Lightweight deep learning fire detection method and system oriented to edge deployment

Abstract

The invention discloses an edge deployment-oriented lightweight deep learning fire detection method and system, relates to the field of computer vision and intelligent perception, and aims to solve the problems that an existing model is complex in structure, high in calculation overhead, depends on high-performance calculation equipment and is difficult to directly deploy on an edge calculation platform. According to the technical key points, the method comprises the following steps: S100, collecting image data of a fire scene, and constructing a diversified flame smoke image data set; s200, a lightweight fire detection network model is constructed, the lightweight fire detection network model comprises a backbone network, a neck network and a head detection network, the backbone network introduces a C3Biform module and is used for extracting multi-scale image features, the neck network is integrated with an AsymptoticFPN module and is used for carrying out interactive fusion on feature maps of different scales, the head detection network adopts a PinwhheelHead structure, and the C3Biform module and the AsymptoticFPN module are used for carrying out interactive fusion on feature maps of different scales; the method is used for improving the perception capability of a rotating target and realizing accurate detection. And S300, deploying the model on an unmanned aerial vehicle platform, and performing fire source space positioning in combination with unmanned aerial vehicle attitude parameters and a camera imaging geometric model.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and intelligent sensing technology, and more specifically, to a lightweight deep learning fire detection method, system, and storage medium for edge deployment. Background Technology

[0002] Fires are characterized by their sudden onset, rapid spread, and wide-ranging impact. Their unpredictability and inherent danger make them a major challenge for several key sectors, including global agricultural production, urban safety management, and forest resource protection. Once a fire breaks out, it often causes numerous casualties and property losses in a very short time, especially in the absence of early warning systems, where the consequences are particularly severe. Therefore, timely and accurate identification of flames or smoke in the early stages of a fire, before it develops into a large-scale open flame, is of paramount practical importance for achieving early warning, gaining crucial time for firefighting, and minimizing disaster losses.

[0003] Traditional fire detection methods often rely on temperature sensors, smoke concentration sensors, or rule-based image feature extraction algorithms for identification. However, these methods often suffer from detection lag, high false alarm rates, and poor adaptability in complex environments, making it difficult to accurately capture changes in the morphology of flames and smoke. In recent years, deep learning-based fire detection technology has gradually emerged, achieving significant improvements in detection accuracy and robustness by automatically extracting key features from images to identify flames and smoke. However, most existing models are complex in structure, computationally expensive, and dependent on high-performance computing equipment, making them difficult to deploy directly on edge computing platforms such as embedded terminals, mobile devices, and drones, thus limiting their promotion and application in real-world scenarios.

[0004] Therefore, there is an urgent need to propose a fire detection method and system that is lightweight, adaptable to edge computing environments, and combines accuracy and robustness to meet the comprehensive requirements of real-time performance, accuracy, and deployment flexibility in practical applications. Summary of the Invention

[0005] The technical problem to be solved by this invention is:

[0006] Existing models are complex in structure, have high computational overhead, rely on high-performance computing devices, and are difficult to deploy directly on edge computing platforms such as embedded terminals, mobile devices, and drones.

[0007] To address this, the present invention provides a lightweight deep learning fire detection method, system, and storage medium for edge deployment. While maintaining detection accuracy, it reduces model computational overhead and enhances the model's deployment capability and real-time response capability on edge computing platforms such as embedded terminals, mobile devices, and drones.

[0008] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0009] This invention provides a lightweight deep learning-based fire detection method for edge deployment, comprising the following steps:

[0010] S100: Collect image data of fire scenes, and enhance the collected image data to construct a diverse dataset of flame and smoke images;

[0011] S200. Construct a lightweight fire detection network model, which includes a backbone network, a neck network, and a head detection network. The backbone network introduces the C3_Biformer module to extract multi-scale image features. The neck network integrates the AsymptoticFPN module to perform interactive fusion of feature maps at different scales. The head detection network adopts the PinwheelHead structure to improve the perception of rotating targets and achieve accurate detection.

[0012] The S300 deploys the trained fire detection network model on the UAV platform, collects video streams in real time during flight to identify fires, and combines the UAV attitude parameters with the camera imaging geometry model to locate the fire source in space.

[0013] Furthermore, the enhancement processing of the acquired image data described in S100 includes: mirroring, contrast transformation, cropping, noise superposition, random filling, and random combination processing of the image.

[0014] Furthermore, the C3_Biformer module in S200 consists of a main branch and a shortcut branch; the main branch includes, in sequence, a convolutional layer, a BiFormer attention module, a bottleneck module, a feature merging unit, and an output convolutional layer; the shortcut branch contains a convolutional layer, which is used to perform a linear transformation on the input features and is fused with the output of the main branch after all calculations in the main branch are completed.

[0015] Furthermore, the functional implementation process of the AsymptoticFPN module described in S200 is as follows:

[0016] Feature maps of multiple scales are extracted from the backbone network; firstly, convolutional enhancement processing is performed on the features of multiple scales to make them fusionable;

[0017] Secondly, multi-scale feature interaction is achieved through cross-layer bidirectional fusion;

[0018] Finally, for the first layer to the first Feature vectors obtained by layer mapping The feature vector is output by fusion using a linear weighting method. Represented as:

[0019]

[0020] In the formula: , , These represent the spatial fusion weights of the three scale features corresponding to this location, and satisfy the normalization condition: .

[0021] Furthermore, the head detection network described in S200 adopts a PinwheelHead structure, consisting of regression and classification branches, and its functional implementation process is as follows:

[0022] In the regression branch, the input features are sequentially processed through two convolutional layers and one 1×1 convolutional layer to generate discrete distance distributions in the four directions of the bounding box; then, a distribution focus regression method is used to decode the discrete distributions into continuous distance values:

[0023]

[0024] In the formula: This represents the upper limit of the number of channels in the discrete regression distribution. for The probability distribution after that, For discrete distance index;

[0025] Based on anchor position and predicted distance , , , Decoding yields the bounding box coordinates:

[0026]

[0027]

[0028]

[0029]

[0030] The classification branch uses a rotation-aware convolution module. First, it performs asymmetric padding on the input features in four directions to form four sets of feature maps with directional offsets. Then, it performs a set of directional convolutions on each directional branch, enabling the network to extract differential features from the horizontal, vertical and other bias directions.

[0031] Then, the outputs of the four-directional convolutions are concatenated along the channel dimension to form enhanced features containing multi-directional structural information. These features are then fused and compressed to obtain a feature map with enhanced orientation. Further high-level semantic information is extracted, and a class prediction for each anchor is generated. Activation yields category confidence:

[0032]

[0033] In the formula: This represents the feature map output by the classification branch. For the process The confidence probability of each target class after activation;

[0034] Finally, the outputs of the regression branch and the classification branch are concatenated along the channel dimension to form the detection results of each feature map layer.

[0035] Furthermore, the step S300, which involves combining the UAV attitude parameters with the camera imaging geometric model to perform spatial localization of the fire source, specifically includes:

[0036] Based on the pixel coordinates of the detected flame or smoke target in the image Calculate the normalized ray direction in the camera coordinate system: Transform the camera coordinate system to the body coordinate system, and then change the direction of the ray after the attitude transformation. Mapping to geographic coordinate system:

[0037]

[0038] In the formula: This is the rotation matrix from the body coordinate system to the geographic coordinate system. This is the rotation matrix from the camera coordinate system to the body coordinate system;

[0039] Combining the position of the drone in the geographic coordinate system The location of the fire source in the geographic coordinate system for:

[0040]

[0041] In the formula: Ground height, Let be the component of the ray in the vertical direction. This refers to the flight altitude.

[0042] Furthermore, step S300 also includes: dynamically correcting the UAV attitude changes by combining the real-time attitude data of the UAV IMU, and filtering the detection positions of consecutive multiple frames using an adaptive confidence weight smoothing algorithm; the adaptive confidence weight smoothing algorithm introduces the detection confidence as a dynamic weight factor on the basis of traditional time-weighted smoothing, and is defined as follows:

[0043]

[0044] In the formula: This is the estimated geographical location of the fire source after adaptive weighted smoothing. The sliding time window length is the number of frames traced backward. For the first Frame fire source geographic coordinates, The detection confidence score of the model in this frame. For time decay weight, This is the time decay coefficient, used to control the weight decay of historical frames.

[0045] This invention provides a lightweight deep learning fire detection system for edge deployment. The system has a program module corresponding to the steps of any of the above-described technical solutions, and executes the steps of the above-described lightweight deep learning fire detection method for edge deployment when running.

[0046] The present invention provides a computer-readable storage medium storing a computer program configured to implement, when invoked by a processor, the steps of the lightweight deep learning fire detection method for edge deployment as described in any of the above technical solutions.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] The fire detection model provided by this invention enhances the network's ability to perceive multi-scale fire features while maintaining lightweight characteristics by introducing the C3_Biformer module. The AsymptoticFPN module strengthens the feature interaction between different levels in the pyramid network by fusing feature maps from different scales. By sequentially extracting and fusing features from different levels, information is gradually transmitted and accumulated, significantly improving the expressive power of the feature maps and detection performance. Simultaneously, a PinwheelHead structure is adopted in the detection head, and a rotation-aware convolutional module is introduced in the classification branch to enhance the network's ability to model the directional changes and pose uncertainties of flames and smoke in complex scenes, effectively improving the recognition accuracy and robustness of rotating, tilted, and irregularly shaped fire targets. The overall model structure is lightweight, with significantly lower parameter count and computational complexity than traditional deep detection models, enabling deployment on edge computing platforms such as embedded terminals, mobile devices, and drones, achieving efficient fire monitoring in low-resource environments.

[0049] The model of this invention has good real-time performance, generalization ability and engineering application prospects, and can be widely used in various fire monitoring scenarios such as urban fire protection, forest fire prevention and industrial monitoring. Attached Figure Description

[0050] Figure 1 This is a structural diagram of the fire detection network model in an embodiment of the present invention;

[0051] Figure 2 This is a partial image effect diagram after data enhancement in an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the C3_Biformer module in an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the BRA module in an embodiment of the present invention;

[0054] Figure 5 This is a schematic diagram of the AsymptoticFPN module in an embodiment of the present invention;

[0055] Figure 6 This is a schematic diagram of the PinwheelConv module in an embodiment of the present invention;

[0056] Figure 7 These are partial flame and smoke images of the test set in this embodiment of the invention;

[0057] Figure 8 This is a partial detection result diagram of the fire detection network model in this embodiment of the invention. Detailed Implementation

[0058] To enable those skilled in the art to better understand the present invention, exemplary embodiments or examples of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments or examples are merely some, not all, of the embodiments or examples of the present invention. All other embodiments or examples obtained by those skilled in the art based on the embodiments or examples of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0059] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0060] Specific Implementation Plan 1: Combining Figure 1 As shown, this invention provides a lightweight deep learning fire detection method for edge deployment, comprising the following steps:

[0061] S100: Collect image data of fire scenes, and enhance the collected image data to construct a diverse dataset of flame and smoke images;

[0062] S200. Construct a lightweight fire detection network model, which includes a backbone network, a neck network, and a head detection network. The backbone network introduces the C3_Biformer module to extract multi-scale image features. The neck network integrates the AsymptoticFPN module to perform interactive fusion of feature maps at different scales. The head detection network adopts the PinwheelHead structure to improve the perception of rotating targets and achieve accurate detection.

[0063] The S300 deploys the trained fire detection network model on the UAV platform, collects video streams in real time during flight to identify fires, and combines the UAV attitude parameters with the camera imaging geometry model to locate the fire source in space.

[0064] Specific Implementation Plan Two: (e.g.) Figure 2 As shown, the enhancement processing of the acquired image data described in S100 includes: image mirroring, contrast transformation, cropping, noise superposition, random filling, and random combination processing. Other aspects of this implementation scheme are the same as in Specific Implementation Scheme One.

[0065] Specific implementation plan three: The backbone network described in S200 introduces the C3_Biformer module, such as... Figure 3As shown, the C3_Biformer module consists of a main branch and a shortcut branch. The main branch includes, in sequence, a convolutional layer, a BiFormer attention module, a bottleneck module, a feature merging unit, and an output convolutional layer. The shortcut branch contains a convolutional layer used to perform a linear transformation on the input features and is fused with the output of the main branch after all calculations in the main branch are completed.

[0066] In the main branch, the first two convolutional layers first perform channel adjustment and local feature extraction on the input features to obtain intermediate features. It then sends it to the BiFormer module. BiFormer internally uses... Figure 4 The BRA sparse attention mechanism shown performs key region selection and global information modeling on the input features. Its calculation process is as follows:

[0067] First, input features The region is divided, and the feature map is divided into Each region contains non-overlapping regions. Each feature vector is represented by a region. The features within all regions are then rearranged according to the region dimension to obtain the reorganized feature tensor. This transformation enables explicit encoding of spatial regions, providing structured input for subsequent BRA sparse attention.

[0068] Based on input features Calculate the query matrix Key matrix Sum matrix :

[0069]

[0070]

[0071]

[0072] In the formula: , , These are the weight matrices for the query, key, and value matrices, respectively.

[0073] Then, the similarity between regions is calculated, and multi-layer feature fusion is performed through the BRA mechanism:

[0074]

[0075]

[0076] In the formula: and These are subsets of the query matrix and the key matrix, respectively. Indicates the degree of matching between the query and the key. An index representing the keys and values ​​most relevant to the query;

[0077] The following formula is used to weight and fuse features:

[0078]

[0079]

[0080] In the formula: and Through The operation is performed on the subtensors selected from the original key and value matrices;

[0081] Finally, through Operations and additional loss functions The final output O is obtained:

[0082]

[0083] After obtaining the BiFormer output, the main branch continues to refine features through the bottleneck module, enhancing cross-channel expressive capabilities. The output of the bottleneck module enters the feature merging unit, where it is fused with the structured features from the preceding path, and the feature output is completed through the last convolutional layer of the main branch.

[0084] After the main branch completes its computation, the shortcut branch generates linear transformation features through a single convolutional layer, which are then concatenated with the final output of the main branch to obtain the final output features of the C3_Biformer module. This structural design complements the local expressive power of convolutional features with the global modeling power of BRA sparse attention, enabling the model to highlight key regions while maintaining computational efficiency and enhancing its ability to represent targets in complex scenes. Other aspects of this implementation scheme are the same as in specific implementation scheme three.

[0085] Specific Implementation Scheme 4: The neck network integrated with the Asymptotic FPN module described in S200 has the following functional implementation process:

[0086] First, such as Figure 5 As shown, feature maps at three scales are extracted from the backbone network, corresponding to L1, L2, and L3 in the figure. The L1 layer, represented by red nodes, has strong semantic understanding but the lowest spatial resolution; the L3 layer, represented by blue nodes, has the richest spatial details but weakest semantic information; the L2 layer, represented by yellow nodes, lies in between, possessing a moderate level of semantic and spatial information. The neck network first performs convolutional enhancement processing on the features at the three scales to make them fusionable.

[0087] Secondly, multi-scale feature interaction is achieved through cross-layer bidirectional fusion. The L3 layer features of the blue nodes are propagated upwards and fused with the L2 layer features of the yellow nodes, allowing the mid-level features to obtain more detailed information. The fused yellow node features continue to propagate upwards and are fused with the L1 layer features of the red nodes, supplementing the high-level semantic features with details. At the same time, the high semantic features of the red nodes are propagated downwards along the reverse path, interacting with the yellow and blue nodes, allowing semantic information to be injected into the mid- and low-level features, thus achieving bidirectional fusion from top to bottom and bottom to top.

[0088] Finally, after multiple cross-scale interactions, the neck network at each spatial location At any point, fused inputs from different scales can be obtained. For the point from the... layer to the first Feature vectors obtained by layer mapping The feature vector is output by fusion using a linear weighting method. Represented as:

[0089]

[0090] In the formula: , , These represent the spatial fusion weights of the three scale features corresponding to this location, and satisfy the normalization condition: .

[0091] Finally, as Figure 5 As shown on the right, the red, yellow, and blue nodes correspond to the fusion results at three scales, which are input into the subsequent prediction head to achieve multi-scale target detection capability, enabling the model to adapt to the detection needs of targets of different sizes simultaneously. This implementation scheme is otherwise the same as specific implementation scheme three.

[0092] Specific Implementation Scheme 5: The head detection network described in S200 adopts a PinwheelHead structure, consisting of a regression branch and a classification branch. The PinwheelHead receives multi-scale feature maps output from the neck network and processes the features at different scales to form a multi-scale feature representation suitable for target detection. Its functional implementation process is as follows:

[0093] In the regression branch, the input features are sequentially processed through two convolutional layers and one 1×1 convolutional layer to generate discrete distance distributions in the left, right, top, and bottom directions of the bounding box. Then, a distribution focus regression method is used to decode the discrete distributions into continuous distance values.

[0094]

[0095] In the formula: This represents the upper limit of the number of channels in the discrete regression distribution. for The probability distribution after that, For discrete distance index;

[0096] Based on anchor position and predicted distance , , , Decoding yields the bounding box coordinates:

[0097]

[0098]

[0099]

[0100]

[0101] In the classification branch, to improve the network's robustness to target rotation and pose changes, a rotation-aware convolutional module, PinwheelConv, is employed. This module leverages orientation-selective convolution to enhance the orientation sensitivity of features, thereby improving the classification branch's ability to discriminate rotating targets. PinwheelConv first performs asymmetric padding on the input features in four directions, forming four sets of feature maps with directional offsets, such as... Figure 6 As shown, (1,0,0,3), (0,3,0,1), (0,1,3,0), and (3,0,1,0) correspond to biases in different directions such as upward, rightward, downward, and leftward. Then, a set of directional convolutions is performed on each directional branch, enabling the network to extract differential features from the horizontal, vertical, and other bias directions.

[0102] Then, the outputs of the four convolutions are concatenated along the channel dimension to form an enhanced feature containing multi-directional structural information. This enhanced feature is then fused and compressed using a 2×2 convolution layer to obtain the directionally enhanced version. Feature maps; the classification branch continues to extract high-level semantic information through depthwise separable convolutions and 1×1 convolutions, and generates class predictions for each anchor through the final 1×1 convolution. Activation yields category confidence:

[0103]

[0104] In the formula: This represents the feature map output by the classification branch. For the process The confidence probability of each target class after activation;

[0105] Finally, the regression branch output and the classification branch output are concatenated along the channel dimension to form the detection results of each layer's feature map. The outputs of each layer are then merged for post-processing, and the top few detection boxes with the highest confidence are selected to improve the detection accuracy and robustness for targets with rotation and pose changes. This implementation scheme is otherwise the same as specific implementation scheme four.

[0106] Specific implementation plan six: Step S300, which involves combining the UAV attitude parameters with the camera imaging geometric model to perform spatial localization of the fire source, specifically includes:

[0107] Based on the pixel coordinates of the detected flame or smoke target in the image Calculate the normalized ray direction in the camera coordinate system: Transform the camera coordinate system to the body coordinate system, and then change the direction of the ray after the attitude transformation. Mapping to geographic coordinate system:

[0108]

[0109] In the formula: This is the rotation matrix from the body coordinate system to the geographic coordinate system, calculated based on the attitude angles of the UAV's IMU. The rotation matrix from the camera coordinate system to the body coordinate system is the extrinsic parameter matrix obtained from camera calibration, used to describe the fixed installation relationship between the camera and the body coordinate system;

[0110] Combining the position of the drone in the geographic coordinate system The location of the fire source in the geographic coordinate system for:

[0111]

[0112] In the formula: Ground height, Let be the component of the ray in the vertical direction. This refers to the flight altitude.

[0113] This implementation plan is otherwise the same as Implementation Plan Five.

[0114] Specific implementation plan seven: Step S300 further includes: combining the real-time attitude data of the UAV IMU to dynamically correct the UAV attitude changes, and using an adaptive confidence weight smoothing algorithm to filter the detection positions of multiple consecutive frames; the adaptive confidence weight smoothing algorithm introduces the detection confidence as a dynamic weight factor on the basis of traditional time-weighted smoothing, and is defined as follows:

[0115]

[0116] In the formula: This is the estimated geographical location of the fire source after adaptive weighted smoothing. The sliding time window length is the number of frames traced backward. For the first Frame fire source geographic coordinates, The detection confidence score of the model in this frame. For time decay weight, This is the time decay coefficient, used to control the weight decay of historical frames. The rest of this implementation scheme is the same as specific implementation scheme six.

[0117] The lightweight deep learning fire detection method (algorithm) proposed in this invention for edge deployment is the underlying technical core of this invention, and various products can be derived based on the algorithm.

[0118] Based on the method proposed in this invention, a lightweight deep learning fire detection system for edge deployment is developed using a programming language. This system has program modules corresponding to the steps of the above-described technical solution, and executes the steps in the above-described lightweight deep learning fire detection method for edge deployment when running.

[0119] The developed system (software) computer program is stored on a computer-readable storage medium, and the computer program is configured to implement the steps of the aforementioned lightweight deep learning fire detection method for edge deployment when invoked by a processor. In other words, the invention is materialized on a carrier, becoming a computer program product.

[0120] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0121] The computational programs (also referred to as programs, software, software applications, or code) of this invention include machine instructions of a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device PLD) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0122] The beneficial effects of the present invention will be described below with reference to specific embodiments.

[0123] Example 1

[0124] The experimental environment was built using the Ubuntu 22.04 operating system and the PyTorch deep learning framework. The PyTorch version was 2.1.2, the CUDA version was 12.8, the Python version was 3.10.8, the CPU was an Intel Xeon Platinum 8352V, and the GPU was an NVIDIA GeForce RTX 4090.

[0125] To ensure optimal model performance, multiple experiments were conducted to verify and determine the most suitable hyperparameter configuration: the number of training iterations was set to 300, and the batch size to 16. To improve the model's convergence speed, the SGD optimizer was selected, with a learning rate of 0.01, a momentum value of 0.937, and a weight decay value of 5 × 10⁻⁶. -4 .

[0126] The experiment uses Precision to represent the percentage of true positives among correctly predicted samples; mAP@0.5 (mean accuracy) measures the model's detection accuracy across different categories, calculated as the average AP for each category when the IoU threshold is 0.5; the number of parameters reflects the model's storage complexity; and FLOPs measure the computational cost of the model during inference, serving as an evaluation metric for the model's detection performance. Figure 7-8 As shown in Table 1, the network model of this invention combines accuracy and real-time performance, and has high application value.

[0127] surface Comparison of mainstream lightweight object detection algorithms

[0128]

[0129] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. An edge-deployment-oriented lightweight deep learning fire detection method, characterized in that, The method comprises the following steps: S100, collecting image data of a fire scene, and performing enhancement processing on the collected image data to construct a diversified fire smoke image data set; S200, constructing a lightweight fire detection network model, the model comprising a backbone network, a neck network and a head detection network, the backbone network introducing a C3_Biformer module for extracting multi-scale image features, the neck network integrating an AsymptoticFPN module for interactive fusion of feature maps of different scales, and the head detection network adopting a PinwheelHead structure for improving the perception ability of the rotating target and realizing accurate detection; S300, deploying the trained fire detection network model on a UAV platform, collecting video streams in real time during flight for fire identification, and combining UAV attitude parameters and camera imaging geometric models for fire source space positioning.

2. The method of claim 1, wherein, The enhancement processing on the collected image data in S100 comprises mirror image, contrast transformation, cutting, noise superposition, random padding processing and random combination processing.

3. The method of claim 2, wherein, The C3_Biformer module in S200 is composed of a main branch and a shortcut branch; the main branch comprises a convolution layer, a convolution layer, a BiFormer attention module, a bottleneck module, a feature merging unit and an output convolution layer in sequence; the shortcut branch contains a convolution layer for linear transformation of input features, and the main branch output is fused after all calculations of the main branch are completed.

4. The method of claim 3, wherein, The function implementation process of the AsymptoticFPN module in S200 is as follows: For the backbone network, multiple scale feature maps are extracted; first, multiple scales of features are subjected to convolution enhancement processing to make them have fusion ability; Secondly, cross-layer bidirectional fusion is realized to realize multi-scale feature interaction; Finally, for the first layer to the first Feature vectors obtained by layer mapping The feature vector is output by fusion using a linear weighting method. Represented as: In the formula: , , respectively represent the spatial fusion weights of three scale features, and satisfy the normalization condition: .

5. The method of claim 4, wherein, The head detection network in S200 adopts a PinwheelHead structure composed of a regression branch and a classification branch, and the function implementation process is as follows: In the regression branch, the input features are subjected to multi-layer convolution to generate discrete distance distribution of four directions of the bounding box; then, the discrete distribution is decoded into continuous distance value by using distribution focal regression method: wherein: is an upper limit on the number of channels for the discrete regression distribution, is is the probability distribution after is a discrete distance index; Based on anchor positions and predicted distances , , , Decoding the bounding box coordinates: The classification branch adopts a rotating perception convolution module, which first performs asymmetric padding on the input features in four directions to form four groups of feature maps with directional offset; then, a group of directional convolution is performed on each directional branch to enable the network to extract differentiated features from horizontal, vertical and other bias directions; Then the outputs of the four direction convolutions are spliced in the channel dimension to form enhanced features containing multi-direction structural information, and then fused and compressed to obtain a feature map after direction enhancement; high-level semantic information is continuously extracted, and class predictions for each anchor are generated, and class confidence is obtained after activation: activation. In the formula: This represents the feature map output by the classification branch. For the process The confidence probability of each target class after activation; Finally, the outputs of the regression branch and the classification branch are spliced in the channel dimension to form the detection result of each layer feature map.

6. The method of claim 5, wherein, In step S300, the fire source space positioning is combined with the UAV attitude parameters and the camera imaging geometric model, specifically: According to the pixel coordinates of the detected flame or smoke object in the image , the normalized ray direction in the camera coordinate system is calculated: , the camera coordinate system is transformed to the body coordinate system, and the ray direction after the pose transformation is mapped to the geographic coordinate system: wherein: Rbody2worldis a rotation matrix from the body coordinate system to the geographical coordinate system, Rcam2bodyis a rotation matrix from the camera coordinate system to the body coordinate system; the position of the drone in the geographical coordinate system the position of the fire point in the geographical coordinate system is: wherein: is the ground level, is the vertical component of the ray, is the flight level.

7. The method of claim 6, wherein, Step S300 further comprises: combining the real-time attitude data of the UAV IMU to dynamically correct the attitude change of the UAV, and using an adaptive confidence weight smoothing algorithm to filter the detection positions of continuous multiple frames; the adaptive confidence weight smoothing algorithm introduces the detection confidence as a dynamic weight factor on the basis of the traditional time weighted smoothing, and is defined as follows: In the formula: is the fire source geographical position estimation value after adaptive weighted smoothing, is the length of the sliding time window, is the fire source geographical position estimation value of the first frame, is the fire source geographical coordinate of the frame, is the detection confidence of the model in the frame, is the time decay weight, is the time decay coefficient.

8. An edge-deployment oriented lightweight deep learning fire detection system, characterized in that, The system has program modules corresponding to the steps of the method of any one of claims 1-7, and when running, the steps of the above edge-oriented lightweight deep learning fire detection method are executed.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program configured to implement the steps of the edge-oriented lightweight deep learning fire detection method of any one of claims 1-7 when called by the processor.

Images