Grassland fire point detection and positioning method based on unmanned aerial vehicle

Abstract

The invention provides a grassland fire point detection and positioning method based on an unmanned aerial vehicle. The method comprises the following steps: Step 1, constructing a scene three-dimensional model by using a multi-overlapping-degree image; 2, detecting a suspicious target by using an optimized grassland fire point detection and positioning YOLOv8s model deployed on the unmanned aerial vehicle, calculating the orientation of a target point in an unmanned aerial vehicle coordinate system, automatically adjusting the flight direction of the unmanned aerial vehicle and the camera direction, and continuously approaching the suspicious target point; 3, the scene three-dimensional model established in the step 1 is used for accurately positioning the suspicious target, and the position # imgabs1 # of the target under the world coordinate system is obtained from the image plane coordinate # imgabs0 # of the target. Based on the optimized YOLOv8s model, steppe smoke and fire points are detected and tracked, and efficient detection and accurate positioning of the steppe fire points are achieved in combination with the live-action three-dimensional model. The optimized YOLOv8s shows higher recognition capability than other models, has 90.82% high-precision positioning when RTK is used, and can effectively recognize early fire points and feed back positioning information to a management department in time.

Description

Technical Field

[0001] The present invention relates to the technical field of fire identification, and in particular to a grassland fire point detection and positioning method based on an unmanned aerial vehicle (UAV). Background Art

[0002] At present, wildfire detection research is mostly focused on forest fires, and related research on grassland fires is relatively incomplete: (1) Current research mostly uses satellites to obtain macroscopic multispectral images and conducts analysis, relying on past data and lacking real-time performance; (2) Compared with the important role played by drones in forest fire detection, grassland fire research mostly ignores the huge potential of drones as an emerging data acquisition platform; (3) The current plan focuses more on processing images after completing routine patrol missions, obtaining fire information, and then calling the platform for continuous reconnaissance of fires, or calculating relevant indexes for early warning management. This has caused great trouble for timely detection of fires and emergency response, and lacks flexibility and real-time performance. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a grassland fire point detection and positioning method based on drones, which identifies suspicious smoke and fire points in aerial images, then determines the direction of the suspicious target and guides the drone to approach for fire reconnaissance, that is, replans the route of the key detection area; fire point positioning refers to using the fire image taken by the drone to perform real-scene three-dimensional reconstruction, combining the identified fire point information, calculating and projecting the exact position of the fire point on the three-dimensional model, and returning the relevant information to the management department, so as to realize the timely discovery and emergency response of grassland fires.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is: A method for detecting and locating grassland fire points based on drones, including: Step 1: Using drones to build a 3D model of the scene using multi-overlapping images; Step 2: Use the optimized grassland fire detection and positioning YOLOv8s model deployed on the drone to detect suspicious targets, including smoke and grassland fire, and locate suspicious targets based on the image coordinates of the suspicious targets. The position of the drone relative to the visible light sensor is used to calculate the target point's position in the drone's coordinate system, and the drone's flight direction and camera direction are automatically adjusted to continuously approach the suspicious target point. Step 3: Use the three-dimensional model of the scene established in Step 1 to accurately locate the suspicious target, and , get the target's world coordinate position .

[0005] The above-mentioned Step 1 process includes using the drone to pre-plan the flight route, the drone system to automatically collect images of the target scene range, and to use multi-overlapping images to construct a three-dimensional model of the scene.

[0006] In the above Step 1, the UAV performs automatic terrain-simulating route planning and terrain-simulating photogrammetry under the guidance of the initial three-dimensional model DEM to capture high-resolution images of the surface.

[0007] The process of building a 3D model of the scene in Step 1 above is as follows: based on multi-view drone images, the SIFT algorithm is used to extract image feature points and establish matching relationships between feature points; with the help of the SfM algorithm, a sparse point cloud is constructed; based on the multi-view stereo matching algorithm, a dense point cloud is generated to increase model details; finally, a grid generation method based on the Poisson equation is used to convert the dense point cloud into a triangular mesh, and the image texture is mapped onto the mesh to generate a real-life 3D model of the detection area.

[0008] In the above Step 3, the precise positioning of the suspicious target is divided into two modes: single-chip positioning and multi-chip positioning.

[0009] When single-chip positioning is used for suspicious targets in Step 3 above, the drone is equipped with real-time dynamic positioning RTK and computing unit, combined with the Ray Tracing algorithm, from the detected target point coordinate Reverse the direction of light in the camera coordinate system:

[0010] in, and is the focal length of the lens, (c x , c y ) is the coordinate of the principal point on the image plane; According to the drone's pose, the light direction is converted to the world coordinate system, and the light parameterization equation is constructed:

[0011]

[0012] in, represents the distance the light travels, Represents the path of light in the world coordinate system, The optical center of the camera for RTK positioning, is the initial rotation matrix. Using the ray tracing algorithm, calculate the nearest intersection point between the ray and the scene model:

[0013] in Represents the scene model, Indicates the intersection distance between the light and the model.

[0014] When using multi-frame positioning for suspicious targets in Step 3 above, the fire point is located by combining multi-frame images with the real-scene 3D model. Multi-view geometric constraints are used to avoid dependence on RTK. Based on a single image, the light direction of the target pixel in the camera coordinate system is calculated and intersected with the 3D model:

[0015] in The ray equation is generated for each image as the intersection distance between the ray and the model:

[0016] Minimize the sum of the squares of the distances from all rays to the target point:

[0017] in, is the light parameter and needs to be Joint optimization; conversion to a system of linear equations , and solve; where:

[0018] The network structure of the YOLOv8s model for grassland fire detection and positioning in Step 2 above includes four parts: input, backbone, neck and prediction. The backbone network is based on CSPDarknet53, combined with CSP (CrossStage Partial) structure, and realizes cross-stage connection of some feature maps through C2f (Cross Stage Partial Fusion with 2 convolutions) module; Neck is based on PANet combined with PAN (Path Aggregation Network) and FPN (feature pyramid networks), and fuses feature maps of different scales through bottom-up and top-down paths; the final prediction result is generated through the convolutional layer, including category probability, confidence and bounding box coordinates, and CIoU_Loss is used as the loss function of the bounding box to optimize the bounding box regression. BCE-loss calculates the classification loss and confidence loss. Finally, non-maximum suppression NMS is used to screen multiple targets and make prediction output.

[0019] The C2f module in the backbone network uses a lightweight network design method GhostNet to generate a set of "intrinsic feature maps" using a small number of standard convolutions. It then applies a series of linear operations to these intrinsic feature maps to generate "ghost feature maps." The intrinsic feature maps and ghost feature maps are concatenated together to form the final output feature map. The bidirectional feature pyramid network BiFPN is used in Neck. Based on PANet, the bidirectional feature pyramid network BiFPN uses bidirectional connections and weighted feature fusion mechanisms to enable bidirectional transmission of features at each scale, dynamically adjust the importance of different features, and better integrate feature information at different scales.

[0020] The above-mentioned grassland fire point detection and positioning YOLOv8s model introduces the channel attention mechanism SE module; and mixes the non-maximum suppression Mixed-NMS method.

[0021] The present invention mentions a method for detecting and locating grassland fire points based on drones, which includes three parts: identification of suspicious targets, target tracking, and fire point positioning. The identification and tracking of suspicious targets refers to the identification of suspicious smoke and fire points in aerial images based on the optimized YOLOv8s model during routine drone inspections, and then the determination of the orientation of the suspicious target and the guidance of the drone to approach for fire reconnaissance, i.e., re-planning of routes in key detection areas; fire point positioning refers to the use of fire images taken by drones to perform real-scene three-dimensional reconstruction, and the calculation and projection of the exact position of the fire point on the three-dimensional model in combination with the identified fire point information, and the return of relevant information to the management department, so as to achieve timely discovery and emergency response of grassland fires. Compared with other grassland fire detection methods, this article has obvious advantages in real-time, accuracy and practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The present invention will be further described below with reference to the accompanying drawings and examples: Figure 1 Schematic diagram of the structure of the channel attention mechanism SE module in Example 1 of the present invention; Figure 2 Schematic diagram of automatic terrain-simulating route planning and terrain-simulating photogrammetry for a UAV in Example 1 of the present invention; Figure 3 This is a schematic diagram of single-chip positioning in fire point positioning in Example 1 of the present invention; Figure 4 This is a schematic diagram of multiple pieces positioning in fire point positioning in Example 1 of the present invention; Figure 5 Schematic diagram of the confusion matrix of the test results in Example 2 of the present invention; Figure 6Schematic diagram of a drone approaching a target in Example 2 of the present invention. DETAILED DESCRIPTION

[0023] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0024] Example 1: A grassland fire point detection and positioning method based on drones, comprising: YOLOv8s model optimization based on grassland fire detection Fire point detection is the core of grassland fire detection and localization using drones. Currently, fire point detection methods are mainly divided into traditional multispectral imagery-based methods and deep learning-based methods. The core concept of multispectral imagery-based fire point detection methods is to extract fire characteristics through techniques such as band ratios. While computationally simple and easy to implement, these methods are prone to false positives or false negatives in complex environments (such as smoke obscuration and vegetation cover). In recent years, neural network-based fire point detection methods have become a research hotspot. Some studies use CNNs to extract deep features of fires, while others employ object detection algorithms to directly locate fire areas in imagery and output bounding boxes. The You Only Look Once (YOLO) series is one of the most well-known and widely used algorithms in the field of object detection. Compared to other algorithms in the YOLO series, YOLOv8 achieves new heights in speed, accuracy, and versatility. YOLOv8s balances speed and accuracy, making it even more useful in drone object detection.

[0025] YOLOv8s model basic architecture YOLOv8s divides the entire network structure into four parts: input, backbone, neck, and prediction. The backbone network is based on CSPDarknet53 and incorporates the Cross Stage Partial (CSP) architecture. The Cross Stage Partial Fusion with 2 convolutions (C2f) module enables cross-stage connections of some feature maps, improving feature extraction efficiency and reducing computational complexity. The neck (PANet) combines the Path Aggregation Network (PAN) with the Feature Pyramid Network (FPN) to fuse feature maps of different scales through bottom-up and top-down approaches, significantly improving the model's ability to detect small objects. Convolutional layers generate the final prediction, including class probabilities, confidence scores, and bounding box coordinates. CIoU_Loss is used as the bounding box loss function to optimize bounding box regression. BCE-loss is used to calculate the classification and confidence losses. Finally, non-maximum suppression (NMS) is used to filter multiple objects and output predictions.

[0026] YOLOv8s model optimization Achieving early grassland fire detection requires a high-performance detection method. This paper optimizes YOLOv8s to improve the accuracy and real-time performance of early grassland fire detection.

[0027] 1) Modifying the C2f Module: GhostNet is a lightweight network design that reduces computation while maintaining high feature expressiveness by generating "ghost" feature maps. The Ghost module is the core component of GhostNet. It first uses a small number of standard convolutions to generate a set of "intrinsic feature maps." These intrinsic feature maps are then subjected to a series of inexpensive linear operations (such as depthwise convolutions) to generate "ghost" feature maps. These intrinsic and ghost feature maps are then concatenated to form the final output feature map. Replacing some convolutions with Ghost modules further compresses the model, reducing computation and parameter count while maintaining feature expressiveness, thereby improving operational efficiency.

[0028] 2) Bidirectional Feature Pyramid Network (BiFPN): BiFPN, or Bidirectional Feature Pyramid Network, is a highly efficient multi-scale feature fusion network. Building on PANet, BiFPN utilizes bidirectional connections and a weighted feature fusion mechanism to enable bidirectional transmission of features at each scale and dynamically adjust the importance of different features, effectively integrating information from different scales. This study used both small flames and smoke as detection targets, given the significant difference in their characteristics. This architecture improves the ability to fuse multi-scale features and process different detection targets, enhancing detection accuracy while reducing the network model size.

[0029] 3) Introducing the Channel Attention Mechanism (SE) module: The Channel Attention Mechanism (CAM) is an attention mechanism used to enhance the feature expression capability of neural networks. In the Squeeze step, the Squeeze-and-Excitation (SE) module compresses the global spatial information of each channel into a channel description vector through global average pooling to capture global context information; in the Excitation step, the weight of each channel is calculated through a lightweight fully connected layer and normalized. Finally, the calculated channel weight is multiplied with the original feature map to achieve feature reweighting. Its structure is as follows: Figure 1As shown in the figure, for the early-stage grassland fire detection task from a drone perspective, it is necessary to enhance the multi-scale feature representation of smoke and small fire points and suppress irrelevant noise. Therefore, the SE module is inserted into the feature fusion module to enhance the multi-scale feature representation capability.

[0030] 4) Mixed Non-Maximum Suppression (Mixed-NMS) Method: YOLOv8s uses traditional NMS (Non-Maximum Suppression) by default, which primarily filters overlapping boxes based on IoU (Intersection over Union) but only considers the overlapping area of the boxes. Mixed-NMS introduces the Skew-NMS and DIoU-NMS methods, which can consider the angle of the target and the distance from the center point of the box. This can better handle the spatial relationship between target boxes and reduce false detections and missed detections. Mixed-NMS considers angle as a parameter to be considered, sorts pre-selected boxes in descending order based on confidence, sets two thresholds, and applies different penalty operations for different DIoU intervals. This can significantly reduce false detections and missed detections in the case of small or tilted objects.

[0031] Drone grassland fire detection and location This paper develops a novel drone-based grassland fire detection and location method based on an optimized YOLOv8s model. This method enables precise positioning of grassland fires using drones. The method consists of three main components: 3D reconstruction of the key detection area, online identification of suspicious targets (including smoke and small fires), and tracking and location of fires.

[0032] A 3D reconstruction of the real scene in the detection area In order to accurately locate the fire point, in addition to utilizing the drone's own observation direction and attitude, a three-dimensional scene model with a certain spatial resolution is also required to meet the positioning requirements. At present, the method of using multi-overlap images to construct a three-dimensional model of the scene based on low-altitude rotor drones is relatively mature. Near-low-altitude photogrammetry uses drones, and the flight route can be planned in advance. The drone system can perform fully automatic image acquisition of the target scene range. This type usually regards the photographic area as a plane area. In grassland environments, the reconstructed scene is relatively simple and open, but sometimes there are scenes with large and complex surface elevation changes. In order to ensure safety and fully automatic close-range flight missions, it is necessary to perform automatic terrain-simulating route planning under the guidance of the initial low-precision three-dimensional model. Such as Figure 2 ,Under the guidance of low-precision DEM, the UAV conducts ,topographic measurement of the target area and captures high-resolution images of the surface.

[0033] In this paper, based on multi-view UAV imagery, the SIFT algorithm is used to extract image feature points and establish matching relationships between feature points. With the help of the SfM algorithm, a sparse point cloud is constructed. Based on the multi-view stereo matching algorithm, a dense point cloud is generated to increase model details. Finally, a mesh generation method based on the Poisson equation is used to convert the dense point cloud into a triangular mesh, and the image texture is mapped onto the mesh to generate a realistic real-scene 3D model of the key detection area.

[0034] B. Online target recognition and tracking based on UAV The present invention deploys the trained YOLOv8s optimization model on the UAV, which can detect suspicious targets (smoke, grassland fire) in a timely manner with high accuracy during the UAV inspection process, and The position of the drone relative to the visible light sensor is used to calculate the orientation of the target point in the drone coordinate system, automatically adjust the flight direction of the drone and the direction of the camera, continuously approach the suspicious target point, and carry out subsequent fire point positioning work.

[0035] C Fire Point Positioning Based on the detected fire point and the high-precision real-scene 3D reconstruction model, the fire point can be accurately located. Image plane coordinates , get the target The world coordinate position Depending on the hardware the drone is equipped with, you can Figure 3 and Figure 4 There are two situations to handle.

[0036] a) Single-chip positioning: When the UAV is equipped with real-time dynamic positioning (RTK) and a high-performance computing unit, combined with the Ray Tracing algorithm, centimeter-level positioning accuracy and robustness in complex environments can be achieved. coordinate Reverse the direction of light in the camera coordinate system:

[0037] in, and is the focal length of the lens, (c x , c y ) is the coordinate of the principal point on the image plane; According to the drone's pose, the light direction is converted to the world coordinate system, and the light parameterization equation is constructed:

[0038]

[0039] in, represents the distance the light travels, Represents the path of light in the world coordinate system, The optical center of the camera for RTK positioning, is the initial rotation matrix. Using the ray tracing algorithm, calculate the nearest intersection point between the ray and the scene model:

[0040] in Represents the scene model, Indicates the intersection distance between the light and the model.

[0041] b) Multi-image positioning: Combine multiple frames of images with the real-world 3D model to locate the fire point. The core of this is to use multi-view geometric constraints to avoid the high-precision reliance on RTK. Based on a single image, the light direction of the target pixel in the camera coordinate system is quickly calculated and intersected with the 3D model:

[0042] in is the intersection distance between the light and the model. Generate the light equation for each image:

[0043] Minimize the sum of the squares of the distances from all rays to the target point:

[0044] in, is the light parameter and needs to be Joint Optimization. Convert to a System of Linear Equations , and solve. Among them:

[0045] Example 2: ATest environment: This experiment used ground-based computing power, namely a personal computer equipped with an NVIDIA RTX 2000Ada 8GB, to train the model. The operating system used was Windows 11, the deep learning framework used was PyTorch 11.3, and Python version 3.8. The training image size was set to 640×640, the batch size was 4, the learning rate was 0.01, and the number of training epochs was 300. The test area was a wasteland near Panlong Avenue in Wuhan, Hubei Province, covering an area of approximately 28,000 square meters. The image data for this experiment was captured by a hexacopter drone equipped with a visible light sensor.

[0046] B data settings: Samples containing early grassland fire features are taken as positive samples, and samples not containing early grassland fire features are taken as negative samples, that is, samples of grassland images with only interference or similar to fire scenes. First, a drone patrol of the grassland is simulated in the test area to obtain multi-perspective early grassland fire videos. Secondly, in order to increase the diversity of the scene, early grassland fire image videos from the perspective of drones are collected from the Internet and public datasets

[28] and frame sampling is performed. A total of 10,405 early grassland fire images (positive samples) are obtained, which are divided into a training set (8,324) and a validation set (2,081) in a ratio of 4:1. In addition, 236 grassland images (negative samples) are collected and put into the validation set. The image labeling tool LabelImg is used to annotate the early grassland fire samples, and flames and smoke are marked as two detection targets respectively. The flame features of the fire are marked with the label "fire", and the smoke features of the fire are marked with the label "smoke". The distribution of various samples in the training set and validation set is shown in Table 1.

[0047] Table 1

[0048] C Experimental Data and Analysis a) Dataset validation test This study made a detailed analysis of the construction of the early grassland fire dataset required for the image-based early grassland fire detection technology based on the YOLOv8s model and built its own dataset. , recall rate and average accuracy The test results are analyzed using the indicators. The verification results are shown in Table 2: Table 2

[0049] According to the test results, it is found that the flame correct detection rate is high when using this model on this data set, and the missed detection rate and false detection rate are relatively low. For smoke, the effect is relatively general, with a missed detection rate of about 0.19 and a false detection rate of about 0.15. Observe the confusion matrix, as shown in Figure 5 , the prediction rate of fire as other objects is relatively low, while the prediction rate of smoke as background is relatively high. This indicates that in grassland environments, smoke is relatively easily confused with the background, leading to missed detections, while fire points are detected relatively accurately. This means that in practical applications, attention should be paid to the possibility of missed smoke detection leading to grassland fires. Smoke features should be combined with other features (such as fire) for detection to reduce the damage caused by missed detections. Furthermore, the significant difference in accuracy between the GoogleNet model and the YOLOv8 series models may be due to their limited feature extraction capabilities and their unsuitability for extracting fine-grained image features.

[0050] b) Target tracking and fire point location test In the experimental area, drones were used to track important targets (smoke and grassland fires). A basic drone flight path was set as the inspection route. Fire points were set along the drone's flight path to test whether the drone could actively approach the suspicious target. The drone's operation was as follows:

[0051] After detecting a suspicious target, the existing real-world 3D model was used to extract the approximate geographic location of the fire point using image orientation parameters. The drone then planned its route based on the thermal power target. This paper conducted 100 target tracking and fire point positioning experiments, using both RTK information and non-RTK information during the positioning process. The test results are shown in Table 3: Table 3

[0052] According to the test results, the overall target tracking success rate was 98%, indicating that the target detection and tracking algorithm performed well in most cases. However, the target was not tracked in 2% of the cases. It was found that when only light smoke was present within the drone's inspection range, it was difficult to detect the target in smoke, which could lead to the target being overlooked. However, when the drone successfully tracked the target, positioning accuracy was significantly improved when using RTK. 90.82% of the tests were in the high-accuracy range (<0.5 meters), of which 15.31% achieved centimeter-level accuracy (<0.1 meters). Without RTK, only 40.81% of the tests were in the high-accuracy range, and 59.19% of the tests had low accuracy (>0.5 meters). However, in grassland environments, this accuracy was sufficient to detect the early danger of grassland fires and initiate timely response, validating the effectiveness of the proposed algorithm.

[0053] D Conclusion This paper proposes a method for early grassland fire detection and precise positioning based on an unmanned aerial vehicle (UAV) platform. Using an optimized YOLOv8s model, grassland smoke and fire points are detected and tracked. Combined with a real-world three-dimensional model, this method achieves efficient detection and precise positioning of grassland fire points. Experimental results show that this method exhibits high detection accuracy and positioning precision in complex grassland environments. The optimized YOLOv8s model demonstrates higher recognition capabilities than other models. It also achieves 90.82% high-precision positioning when using RTK, effectively identifying early fire points and promptly feeding back positioning information to management departments. Overall, this method provides a feasible solution for early warning and rapid response to grassland fires. Future work will focus on improving the accuracy of early grassland fire feature recognition and positioning.

Claims

1. A method for detecting and locating grassland fire points based on drones, characterized in that: include: Step 1: Using drones to build a 3D model of the scene using multi-overlapping images; Step 2: Use the optimized grassland fire detection and positioning YOLOv8s model deployed on the drone to detect suspicious targets, including smoke and grassland fire, and locate suspicious targets based on the image coordinates of the suspicious targets. The position of the drone relative to the visible light sensor is used to calculate the target point's position in the drone's coordinate system, and the drone's flight direction and camera direction are automatically adjusted to continuously approach the suspicious target point. Step 3: Use the three-dimensional model of the scene established in Step 1 to accurately locate the suspicious target, and , get the target's world coordinate position .

2. The grassland fire point detection and positioning method based on drone according to claim 1 is characterized in that: The process of Step 1 includes using a drone to pre-plan a flight route, the drone system to automatically collect images of the target scene range, and to construct a three-dimensional model of the scene using multi-overlapping images.

3. The grassland fire point detection and positioning method based on drone according to claim 2 is characterized in that: In the aforementioned Step 1, the UAV performs automatic terrain-simulating route planning and terrain-simulating photogrammetry under the guidance of the initial three-dimensional model DEM to capture high-resolution images of the ground surface.

4. The method for detecting and locating grassland fire points based on drones according to claim 3, characterized in that: The process of constructing the three-dimensional model of the scene in Step 1 is as follows: based on multi-view UAV images, the SIFT algorithm is used to extract image feature points and establish matching relationships between feature points; with the help of the SfM algorithm, a sparse point cloud is constructed; Based on the multi-view stereo matching algorithm, a dense point cloud is generated to increase model details; finally, a mesh generation method based on the Poisson equation is used to convert the dense point cloud into a triangular mesh, and the image texture is mapped onto the mesh to generate a real-life 3D model of the detection area.

5. The grassland fire point detection and positioning method based on drone according to claim 1 is characterized in that: The precise positioning of the suspicious target in Step 3 is divided into two modes: single-chip positioning and multi-chip positioning.

6. The method for detecting and locating grassland fire points based on drones according to claim 5, characterized in that: When the single chip positioning is used for the suspicious target in Step 3, the UAV is equipped with real-time dynamic positioning RTK and computing unit, combined with the Ray Tracing algorithm, from the detected target point coordinate Reverse the direction of light in the camera coordinate system: in, and is the focal length of the lens, (c x , c y ) is the coordinate of the principal point on the image plane; According to the drone's pose, the light direction is converted to the world coordinate system, and the light parameterization equation is constructed: in, represents the distance the light travels, Represents the path of light in the world coordinate system, The optical center of the camera for RTK positioning, is the initial rotation matrix; use the ray tracing algorithm to calculate the nearest intersection point between the ray and the scene model: in Represents the scene model, Indicates the intersection distance between the light and the model.

7. The grassland fire point detection and positioning method based on drone according to claim 5 is characterized in that: When using multi-frame positioning for suspicious targets in Step 3, the fire point is located by combining multi-frame images with the real-scene 3D model. Multi-view geometric constraints are used to avoid dependence on RTK. Based on a single image, the light direction of the target pixel in the camera coordinate system is calculated and intersected with the 3D model: in The ray equation is generated for each image as the intersection distance between the ray and the model: Minimize the sum of the squares of the distances from all rays to the target point: in, is the light parameter and needs to be Joint optimization; conversion to a system of linear equations , and solve; where: 。 8. The method for detecting and locating grassland fire points based on drones according to claim 1, characterized in that: The network structure of the YOLOv8s model for grassland fire detection and positioning in the Step 2 includes four parts: input, backbone, neck and prediction. The backbone network is based on CSPDarknet53, combined with CSP (Cross Stage Partial) structure, and the cross-stage connection of some feature maps is realized through the C2f (Cross Stage Partial Fusion with 2 convolutions) module; the neck is based on PANet combined with PAN (Path Aggregation Network) and FPN (feature pyramid networks), and feature maps of different scales are fused through bottom-up and top-down paths; the final prediction result is generated through the convolutional layer, including category probability, confidence and bounding box coordinates, and CIoU_Loss is used as the loss function of the bounding box to optimize the bounding box regression, BCE-loss calculates the classification loss and confidence loss, and finally non-maximum suppression NMS is used to screen multiple targets and predict the output.

9. The method for detecting and locating grassland fire points based on drones according to claim 8, characterized in that: The C2f module in the backbone network uses standard convolution to generate a set of "intrinsic feature maps" through the lightweight network design method GhostNet; then applies linear operations to these intrinsic feature maps to generate "ghost feature maps"; the intrinsic feature maps and ghost feature maps are spliced together to form the final output feature map; The bidirectional feature pyramid network BiFPN is used in Neck. Based on PANet, the bidirectional feature pyramid network BiFPN uses bidirectional connections and weighted feature fusion mechanisms to enable bidirectional transmission of features at each scale, dynamically adjust the importance of different features, and better integrate feature information at different scales.

10. The method for detecting and locating grassland fire points based on drones according to claim 8, characterized in that: The grassland fire point detection and positioning YOLOv8s model introduces a channel attention mechanism SE module; and a mixed non-maximum suppression Mixed-NMS method.

Images