Dual-light information fusion fire detection method suitable for inspection unmanned aerial vehicle end

Abstract

The invention discloses a dual-light information fusion fire detection method suitable for an inspection unmanned aerial vehicle end, and the method comprises the steps: building an unmanned aerial vehicle inspection fire data set through collecting the existing public fire data information; the fire recognition capability of the model is improved by improving a YOLO11 network model framework; infrared information fire identification information is determined by combining infrared fire energy characteristics and utilizing a self-adaptive threshold value and a minimum connected region mode, and then an energy-space-time multi-source information judgment method is designed; the fire disaster situation is further judged by utilizing the judgment criterion, the fire occurrence position is determined, the fire detection accuracy is improved, and the omission ratio and the false detection ratio are reduced. Compared with the prior art, the method has the characteristics of adaptability to complex environment change, day and night detection, manpower saving, low requirement on equipment deployment and the like.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and UAV inspection fire detection, specifically to a dual-light information fusion fire detection strategy based on a UAV platform, which is applicable to target detection and localization of flames and smoke in complex environments. Background Technology

[0002] Fire is a major hidden danger threatening life and property. Its destructive power increases exponentially with the burning time. In the early stages, a small flame can develop into a disaster with thick smoke and out-of-control fire in just a few minutes. Fire detection is the "first line of defense" to stop this process. Therefore, effective and reliable fire detection is of great significance for maintaining and protecting social safety.

[0003] In existing technologies, fire monitoring can be categorized by monitoring platform into ground patrols, fixed monitoring stations, and satellite imagery. Ground patrols and fixed monitoring stations are suitable for small-scale detection and rely on manual monitoring. Satellite imagery can cover large areas, but it is susceptible to weather conditions, and satellite observations are periodic, making real-time monitoring and effective control of fire development impossible. Faced with complex and wide-ranging geographical environments, traditional monitoring methods cannot meet the demands for real-time performance, accuracy, and high coverage.

[0004] The development of unmanned systems has brought new solutions to fire inspection, suitable for complex terrain scenarios, and can play an important role in areas where traditional monitoring methods are insufficient. At the same time, the remote monitoring system formed by drones equipped with sensors can meet the needs of real-time monitoring over a wide area.

[0005] While drone-based fire monitoring can utilize a single camera mounted on a drone to form a remote monitoring system, meeting the needs of large-scale monitoring, it faces several challenges: First, current monitoring systems primarily rely on human visual judgment for fire identification, consuming significant manpower and posing risks of visual fatigue and subjective misjudgment. Second, single-sensor detection methods struggle to guarantee high accuracy. Visible light cameras can capture flame shape and color features during the day, but significant errors occur at night or in foggy conditions. Infrared cameras can identify fires by capturing flame energy characteristics, enabling fire monitoring at night or in foggy conditions, but their data is susceptible to interference from ambient temperature. Third, traditional deep learning networks are insufficient for extracting features from small fire sources, easily leading to missed detections and failing to meet fire inspection needs. Furthermore, their large model parameters make them unsuitable for portability to mobile devices for drone inspection. Therefore, an effective solution is urgently needed to address these issues. Summary of the Invention

[0006] The purpose of this invention is to provide a dual-light information fusion fire detection method suitable for inspection drones. It constructs a drone-based fire inspection dataset by collecting existing publicly available fire data; improves the YOLO11 network model framework to enhance the model's fire identification capability; and combines infrared fire energy characteristics with adaptive thresholds and minimum connectivity regions to determine infrared fire identification information, thereby designing an energy-spatiotemporal multi-source information discrimination method. Using the discrimination criteria, the fire situation is further assessed, the fire location is determined, and the accuracy of fire detection is improved, reducing the false negative and false positive rates. This invention is characterized by its adaptability to complex environmental changes, day and night detection capabilities, labor savings, and low requirements for deployment equipment.

[0007] The technical solution to achieve the purpose of this invention is as follows: (This can be omitted for now)

[0008] Compared with existing technologies, the significant advantages of this invention are: (The innovation lies in fusing infrared and visible light information to propose an energy-spatiotemporal determination criterion that improves the reliability of fire detection.)

[0009] (1) By combining dual light information, collaborative fire detection is carried out from the spatial and temporal dimensions. The high-precision detection of deep learning network model and the infrared feature information of fire are combined to realize the judgment of dual infrared and visible light information. A dual light information fusion fire detection method suitable for inspection drones is constructed to improve the accuracy of fire detection and meet the needs of multiple application scenarios in different times and environments.

[0010] (2) By fusing visible light and infrared detection information, dual judgment of fire detection can be achieved, improving detection accuracy. Existing detection methods mostly use a single sensor (visible light), which makes it difficult to guarantee high accuracy in fire detection. By fusing visible light and infrared detection information to achieve dual-light detection, the advantages of both can be fully combined to achieve all-day detection.

[0011] (3) By improving the YOLO11 network architecture, adding a small target detection head and attention mechanism, the detection accuracy of YOLO11 for fire targets is improved. This improves the traditional deep learning network's insufficient feature extraction of small target fire source targets, which easily leads to missed detection. By making the model lightweight, it is possible to port the model to mobile devices, meet the needs of rapid detection, and has low energy consumption, making it suitable for UAV platforms.

[0012] (4) Creatively proposes an energy-spatiotemporal criterion. By coupling infrared and visible light fire detection information, a judgment coefficient is set, and multiple judgments are made on the occurrence of fire from a spatial perspective based on the judgment threshold. Using multi-frame detection data from the sensor, multiple judgments are made on the fire from a temporal perspective, which effectively improves the accuracy of fire detection. Attached Figure Description

[0013] Figure 1 A dual-light information fusion fire detection method for inspection drones;

[0014] Figure 2 Flowchart of the fire detection method using dual-light information fusion at the inspection drone end;

[0015] Figure 3 Flowchart for determining and processing energy-spatiotemporal multi-source information;

[0016] Figure 4 Fire detection simulation results, including (a) visible light detection results, (b) infrared detection results, and (c) superimposed infrared and visible light results. Detailed Implementation

[0017] The terminology used in this invention is for illustrative purposes only and is not intended to limit the invention. The following description, in conjunction with the appendix, further clarifies this concept. Figure 1 -5. Some embodiments of the present invention will be described in detail below.

[0018] The fire detection algorithm provided in this embodiment employs a dual-light information fusion approach to improve fire detection accuracy. Specifically, it monitors for fires by combining visible light and infrared images of the monitoring area collected by a drone, outputting marked locations, and using an energy-spatiotemporal multi-source information discrimination method. Weighting coefficients for the dual-light information are determined to obtain a fire determination score, which then determines whether a fire has occurred. Using onboard navigation information, the location of the fire is determined, and the detection results are transmitted to the command center for timely and appropriate responses to avoid excessive economic losses.

[0019] Figure 1 This specification demonstrates a dual-light information fusion fire detection method for inspection drones, based on an embodiment thereof. Figure 1 A detailed description is provided, including the following steps:

[0020] Step 1: The fire inspection dataset needs to be acquired synchronously using infrared and visible light sensors. In this embodiment, the dataset is obtained by filtering publicly available datasets, and its specific characteristics are as follows:

[0021] The dataset contains two target types: Flame and Smoke. A certain proportion of negative samples, such as lights and clouds, are mixed in to assess the algorithm's reliability and detection accuracy.

[0022] The dataset is from the perspective of a drone, and includes both infrared and visible light data within the same monitoring area. The monitoring area here refers to the area observed by the drone within its field of view at a certain moment.

[0023] The dataset includes lighting scenarios such as daytime, nighttime, cloudy, and sunny days to improve the algorithm's adaptability to different lighting conditions.

[0024] The dataset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The training set is used to train the neural network parameters, the validation set is used to monitor when training stops, and the test set is used to evaluate the algorithm's performance to ensure its effectiveness.

[0025] Step 2: Using the YOLO11 basic network architecture, integrating the attention mechanism and small target detection head, a dedicated inspection drone fire detection network suitable for drone inspection is constructed.

[0026] Step 2.1: Construct a fire detection network model for inspection drones.

[0027] The original YOLO11 network architecture uses an optimized backbone network, increasing the model's feature extraction capabilities. By introducing convolutional modules (C3k2 and SPPF), the model's expressive power is improved while reducing the number of parameters. A C2PSA module is used in the neck region to optimize the model's feature fusion capabilities. This invention further introduces the EfficientChannel Attention (ECA) mechanism to enhance the network's ability to extract key features. By adding a small target detection head, the model's ability to detect small targets in flames and smoke is improved. Furthermore, data augmentation of the dataset simulates different lighting conditions, improving the model's applicability to various lighting conditions and enabling timely and accurate fire detection.

[0028] During inspection operations, drones need to navigate complex and diverse environments, such as forests and built-up areas. The manifestation of fires, the characteristics of smoke and flames, and lighting conditions vary depending on the scenario. In forest inspections, fires may occur within dense vegetation, easily obscuring smoke and flames; in built-up areas, the presence of reflected light and shadows from surrounding buildings can negatively impact visibility. Therefore, the model needs to possess detection capabilities for these specific scenarios, accurately identifying minute flame and smoke targets against complex backgrounds.

[0029] (1) Introducing attention mechanism to enhance key feature extraction

[0030] A combination of channel attention and spatial attention is employed. The channel attention mechanism adaptively adjusts the feature responses of different channels, highlighting those crucial for fire detection. For example, in fire detection, features such as flame color and temperature are prominently represented in specific channels; channel attention can enhance the weights of these channels. The spatial attention mechanism, on the other hand, focuses on key fire-related regions in the image, ignoring irrelevant background information.

[0031] An attention module, ECA (Efficient Channel Attention), is inserted after the convolutional layers of the network. This adaptively enhances key fire feature channels (such as the red flame channel and smoke texture channel). For spatial attention, a convolution-based spatial attention module is used. It performs max pooling and average pooling on the feature map along the channel dimension, then concatenates the two pooling results, and generates a spatial attention map through convolution. Element-wise multiplication of the feature map further enhances the feature representation of key regions.

[0032] ;

[0033] Regional features extracted by attention mechanism

[0034] : Enhance the channel features of the region and generate a channel weight vector

[0035] : Intermediate feature map after enhancing key channels;

[0036] Channel-level weighted average;

[0037] : For intermediate features Apply spatial attention convolution;

[0038] (2) Add a small target detection head to improve the detection capability of tiny targets.

[0039] The design of the small target detection head incorporates a new high-resolution branch (resolution × 4) and integrates shallow features.

[0040] In UAV inspections, flame / smoke targets are typically smaller than 1% of the image size (e.g., a target occupies only 5×5 pixels in a 50×50 pixel image). After downsampling with conventional detection heads (such as YOLO's 1 / 32 scale branch), small targets may only have 1-2 pixels remaining, failing to preserve texture (e.g., the Sobel feature of flame edges). In low-resolution feature maps, small targets are difficult to distinguish from noise (e.g., smoke and clouds).

[0041] A specialized small target detection head is designed, employing a multi-scale feature fusion method. Features are extracted from both shallow and deep feature maps of the network. Shallow feature maps have higher resolution and can retain more detailed information, making them suitable for detecting small targets; deep feature maps have stronger semantic information, providing more accurate classification and localization information. The shallow and deep feature maps are then fused, and a new feature map is generated through convolution operations for small target detection.

[0042] Feature fusion formula

[0043]

[0044] : Features output by the small target detection head;

[0045] Perform a 1x1 convolution operation;

[0046] Shallow high-resolution features (size) Number of channels )

[0047] Deep low-resolution features (size) Number of channels )

[0048] Element-wise addition (first aligned by upsampling) size).

[0049] Adding small target detection heads can significantly improve the model's ability to detect these tiny targets, enabling early warning of fires and providing valuable time for timely firefighting measures.

[0050] (3) Improve the model’s applicability to different lighting conditions.

[0051] During the training phase, the training data undergoes diverse lighting enhancement processing. This includes adjusting parameters such as image brightness, contrast, saturation, and hue to simulate inspection scenarios under different lighting conditions. For example, inspections may encounter low-light conditions such as early morning and evening, as well as strong direct sunlight at midday. Data augmentation allows the model to learn the characteristics of flames and smoke under various lighting conditions.

[0052] Add illumination normalization processing

[0053] Before the model input layer, an illumination normalization module is added. This module preprocesses the illumination of the input image, transforming images under different lighting conditions into a relatively uniform lighting environment. For example, histogram-based or adaptive histogram equalization methods can be used to enhance image contrast and improve the model's robustness to illumination changes.

[0054] Step 2.2: Train the fire detection network model of the inspection drone.

[0055] Using the constructed fire detection network model of the inspection drone, the training set partitioned in step 1 is used as the model input data to optimize the network parameters of the fire detection network model of the inspection drone. By continuously adjusting the parameters of the fire detection network model of the inspection drone, the loss function value of the model on the training set is continuously reduced. After multiple iterations of training, a well-trained model is obtained. The validation set partitioned in step 1 is input into the trained dedicated fire detection model of the inspection drone for validation. When the performance index of the model on the validation set no longer improves, training is stopped, and the optimized fire detection network model of the inspection drone is obtained. The test set partitioned in step 1 is input into the optimized fire detection network model of the inspection drone for testing. The test set images are input into the model, and the model generates prediction results. The detection accuracy and robustness of the model on the test set are quantified by calculating indicators to obtain the optimal model.

[0056] Step 2.3: Lightweighting of the inspection drone fire detection network model

[0057] Considering the size and performance of the fire detection network model for inspection drones, quantization and pruning techniques are used to lightweight the trained network, thereby meeting the application requirements of drone platforms. Quantization converts parameters in the neural network (such as weights and activation values) from high precision (e.g., 32-bit floating-point numbers) to low precision (e.g., 8-bit integers), thereby reducing model size and improving the computational efficiency of the neural network. Considering the loss of precision and the impact of quantization errors, this invention employs a hybrid quantization method to ensure model accuracy. Pruning in neural networks refers to removing unnecessary connections or nodes from the trained neural network to reduce model size, accelerate inference, and reduce overfitting. This invention prunes the network after training by adjusting pruning parameters.

[0058] The unprocessed weight tensor, consisting of floating-point weights, is shown below:

[0059]

[0060] The processed weight tensor is

[0061]

[0062] In the formula, S is the scaling factor.

[0063] b is the target bit width (e.g., when the convolutional layer uses 8 bits, b=8).

[0064] : Rounding to the nearest integer

[0065] clip: Limits the value Within range

[0066] Pruning is performed on the network after training is completed by adjusting the pruning parameters.

[0067] (1) Calculate the L2 norm of the channel;

[0068] Let the weight matrix be

[0069] The importance score of a single output channel is

[0070]

[0071] Sort by importance score Sort in descending order to generate an index sequence .

[0072] (2) Define the target pruning ratio r

[0073] Calculate the number of reserved channels

[0074] The channels to be retained are determined based on the number of channels retained and the ranking results of importance scores.

[0075] Step 3: The fire area identification method based on energy characteristics determines the infrared information fire identification method by designing an adaptive threshold and a minimum connected area.

[0076] First, a static thresholding method is used for initial segmentation. Then, all pixels in the image are traversed, and the average gray value of the neighborhood of pixels identified as fire zones is calculated and compared with the static threshold. If the average gray value is large, the threshold for that pixel is lowered, making it easier to be identified as a bright spot during thresholding. If the average gray value is small, the segmentation threshold for that point is raised accordingly, making it easier to be identified as a dark spot. That is, points located in dark areas are more likely to be identified as dark spots, and points located in bright areas are more likely to be identified as bright spots. Compared with the general thresholding method, this segmentation method incorporates the regional gray value distribution characteristics of the image, making it easier to segment complete targets while avoiding noise interference.

[0077] Then, based on image morphology processing, erosion and dilation operations are performed on the detection area to ensure the connectivity of the area.

[0078] Step 4, the fire detection information combined with that from Step 2 and Step 3, such as... Figure 4 As shown, an energy-spatiotemporal multi-source information discrimination method is designed. Using the discrimination criteria, the fire situation is further assessed to determine whether an alarm response is necessary. The energy-spatiotemporal multi-source information discrimination method significantly improves the accuracy of fire detection and reduces the false alarm and missed detection rates. The specific steps are as follows:

[0079] Step 4.1 Based on the fire detection results, if both infrared and visible light detection results indicate a fire, proceed directly to Step 4.2 for multi-frame detection verification. If both infrared and visible light detection results indicate no fire, skip Step 4 and continue the inspection. If one of the infrared and visible light detections indicates a fire and the other indicates no fire, then the fire detection information obtained in Steps 2 and 3 is weighted according to different environmental conditions. In night mode or when lighting conditions are insufficient, the weight of infrared image detection results is greater than that of visible light detection results. In daytime mode or when lighting conditions are sufficient, the weight of visible light detection results is greater than that of infrared detection results. The dual-light detection results are weighted to obtain a fire determination coefficient, and the fire is determined from a spatial dimension.

[0080] Target detection model box , where x min y min x is the coordinate of the bottom left corner of the detection model box of the target. max y max The coordinates of the upper right corner of the detection model box for the target.

[0081] Step 4.2 Utilize spatial judgment methods to evaluate multiple frames of data, identifying areas with fire detection scores exceeding a set threshold as the final fire detection result and triggering a fire alarm. This achieves fire detection from a temporal perspective.

[0082] Get the center pixel coordinates of the current detection box The calculation method is as follows:

[0083]

[0084] Where x min y min x is the coordinate of the bottom left corner of the detection model box of the target. max y max The coordinates of the upper right corner of the detection model box for the target.

[0085] Step 5: Combining the judgment results from Step 4, if a fire is triggered, the location of the fire is determined using dual-light information and airborne navigation information, triggering the alarm mechanism and transmitting the results back to the command center, thus completing the fire detection. The main principle is to determine the pixel coordinates of the target in the camera, then use the distance ratio to determine the target's relative coordinates to the camera in the real world, thereby using the camera's three-dimensional coordinate information to determine the target's three-dimensional coordinates in the real world. The specific steps are as follows:

[0086] (1) Construct a coordinate system, calibrate the relative positions of the camera and the UAV center, obtain calibration parameters, and process and determine the coordinates of the target pixel T in the camera coordinate system C after acquiring the image through the camera. , , );

[0087] (2) Obtain the UAV's location information in the Global Positioning System (GPS) u B u H u );L u For the longitude and B of the drone u For the latitude and H of the drone u This refers to the flight altitude of the drone.

[0088] (3) Coordinate system transformation

[0089] After obtaining the target's coordinates in camera coordinate system C, a coordinate transformation matrix is ​​established based on the geometric transformation relationships between the coordinate systems. The coordinates of the target image point in different coordinate systems are calculated step by step, ultimately determining its position in the geocentric rectangular coordinate system E. The formula is a combined calculation formula from camera coordinate system C to geocentric rectangular coordinate system E, where ( , , ), ( , , ) are the coordinates of the target in the geocentric rectangular coordinate system E and the camera coordinate system C, respectively, and A, B, and C are coordinate transformation matrices.

[0090]

[0091]

[0092]

[0093]

[0094] In the formula For the longitude of the drone, for , ( , , ( ) represents the coordinates of the UAV in the geocentric rectangular coordinate system E, and the attitude yaw angle of the mounted sensor in the geographic coordinate system. Pitch angle and roll angle , This refers to the roll angle of the gimbal camera. The elevation angle of the gimbal camera.

[0095] Based on the latitude, longitude and altitude information of the drone , , ), L u For the longitude and B of the droneu For the latitude and H of the drone u The following formula is used to calculate the coordinates of the UAV in the geocentric rectangular coordinate system E to determine its flight altitude. , , );

[0096]

[0097] In the formula, N represents the radius of curvature of the ramusoidal circle, and e represents the first eccentricity.

[0098] Example

[0099] Based on existing public datasets, a fire detection network model for inspection drones was trained, and the model was used to identify fires using visible light data. The results are as follows: Figure 4 As shown in (a), the smoke in the detection area is selected using a model target detection box. Energy feature discrimination is performed on the infrared data of the unified detection area, and areas identified as fires are highlighted in red. Figure 4 As shown in (b), the two images are finally superimposed for discrimination, as follows: Figure 4 As shown in (c), a fire has been identified.

[0100] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A dual-light information fusion fire detection method suitable for a patrol unmanned aerial vehicle terminal, characterized in that, The specific steps are: Step 1: Establish a fire data set for unmanned aerial vehicle inspection, and label the relevant content of fire detection in the visible light image and infrared data set to provide a data basis for subsequent information fusion, and proceed to Step 2; Step 2: Based on the features of the visible light image information in the fire detection data set, based on the original framework of YOLO11, integrate attention mechanism and small target detection head to build a fire detection network model for inspection unmanned aerial vehicle, realize the detection of fire content in visible light image, further label the position information in the image, and proceed to Step 3; Step 3: Based on the features of infrared information in the fire detection data set, design a fire area recognition method based on energy features, determine the infrared information fire recognition method by designing adaptive threshold and minimum connected region; Realize the detection of fire content in infrared information, further label the position information, and proceed to Step 4; Step 4: Combine the position information of Step 2 and Step 3, design an energy-spatiotemporal multi-source information judgment method; Based on dual light information, design the weight coefficient of dual light information to get the fire judgment score; Combine multiple frames of information, when the fire judgment score is greater than the set threshold, trigger the fire alarm, and proceed to Step 5; Step 5: Combine the judgment results of Step 4, if the fire alarm is triggered; Use dual light information and airborne navigation information to determine the location of the fire, trigger the alarm mechanism, and return the results to the command center to complete the fire detection.

2. The dual light information fusion fire detection method suitable for the end of the inspection unmanned aerial vehicle according to claim 1, characterized in that, In Step 1, the target type of the data set is divided into two categories: flame and smoke, and a certain proportion of negative samples is mixed into the data set to judge the reliability of the algorithm and evaluate the detection accuracy of the algorithm; The data set is from the perspective of the unmanned aerial vehicle, and there are infrared data and visible light data in the same monitoring area. The monitoring area here refers to the area observed by the unmanned aerial vehicle within its field of view at a certain moment; The light scene of the data set includes day, night, cloudy and sunny, which is used to improve the adaptability of the algorithm to different light conditions; The data set is divided into training set, validation set and test set according to the ratio of 8:1:

1. The training set is used for neural network parameter training, the validation set is used for monitoring the stopping of training, and the test set is used for evaluating the performance of the algorithm to ensure the effectiveness of the algorithm. 3.The dual-light information fusion fire detection method suitable for the end of a patrol unmanned aerial vehicle according to claim 1, characterized in that, In Step 2, in order to build a fire detection network model suitable for inspection unmanned aerial vehicle, (1) Introduce attention mechanism to enhance key feature extraction The combination of channel attention and spatial attention is selected, and the attention module ECA (Efficient Channel Attention) is inserted after the convolution layer of the network to adaptively enhance the key feature channel of fire; For spatial attention, a spatial attention module based on convolution is used, which combines the maximum and average pooling results in the channel dimension of the feature map, and generates a spatial attention map through convolution operation to enhance the feature expression of the key region; ； : attention mechanism extracts region features; : region is enhanced by channel feature, and channel weight vector is generated; : enhanced intermediate feature map after key channel enhancement : channel level weight weighting; : on intermediate features applying spatial attention convolutions; (2) Increase small target detection head to improve the detection ability of small target; A small target detection head is designed, and a multi-scale feature fusion method is adopted; features are extracted from the shallow and deep feature maps of the network respectively, the shallow and deep feature maps are fused, and a new feature map is generated through convolution operation for small target detection; The feature fusion formula is: ； : small target detection head output features; : perform a 1x1 convolution operation; : shallow high resolution features (size , number of channels ); : deep low resolution features (size , number of channels ); : Element-wise addition (first aligning the dimensions by up-sampling Upsample ) The small target detection head is added to improve the detection ability of the model for these small targets, realize early warning of fire, and provide valuable time for timely fire extinguishing measures; (3) Improve the applicability of the model to different light conditions: Before the input layer of the model, a light normalization module is added; The addition of the light normalization module preprocesses the light of the input image, and converts the images under different light conditions to a relatively unified light environment. 4.The dual-light information fusion fire detection method suitable for the end of a patrol unmanned aerial vehicle according to claim 1, characterized in that, In step 2, the lightweight processing of the inspection unmanned aerial vehicle fire detection network model, The weight tensor before processing is a floating-point weight as follows: ； The weight tensor after processing is: ； In the formula, S is a scaling factor, ; b is a target bit width, to round off, clip is to limit a value to be within a range range; Pruning is performed after the training of the unmanned aerial vehicle fire detection network model, and the pruning parameters are adjusted to achieve it; (1) Calculate the L2 norm of the channel; Let the weight matrix ; The importance score of a single output channel is: ； Ranking according to importance score Descending order, generating index sequence ; (2) Define the target pruning ratio r: Computing the number of channels to retain ; determining the channels to retain based on the ranking of the channels according to the number of channels to retain and the importance scores. 5.The dual-light information fusion fire detection method suitable for the end of a patrol unmanned aerial vehicle according to claim 1, characterized in that, In step 3, the static threshold segmentation method is used for initial segmentation, and then all pixel points of the image are traversed to calculate the neighborhood gray mean value of the pixel points judged as fire area in image segmentation, and compared with the static threshold. If the gray mean value is large, the threshold value of the pixel point will be lowered, and the pixel point will be judged as a bright point in threshold segmentation. If the gray mean value is small, the segmentation threshold of the pixel point will be increased accordingly, so that the pixel point will be judged as a dark point. 6.The dual-light information fusion fire detection method suitable for the end of a patrol unmanned aerial vehicle according to claim 1, characterized in that, The specific steps in step 4 are: Step 4.1 Energy-Spatiotemporal Multi-Source Information Judgment Method: According to the judgment result of fire detection, if the infrared and visible light detection results both judge that there is fire, then directly enter step 4.2 for multi-frame detection verification. If the infrared and visible light detection results both judge that there is no fire, then skip step 4 and continue inspection. If one of the infrared and visible light judges that there is fire and the other judges that there is no fire, then based on the fire detection information obtained through steps 2 and 3, set the weight value under different environmental conditions. In the night mode or insufficient light condition, the infrared image detection result is more important than the visible light detection result. In the daytime mode or sufficient light condition, the visible light detection result is more important than the infrared detection result. The double light detection results are weighted and calculated to obtain the fire judgment coefficient, and the fire is judged from the spatial dimension; Target detection model frame wherein x min , y min is the left lower corner coordinate of the target detection model frame, x max , y max is the right upper corner coordinate of the target detection model frame; Step 4.2, use the spatial judgment method to judge the multi-frame data, and judge the area with a fire judgment score greater than the set threshold as the final fire detection result, and perform fire alarm; Realize fire judgment from the time dimension; obtaining the center pixel coordinate in the current detection frame which is calculated as follows: ； wherein x min , y min is the left lower corner coordinate of the detection model frame of the target, x max , y max is the right upper corner coordinate of the detection model frame of the target.

7. The dual light information fusion fire detection method suitable for the end of the patrol unmanned aerial vehicle according to claim 1, characterized in that, In step 5, the specific steps of determining the location of the fire are as follows: confirm the pixel coordinates of the target in the camera, then determine the relative coordinates of the target in the real world by distance ratio, and then convert the three-dimensional coordinates of the target in the real world by using the three-dimensional coordinate information of the camera: (1) Construct a coordinate system, calibrate the relative position of the camera and the center of the unmanned aerial vehicle, obtain the calibration parameters, process and determine the coordinates of the to-be-positioned target pixel T in the camera coordinate system C after obtaining the image by the camera , , ); (2) acquiring global positioning system position information (L u , B u , H u ) of the unmanned aerial vehicle U; L u is the longitude of the unmanned aerial vehicle, B u is the latitude of the unmanned aerial vehicle, and H u is the flight height of the unmanned aerial vehicle; (3) Coordinate system conversion: After the coordinates of the target in the camera coordinate system C are obtained, the coordinates of the target image point in different coordinate systems are calculated step by step according to the geometric conversion formula of the coordinate system, and finally the position of the target in the geocentric space rectangular system E is determined. 8.The dual-light information fusion fire detection method suitable for the end of a patrol unmanned aerial vehicle according to claim 7, characterized in that, In step 5, after obtaining the target's coordinates in camera coordinate system C, a coordinate transformation matrix is ​​established based on the geometric transformation relationships between the coordinate systems. The coordinates of the target image point in different coordinate systems are calculated step by step, and finally its position in the geocentric rectangular coordinate system E is determined. The formula is a combined calculation formula from camera coordinate system C to geocentric rectangular coordinate system E, where ( , , ), ( , , ) are the coordinates of target T in the geocentric rectangular coordinate system E and the camera coordinate system C, respectively, and A, B, and C are coordinate transformation matrices; ； In the formula is the longitude of the UAV, is , is the latitude of the UAV, , , is the coordinate of the UAV in the geocentric rectangular coordinate system E, the attitude yaw angle of the loaded sensor in the geographic coordinate system , the attitude pitch angle and the attitude roll angle , is the roll angle of the gimbal camera, is the pitch angle of the gimbal camera; Based on the longitude and latitude and height information of the UAV (L u , B u , H u ), the coordinates of the UAV in the geocentric rectangular coordinate system E are calculated by using the following formula: 、 、 ); ； In the formula, N represents the curvature radius of the prime vertical, and e represents the first eccentricity.

Images