Visible light forest fire identification and positioning method of low-precision holder

Abstract

The invention relates to a forest fire recognition and positioning method, in particular to a visible light forest fire recognition and positioning method of a low-precision holder, which comprises the following steps: data acquisition and preprocessing: constructing a forest fire data set; the method comprises the steps of network model construction, model training and optimization, and model verification and application. On the basis of a deep learning theory, a fire point target is extracted and discriminated by combining strong feature extraction and nonlinear fitting capability of a neural network, and then an ignition point is identified. A flame target is highlighted in combination with an SAM space attention mechanism and a focus loss function, environmental interference is reduced, and positive and negative samples are balanced. In order to solve the problem of positioning errors caused by inaccurate horizontal correction and direction correction and low precision of a holder during holder installation in a fire point positioning task, the yaw and pitch values of the holder are subjected to position compensation so as to realize accurate positioning of a forest fire.

Description

Technical Field

[0001] This invention relates to a method for identifying and locating wildfires, and more particularly to a method for identifying and locating wildfires using visible light with a low-precision gimbal. Background Technology

[0002] With the intensification of global climate change, the frequency and severity of forest fires are on the rise. Wildfires not only destroy forest resources but also threaten human lives and property, causing serious impacts on the ecological environment. Accurate identification and location of wildfires can not only provide early warnings but also assist rescue forces in quickly locating the fire source and developing scientific firefighting plans after a fire breaks out, thereby minimizing the losses caused by the fire.

[0003] Existing methods for identifying and locating wildfires have some drawbacks: 1. The positioning accuracy is greatly affected by the horizontal and vertical correction of the pan-tilt unit and the accuracy of the pan-tilt unit itself, and it is difficult to locate fire points at a distance [1]. 2. A large number of sensors need to be deployed, which is costly and complex to maintain. The accuracy of the measurement will be reduced by environmental factors such as dust and water vapor [2]. 3. Using manual video monitoring can easily cause eye fatigue, and the fire in the video is not easy to detect, which can easily lead to missed reports [3]. 4. It is greatly affected by the weather and has a high cost [4].

[0004] Peng Daqian, Huang Shengying, Du Haitao, et al. A method, system and storage medium for locating forest fires [P]. Guangdong Province: CN202410024120.9, 2024-05-24.

[0005] [Cai Yu, Zuo Chunting, Wang Jinlong, et al. A method for fire point location correction based on laser ranging and GIS map [P]. Beijing: CN202111506418.6, 2022-04-01.]

[0006] Remote video surveillance positioning geographic information system software. Hubei Province, Wuhan University, 2011-01-01.

[0007] Cai Sijie. A method, device and system for fire point localization based on UAV [P]. Zhejiang Province: CN201910435590.3, 2020-11-24. Summary of the Invention

[0008] This disclosure provides a low-precision gimbal-based method for visible light wildfire identification and localization. Addressing the limitations of traditional wildfire prevention methods in responding to weather and environmental changes, as well as the issues of high manpower requirements and difficulty in locating fire sources, this method utilizes deep learning theory, combining the powerful feature extraction and nonlinear fitting capabilities of neural networks to extract, discriminate, and identify fire points. Furthermore, considering the characteristics of early-stage forest fires—small flame targets, complex backgrounds, and indistinct features—the method incorporates the SAM spatial attention mechanism and a focus loss function to highlight flame targets, reduce environmental interference, and balance positive and negative samples. To address the positioning errors caused by inaccurate horizontal and directional corrections during gimbal installation and low gimbal precision, position compensation is applied to the gimbal's yaw and pitch values ​​to achieve accurate wildfire location.

[0009] A method for identifying wildfires using visible light with a low-precision gimbal, comprising the following steps:

[0010] S1. Data Acquisition and Preprocessing: Collect forest fire videos and perform frame extraction to obtain wildfire image data. Use professional annotation tools to label flames and smoke targets in the images to construct a wildfire dataset.

[0011] S2. Dataset Partitioning: The collected dataset is divided into training set, validation set, and test set. The training set is used to build the model, the validation set is used for model parameter tuning and optimization, and the test set is used to evaluate the final performance of the model. The ratio of training set to test set to validation set is usually 8:1:1.

[0012] S3. Neural Network Model Construction: Constructing a wildfire identification model based on convolutional neural networks;

[0013] S4. Model Training and Optimization: Train the model using the training set and adjust the model's hyperparameters using the validation set to avoid overfitting and underfitting; use optimization algorithms to iteratively adjust the model's weight parameters to minimize the loss value.

[0014] S5. Model Validation and Application: In practical applications, cameras are used to monitor forests around the clock, trained neural network models are used to infer and judge video streams, and wildfire target recognition algorithms are used to detect whether there are fire points in the images.

[0015] Furthermore, in S1, mosaic data augmentation is used to expand the dataset, further improving the model's generalization ability and increasing the number of small flame targets in the dataset, thus supporting the early identification of flames.

[0016] Furthermore, in S5, the wildfire target recognition algorithm uses the YOLOv8 algorithm as its model structure. To address the difficulty of detecting early-stage flames, the YOLOv8 algorithm is optimized by integrating the SAM spatial self-attention mechanism and a focus loss function for background suppression and positive / negative sample equalization. To improve the recognition capability of early-stage fire targets, the algorithm is optimized by introducing SAM spatial attention to reduce the influence of background clutter and other interference factors during detection and to more accurately locate the flame target region in the image. For the input image F∈RH×W×C, the spatial attention is calculated as follows:

[0017]

[0018] Where M S (F) represents the spatial attention map, σ represents the Sigmoid function, and f 7×7 This indicates a convolution operation with a 7×7 kernel, AvgPool represents average pooling, and MaxPool represents max pooling.

[0019] The model's input layer receives the image data to be tested, fixing the image size to 640×640, and then extracts features from the input image through a feature extraction network. The SPPF module performs max pooling operations on the input features with four different pooling kernel sizes, and concatenates the pooling results along the channel dimension. After fixing the size of the input features, it is fed into the detection head of the algorithm for prediction, realizing the fusion of multiple receptive fields and making full use of contextual semantic information. The SAM module mainly consists of a concat channel concatenation layer, a max pooling layer, and an average pooling layer, which enhances effective features, suppresses background interference, and focuses on the flame target region in the image. The Detect layer integrates the target features extracted by the model to perform category prediction and bounding box regression, and removes overlapping detection boxes using the NMS non-maximum suppression algorithm.

[0020] By leveraging the Focal Loss function, the model can focus the network's learning on the difficult-to-distinguish early flame targets, enabling the network to update gradients in the correct direction.

[0021] The Focal Loss function is a modification of the cross-entropy loss function, and its formula is shown below:

[0022]

[0023] As shown in the equation, the Focal Loss function adds a parameter factor γ to the cross-entropy formula. When γ = 0, Focal Loss follows the cross-entropy formula. When γ > 0, the loss function places greater emphasis on the loss calculation for difficult-to-separate samples and neglects the loss calculation for easily separable samples. Furthermore, a balancing factor α is added to address the imbalance between the number of positive and negative samples.

[0024] A method for locating wildfires using visible light with a low-precision gimbal, comprising the following steps:

[0025] Step 1: The wildfire target recognition algorithm analyzes the video image and detects the presence of a flame target in the image, and obtains the yaw and pitch values ​​of the gimbal at this time;

[0026] Step 2: Since the spatial position and altitude information of the equipment gimbal are known, the distance and altitude information of the fire point from the equipment gimbal can be calculated based on the yaw and pitch values ​​of the gimbal at this time. Then, by selecting the first collision point with a line of sight through the visibility analysis and combining it with the elevation map, the location of the fire area can be calculated to achieve localization.

[0027] Step 3: The fire point location information obtained at this time has a certain error. This error is related to the accuracy of horizontal and directional correction during gimbal installation and the accuracy of the gimbal itself. In addition, when locating a fire point five kilometers away, slight differences in the yaw and pitch values ​​of the gimbal will cause a certain positioning error. The accuracy problem is solved by matrix correction method.

[0028] Furthermore, the matrix correction method is as follows:

[0029] A. Using the yaw angle of the gimbal as a variable, construct a correction matrix every 5° from 0° to 360°.

[0030] B. Construct the YOLT remote sensing image target recognition algorithm to identify houses and towers in the remote sensing image and record the GPS position of the target; if no target is identified, return to step 1 to construct the next 5° correction matrix for gimbal yaw;

[0031] C. Using the UniDetector algorithm, identify houses and towers within a 5° range at the same angle. Calculate the GPS position of the target by combining visibility analysis with a coarse positioning method based on the elevation map. At this point, the yaw and pitch values ​​of the gimbal are the yaw and pitch values ​​before correction. If no target is identified, return to step 1 to construct the next 5° correction matrix for the gimbal yaw.

[0032] D. The GPS location information returned in the remote sensing image is accurate GPS information, while the GPS location information predicted by the system may have some errors due to various factors. At this time, adjust the yaw and pitch values ​​of the gimbal so that the predicted GPS location information is consistent with the GPS location information returned in the remote sensing image, and record the yaw and pitch values ​​of the gimbal at this time as the true yaw and pitch values ​​of the gimbal;

[0033] E. When four or more building or tower targets are identified within a 5° yaw range of the gimbal, the homography matrix can be used to find the correspondence between the uncorrected yaw and pitch values ​​and the actual gimbal yaw and pitch values, and to construct a transformation matrix; the formula is as follows:

[0034]

[0035] Where Y′ is the gimbal yaw value before correction, P′ is the gimbal pitch value before correction, Y is the gimbal yaw value after correction, P is the gimbal pitch value after correction, and H is the gimbal pitch value after correction. 3×3 This is the transformation matrix; if no four houses or towers are identified within a 5° range, return to step 1 to construct the next 5° correction matrix for gimbal yaw.

[0036] F. Repeat the above steps to construct the correction matrix for the gimbal yaw range from 0° to 360° at 5° intervals. After the position correction matrix is ​​constructed, the uncorrected gimbal yaw and pitch can be corrected in actual applications. Then, the corrected precise positioning information can be obtained by combining the visibility analysis with the elevation map.

[0037] G. Verify the effectiveness of the proposed position correction method: Conduct fire point location tests at distances of 1, 3, 5, 7, 10, 15, and 18 kilometers from the device.

[0038] To address the limitations of traditional wildfire prevention methods in responding to weather and environmental changes, as well as the problems of high manpower requirements and difficulty in locating fire sources, this study uses deep learning theory as a foundation and combines the powerful feature extraction and nonlinear fitting capabilities of neural networks to extract, judge, and identify fire points.

[0039] In addition, considering the characteristics of early-stage forest fires, such as small flame targets, complex backgrounds, and indistinct features, the study combines the SAM spatial attention mechanism and the focus loss function to highlight flame targets, reduce environmental interference, and balance positive and negative samples.

[0040] To address the positioning errors caused by inaccurate horizontal and directional corrections and low gimbal accuracy during gimbal installation in fire location missions, position compensation is performed on the yaw and pitch values ​​of the gimbal to achieve accurate wildfire location. Attached Figure Description

[0041] Figure 1 Wildfire target identification process;

[0042] Figure 2 The structure of the wildfire identification algorithm model;

[0043] Figure 3 The structure of the SAM spatial attention mechanism;

[0044] Figure 4 Results of wildfire target identification;

[0045] Figure 5 This is a schematic diagram showing the distance relationship between the gimbal and the firing point;

[0046] Figure 6 A flowchart of a coarse positioning method combining visibility analysis and elevation maps;

[0047] Figure 7 Flowchart for constructing the position correction matrix;

[0048] Figure 8 This is a schematic diagram of the actual location.

[0049] Figure 9 This table compares the errors before and after correction at different distances. Detailed Implementation

[0050] This method combines deep learning technology and requires building a neural network model for wildfire target detection. The model construction steps are as follows:

[0051] 1. Data Acquisition and Preprocessing: Forest fire videos were collected and frame-by-frame extraction was performed to obtain wildfire image data. Flames and smoke targets in the images were labeled using professional annotation tools to construct a wildfire dataset. Furthermore, to further improve the model's generalization ability, mosaic data augmentation was used to expand the dataset, increasing the number of small flame targets and supporting early flame identification.

[0052] 2. Dataset Partitioning: The collected dataset is divided into training, validation, and test sets. The training set is used to build the model, the validation set is used for model parameter tuning and optimization, and the test set is used to evaluate the final performance of the model. Typically, the ratio of training set to test set to validation set is 8:1:1.

[0053] 3. Construction of Neural Network Model: Since the local connectivity, weight sharing, and invariance of translation and spatial transformation of convolutional neural networks are very suitable for processing image data, a wildfire identification model is constructed based on convolutional neural networks.

[0054] 4. Model Training and Optimization: Train the model using the training set and adjust its hyperparameters using the validation set to avoid overfitting and underfitting. Utilize optimization algorithms, such as stochastic gradient descent (SGD, Adam), to iteratively adjust the model's weight parameters and minimize the loss value.

[0055] 5. Model Validation and Application: In practical applications, cameras are used to monitor the forest around the clock. A trained neural network model is used to infer and judge the video stream to detect whether there are fire points in the images. The detection process of the wildfire target recognition algorithm is as follows: Figure 1 As shown.

[0056] The model structure uses the YOLOv8 algorithm, and optimizes it to address the difficulty of detecting early-stage flames. It integrates the SAM spatial self-attention mechanism and a focus loss function for background suppression and positive / negative sample equalization. The model structure is shown in Figure 2. To improve the identification capability of early-stage fire targets, the algorithm is optimized by introducing SAM spatial attention to reduce the influence of background clutter and other interference factors during detection and to more accurately locate the flame target region in the image. For the input image F∈RH×W×C, the spatial attention is calculated as follows:

[0057] Where M S (F) represents the spatial attention map, σ represents the Sigmoid function, and f 7×7 This indicates a convolution operation with a 7×7 kernel. `AvgPool` represents average pooling, and `MaxPool` represents max pooling. The SAM spatial attention structure diagram is shown below. Figure 3 As shown.

[0058] The model's input layer receives the image data to be tested, fixing the image size to 640×640, and then extracts features from the input image through a feature extraction network. The SPPF module performs max pooling operations on the input features with four different pooling kernel sizes, and concatenates the pooling results along the channel dimension. After fixing the input feature size, it is fed into the detection head of the algorithm for prediction, realizing multi-receptive field fusion and making full use of contextual semantic information. The SAM module mainly consists of a concat channel concatenation layer, a max pooling layer, and an average pooling layer, which enhances effective features, suppresses background interference, and focuses on the flame target region in the image. The Detect layer integrates the target features extracted by the model to perform class prediction and bounding box regression, and removes overlapping detection boxes using the NMS non-maximum suppression algorithm.

[0059] The imbalance between positive and negative samples significantly hinders the optimization process of the model. Since most of the samples are easily distinguishable negative samples (background), the model cannot fully learn the information of the test target that belongs to the positive samples during training. On the other hand, a large number of easily distinguishable background negative samples will contribute significantly to the network's loss, thus dominating the gradient update direction and obscuring important positive sample information.

[0060] The Focal Loss function prioritizes difficult-to-distinguish positive samples, assigning a larger loss to these and reducing the weight of easily distinguishable negative background samples. For early-stage forest fire target detection, the imbalance between positive and negative samples is even more severe. Therefore, combining the Focal Loss function with the algorithm is an effective strategy to improve the performance of early-stage forest fire target detection.

[0061] By utilizing the Focal Loss function, the model can focus its learning on the difficult-to-distinguish early flame targets, guiding the network to update gradients in the correct direction. The Focal Loss function is a modification of the cross-entropy loss function, and its formula is shown below:

[0062]

[0063] As shown in the equation, the Focal Loss function adds a parameter factor γ to the cross-entropy formula. When γ = 0, Focal Loss follows the cross-entropy formula. When γ > 0, the loss function places greater emphasis on the loss calculation for difficult-to-classify samples and neglects the loss calculation for easily classified samples. Furthermore, a balance factor α is added to address the imbalance between the number of positive and negative samples. The results of wildfire target identification are shown below. Figure 4 As shown.

[0064] When a fire is detected, the latitude and longitude of the fire point are located by combining the yaw and pitch values ​​of the pan-tilt unit with visibility analysis and elevation maps. Furthermore, addressing issues such as equipment installation errors, low accuracy of the pan-tilt unit itself, and significant location errors for fires more than 5 kilometers away, a position matrix correction method is proposed to correct the pan-tilt unit's position, thereby achieving precise location of the wildfire through visibility analysis and elevation maps.

[0065] The steps are as follows:

[0066] 1. The wildfire target recognition algorithm analyzes video images and detects the presence of flame targets in the image, then obtains the gimbal yaw and pitch values ​​at that time.

[0067] 2. Given the known spatial location and altitude of the equipment's gimbal, the distance and altitude of the fire point from the gimbal can be calculated based on the current yaw and pitch values. Then, by using visibility analysis to select the first visible collision point and combining it with the elevation map, the location of the fire area can be determined, thus achieving localization. The relationship between the gimbal and the fire point is as follows: Figure 5 As shown, the flowchart of the coarse positioning method combining visibility analysis and elevation map is as follows: Figure 6 As shown.

[0068] 3. The fire point location information obtained at this time has a certain degree of error. This error is related to the accuracy of the horizontal and directional corrections during gimbal installation, as well as the accuracy of the gimbal itself. In addition, when locating a fire point five kilometers away, even slight differences in the yaw and pitch values ​​of the gimbal will cause a certain positioning error.

[0069] To address the issue of accurate fire point positioning, a position matrix correction method is proposed. This method can correct the yaw and pitch values ​​of the gimbal. Then, the corrected yaw and pitch values ​​of the gimbal are used as input, and the first line-of-sight collision point is found through line-of-sight analysis. Combined with the elevation map, the precise fire point position information is obtained.

[0070] The process of constructing the position correction matrix is ​​as follows: Figure 7 As shown, the specific process is as follows:

[0071] 1. Using the yaw angle of the gimbal as a variable, construct a correction matrix every 5° interval from 0° to 360°.

[0072] 2. Construct the YOLT remote sensing image target recognition algorithm to identify houses and towers in the remote sensing image and record their GPS positions. If no target is identified, return to step 1 to construct the next 5° correction matrix for the gimbal yaw.

[0073] 3. Using the UniDetector algorithm, identify buildings and towers within a 5° range at the same angle. Calculate the target's GPS position using a coarse positioning method combining line-of-sight analysis and elevation maps. At this point, the gimbal's yaw and pitch values ​​are the uncorrected values. If no target is identified, return to step 1 to construct the next 5° yaw correction matrix for the gimbal.

[0074] 4. The target GPS location information returned in the remote sensing image is accurate GPS information, while the GPS location information predicted by the system may have some errors due to various factors. At this time, adjust the yaw and pitch values ​​of the gimbal so that the predicted GPS location information is consistent with the GPS location information returned in the remote sensing image, and record the gimbal yaw and pitch values ​​at this time as the true gimbal yaw and pitch values.

[0075] 5. When the gimbal identifies four or more building or tower targets within a 5° yaw range, a homography matrix can be used to find the correspondence between the uncorrected yaw and pitch values ​​and the actual gimbal yaw and pitch values, constructing a transformation matrix. The formula is as follows:

[0076]

[0077] Where Y′ is the gimbal yaw value before correction, P′ is the gimbal pitch value before correction, Y is the gimbal yaw value after correction, P is the gimbal pitch value after correction, and H is the gimbal pitch value after correction.3×3 This is the transformation matrix. If no four houses or towers are identified within a 5° range, return to step 1 to construct the next 5° correction matrix for gimbal yaw.

[0078] 6. Repeat the above steps to construct the correction matrix for the gimbal yaw range from 0° to 360° at 5° intervals.

[0079] After constructing the position correction matrix, uncorrected gimbal yaw and pitch can be corrected in practical applications. Then, accurate positioning information is obtained by combining line-of-sight analysis with an elevation map. Actual positioning results are as follows: Figure 8 As shown.

[0080] 7. To verify the effectiveness of the proposed position correction method, fire point location tests were conducted at distances of 1, 3, 5, 7, 10, 15, and 18 kilometers from the equipment. Figure 9 The data records a comparison between the error without using the position matrix correction method and the error after correction. Figure 9 It can be seen that the positioning error without using the position matrix correction method is relatively large, and the positioning error is extremely large when the distance between the fire point and the equipment exceeds 5 kilometers. The error after correction is effectively reduced compared to before correction. In addition, accurate positioning can still be achieved when facing a fire point more than 5 kilometers away.

[0081] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for identifying wildfires using visible light with a low-precision gimbal, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing: Collect forest fire videos and perform frame extraction to obtain wildfire image data; use professional annotation tools to annotate flames and smoke targets in the images to construct a wildfire dataset. S2. Dataset partitioning: Divide the collected dataset into training set, validation set and test set; The training set is used to build the model, the validation set is used for model parameter tuning and optimization, and the test set is used to evaluate the final performance of the model. Typically, the ratio of the training set to the test set and validation set is 8:1:

1. S3. Neural Network Model Construction: Constructing a wildfire identification model based on convolutional neural networks; S4. Model Training and Optimization: Train the model using the training set and adjust the model's hyperparameters using the validation set to avoid overfitting and underfitting; use optimization algorithms to iteratively adjust the model's weight parameters to minimize the loss value. S5. Model Validation and Application: In practical applications, cameras are used to monitor forests around the clock, trained neural network models are used to infer and judge video streams, and wildfire target recognition algorithms are used to detect whether there are fire points in the images.

2. The visible light wildfire identification method using a low-precision gimbal according to claim 1, characterized in that, In step S1, mosaic data augmentation is used to expand the dataset, further improving the generalization of the model and increasing the number of small flame targets in the dataset, thus providing support for the early identification of flames.

3. The visible light wildfire identification method using a low-precision gimbal according to claim 1, characterized in that, In S5, the wildfire target recognition algorithm uses the YOLOv8 algorithm as its model structure. To address the difficulty of detecting early-stage flames, the YOLOv8 algorithm is optimized by integrating the SAM spatial self-attention mechanism and a focus loss function for background suppression and positive / negative sample equalization. To improve the recognition capability of early-stage fire targets, the algorithm is optimized by introducing SAM spatial attention to reduce the influence of background clutter and other interference factors during detection and to more accurately locate the flame target region in the image. For the input image F∈RH×W×C, the spatial attention is calculated as follows: Where M S (F) represents the spatial attention map, σ represents the Sigmoid function, and f 7×7 This indicates a convolution operation with a 7×7 kernel, AvgPool represents average pooling, and MaxPool represents max pooling. The model's input layer receives the image data to be tested, fixes the image size to 640×640, and then extracts features from the input image through a feature extraction network. SP The PF module performs max pooling on the input features using four different pooling kernel sizes, and concatenates the pooling results along the channel dimension. The input features are then fed into the detection head of the algorithm for prediction, achieving multi-receptive field fusion and fully utilizing contextual semantic information. The SAM module mainly consists of a concat channel concatenation layer, a max pooling layer, and an average pooling layer, enhancing effective features, suppressing background interference, and focusing on the flame target region in the image. The Detect layer integrates the target features extracted by the model for category prediction and bounding box regression, and removes overlapping detection boxes using the NMS non-maximum suppression algorithm. By utilizing the FocalLoss function, the model can shift the focus of network learning to the difficult-to-distinguish early flame targets, enabling the network to update gradients in the correct direction. The Focal Loss function is a modification of the cross-entropy loss function, and its formula is shown below: As shown in the formula, the Focal Loss function adds a parameter factor γ to the cross-entropy formula. When γ = 0, Focal Loss is the cross-entropy formula. When γ > 0, the loss function will pay more attention to the loss calculation of hard-to-separate samples and ignore the loss calculation of easy-to-separate samples. In addition, a balance factor α is added to adjust the imbalance of positive and negative sample numbers.

4. A method for locating wildfires using visible light with a low-precision gimbal, characterized in that, Includes the following steps: Step 1: The wildfire target recognition algorithm analyzes the video image and detects the presence of a flame target in the image, and obtains the yaw and pitch values ​​of the gimbal at this time; Step 2: Since the spatial position and altitude information of the equipment gimbal are known, the distance and altitude information of the fire point from the equipment gimbal can be calculated based on the yaw and pitch values ​​of the gimbal at this time. Then, by selecting the first collision point with a line of sight through the visibility analysis and combining it with the elevation map, the location of the fire area can be calculated to achieve localization. Step 3: The fire point location information obtained at this time has a certain error. This error is related to the accuracy of horizontal and directional correction during gimbal installation and the accuracy of the gimbal itself. In addition, when locating a fire point five kilometers away, slight differences in the yaw and pitch values ​​of the gimbal will cause a certain positioning error. The accuracy problem is solved by matrix correction method.

5. The visible light wildfire positioning method using a low-precision gimbal according to claim 4, characterized in that, The matrix correction method is as follows: A. Using the yaw angle of the gimbal as a variable, construct a correction matrix every 5° from 0° to 360°. B. Construct the YOLT remote sensing image target recognition algorithm to identify houses and towers in the remote sensing image and record the GPS position of the target; if no target is identified, return to step 1 to construct the next 5° correction matrix for gimbal yaw; C. Using the UniDetector algorithm, identify houses and towers within a 5° range at the same angle. Calculate the GPS position of the target by combining visibility analysis with a coarse positioning method based on the elevation map. At this point, the yaw and pitch values ​​of the gimbal are the yaw and pitch values ​​before correction. If no target is identified, return to step 1 to construct the next 5° correction matrix for the gimbal yaw. D. The GPS location information of the target returned in the remote sensing image is accurate GPS information. The GPS location information predicted by the system will have certain errors due to various factors. At this time, adjust the yaw and pitch values ​​of the gimbal so that the predicted GPS location information is consistent with the GPS location information returned in the remote sensing image. Record the yaw and pitch values ​​of the gimbal at this time as the true yaw and pitch values ​​of the gimbal. E. When four or more building or tower targets are identified within a 5° yaw range of the gimbal, the homography matrix can be used to find the correspondence between the uncorrected yaw and pitch values ​​and the actual gimbal yaw and pitch values, and to construct a transformation matrix; the formula is as follows: Where Y′ is the gimbal yaw value before correction, P′ is the gimbal pitch value before correction, Y is the gimbal yaw value after correction, P is the gimbal pitch value after correction, and H is the gimbal pitch value after correction. 3×3 This is the transformation matrix; if no four houses or towers are identified within a 5° range, return to step 1 to construct the next 5° correction matrix for gimbal yaw. F. Repeat the above steps to construct the correction matrix for the gimbal yaw range from 0° to 360° at 5° intervals. After the position correction matrix is ​​constructed, the uncorrected gimbal yaw and pitch can be corrected in actual applications. Then, the corrected precise positioning information can be obtained by combining the visibility analysis with the elevation map. G. Verify the effectiveness of the proposed position correction method: Conduct fire point location tests at distances of 1, 3, 5, 7, 10, 15, and 18 kilometers from the device.

Images