Intelligent forest fire detection algorithm based on infrared vision of unmanned aerial vehicle

Abstract

According to the forest fire intelligent detection algorithm based on the unmanned aerial vehicle infrared vision, the unmanned aerial vehicle technology and infrared thermal imaging are combined, and high-precision real-time identification of a fire source in a complex forest environment is achieved. Comprising the following steps: S100, acquiring forest infrared image data in real time by using an infrared thermal imaging camera carried by an unmanned aerial vehicle; s200, carrying out gray scale correction, noise suppression and enhancement processing on the image; s300, generating six groups of anchor frame parameters according to forest fire point infrared image size distribution clustering; s400, inputting the preprocessed image into a YOLOv7 model, and obtaining position coordinates and area information of a fire point through feature extraction, multi-scale fusion and prediction output; and S500, fusing the fire point information with the GPS positioning data of the unmanned aerial vehicle, generating a monitoring report containing the position of the fire point and the preliminary evaluation of the fire behavior, pushing the monitoring report to a ground monitoring center, and triggering an alarm. The system effectively solves the problems that traditional monitoring is lagged in response, limited in coverage, prone to interference and the like, and facilitates early-stage accurate detection and rapid decision making of forest fire.

Description

Technical Field

[0001] This invention relates to the field of forest fire monitoring technology, specifically to a forest fire intelligent detection algorithm based on UAV infrared vision, which is suitable for early warning and real-time monitoring of fires in large forest areas. Background Technology

[0002] Forests are vital ecological resources, and forest fire prevention is of paramount importance. Traditional forest fire monitoring methods mainly include manual patrols, lookout tower observations, and satellite remote sensing, but these methods have significant drawbacks: manual patrols have limited coverage, are labor-intensive, and are difficult to conduct in complex terrain or in adverse weather conditions; lookout tower observations are affected by terrain obstructions, resulting in numerous blind spots; satellite remote sensing suffers from low temporal resolution and slow response, making it difficult to detect early small fires, and is easily affected by weather conditions such as clouds. With the development of drone technology, the application of drones equipped with visible light cameras for forest fire monitoring is gradually increasing, but the detection effectiveness of visible light monitoring drops significantly at night, in smoky weather, or when fire points are obscured by vegetation. Infrared thermal imaging technology can capture the thermal radiation signals of objects, is unaffected by lighting conditions and smoke, and can effectively identify early hidden fire points. Therefore, monitoring methods combining drones with infrared vision have become a research hotspot.

[0003] Currently, fire detection based on UAV infrared vision mostly employs traditional image processing algorithms (such as threshold segmentation and edge detection) or conventional target detection algorithms (such as YOLOv3 and Faster R-CNN). Traditional image processing algorithms have poor environmental adaptability and are easily interfered with by complex backgrounds (such as rocks and high-temperature objects), resulting in low detection accuracy. Conventional target detection algorithms, when processing infrared images, suffer from problems such as missed detection of small-sized fire points and insufficient real-time detection due to the lack of obvious fire point features and large size differences. YOLOv7, as a current advanced target detection algorithm, has the advantages of fast detection speed and high accuracy, but when directly applied to infrared detection of forest fires, it still suffers from problems such as low anchor frame matching degree and insufficient targeted feature extraction, making it difficult to meet the needs of accurate early detection of forest fires.

[0004] Therefore, there is an urgent need for an optimized intelligent detection algorithm that combines UAV infrared vision with the YOLOv7 algorithm to improve the accuracy, real-time performance, and anti-interference capabilities of forest fire detection, and to address the shortcomings of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent forest fire detection algorithm based on UAV infrared vision, to solve the problems of slow response, low detection accuracy, and weak anti-interference ability in existing forest fire monitoring, and to achieve early and accurate detection and real-time warning of forest fires. The method includes the following steps:

[0006] S100: Uses an infrared thermal imaging camera mounted on a drone to collect real-time infrared image data of forests;

[0007] S200: Performs grayscale correction, noise suppression, and enhancement processing on the image;

[0008] S300: Generates 6 sets of anchor frame parameters by clustering based on the size distribution of infrared images of forest fire points;

[0009] S400: Input the preprocessed image into the YOLOv7 model, and obtain the coordinates and area information of the fire point location through feature extraction, multi-scale fusion and prediction output;

[0010] S500: It integrates fire information with UAV GPS positioning data to generate a monitoring report containing the fire location and preliminary fire assessment, which is then pushed to the ground monitoring center and triggers an alarm.

[0011] In step S100 of this invention, the selected infrared thermal imaging camera is equipped with a high-sensitivity infrared detector with a detection wavelength range of 8-14μm. It can capture temperature differences exceeding 0.1℃ in forest environments, clearly identifying the temperature boundary between the fire point and surrounding vegetation even in low-visibility conditions such as dense smoke, fog, or nighttime, avoiding the omission of fire points due to environmental obstruction by traditional visible light cameras. The drone adopts a multi-rotor or fixed-wing hybrid configuration and is equipped with a high-precision gimbal stabilization system, achieving an attitude control accuracy of ±0.01, ensuring the infrared camera maintains a stable shooting angle during flight. The resolution of the acquired images is no less than 640×512 pixels, and the frame rate is maintained at 25 frames per second, meeting both real-time requirements and ensuring clear presentation of fire point details in the images. The acquired infrared image data is temporarily stored through the drone's built-in local cache module, while a real-time transmission link is activated to prevent data loss due to sudden disconnections, providing a continuous and complete data source for subsequent processing.

[0012] In step S200 of this invention, the acquired infrared image undergoes grayscale correction, noise suppression, and image enhancement processing to eliminate environmental interference and equipment errors, and improve the distinction between the fire point area and the background. Grayscale correction determines the minimum and maximum grayscale values ​​of the image by statistically analyzing the grayscale histogram distribution of a single frame of infrared image, and dynamically adjusts the grayscale mapping interval to disperse the concentrated grayscale values ​​in the original image to the full grayscale range of 0-255. Noise suppression employs a median filtering algorithm to filter out salt-and-pepper noise and Gaussian noise generated by detector thermal noise and electromagnetic interference in the infrared image. Image enhancement uses a multi-scale Retinex algorithm to decompose the infrared image into a reflection component and an illumination component. Addressing the uneven illumination problem caused by tree canopy occlusion and terrain undulations in forests, the algorithm smooths the illumination component using Gaussian filtering to eliminate local differences in brightness; simultaneously, it adaptively adjusts the gain of the reflection component, dynamically increasing the gain coefficient based on the grayscale value of the fire point area to enhance the contour details of the fire point, while maintaining a low gain for the background area to avoid excessive amplification of background noise.

[0013] In step S300 of this invention, based on the size distribution of forest fire infrared images, six sets of anchor box parameters are clustered to optimize the adaptability of the YOLOv7 model to forest fire detection. First, a sample library containing more than 10,000 forest fire infrared images is constructed, covering different fire scales (small fire, medium fire, large fire), different forest types (coniferous forest, broadleaf forest, mixed forest), and different environmental conditions (daytime, nighttime, smoky weather). The bounding box information (x, y coordinates, width, and height) of the fire target in each image is accurately annotated using a manual annotation tool. Subsequently, the K-means clustering algorithm was used to perform cluster analysis on the aspect ratio of all bounding boxes. The clustering results were iteratively optimized with the objective function of maximizing the average intersection-union ratio (IoU) between the cluster center and the sample bounding box. Finally, 6 sets of anchor box parameters adapted to the size distribution of forest fire spots were generated, with aspect ratios ranging from 1:0.8 to 1:1.2 and sizes ranging from 16×16 to 128×128 pixels, corresponding to fire spot targets of different sizes such as "small sparks, small flames, medium flames, and large flames".

[0014] In step S400 of this invention, the preprocessed infrared image is input into the YOLOv7 model. Through feature extraction, multi-scale fusion, and prediction output, the location coordinates, confidence level, and area information of the fire point are obtained. Prediction heads are set up on the fused three-scale feature maps to detect fire points. Each prediction head includes a classification branch. The model uses the CIoU (Complete IoU) loss function to calculate the difference between the predicted bounding box and the ground truth bounding box, and continuously optimizes the parameters through backpropagation to gradually improve the prediction accuracy.

[0015] In step S500 of this invention, the high-precision GPS module (positioning accuracy ≤ 1 meter) carried by the UAV outputs the latitude and longitude coordinates (Lng, Lat) of the flight position in real time. Using the UAV's gimbal attitude data (heading angle, pitch angle, roll angle) and the intrinsic parameters of the infrared camera (focal length, principal point coordinates), a mapping relationship between the image pixel coordinate system and the geographic coordinate system is established: based on the pixel coordinates (x1, y1) and (x2, y2) of the fire point in the image, the pixel coordinates ((x1+x2) / 2, (y1+y2) / 2) of the center point of the fire point's bounding box are calculated, and then converted into the corresponding geographic latitude and longitude coordinates using a coordinate transformation formula, thus achieving the location positioning from the "image fire point" to the "real-world fire point". Attached Figure Description

[0016] Figure 1 This is an overall flowchart of a forest fire intelligent detection algorithm based on UAV infrared vision provided in one embodiment of the present disclosure;

[0017] Figure 2 This is a flowchart of grayscale correction provided in one embodiment of this disclosure;

[0018] Figure 3 This is a flowchart of median filtering for noise reduction provided in one embodiment of this disclosure.

[0019] Figure 4 This is a flowchart of a multi-scale Retinex image enhancement algorithm provided in one embodiment of this disclosure.

[0020] Figure 5 This is a flowchart of the K-means clustering algorithm provided in one embodiment of this disclosure.

[0021] Figure 6 This is a YOLOv7 model network structure diagram provided in one embodiment of this disclosure. Detailed Implementation

[0022] Forest fires are characterized by their suddenness, destructiveness, and difficulty in extinguishing. They not only burn vast amounts of timber resources and damage the ecological environment but also threaten the lives and property of nearby residents. Traditional forest fire detection methods, such as manual patrols and observation towers, suffer from drawbacks such as low efficiency, limited coverage, and susceptibility to terrain and weather conditions. Utilizing drones equipped with infrared vision devices for intelligent forest fire detection aims to overcome the limitations of traditional methods, achieving rapid, comprehensive, and accurate monitoring of forest areas. This will enable timely detection of fire hazards and early-stage fires, providing strong support for early fire suppression and reducing forest fire losses. Figure 1 The diagram shows the overall framework of the model. The method includes the following steps:

[0023] S100: Uses an infrared thermal imaging camera mounted on a drone to collect real-time infrared image data of forests;

[0024] S200: Performs grayscale correction, noise suppression, and enhancement processing on the image;

[0025] S300: Generates 6 sets of anchor frame parameters by clustering based on the size distribution of infrared images of forest fire points;

[0026] S400: Input the preprocessed image into the YOLOv7 model, and obtain the coordinates and area information of the fire point location through feature extraction, multi-scale fusion and prediction output;

[0027] S500: It integrates fire information with UAV GPS positioning data to generate a monitoring report containing the fire location and preliminary fire assessment, which is then pushed to the ground monitoring center and triggers an alarm.

[0028] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0029] Reference Figure 1 The drone in question is a DJI M300 RTK multi-rotor drone equipped with a stabilization gimbal and supports RTK positioning to ensure flight stability and positioning accuracy. The infrared thermal imaging camera is a FLIR Vue TZ2 dual-band infrared camera, supporting 8-14μm long-wave infrared detection, with a sensitivity ≤50mK (0.05℃ temperature difference resolution), a frame rate of 25Hz, and a resolution of 640×512 pixels. It also features automatic non-uniformity correction (NUC) to reduce the impact of detector temperature drift.

[0030] Reference Figure 2 The aforementioned grayscale correction addresses image grayscale distortion caused by inconsistent detector response and optical system attenuation in infrared thermal imaging cameras. First, a standard grayscale chart is captured using an infrared thermal imaging camera to obtain the actual pixel grayscale outputs corresponding to different grayscale values. Then, a grayscale correction mapping function is established, using the standard grayscale value as a benchmark, to correct the grayscale value of each pixel in the original infrared image, making the image grayscale distribution more closely match the temperature distribution of the actual scene. Specifically, if the ideal image is... The actual image is The non-uniformity is determined by the scaling factor. If this is caused, then the relationship is as shown in equation (1):

[0031] (1)

[0032] Using a uniform grayscale field (such as a constant C) for imaging, the actual output is measured. ,but:

[0033] (2)

[0034] For any image After correction, it becomes:

[0035] (3)

[0036] The formula for correcting nonlinear grayscale distribution (such as temperature drift in infrared images) is:

[0037] (4)

[0038] in, For correction factor (usually) Brighten the shadows, Darken the highlights.

[0039] Reference Figure 3 The median filtering method described above suppresses noise in the corrected infrared image. Median filtering selects the median grayscale value of each pixel's neighborhood as the new grayscale value for that pixel. This effectively suppresses common noises in infrared images, such as salt-and-pepper noise and impulse noise, while also preserving edge information of fire points and avoiding blurring of fire features during filtering. The size of the filtering window can be adjusted according to the image noise intensity; typically, a 3×3 or 5×5 window size is chosen to achieve a balance between noise suppression and image detail preservation.

[0040] Let the pixel grayscale value within the window be... After sorting, it is The median filter output is:

[0041] (5)

[0042] When n is odd, the calculation is performed according to formula (5). When n is odd, the average of the two middle values ​​is taken.

[0043] Reference Figure 4 The multi-scale Retinex algorithm, combining multi-scale Gaussian filtering with weighted fusion techniques, achieves a balance between dynamic range compression, detail enhancement, and color constancy, thereby improving the visual quality of fire images in complex scenes such as low light and smoke obstruction. The basic principle is as follows:

[0044] Retinex's theory posits that the brightness of an image perceived by the human eye is determined by ambient light (incident light component). ) and the object's reflection properties (reflection component) This is jointly determined, and mathematically expressed as:

[0045] (6)

[0046] in, For the original image at pixels The grayscale value at that location; The reflection component represents the inherent properties of the object itself; The illumination component reflects the global illumination information in the scene.

[0047] Multi-scale Retinex introduces a multi-scale Gaussian filtering and weighted fusion mechanism to apply Gaussian filtering at different scales to the RGB channels of the original image, generating multiple blurred versions. The Gaussian filtering formula is as follows:

[0048] (7)

[0049] in, Let be the standard deviation of the i-th scale.

[0050] Logarithmic transformation converts multiplication into addition, facilitating the separation of illumination and reflection components, specifically expressed as:

[0051] (8)

[0052] Where * represents the convolution operation.

[0053] Reflection components at each scale By weight Linear superposition generates the final reflection component. Specifically, it is expressed as:

[0054] (9)

[0055] in, The reflection component is the logarithmic difference between the original image and the illumination component. This represents the illumination component.

[0056] Reference Figure 5 The K-means clustering algorithm described above divides fire point bounding boxes into six clusters based on size features using a distance metric based on the intersection-over-union (IoU). The center size of each cluster is directly used as the anchor box parameter, thereby improving the model's detection accuracy for multi-scale fire points. Its basic principle is as follows:

[0057] Forest fire image dataset, with bounding boxes labeled for all fire points. Each bounding box is composed of Represent (center coordinates, width, height). Randomly select K=6 bounding boxes as initial cluster centers. Each center is .

[0058] Assign bounding boxes, for each bounding box Calculate its relationship with all cluster centers. Based on the distance, assign it to the cluster with the smallest distance. The specific process is as follows:

[0059] For two bounding boxes and IoU is the ratio of the area of ​​their intersection to the area of ​​their union:

[0060] (10)

[0061] in, Represents the area of ​​the union; This represents the area of ​​the intersection.

[0062] Using Euclidean distance (such as the difference between the width and height of the bounding box), distance Defined as:

[0063] (11)

[0064] Update cluster centers for each cluster. Calculate the mean width and height of all bounding boxes and use them as the new cluster centers. :

[0065] (12)

[0066] in, For clusters The number of bounding boxes in the data.

[0067] This continues until the change in cluster centers is less than a threshold or the maximum number of iterations (e.g., 100) is reached. The final result is 6 cluster centers. This refers to the anchor frame parameters adapted to the size distribution of forest fire spots, with each set of parameters... Corresponding to fire targets of different sizes.

[0068] Reference Figure 6 The YOLOv7 model network described above achieves accurate identification and information extraction of fire targets through feature extraction, multi-scale fusion, and prediction output. The YOLOv7 feature extraction module adopts a hierarchical architecture and generates multi-scale feature maps through downsampling and feature enhancement, providing rich semantic and spatial information for subsequent detection.

[0069] In the feature extraction stage, the model employs a hierarchical backbone network, initially capturing features based on the CBS module. Subsequently, the multi-branch parallel structure of the E-ELAN module enhances feature diversity, and spatial pyramid pooling of the SPPCSPC structure expands the receptive field to adapt to dynamic changes in fire point size. This series of operations generates multi-scale feature maps containing rich semantic and spatial information, laying a solid foundation for subsequent detection. Moving to the multi-scale fusion stage, YOLOv7 innovatively integrates the FPN and PAN networks. The FPN transmits high-level semantic information through a top-down path, strengthening the classification ability of low-level features; the PAN aggregates low-level spatial details through a bottom-up path, improving the localization accuracy of high-level features. Their synergistic effect allows the model to simultaneously capture both global semantics and local details of fire points, significantly improving detection performance, especially for small-scale fire points and those in complex backgrounds. In the prediction output stage, the model employs a decoupled head design, separating classification and regression tasks to avoid interference between tasks and improve prediction accuracy. Simultaneously, a dynamic anchor box allocation strategy flexibly matches anchor boxes based on the actual size of the fire points, reducing missed detections. By combining multi-task loss function optimization, the model outputs the fire point category, bounding box coordinates, and confidence score, achieving accurate identification and information extraction. The entire process is efficient and compact, significantly improving the accuracy and robustness of fire point detection while maintaining real-time performance, providing strong technical support for applications such as forest fire early warning.

[0070] Referring to Table 1, which presents parameter information for the DJI M300 RTK multirotor drone provided in one embodiment of this disclosure, the table shows the parameter information of the DJI M300 RTK multirotor drone used. This model adopts a quadcopter design with a maximum takeoff weight of no more than 9 kg, ensuring a balance between flexibility and payload capacity. Its maximum horizontal flight speed reaches 23 m / s, and its wind resistance reaches 12 m / s (Force 5 wind), adapting to complex environmental operations. Equipped with an RTK positioning system, its horizontal accuracy is ±1 cm + 1 ppm, and its vertical accuracy is ±1.5 cm + 1 ppm. A single battery provides approximately 55 minutes of flight time, and it can be equipped with dual gimbals to meet multi-mission requirements.

[0071] Table 1

[0072] Parameter categories Specific parameters model DJI M300 RTK Number of axes Quadcopter (Multirotor) Maximum takeoff weight ≤9kg (including battery and load) Maximum horizontal flight speed 23 m / s (near sea level, windless environment) Maximum speed of ascent 6 m / s Maximum descent speed 7 m / s Hovering accuracy Vertical: ±0.5 m (when visual positioning is working normally); Horizontal: ±1.5 m Flight time A single battery lasts for approximately 55 minutes. Operating ambient temperature -20℃ to 50℃ RTK positioning accuracy Horizontal: ±1 cm + 1 ppm; Vertical: ±1.5 cm + 1 ppm Image transmission system OcuSync 3.0 supports 1080p / 60fps high-definition image transmission with a maximum transmission distance of 15 km. Wind resistance ≤12 m / s (Force 5 wind) Load interface Supports dual gimbal mounting

[0073] Referring to Table 2, which presents parameter information for the FLIR Vue TZ2 dual-band infrared camera provided in one embodiment of this disclosure, the camera integrates visible light and long-wave infrared (7.5–13.5 μm) dual sensors, with an infrared resolution of 640×512 pixels and a NETD sensitivity ≤50 mK, enabling precise capture of weak thermal signals. The visible light module uses a 1 / 2.8-inch CMOS sensor and outputs 1920×1080 Full HD images.

[0074] Table 2

[0075] Parameter categories Specific parameters model FLIR Vue TZ2 Band type Dual-band (visible light + infrared) Infrared resolution 640×512 pixels Infrared spectral range 7.5–13.5 μm (long-wave infrared) Visible light sensor 1 / 2.8-inch CMOS sensor with approximately 2 million effective pixels. Visible light resolution 1920×1080 pixels (Full HD) Field of view (FOV) Infrared: 45° (horizontal) × 37° (vertical); Visible light: 69° (horizontal) × 42° (vertical) Image output Simultaneously outputs infrared and visible light images (which can be overlaid or displayed in split-screen mode). Data Interface 10-pin interface (supports power, video output and serial communication) Operating temperature range -20℃ to 50℃ weight Approximately 375 g (excluding lens cover)

[0076] Referring to Table 3, Table 3 compares the defect detection results of the proposed method with other traditional detection methods in one embodiment of this disclosure. In another embodiment, the proposed method is compared with traditional sensor detection, satellite remote sensing inspection, and fixed camera detection methods. The intelligent forest fire detection algorithm based on UAV infrared vision designed in this invention outperforms traditional methods in terms of detection accuracy. The real-time performance (51 frames / second) and anti-interference capability of UAV infrared vision make it a core technology for early detection and prevention of forest fires, especially suitable for forest areas with complex terrain and poor communication, providing key data support for emergency response.

[0077] Table 3

[0078] Fire detection methods Accuracy (%) Traditional sensor detection 75%-85%. Satellite remote sensing detection 60%-70% Fixed camera visual inspection 80%-88% Infrared vision for drones 93%-98%

[0079] Although the foregoing description, in conjunction with the accompanying drawings, illustrates specific embodiments of the intelligent forest fire detection algorithm based on UAV infrared vision of the present invention, it should be clarified that the scope of protection of the present invention is not limited to the detailed design of the described embodiments. These embodiments are merely illustrative examples of typical application scenarios; their core lies in improving the accuracy of fire detection by combining UAVs with infrared vision and the YOLOv7 detection model.

Claims

1. An unmanned aerial vehicle infrared vision-based forest fire intelligent detection algorithm, which combines infrared thermal imaging technology and YOLOV7 target detection mechanism to realize high-precision and real-time identification of fire sources in complex forest environments. S100: real-time collection of forest infrared image data by an infrared thermal imaging camera carried by a UAV; S2 00: gray scale correction, noise suppression and enhancement processing of the image; S3 00: generation of 6 groups of anchor box parameters according to the clustering of the size distribution of the forest fire point infrared image; S400: input of the preprocessed image into a YOLOv7 model, feature extraction, multi-scale fusion and prediction output, and acquisition of fire point position coordinates and area information; S5 00: fusion of the fire point information and the UAV GPS positioning data to generate a monitoring report containing the fire point position and preliminary assessment of the fire, and push to the ground monitoring center and trigger an alarm. The present application solves the problems of response lag, limited coverage and large environmental interference in traditional forest fire monitoring, realizes early and accurate detection of forest fires, provides fast decision support for forest fire prevention and rescue, and has the advantages of high detection efficiency, strong anti-interference ability and flexible deployment.

2. The forest fire intelligent detection algorithm based on UAV infrared vision of claim 1, wherein, The gray scale correction in step S200 adopts an adaptive gray scale stretching algorithm, dynamically adjusts the gray scale mapping interval by statistically analyzing the gray scale histogram distribution of the infrared image, and improves the gray scale contrast of the fire point area and the background area by more than 30%. 3.The forest fire intelligent detection algorithm based on UAV infrared vision of claim 1, wherein, The noise suppression in step S200 adopts a median filter algorithm, sets a 3*3 filter window, and performs hierarchical suppression on the salt and pepper noise and Gaussian noise in the infrared image, with a noise removal rate of not less than 92% while retaining the edge features of the fire point.

4. The forest fire intelligent detection algorithm based on UAV infrared vision of claim 1, wherein, The image enhancement in step S200 adopts a multi-scale Retinex algorithm, decomposes the reflection component and the illumination component of the infrared image, adjusts the reflection component with adaptive gain, enhances the detail information of the fire point area, and suppresses the uneven illumination interference of the forest background. 5.The forest fire intelligent detection algorithm based on UAV infrared vision of claim 1, wherein, The specific process of the re-clustering to generate 6 groups of anchor box parameters in step S300 is as follows: 10000+ forest infrared sample images containing different scale fire points are collected, the boundary box information of the fire point target is labeled, the K-means clustering algorithm is used for clustering analysis of the width-height ratio of the boundary box, and 6 groups of anchor box parameters adaptive to the size distribution of the forest fire point are generated, the width-height ratio range of the anchor box parameters is 1:0.8~1:1.2, and the size range is 16*16~128*128 pixels. 6.The unmanned aerial vehicle (UAV) infrared vision-based forest fire intelligent detection algorithm according to claim 1, wherein, The feature extraction process of the YOLOv7 model in step S400 adopts CSPDarknet53 as the backbone network, performs multi-scale feature extraction on the infrared image through 5 residual block groups, generates 8 times, 16 times and 32 times down-sampled feature maps, and the feature maps correspond to the detection of small, medium and large fire point targets.

7. The unmanned aerial vehicle infrared vision-based forest fire intelligent detection algorithm according to claim 1, characterized in that, The multi-scale fusion in step S400 adopts a PANet (Path Aggregation Network) structure, performs cross-scale fusion on the high, medium and low resolution feature maps output by the backbone network, realizes the complementation of feature information through up-sampling and convolution operation, improves the detection accuracy of small size fire point targets, and makes the minimum recognition size of the fire point detection reach 16*16 pixels. 8.The unmanned aerial vehicle (UAV) infrared vision-based forest fire intelligent detection algorithm according to claim 1, wherein, The fire preliminary assessment in step S500 is comprehensively judged by the area information of the fire point, the confidence and the gray mean value of the infrared image, and is specifically divided into three levels: small fire (area < 300 pixel², confidence ≥ 0.8, and gray mean value < 200), medium fire (300 pixel² ≤ area < 1000 pixel², confidence ≥ 0.85, and 200 ≤ gray mean value < 230), and large fire (area ≥ 1000 pixel², confidence ≥ 0.9, and gray mean value ≥ 230). 9.The unmanned aerial vehicle (UAV) infrared vision-based forest fire intelligent detection algorithm according to claim 1, wherein, The wireless communication module in step S500 adopts a 5G+LoRa dual-mode communication mode. When the unmanned aerial vehicle is in a forest signal coverage good area, the real-time push of the monitoring report is realized through the 5G network, and the push delay is ≤ 1s. When it is in a weak signal area, it is automatically switched to the LoRa communication mode, thereby guaranteeing the continuity of data transmission, and the communication distance is ≥ 3km. 10.The unmanned aerial vehicle (UAV) infrared vision-based forest fire intelligent detection algorithm according to claim 1, wherein, The alarm mechanism in step S500 includes an audible and light alarm and a short message alarm. After the ground monitoring center receives the monitoring report, the audible and light alarm device is triggered immediately, and a short message alarm information containing the GPS coordinates of the fire point, the fire level, etc. is sent to the preset forest fire prevention person's mobile terminal, and the alarm response time is ≤ 2s.

Images