A method and system for low-altitude fire monitoring by a drone

Abstract

This application relates to a method and system for monitoring low-altitude fires using unmanned aerial vehicles (UAVs). The method includes: collecting low-altitude imaging data within a monitoring area using a UAV; identifying fires in the monitoring area based on the low-altitude imaging data using a trained improved target detection model, obtaining identification results; wherein the improved target detection model includes an intra-group self-attention and convolutional hybrid module and a multi-scale feature fusion module; and triggering fire grading early warning and fire simulation prediction based on the identification results. This application achieves a shared parallel mechanism based on hybrid modules, improving the model's computational efficiency. Furthermore, the four-branch fusion and splicing mechanism of the fusion module ensures the model's detection accuracy, effectively reducing false alarms and false negatives, and enhancing robust identification capabilities for flames and smoke of different sizes and under different background conditions.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) monitoring technology, and in particular to a method and system for monitoring low-altitude fires using UAVs. Background Technology

[0002] With global climate change and frequent extreme weather events, forest fires pose an increasingly severe threat to human life and property, the ecological environment, and economic and social development. Smoke and flames are the most prominent visual features in the early stages of a fire, and their rapid and accurate identification is the core technological foundation for early detection. However, due to the variable color and irregular shape of smoke from the perspective of low-altitude drones, and its susceptibility to interference from environmental factors such as light, fog, and dust, traditional detection methods generally suffer from weak generalization ability, high response delay, and high false alarm rate, making it difficult to meet the practical application needs in complex natural scenarios.

[0003] Currently, although various real-time target detection algorithms have been applied to UAV fire detection, problems such as insufficient detection accuracy and frequent false alarms and missed alarms still exist when faced with fine targets, multi-scale changes, and complex background interference. In particular, flames have unique color distribution and dynamic texture characteristics, which can easily be confused with sunlight reflection, ground objects, or clouds and fog, further increasing the difficulty of identification.

[0004] Currently, no effective solution has been proposed for improving the accuracy of drone fire monitoring in related technologies. Summary of the Invention

[0005] This application provides a method and system for monitoring low-altitude fires using unmanned aerial vehicles (UAVs), which at least addresses the problem of how to improve the accuracy of UAV fire monitoring in related technologies.

[0006] In a first aspect, embodiments of this application provide a method for monitoring low-altitude fires using unmanned aerial vehicles (UAVs), the method comprising: Low-altitude imaging data of the monitoring area is collected by drones; Based on the low-altitude imaging data, a fire is identified in the monitoring area using a trained improved target detection model, and the identification result is obtained. The improved target detection model includes an intra-group self-attention and convolution hybrid module and a multi-scale feature fusion module. The self-attention and convolution hybrid module within the group includes a convolution operator and a self-attention operator that share an initial feature projection, and the convolution operation and self-attention operation of the operator are executed in parallel before feature fusion. The multi-scale feature fusion module has a multi-scale feature fusion branch, a deformable convolution branch, a bilinear enhanced multi-head self-attention mechanism branch, and a pyramid pooling branch. The features output by the four branches are fused and then concatenated with residual links. Based on the identification results, fire grading early warning and fire simulation prediction are triggered.

[0007] In some embodiments, the improved target detection model further includes a dynamic multi-scale spatial attenuation module; The dynamic multi-scale spatial decay module introduces a multi-scale decay matrix to enable the model to adapt to different distance dependencies, and introduces hierarchical multi-scale decomposition to enable the model to capture diagonal direction information.

[0008] In some embodiments, the improved object detection model further includes a channel-enhanced attention mechanism module; The channel-enhanced attention mechanism module contains three parallel convolutional channels. It learns the correlation between channels through a self-attention mechanism to dynamically adjust the importance weights of different channels, thereby reducing the loss of spatial information caused by pooling.

[0009] In some embodiments, the improved object detection model further includes a multi-scale dynamic grouping attention module; The multi-scale dynamic grouping attention module extracts features at multiple scales and compresses spatial information through global average pooling, thereby reducing the computational complexity of the model.

[0010] In some embodiments, collecting low-altitude imaging data within the monitoring area using a drone includes: Based on the collaborative operation of fixed UAV airports and UAVs, low-altitude imaging data is collected by UAVs within the monitoring area. The low-altitude imaging data includes visible light imaging data, infrared thermal imaging data, and depth imaging data.

[0011] In some embodiments, the method includes: Based on the low-altitude imaging data collected by the UAV, the Neural Radiation Field (NeRF) technology is used to model the three-dimensional scene of the monitoring area, and combined with the digital elevation model, a three-dimensional geographic information model of the monitoring area is constructed.

[0012] In some embodiments, the method includes: Environmental data within the monitoring area is collected synchronously through a ground-based IoT sensor network. This environmental data includes temperature and humidity data, as well as wind speed and direction data.

[0013] In some embodiments, triggering fire simulation prediction based on the identification results includes: Based on the identification results, if the identification results indicate that a fire point has been detected within the monitoring area, then after triggering a level-two fire warning, the fire spread path and speed will be dynamically simulated in real time based on the three-dimensional geographic information model of the monitoring area, the temperature and humidity data, and the wind speed and direction data to provide decision support for the deployment of rescue forces.

[0014] In some embodiments, triggering a fire hazard level warning based on the identification result includes: Based on the identification results, if the identification results indicate that abnormal temperature or smoke is detected, a level one fire warning is triggered to drive the drone to patrol and verify the fire point in the monitoring area. If the identification result indicates that a fire point has been detected within the monitoring area, a level-two fire warning is triggered to make a preliminary judgment on the fire based on the abnormal temperature area and the identification frame, and to push alarm information to the on-duty personnel. After the on-duty personnel confirm the existence of the alarm information, a level-three fire warning is triggered to coordinate with fire and emergency departments to activate the emergency response plan.

[0015] In a second aspect, embodiments of this application provide a low-altitude fire monitoring system for unmanned aerial vehicles (UAVs), the system being used to perform the method described in the first aspect above, the system including a data acquisition module, an identification module, and an early warning module; The acquisition module is used to acquire low-altitude imaging data within the monitoring area via a drone; The identification module is used to identify fires in the monitoring area based on the low-altitude imaging data using a trained improved target detection model, and obtain the identification result. The improved target detection model includes an intra-group self-attention and convolution hybrid module and a multi-scale feature fusion module. The early warning module is used to trigger fire level early warning and fire simulation prediction based on the identification results.

[0016] Compared to related technologies, this application provides a method and system for monitoring low-altitude fires using unmanned aerial vehicles (UAVs). The method involves collecting low-altitude imaging data within a monitoring area using a UAV. Based on this imaging data, a trained improved target detection model is used to identify fires in the monitoring area, yielding identification results. The model's intra-group self-attention and convolution hybrid module includes convolution operators and self-attention operators sharing initial feature projections. The convolution and self-attention operations are executed in parallel before feature fusion. The multi-scale feature fusion module has multi-scale feature fusion branches, deformable convolution branches, bilinear enhanced multi-head self-attention mechanism branches, and pyramid pooling branches. The features output from the four branches are fused and then concatenated with residual links. Based on the identification results, fire grading warnings and fire simulation predictions are triggered. This achieves a shared parallel mechanism based on the hybrid module, improving the model's computational efficiency. The four-branch fusion and concatenation mechanism of the fusion module ensures the model's detection accuracy, effectively reducing false alarms and false negatives, and enhancing robust identification capabilities for flames and smoke of different sizes and background conditions. This solves the problem of improving the accuracy of UAV fire monitoring. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a method for monitoring low-altitude fires using unmanned aerial vehicles (UAVs) according to an embodiment of this application. Figure 2 This is a schematic diagram of the structure of an improved target detection model according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of the intra-group self-attention and convolution hybrid module according to an embodiment of this application; Figure 4 This is a schematic diagram of the channel-enhanced attention mechanism module according to an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a multi-scale dynamic grouping attention module according to an embodiment of this application; Figure 6 This is a schematic diagram of the structure of the multi-scale feature fusion module according to an embodiment of this application; Figure 7 This is a schematic diagram of the structure of the bilinear enhanced multi-head self-attention mechanism branch according to an embodiment of this application; Figure 8 This is a schematic diagram of the structure of the dynamic multi-scale spatial attenuation module according to an embodiment of this application; Figure 9 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0019] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0020] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0021] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0022] This application provides a method for monitoring low-altitude fires using unmanned aerial vehicles (UAVs). Figure 1 This is a flowchart illustrating the low-altitude fire monitoring method for unmanned aerial vehicles (UAVs) according to an embodiment of this application, as shown below. Figure 1 As shown, the method includes: Step S1: Collect low-altitude imaging data of the monitoring area using a drone; Specifically, step S1 involves the collaborative operation of a fixed UAV airport and UAVs, where UAVs collect low-altitude imaging data within the monitoring area. The low-altitude imaging data includes visible light imaging data, infrared thermal imaging data, and depth imaging data. Based on low-altitude imaging data collected by UAVs, NeRF technology is used to model the three-dimensional scene of the monitoring area, and combined with digital elevation model to construct a three-dimensional geographic information model of the monitoring area.

[0023] It should be noted that the data source scheduling involves coordinated operations between fixed UAV airfields and individual UAVs, forming a complementary mechanism of macro-level patrols and micro-level reconnaissance. The UAVs are equipped with high-definition visible light cameras and infrared thermal imagers (temperature sensitivity ±0.5℃), while a ground-based IoT sensor network simultaneously collects environmental data (temperature and humidity data, wind speed and direction data). Data backhaul and synchronization: Relying on 5G-A private networks or low-altitude intelligent networks, monitoring data (2K high-definition video streams, thermal imaging data, depth imaging data, and sensor data) are backhauled to edge computing nodes or command centers in real time with low latency and high bandwidth. NeRF (Neural Radiation Field) technology is used to create 3D scene models of forest areas, and a high-precision geographic information base is constructed by combining it with a digital elevation model (DEM).

[0024] Step S2: Based on low-altitude imaging data, fire identification is performed in the monitored area using a trained improved target detection model to obtain the identification results; It should be noted that the improved target detection model for intelligent flame and smoke recognition (edge / cloud collaborative computing), especially for small and low-contrast targets, achieves higher recognition accuracy and positioning precision. Specifically, by utilizing lightweight model design and pruning techniques, combined with 5G-A and edge computing, second-level fire identification and minute-level emergency response are achieved; an integrated air-ground monitoring network of "perception-transmission-analysis-decision-action" is constructed, realizing intelligent early warning and multi-terminal collaboration; the lightweight model is easy to deploy on resource-constrained UAV edge computing modules, reducing system deployment and maintenance costs.

[0025] Step S3: Based on the identification results, trigger fire level warning and fire simulation prediction.

[0026] Specifically, in step S3, based on the identification results, if the identification results indicate that abnormal temperature or smoke has been detected, a level one fire warning is triggered to drive the drone to patrol and verify the fire point within the monitoring area. If the identification result indicates that a fire point has been detected within the monitoring area, a level-two fire warning will be triggered to make a preliminary judgment on the fire based on the abnormal temperature area and the identification frame, and to push alarm information to the on-duty personnel. After the on-duty personnel confirm the existence of an alarm, a level-three fire warning is triggered to coordinate with fire and emergency departments to activate the emergency response plan. In addition, after a level-two fire warning is triggered, the fire spread path and speed are simulated in real time based on a three-dimensional geographic information model of the monitored area, temperature and humidity data, and wind speed and direction data to provide decision support for the deployment of rescue forces.

[0027] It should be noted that after the target detection model identifies a potential fire point or smoke, the system automatically activates a three-level early warning mechanism by combining thermal imaging temperature data and visible light analysis results. Specifically, Level 1 warning (potential risk): Sensors detect abnormal temperature or smoke, triggering automatic system verification; Level 2 warning (fire confirmation): The AI ​​model confirms the fire point, makes a preliminary judgment on the fire intensity based on the temperature range and the recognition frame, and pushes alarm information through multiple channels such as WeChat groups, SMS, and telephone; Level 3 warning (emergency response): After confirmation by on-duty personnel, the emergency plan is activated, coordinating with fire and emergency departments, and drones continuously monitor and provide real-time images. Furthermore, based on the constructed 3D scene model and environmental data, the system dynamically simulates the fire spread path and speed, providing decision support for the deployment of rescue forces; the command center, through the low-altitude emergency command platform, dispatches multiple terminals such as drones, individual equipment, command vehicles, and even robotic dogs for collaborative operations. Among them, drones can execute "fire re-inspection routes" to confirm the on-site situation, forming a complete closed-loop management from discovery, early warning, handling to verification.

[0028] It should be further explained that the Level 1 warning (potential risk - automatic verification) is triggered when a single high-confidence indicator or multiple low-confidence indicators occur simultaneously. For example, triggering conditions include: thermal imaging detecting an abnormal increase in local temperature (but not reaching the open flame threshold) and the smoke recognition model providing a low confidence level (e.g., 0.3-0.4), or a ground wind sensor showing a sudden increase in wind speed. System action: Instead of immediately alerting humans, the system automatically dispatches the nearest drone to change its flight path, conduct close-range reconnaissance and multi-angle filming of the abnormal area, and initiates real-time video stream analysis to obtain more evidence. This process requires no human intervention, achieving "silent escalation of reconnaissance." Level 2 Warning (Confirmed Fire - Automatic Alarm): Triggered when multiple sources of evidence form a high-confidence causal chain. For example: Triggering conditions: The visible light AI model identifies a flame (confidence > 0.7), the thermal imaging temperature at the same location continues to rise rapidly, and the review video from the Level 1 warning confirms the dynamic flame characteristics. System Action: The system automatically determines that a fire has been confirmed and automatically executes the pre-planned alarm. It not only notifies on-duty personnel via SMS and app push notifications but also automatically extracts structured information such as fire coordinates, initial spread direction, surrounding terrain, and nearest water source / road, generating a preliminary report which is also pushed out. Simultaneously, it automatically creates an emergency task group and calls the pre-set on-site supervisor's phone number. Level 3 Early Warning (Emergency Response - Resource Coordination): Based on Level 2 early warning, it intelligently upgrades by combining digital contingency plans and resource status. Triggering conditions: The Level 2 early warning is manually confirmed, or the system determines through video analysis that the fire is rapidly expanding (e.g., the rate of increase in burned area exceeds a threshold), or meteorological data predicts strong winds within the next half hour. System actions: The system automatically activates the matching contingency plan in the digital emergency plan library. It automatically sends standardized coordination requests to the command systems of fire, emergency, and forestry departments; it automatically dispatches multiple drones to form a formation to perform tasks such as fire monitoring, surrounding area search, and evacuation route inspection; it automatically generates optimal rescue force delivery route suggestions; and it automatically connects to the emergency command screen, pushing all real-time images and data layers.

[0029] Furthermore, it should be noted that the UAV low-altitude fire monitoring method proposed in this invention has good generalization ability, and its technical process can be flexibly extended to other similar low-altitude target detection scenarios, such as aerial photography monitoring and small aircraft identification in emergency rescue. In practical applications, the process steps or environmental parameters can be adaptively adjusted according to specific scenario requirements, thereby better meeting diverse field application needs. Specifically: The emergency management and forest fire prevention industry is the primary application industry of this application. The system directly serves agencies such as the Forestry and Grassland Administration, the Ministry of Emergency Management, and the Fire and Rescue Bureau, and is used to solve the pain points of "difficult detection, slow confirmation, and delayed response" in traditional forest fire prevention. It represents the industry's digital transformation direction from "human-based prevention" to "technology-based prevention". Low-altitude economy and drone industry: It involves the deployment and operation of drones (as flight platforms), fixed drone airports (as infrastructure), and extended applications of low-altitude logistics (such as emergency material delivery), and is an important technological support for promoting the development of the low-altitude economy. Smart Cities and Public Safety: In addition to forests, this technology can be extended to urban fire protection (high-rise buildings, chemical industrial parks), monitoring of key infrastructure (substations, oil and gas pipelines), security for large-scale events, and other scenarios. It is a key component in building a smart city's "integrated air-space-ground" sensing network and emergency response system. In the environmental protection and ecological monitoring industry, it can be used to monitor the ecological status of nature reserves, promptly detect fires caused by natural or human factors, and protect biodiversity. Furthermore, after a fire occurs, it can also be used to assess the burned area and ecological damage.

[0030] This application's embodiments, based on an integrated air-space-ground design, integrate advanced computer vision algorithms and low-altitude communication technology, significantly improving the early detection capability and emergency response efficiency of forest fires. Specific technical effects include: 1. High-precision flame and smoke detection: Employing an improved algorithm, by introducing an attention mechanism, a lightweight neck network, and optimizing the loss function, accurate identification of flames and smoke is achieved. 2. Real-time low-altitude monitoring: The system utilizes UAVs equipped with infrared thermal imaging and visible light cameras, combined with a 5G private network to achieve real-time high-definition video transmission. AI models are deployed on edge computing nodes for second-level fire situation analysis, greatly shortening response time. 3. Lightweight and efficient deployment: The detection model has been pruned and optimized, making it suitable for deployment on resource-constrained edge devices. This enables the system to operate stably in remote forest areas and supports long-term automated inspections. 4. Intelligent early warning and collaborative response: The system establishes a three-level early warning mechanism, automating the entire process from risk perception to emergency response. Through a low-altitude emergency command platform, data from multiple departments such as fire protection and forestry are integrated, enabling multi-terminal collaboration among UAVs, individual soldier equipment, and other terminals, improving overall prevention and control capabilities. 5. Scalability: The modular system design supports rapid iteration and domestic adaptation, while low-code operation lowers the barrier to entry and facilitates deployment to different application scenarios. In summary, this invention, through technological innovation, has significant advantages in detection accuracy, real-time performance, lightweight design, and system integration, providing an efficient and reliable solution for forest fire prevention and control.

[0031] In some specific embodiments, Figure 2 This is a schematic diagram of the structure of the improved target detection model according to an embodiment of this application, as shown below. Figure 2 As shown, the improved target detection model in the above embodiments includes an intra-group self-attention and convolution hybrid module, a multi-scale feature fusion module, a dynamic multi-scale spatial decay module, a channel-enhanced attention mechanism module, and a multi-scale dynamic grouping attention module. For the hybrid module of intra-group self-attention and convolution in the model Figure 3 This is a schematic diagram of the structure of the intra-group self-attention and convolution hybrid module according to an embodiment of this application, as shown below. Figure 3 As shown, the in-group self-attention and convolution hybrid module contains convolution operators and self-attention operators that share the initial feature projection, and the convolution and self-attention operations of the operators are executed in parallel before feature fusion; specifically: Within each content group of the hybrid self-attention and convolution module, convolutional operations (capturing local details and structure) and self-attention operations (modeling long-range dependencies within the group) are performed in parallel, and the results are fused. This provides richer and more robust feature interactions than self-attention alone.

[0032] like Figure 3 As shown, the first step (sharing initial feature projection): the input feature map The sequence (B, H, W, C) is rearranged into (B, N, C), where N = H * W. The similarity between the sequence and the cluster centers (after LayerNorm) is calculated to obtain the affinity matrix A. The cluster centers are then weighted and summed using the affinity matrix to obtain aggregated similar token features, which are then rearranged into a 4D form (B, H, W, C). The second step (parallel execution of convolution and self-attention operations): The aggregated features are processed using in-group self-attention and convolution operators respectively, then LayerNorm is applied and the features are rearranged back into a 4D form. The third step: The features are rearranged into a sequence, and cross-attention (actually self-attention is used here because Query, Key, and Value are the same input) and residual connections are used. Then LayerNorm is applied, where the convolution operator path captures local structure, the attention operator path models long-range dependencies, C represents the feature dimension (number of channels) of each token center, determining the richness of information it can express, and S represents the number of centers, determining the granularity of the model's "abstraction" or "induction" of image content.

[0033] It should be noted that the hybrid mechanism allows the network to possess both the inductive bias of convolution (which is beneficial for capturing local textures and edges) and the dynamic global modeling capability of attention, theoretically generating more realistic high-frequency details. By sharing the first-stage convolutional projection, it is more efficient than concatenating a single convolution and an attention module. Furthermore, the introduction of convolutional operations can provide a smoother optimization path for attention-based modules, which is helpful for model training.

[0034] Further explanation is needed regarding the advantages of fusing convolution and attention: the intra-group self-attention and convolution hybrid module efficiently combines the locality and translation invariance of convolution with the global modeling and dynamic weight allocation capabilities of attention by sharing the first-stage projection. This allows the model to capture both local details and global context simultaneously, which is beneficial for restoring high-frequency details and overall structure in image super-resolution tasks. Flexible structure: It can be configured according to task requirements and computational resources. This module allows the intra-group self-attention and convolution hybrid to be applied at different locations (intra-group, inter-group, final enhancement), providing greater structural flexibility. Content-aware aggregation: Learnable cluster centers perform content-aware aggregation of tokens, enabling the model to group semantically similar tokens together, thus more effectively modeling long-range dependencies. Computational efficiency: The intra-group self-attention and convolution hybrid reduces computation by sharing projection, resulting in higher computational efficiency compared to performing convolution and attention operations separately. Scalability: This design allows for further exploration of convolution and attention fusion methods, such as adjusting fusion weights, using different convolution kernel sizes, and the number of attention heads, providing a foundation for subsequent research.

[0035] In some specific embodiments, Figure 4 This is a schematic diagram of the channel-enhanced attention mechanism module according to an embodiment of this application, as shown below. Figure 4 As shown, the channel-enhanced attention mechanism module contains three parallel convolutional channels. It learns the correlation between channels through a self-attention mechanism to dynamically adjust the importance weights of different channels, reducing spatial information loss caused by pooling. Specifically: like Figure 4 As shown, the three parallel convolutional channels are: FV (Value convolution) – generating value features, FQ (Query convolution) – generating query features, and FK (Key convolution) – generating key features. The self-attention calculation process is as follows: similarity tensor calculation (multiplication of FQ and FK matrices → channel similarity matrix); weight normalization (spatial dimension aggregation → Softmax → channel attention weights); attention weighting (weights * FV → enhanced feature representation); residual connection (attention weighted result + original input); batch normalization (stabilizing the training process); activation function (ReLU activation, introducing nonlinearity).

[0036] It should be noted that the channel-enhanced attention mechanism module implements the redistribution of channel weights: it learns the correlation between channels through a self-attention mechanism and dynamically adjusts the importance of different channels, rather than relying on the convolution weights in a fixed way; it also alleviates the spatial information loss caused by pooling: the module reconstructs the feature relationship in the channel dimension, making up for the problem that traditional CNNs need to increase the number of channels after pooling to compensate for information loss.

[0037] In some specific embodiments, Figure 5 This is a schematic diagram of the structure of the multi-scale dynamic grouping attention module according to an embodiment of this application, such as... Figure 5 As shown, the multi-scale dynamic grouping attention module extracts features at multiple scales and compresses spatial information through global average pooling, reducing the computational complexity of the model; specifically: The multi-scale dynamic grouping attention module achieves a more powerful multi-scale dynamic grouping attention by introducing four major improvements: multi-scale feature extraction, dynamic temperature parameters, hierarchical attention, and spatial-channel collaboration. This design enhances spatial awareness and improves computational efficiency while maintaining the advantages of channel relationship modeling, making it suitable for various computer vision tasks, especially showing significant advantages on resource-constrained mobile devices. The module can be used as a plug-and-play component to replace standard convolutional or attention modules, achieving performance improvements in various network architectures and providing a new technical path for lightweight, high-performance vision models. The calculation of the similarity tensor involves high-dimensional matrix multiplication. This module uses global average pooling to compress spatial information, reducing the computational complexity to O(BC). 2+BCH*W) significantly reduces computational cost. Attention weights focus more on channel relationships: through global average pooling, we extract global information for each channel and then calculate the attention weights between channels, avoiding spatial interference with channel attention; residual connections are preserved to ensure gradient flow. The complete formula system is as follows:

[0038] The input is Multi-scale feature extraction of Q, K, and V; dynamic temperature parameters Hierarchical attention calculation, intra-group attention The number of groups is 8; inter-group attention Spatial attention The final output is Y. This indicates channel-dimension multiplication. This represents spatial dimension multiplication.

[0039] In some specific embodiments, Figure 6 This is a schematic diagram of the structure of the multi-scale feature fusion module according to an embodiment of this application, such as... Figure 6 As shown, the multi-scale feature fusion module has a multi-scale feature fusion branch, a deformable convolution branch, a bilinear enhanced multi-head self-attention mechanism branch, and a pyramid pooling branch. The features output from these four branches are fused and then concatenated with residual connections. Specifically: The following uses mathematical formulas to illustrate the improvements and innovations of the multi-scale feature fusion module: First, the bilinear enhanced multi-head self-attention mechanism branch:

[0040] in, It is a learnable bilinear weight matrix, d = C / h, where h is the number of attention heads. These represent the query, key, and value corresponding to the i-th attention head, respectively.

[0041] Multi-head output fusion:

[0042] Given input features and position encoding The location-enhanced features are calculated as follows: Interpolate represents the bilinear interpolation operation.

[0043] Channel attention weights are obtained through a fully connected layer and an activation function:

[0044] Where GAP represents global average pooling, and δ is the ReLU activation function. It is the Sigmoid function. , r is the reduction ratio.

[0045] Then, the feature map is multiplied element-wise with the channel attention weights to obtain the channel-enhanced features as follows ( (represents element-wise multiplication)

[0046] It should be noted that, Figure 7 This is a schematic diagram of the structure of the bilinear enhanced multi-head self-attention mechanism branch according to an embodiment of this application, as shown below. Figure 7 As shown, the input feature map is in the format [B, C, H, W], where B is the batch size, C is the number of channels, and H and W are the spatial dimensions. Channel attention branch: Channel statistics are obtained through global average pooling; two 1x1 convolutional layers learn the dependencies between channels; channel weights are generated using Sigmoid; and weighted channel-wise with the input feature map. Multi-head bilinear attention branch: Query / key / value transformation: three independent convolutional layers generate Q, K, and V; Positional encoding: spatial positional information is added to Q and K; Bilinear attention mechanism: a bilinear weight matrix is ​​introduced to enhance expressive power; spatial positional relationships are considered when calculating attention scores; Softmax normalization generates attention weights; Multi-head parallelism: multiple attention heads are computed in parallel to capture different feature relationships; Output fusion: multi-head outputs are concatenated and integrated through fusion convolution; Two-branch fusion: the outputs of two branches are fused by addition or concatenation. Its main improvements include: dual-branch structure: simultaneously capturing channel attention and spatial attention; position awareness: explicitly adding position encoding in attention calculation; bilinear enhancement: enhancing feature interaction capabilities through bilinear transformation; multi-head mechanism: parallel multi-attention heads capturing diverse feature relationships; lightweight design: using 1*1 convolution to maintain computational efficiency.

[0047] Second, deformable convolution branch. Among these, offset prediction, for position p, the offset... The prediction is as follows (where...) (This is an offset prediction convolutional layer)

[0048] The prediction formula for the modulation factor m in the deformable convolution mechanism is as follows (the modulation factor m is predicted by another convolutional layer). Prediction, obtained through the Sigmoid function and scaling):

[0049] Deformable convolution operation, the deformable convolution output formula is as follows (where K is the convolution kernel size, It is a predefined sampling location. (These are convolution weights)

[0050] Third, pyramid pooling branch (multi-scale pooling). Let the input features... Given a set of pooling scales For each scale First, adaptive average pooling is performed, followed by a 1*1 convolutional layer f. i Dimensionality reduction is performed. The feature formula for the i-th scale is as follows:

[0051] Fourth, the multi-scale feature fusion branch. Upsampling restores the original size: Final Fusion: , where g is the fused convolutional layer.

[0052] Multi-scale feature extraction in cross-scale feature fusion. Multi-scale features are extracted in parallel using convolutional kernels of different sizes: , , , Where f{k*k} represents k*k convolution, and f{dilated} represents dilated convolution.

[0053] An adaptive gating mechanism in cross-scale feature fusion. First, multi-scale features are concatenated and then subjected to global average pooling, followed by a gating network. (e.g., small fully connected networks) generate gating weights The formula is as follows:

[0054] This is then used to weight the features at each scale: , , , Finally, the weighted features are concatenated and merged: .

[0055] Final module output (using learnable residual connections): Where γ and β are learnable scaling and bias parameters, It is obtained by 1*1 pointwise convolution projection.

[0056] In some of these embodiments, Figure 8 This is a schematic diagram of the structure of the dynamic multi-scale spatial attenuation module according to an embodiment of this application, as shown below. Figure 8As shown, the dynamic multi-scale spatial decay module introduces a multi-scale decay matrix to enable the model to adapt to different distance dependencies, and introduces hierarchical multi-scale decomposition to enable the model to capture diagonal direction information; specifically: By introducing decay matrices at different scales s, the model focuses on local details at small scales and captures long-range dependencies at large scales, enabling it to adapt to dependencies at different distances. The formula is shown below:

[0057] Among them, the attenuation factor γ and the distance scaling parameter α are dynamically calculated based on query-key pairs, making the spatial prior more adaptive; level and vertical Different scaling parameters are used in the direction (anisotropic attenuation); This represents the value of the spatial decay matrix at scale s at positions i and j (the degree of attention decay of position j on position i). This represents the global decay factor at scale s. Control the overall attenuation intensity at this scale; This represents the anisotropic distance scaling parameter. >0 controls distance sensitivity in both horizontal and vertical directions; Represents the activation function, with dimension . Ensure that the scaling parameter is positive; and Represents a learnable weight matrix, and a linear transformation generates dynamic parameters; This represents global average pooling, which reduces dimensionality and aggregates global information. This indicates information about the merged query and the key.

[0058] Hierarchical multi-scale decomposition is introduced to address the potential loss of diagonal direction information, as shown in the following formula ( Predicting fine scales, Predicting roughness scale):

[0059] Among these features are: bidirectional decomposition order (height first, then width; width first, then height, capturing information flow in different directions); and multi-scale value projection (using different value projections at different scales). Extract scale-specific features; complementary decomposition paths (reduce information bias caused by a single decomposition order). This formula is the core operator of directional decomposition attention. D represents the direction, specifically the direction along the coordinate axis during attention calculation. There are two specific values ​​for D: when D=H, it is the height direction (vertical direction), calculated along the height axis of the feature map; when D=W, it is the width direction (horizontal direction), calculated along the width axis of the feature map. A represents the intermediate feature being processed, which can be the original Q or the output of the previous attention step.

[0060] Furthermore, the dynamic multi-scale spatial attenuation module also includes adaptive feature fusion. The formula is shown below:

[0061] Where G represents the adaptive gating weights, controlling the fusion ratio of the two-scale attention. σ represents the sigmoid activation function. Given a gating weight matrix, a fusion strategy is learned. This represents the gated bias vector, providing the bias term. Channel dimensions are spliced ​​together, and multi-scale features are merged. This indicates element-wise multiplication and element-level weighted fusion. The attention output after fusion. λ represents the interaction term coefficient, which controls the intensity of multi-scale feature interactions.

[0062] Enhanced local location encoding (supplementing local feature space information). The depthwise separable convolution is extended to a deformable convolution enhancement, as shown in the following formula:

[0063] Here, α represents the deformable convolution weight coefficient, a learnable scalar (initial value 0.5), which balances the contributions of the two convolutions. This represents a deformable convolutional offset field that predicts the positional offset of each sampling point. It is an offset field predicted from Q and K by a small network, enabling local enhancement to adaptively focus on important regions. This represents deformable convolution, same-dimensional mapping, and local features with adaptive receptive field.

[0064] This application provides a specific embodiment of a training method for an object detection model. This training method is applicable to the improved object detection model in the above embodiments, and the specific training process is as follows: Network setup and training: Environment Configuration Specifications. The training environment is built on a high-performance computing platform. The computing hardware uses an NVIDIA GeForce RTX 4090 or equivalent graphics card with at least 24GB of VRAM to ensure support for large-scale batch data processing and mixed-precision training. The software stack is based on the Ubuntu 22.04 LTS operating system, equipped with the PyTorch 2.0.0 or later deep learning framework, and CUDA 11.8 toolkit to fully utilize the GPU's parallel computing capabilities. Python version 3.9 or later is recommended, and project dependencies should be managed through virtual environments.

[0065] Core training hyperparameters. Model optimization employs stochastic gradient descent (SGD) with a momentum parameter set to 0.937 and a weight decay of 5e-4 for regularization. The learning rate strategy starts at 0.01 and includes a three-round linear warm-up phase during the initial training phase, gradually increasing the learning rate from 0.001 to the initial value to ensure training stability. The model is trained for a total of 300 rounds with a batch size of 32, which can be dynamically adjusted based on actual GPU memory usage. The network input images are uniformly scaled to 640x640 resolution; this resolution can be increased appropriately to address the typical characteristics of small, distant targets in low-altitude scenes. The positive and negative sample matching mechanism uses an advanced Task-Aligned Assigner to replace the traditional prior anchor box-based method, thereby improving the alignment accuracy between the learned target and the detection task.

[0066] Training Strategy and Optimization. The learning rate scheduling primarily employs cosine annealing to smoothly decrease the learning rate throughout the training cycle. Alternatively, a segmented decrease strategy can be configured, for example, reducing the learning rate to 90% of the previous stage after every 50 training iterations. To suppress overfitting, in addition to weight decay, Dropout layers can be introduced into the fully connected layers of the network when the training dataset size is limited, with a recommended dropout rate range of 0.2 to 0.5. To further ensure numerical stability and efficiency during training, Automatic Mixed Precision (AMP) is enabled to accelerate computation, and a gradient pruning mechanism with a global gradient norm threshold of 10.0 is set.

[0067] This configuration is specifically designed and optimized for training and optimizing flame and smoke detection models from the perspective of low-altitude UAVs, emphasizing a balance between model accuracy and training convergence efficiency. In practical applications, users can fine-tune hyperparameters such as batch size, learning rate, and regularization strength according to the distribution characteristics of a specific dataset and available hardware resources.

[0068] Model Functionality: This model focuses on early fire warning and monitoring from the perspective of low-altitude drones. It achieves real-time analysis of continuous low-altitude visual data by accessing and simultaneously processing multi-source video streams (including drone aerial photography and low-altitude fixed monitoring). The system is built upon a deep convolutional neural network and a Transformer architecture. This network has been specifically trained and optimized using large-scale low-altitude flame and smoke samples, enabling efficient and accurate parsing of each frame of the input video.

[0069] The core function lies in the intelligent identification and localization of flames and smoke appearing in the image: the model can not only accurately mark their two-dimensional spatial positions in real time with bounding boxes, but also simultaneously output a confidence score for each detected target, providing a quantitative basis for the credibility of the results. In addition, the system can continuously track and judge the behavior of suspected targets through inter-frame correlation analysis.

[0070] Model iteration: 1. Incremental dataset construction and sample classification First, to address the performance bottlenecks exposed by the current model during actual deployment or validation, a multimodal, multi-scenario incremental dataset is systematically constructed. This dataset covers typical environments (such as lighting changes, occlusion, and complex backgrounds) and edge conditions (such as low resolution and motion blur) through scene sampling, ensuring that data diversity meets the generalization requirements of the model.

[0071] Based on the constructed incremental dataset, forward inference of the model is performed, and refined sample classification and annotation are carried out for each image by combining the evaluation criteria of the object detection task (such as IoU threshold and confidence score): Hard Positives: Samples with an IoU ≥ 0.7 between the target region and the ground truth and a model detection confidence ≥ 0.8; False Positives (FPs): Background regions that the model misclassifies as targets (with IoU < 0.3 with all true labels and confidence ≥ 0.7); False Negatives (FNs): Regions where the target actually exists but the model does not output a detection box (IoU ≥ 0.5 with the ground truth label and the model does not trigger detection). True negative samples (TNs): Background regions in an image where there are no targets and the model does not output detection boxes.

[0072] 2. Difficult Case Mining and Data Augmentation Strategies Based on the classification results, implement differentiated sample utilization strategies: Negative sample augmentation: False positive samples (FPs) are included in the negative sample library. Through geometric transformation (random rotation ±15°, scaling 0.8-1.2 times), color perturbation (brightness ±20%, contrast ±15%) and blurring (Gaussian blur σ=1-3), a variety of negative samples are generated to alleviate the overfitting of the model to background noise. Positive sample completion: Fine-grained annotation of missed samples (FNs) is carried out using a two-stage annotation method. First, the target bounding box is manually corrected using a high-precision annotation tool (such as LabelImg++) to ensure that the IoU between the annotation box and the real label is ≥0.9. Second, for small targets or low-contrast targets, super-resolution reconstruction (such as ESRGAN) is introduced to enhance the visibility of the target and improve the annotation quality.

[0073] Meanwhile, data augmentation is performed on valid positive samples (including explicit positive samples and supplementary missed samples). In addition to basic geometric transformations, additional mosaic augmentation (4-image stitching), MixUp (image and label mixing) and CutOut (random occlusion) are added to force the model to learn robust feature representations of the target in different contexts.

[0074] 3. Incremental Model Training and Parameter Optimization Using the original model as the initial weights, incremental training is performed based on the incrementally constructed positive sample set (explicit positive samples + supplemented missed samples) and the augmented negative sample set. During training, a dynamic learning rate strategy (such as Cosine Annealing with Warmup) is adopted. The initial learning rate is set to 1e-4, and differentiated learning rates are set for different network modules (such as the backbone network learning rate of 1e-5 and the detection head learning rate of 5e-5).

[0075] To address the issue of missed detections, the design of the loss function was optimized: Focal Loss was introduced to replace the traditional cross-entropy. By adjusting the focusing parameter α=0.75 and the modulation factor γ=2, the loss weight of easily distinguishable samples (high-confidence positive samples) was reduced, and the learning of difficult-to-distinguish samples (low-confidence positive samples and missed detection samples) was focused. At the same time, GIoU Loss was added to the detection head to replace IoULoss, thereby optimizing the localization accuracy of bounding box regression.

[0076] 4. Multi-dimensional performance evaluation and iterative closed loop After training, the model performance is evaluated using a combination of quantitative metrics and qualitative analysis: Quantitative evaluation: Using a public dataset evaluation protocol, we calculated the core metrics for object detection—mean precision (mAP@0.5:0.95), recall (Recall@100), precision (Precision@0.5), and FPS (frames per second)—and compared the performance differences between the new and old models. Qualitative analysis: Select test set images from scenarios with high rates of missed detection (such as low light at night, rainy or foggy weather), and compare the detection box coverage and localization accuracy of the new and old models using visualization tools (such as TensorBoard) to identify the remaining defects in the localization model (such as missed detection of specific categories or small targets).

[0077] If the new model does not meet the preset performance threshold (e.g., mAP@0.5≥90% or Recall≥95%), then enter the iterative optimization loop: Data level: Supplement difficult case data based on the evaluation results (e.g., collect extended datasets under the same scene and lighting conditions for missed detection scenarios of a certain type of target). At the algorithm level: adjust the network architecture (such as increasing the number of backbone network layers, replacing with a more efficient detector head), optimize the loss function (such as introducing CIoU Loss to replace GIoU Loss) or introduce regularization techniques (such as Dropout, weight decay) to alleviate overfitting; At the training strategy level: Try more complex training techniques (such as knowledge distillation, using a large model to guide a small model), mixed precision training (FP16 / FP32 mixed computation) or distributed training (multi-GPU parallel acceleration) to improve training efficiency and model generalization.

[0078] Repeat the above process until the model achieves optimal performance within the existing data domain and meets the real-time and accuracy requirements in actual deployment scenarios, and finally outputs the iteratively optimized target detection model.

[0079] It should be noted that the steps shown in the above process or in the flowchart of the accompanying figures can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0080] This application provides a drone-based low-altitude fire monitoring system. The system is used to execute the method provided in the above embodiments, and includes a data acquisition module, an identification module, and an early warning module. The data acquisition module is used to collect low-altitude imaging data within the monitoring area via drones; The identification module is used to identify fires in the monitoring area based on low-altitude imaging data and a trained improved target detection model to obtain the identification results. The improved target detection model includes an intra-group self-attention and convolution hybrid module and a multi-scale feature fusion module. The early warning module is used to trigger fire level early warning and fire simulation prediction based on the identification results.

[0081] This embodiment provides an electronic device including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0082] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0083] Optionally, the electronic device may further include a processor, memory, network interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for monitoring low-altitude fires using an unmanned aerial vehicle (UAV). The display screen may be an LCD screen or an e-ink screen. The input device may be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0084] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0085] Furthermore, in conjunction with the UAV low-altitude fire monitoring method in the above embodiments, this application embodiment can provide a storage medium for implementation. This storage medium stores a computer program; when executed by a processor, the computer program implements any of the UAV low-altitude fire monitoring methods in the above embodiments.

[0086] In one embodiment, Figure 9 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application, such as... Figure 9 As shown, an electronic device is provided, which can be a server, and its internal structure diagram can be as follows. Figure 9As shown, the electronic device includes a processor, a network interface, internal memory, and non-volatile memory connected via an internal bus. The non-volatile memory stores an operating system, computer programs, and a database. The processor provides computing and control capabilities, the network interface communicates with external terminals via a network, the internal memory provides an environment for the operation of the operating system and computer programs, the computer programs are executed by the processor to implement a method for monitoring low-altitude fires using a drone, and the database stores data.

[0087] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. A specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0088] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0089] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0090] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for monitoring low-altitude fires using unmanned aerial vehicles (UAVs), characterized in that, The method includes: Low-altitude imaging data of the monitoring area is collected by drones; Based on the low-altitude imaging data, a fire is identified in the monitoring area using a trained improved target detection model, and the identification result is obtained. The improved target detection model includes an intra-group self-attention and convolution hybrid module and a multi-scale feature fusion module. The self-attention and convolution hybrid module within the group includes a convolution operator and a self-attention operator that share an initial feature projection, and the convolution operation and self-attention operation of the operator are executed in parallel before feature fusion. The multi-scale feature fusion module has a multi-scale feature fusion branch, a deformable convolution branch, a bilinear enhanced multi-head self-attention mechanism branch, and a pyramid pooling branch. The features output by the four branches are fused and then concatenated with residual links. Based on the identification results, fire grading early warning and fire simulation prediction are triggered.

2. The method according to claim 1, characterized in that, The improved target detection model also includes a dynamic multi-scale spatial attenuation module; The dynamic multi-scale spatial decay module introduces a multi-scale decay matrix to enable the model to adapt to different distance dependencies, and introduces hierarchical multi-scale decomposition to enable the model to capture diagonal direction information.

3. The method according to claim 1, characterized in that, The improved target detection model also includes a channel-enhanced attention mechanism module; The channel-enhanced attention mechanism module contains three parallel convolutional channels. It learns the correlation between channels through a self-attention mechanism to dynamically adjust the importance weights of different channels, thereby reducing the loss of spatial information caused by pooling.

4. The method according to claim 1, characterized in that, The improved object detection model also includes a multi-scale dynamic grouping attention module; The multi-scale dynamic grouping attention module extracts features at multiple scales and compresses spatial information through global average pooling, thereby reducing the computational complexity of the model.

5. The method according to claim 1, characterized in that, Low-altitude imaging data collected by drones within the monitoring area includes: Based on the collaborative operation of fixed UAV airports and UAVs, low-altitude imaging data is collected by UAVs within the monitoring area. The low-altitude imaging data includes visible light imaging data, infrared thermal imaging data, and depth imaging data.

6. The method according to claim 5, characterized in that, The method includes: Based on the low-altitude imaging data collected by the UAV, the Neural Radiation Field (NeRF) technology is used to model the three-dimensional scene of the monitoring area, and combined with the digital elevation model, a three-dimensional geographic information model of the monitoring area is constructed.

7. The method according to claim 6, characterized in that, The method includes: Environmental data within the monitoring area is collected synchronously through a ground-based IoT sensor network. This environmental data includes temperature and humidity data, as well as wind speed and direction data.

8. The method according to claim 7, characterized in that, Based on the identification results, triggering fire simulation prediction includes: Based on the identification results, if the identification results indicate that a fire point has been detected within the monitoring area, then after triggering a level-two fire warning, the fire spread path and speed will be dynamically simulated in real time based on the three-dimensional geographic information model of the monitoring area, the temperature and humidity data, and the wind speed and direction data to provide decision support for the deployment of rescue forces.

9. The method according to claim 1, characterized in that, Based on the identification results, triggering a fire hazard level warning includes: Based on the identification results, if the identification results indicate that abnormal temperature or smoke is detected, a level one fire warning is triggered to drive the drone to patrol and verify the fire point in the monitoring area. If the identification result indicates that a fire point has been detected within the monitoring area, a level-two fire warning is triggered to make a preliminary judgment on the fire based on the abnormal temperature area and the identification frame, and to push alarm information to the on-duty personnel. After the on-duty personnel confirm the existence of the alarm information, a level-three fire warning is triggered to coordinate with fire and emergency departments to activate the emergency response plan.

10. A low-altitude fire monitoring system for unmanned aerial vehicles (UAVs), characterized in that, The system is used to perform the method according to any one of claims 1 to 9, and the system includes a data acquisition module, an identification module, and an early warning module; The acquisition module is used to acquire low-altitude imaging data within the monitoring area via a drone; The identification module is used to identify fires in the monitoring area based on the low-altitude imaging data using a trained improved target detection model, and obtain the identification result. The improved target detection model includes an intra-group self-attention and convolution hybrid module and a multi-scale feature fusion module. The early warning module is used to trigger fire level early warning and fire simulation prediction based on the identification results.

Images