructed, comprehensively considering three major indicators: rescue time cost, personnel safety risk coefficient, and resource utilization rate, to optimally allocate rescue resources. The resource scheduling optimization scheme is shown in Table 6.

[0136] Table 6 Resource Scheduling Optimization Scheme

[0137]

[0138] 3) Multi-terminal (5 vehicle-mounted command terminals, 30 mobile rescue terminals, and 1 cloud-based command center) enables real-time synchronization of the sand table status, supporting over 80 users to collaboratively annotate simultaneously. Commanders can delineate cordoned-off areas, mark temporary shelter locations, and mark dangerous obstacles on the sand table, with operation latency controlled within 18 milliseconds. When the sand table detects that the building collapse risk factor exceeds the threshold, the system automatically triggers an alarm work order, pushing it to the on-site rescue command center and the fire brigade command center, tracking the work order's progress in real time until the risk is eliminated.

[0139] Through the full-process application of this system, the rescue time for this high-rise building fire was shortened by 40% compared to the traditional command mode. All trapped personnel were successfully rescued, no secondary disasters occurred, and rescue efficiency and safety were significantly improved.

[0140] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A general-purpose emergency panoramic command and intelligent decision system, characterized in that, include: The air-space-ground collaborative sensing network unit is used to acquire multi-source data of the entire disaster site, generate a panoramic digital land surface model and a vehicle surrounding environment model based on the multi-source data, and realize automatic obstacle avoidance and route optimization using a dynamic path planning algorithm. This includes: randomly expanding exploration through a preset adaptive step size, while simultaneously detecting in real time whether the expanded path collides with obstacles in the vehicle surrounding environment model; adaptively gridding modeling of the barrier-free area, and using a search algorithm to quickly search for an effective path connecting the starting point and the target point within the feasible area; this includes: adaptively gridding modeling of the barrier-free area, initializing a candidate path set, randomly generating multiple initial path schemes within the feasible area and evaluating their path lengths, initializing the parameters and maximum number of iterations of the search algorithm; selecting an existing path from the candidate path set, and locally optimizing and adjusting the node sequence of the existing path according to a preset adjustment probability; the formula for local optimization and adjustment is: ； In the formula, P a represents the coordinates of the i-th node in the node sequence of the existing path; a represents the coordinates of the i-th node in the node sequence of the existing path after the local optimization adjustment; a ε represents a very small positive number; η represents a random disturbance intensity coefficient; λ represents a curvature optimization intensity coefficient; δ represents a Gaussian distribution; σ 2 represents a variance; ρ represents a preset adjustment probability; r 1, r 2 both represent random numbers;​​ The feasibility of the existing path after local optimization is verified to ensure that the path is completely within the barrier-free area and meets the minimum turning radius, thus obtaining a new path. The quality of the new path is compared with the original paths in the candidate path set. If the new path is the best, the worst path in the candidate path set is replaced, and the candidate path set is updated. If the maximum number of iterations is reached, the best path in the candidate path set is output as the effective path connecting the starting point and the target point. Otherwise, the path optimization and update process continues to perform node redundancy removal and segmentation on the effective path, and a high-order Bézier curve equation is used to achieve path smoothing to improve vehicle tracking performance. Combined with real-time data from the vehicle's surrounding environment model, a preset evaluation function is used to evaluate and optimize the effective path from multiple dimensions to ensure the safety and feasibility of the effective path. The multi-source data fusion and intelligent analysis unit is used to perform spatiotemporal alignment and cross-modal fusion of multi-source data using spatiotemporal alignment algorithms, and to complete disaster target identification and scene semantic segmentation through anomaly target detection model and panoramic semantic segmentation model, generating interactive semantic maps and multi-dimensional feature vectors. The dynamic sand table and collaborative decision-making unit are used to construct a three-dimensional intelligent sand table based on semantic maps and multi-dimensional feature vectors, and overlay a dynamic data layer. Through contingency plan generation and resource scheduling, decision-making schemes are formed to achieve multi-terminal sand table status synchronization, collaborative annotation, and automatic triggering and push of alarm work orders.

2. The general-purpose emergency panoramic command and intelligent decision-making system according to claim 1, characterized in that, The air-space-ground collaborative sensing network unit includes: The UAV swarm full-domain scanning module is used to scan the entire disaster site using five-lens oblique photography, infrared thermal imaging and multispectral sensors carried by the UAV, in order to generate a panoramic digital surface model and identify thermal anomaly targets and chemical pollutant outlines. The vehicle-mounted command vehicle's environmental perception module integrates lidar, surround-view cameras, and weather sensors to build a real-time model of the vehicle's surrounding environment and run dynamic path planning algorithms to achieve automatic obstacle avoidance and route optimization based on changes in the on-site environment. The air-space-ground collaborative sensing network unit is connected to the multi-source data fusion and intelligent analysis unit and the dynamic sand table and collaborative decision-making unit, while the UAV cluster full-domain scanning module and the vehicle-mounted command vehicle environmental perception module are connected.

3. The general-purpose emergency panoramic command and intelligent decision-making system according to claim 2, characterized in that, The integrated LiDAR, surround-view camera, and weather sensor construct a real-time model of the vehicle's surrounding environment and run a dynamic path planning algorithm to achieve automatic obstacle avoidance and route optimization based on changes in the environment. Data from lidar, surround-view cameras, and weather sensors are acquired respectively. A model of the vehicle's surrounding environment is built in real time based on the acquired data. The parameters of the dynamic path planning algorithm are initialized, and the maximum number of iterations is set. If the maximum number of iterations is reached, the final optimized path is executed, and automatic obstacle avoidance and route optimization are implemented based on changes in the on-site environment; otherwise, iterative optimization continues.

4. The general-purpose emergency panoramic command and intelligent decision-making system according to claim 1, characterized in that, The multi-source data fusion and intelligent analysis unit includes: The multimodal data fusion engine module is used to unify UAV, vehicle sensor and geographic information data into the same spatiotemporal coordinate system through spatiotemporal alignment algorithm, and to use graph neural network to extract cross-modal features to construct multi-dimensional feature vectors of disaster scenarios; The panoramic image anomaly recognition module is used to identify disaster targets based on the fused multidimensional feature vectors using the YOLOv8 model, and to perform semantic segmentation of disaster scene elements using a panoramic semantic segmentation model, ultimately generating an interactive semantic map. The multimodal data fusion engine module and the panoramic image anomaly recognition module are connected.

5. The general-purpose emergency panoramic command and intelligent decision-making system according to claim 4, characterized in that, The process of unifying UAV, vehicle-mounted sensor, and geographic information data into the same spatiotemporal coordinate system using a spatiotemporal alignment algorithm, and then using graph neural networks for cross-modal feature extraction to construct a multidimensional feature vector for the disaster scenario includes: Establish a unified spatiotemporal coordinate system to perform spatiotemporal registration of UAV imagery, vehicle-mounted sensor data cloud, and geographic information data; Establish spatial correspondences between different modal data, and map visible light image pixels, infrared hot spot regions, and lidar point clouds together; The registered UAV imagery, vehicle sensor data cloud, and geographic information data are used to construct a graph neural network input structure, with spatial location as nodes and intermodal feature associations as edges, to form a topological graph representation of the disaster scene; By leveraging the message passing mechanism of graph neural networks, multimodal feature information within the neighborhood of nodes is aggregated to learn a deep fusion representation of visible light texture, infrared thermal radiation, and lidar geometric features; By iteratively optimizing the node feature representation through multi-layer graph convolution operations, feature conflicts are eliminated and complementary information is enhanced, ultimately constructing a multi-dimensional feature vector for the disaster scenario.

6. The general-purpose emergency panoramic command and intelligent decision-making system according to claim 4, characterized in that, The process involves identifying disaster targets using the YOLOv8 model based on the fused multidimensional feature vectors, and semantically segmenting disaster scene elements using a panoramic semantic segmentation model to ultimately generate an interactive semantic map. The YOLOv8 model was iteratively optimized using the few-sample incremental learning method to obtain the optimized YOLOv8 model. The fused multidimensional feature vectors are then input into the optimized YOLOv8 model and the panoramic semantic segmentation model, respectively. Multi-scale features are extracted using the feature pyramid network of the optimized YOLOv8 model to complete the identification of preset disaster targets; The hierarchical attention mechanism of the panoramic semantic segmentation model is used to parse the global contextual relationships and obtain pixel-level scene understanding; The spatial resolution of the feature map is restored by the segmentation decoder, and the segmentation mask of the disaster scene elements is output. By integrating target recognition and semantic segmentation results and overlaying geographic information, an interactive semantic map is generated.

7. The general-purpose emergency panoramic command and intelligent decision-making system according to claim 6, characterized in that, The iterative optimization of the YOLOv8 model using the few-sample incremental learning method yields the following optimized YOLOv8 model: Initialize the parameters of the few-sample incremental learning algorithm and set the maximum number of iterations. Perform a preliminary evaluation of the YOLOv8 model based on the initial few-sample dataset and select a set of candidate model parameters to be optimized based on the detection accuracy. If the current YOLOv8 model accuracy meets the preset threshold, the optimization result is output directly; otherwise, the incremental learning process is entered. The training dataset is constructed using newly added small sample data, and positive and negative examples are labeled in the samples to form the labeled data required for incremental learning. If it is the first incremental learning, a basic classifier is built; otherwise, the parameters of the existing classifier are updated incrementally using the newly labeled data. The updated classifier is used to filter the parameter space of the YOLOv8 model, the correlation of the YOLOv8 model parameters is calculated, and the optimal parameter combination is selected to update the YOLOv8 model. Repeat the incremental learning process until the maximum number of iterations is reached, and finally output the optimized YOLOv8 model.

8. The general-purpose emergency panoramic command and intelligent decision-making system according to claim 1, characterized in that, The dynamic sandbox and collaborative decision-making unit include: The 3D intelligent sand table construction module is used to develop real-time rendering technology based on game engines. It integrates semantic maps and multi-dimensional feature vectors to construct a 3D sand table, and overlays dynamic data layers of disaster situation, resource distribution and risk warning to achieve panoramic roaming and multi-scale detailed display. The intelligent decision engine module is used to automatically match historical case databases based on disaster type and level to generate graded response plans, and optimize rescue routes and material allocation through operations research models to form a decision plan that takes into account time cost, risk coefficient and resource utilization. The cross-platform collaboration module is used to realize real-time synchronization of the sandbox status across multiple terminals, support collaborative annotation operations by users, and automatically trigger alarm work order push and processing progress tracking when abnormal events occur. The three-dimensional intelligent sand table construction module is connected to the intelligent decision engine module and the cross-platform collaboration module.

Images