Fire accurate detection positioning system and method facing vertical rod fixed monocular vision

Abstract

The application discloses a fire accurate detection positioning system and method for fixed monocular vision of a vertical rod, and relates to the technical field of computer vision, which comprises the following steps: continuously collecting images of a monitoring area by a camera fixed at the top of the vertical rod, acquiring a continuous multi-frame monitoring image sequence, and respectively adding area identification codes to different distance areas; performing real-time fire detection on the collected images; performing space-time consistency judgment on the detection results of the continuous multi-frames; calculating the actual geographic position coordinates of a fire point; transmitting a fire information packet to a local server; and performing secondary verification on the received fire information at the local server side. The application has the advantages that, by means of partition preprocessing and an adaptive multi-resolution detection mechanism, the scale difference problem of near and far targets in a large-scale monitoring scene is effectively solved, the real-time detection capability of the edge device is maintained while being light-weighted, and the recognition precision of small-size and far-away fire is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to a precise fire detection and positioning system and method for fixed monocular vision on poles. Background Technology

[0002] Currently, outdoor fire monitoring based on computer vision mainly relies on two technical approaches: one is to use thermal imaging cameras, which can sense heat sources, but the equipment is expensive, difficult to deploy on a large scale, and is easily affected by environmental heat sources, leading to false alarms; the other is to use visible light binocular vision systems, which calculate distance through stereo vision, but the calibration process is complex, has high requirements for installation foundation and environmental stability, and also faces cost pressures.

[0003] Existing monocular vision fire detection methods still face many challenges: On the one hand, the monitoring scene has a large depth, and the scale difference between near and far targets in the image is significant. A single detection model is difficult to maintain both high accuracy and real-time performance, which can easily lead to missed detection of small or blurry fires. On the other hand, existing methods mostly rely on single-frame image analysis, which is susceptible to false alarms due to changes in light and shadow and interference from similar colored objects. Moreover, most systems lack accurate geolocation and multiple verification mechanisms, making it difficult to meet the need for accurate acquisition of key information such as the location and extent of fire points in emergency response, thus limiting their reliability in actual deployment. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a precise fire detection and positioning system and method based on fixed monocular vision for poles. This technical solution resolves at least one of the technical problems mentioned in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for accurate fire detection and location using fixed monocular vision on poles includes: The camera, fixed to the top of the pole, continuously captures images of the monitored area, obtains a sequence of multiple consecutive monitoring images, preprocesses the monitoring images according to their distance from the camera, and adds area identification codes to areas at different distances. Using a lightweight deep learning model deployed on an edge computing device, real-time fire detection is performed on the collected images, and the detection results, including suspected fire areas and confidence levels, are output. The lightweight deep learning model calls different resolution branches to identify and predict the monitored images based on the area identification code. The spatiotemporal consistency of the detection results in multiple consecutive frames is judged, and if the set conditions are met, it is confirmed as a valid fire event. Based on the known height of the pole, camera intrinsic parameters, and the location of the effective fire event in the image, the actual geographical coordinates of the fire point are calculated using a geometric model. Fire information packets containing time, geographical location, and image data are encrypted and transmitted to the local server via a local area network. On the local server, a different algorithm than that used on the edge computing device is used to perform secondary verification of the received fire information, and a decision is made on whether to trigger an alarm based on the verification result.

[0006] Preferably, the lightweight deep learning model is trained based on the YOLOv8 architecture. The lightweight deep learning model sets up a P2 high-resolution detection branch, a P3 detection branch corresponding to traditional 8× downsampling, a P4 detection branch corresponding to traditional 16× downsampling, and a P5 detection branch corresponding to traditional 32× downsampling in the neck network. For the image region with the remote area identification code, the P2 high-resolution detection branch, the P4 detection branch, and the P5 detection branch are used for prediction and recognition. For image regions with near-field identification codes, P3, P4, and P5 detection branches are used for prediction and recognition.

[0007] Preferably, the feature map construction process of the P2 high-resolution detection branch is as follows: The traditional 8× downsampled P3 layer features output by the backbone network are upsampled by 2× and then concatenated with the traditional 4× downsampled P2 layer shallow features output by the backbone network to construct a high-resolution feature map that retains more fine-grained texture.

[0008] Preferably, the feature map construction process of the P2 high-resolution detection branch also includes calibrating the stitched high-resolution feature map through a two-dimensional space-channel interactive calibration mechanism.

[0009] Preferably, the step of performing a spatiotemporal consistency judgment on the detection results of multiple consecutive frames, and confirming a valid fire event if the set conditions are met, specifically includes: When a suspected fire area appears in a certain frame of an image, the frame and at least two consecutive frames acquired after it are cached as an image sequence for spatiotemporal consistency judgment. If all adjacent images in the spatiotemporal consistency judgment image sequence satisfy the condition that the overlap of the suspected fire area is higher than the overlap threshold and the confidence level of each frame image is higher than the confidence threshold, then it is judged as a valid fire event.

[0010] Preferably, the step of calculating the actual geographical coordinates of the fire point using a geometric model based on the known height of the pole, camera intrinsic parameters, and the location of the effective fire event in the image specifically includes: Capture camera pixel size and focal length f; Establish a ground projection coordinate system with the center of the bottom of the pole as the origin O, and the X-axis pointing east, the Y-axis pointing north, and the Z-axis pointing upward; Let the perpendicular distance from the optical center C of the camera to the origin O be... ; Record the latitude and longitude of point C. , Here are the longitude coordinates of point C. Here are the latitude coordinates of point C; The center pixel of the suspected fire area of ​​an effective fire event Convert to ground projection coordinates , The X-axis coordinate of the center of the suspected fire area in the ground projection coordinate system for a valid fire event. The Y-axis coordinate of the center of the suspected fire area of ​​an effective fire event in the ground projection coordinate system; Combining the latitude and longitude of point C The ENU→LLA algorithm is used to... Convert to fire point latitude and longitude ( , ), The longitude coordinates of the center of the suspected fire area for an effective fire incident. Latitude coordinates of the center of the suspected fire area of ​​a valid fire incident; The same method is used to calculate the latitude and longitude of the four corner points of the suspected fire area of ​​the effective fire event, and the suspected fire area of ​​the effective fire event is calculated based on the latitude and longitude of the four corner points.

[0011] Furthermore, a precise fire detection and positioning system based on fixed-pole monocular vision is proposed to achieve the aforementioned precise fire detection and positioning method based on fixed-pole monocular vision, including: The image acquisition and preprocessing module is used to continuously acquire images of the monitoring area through a camera fixed to the top of the pole, divide the acquired monitoring images into regions according to their distance from the camera, and attach region identification codes to regions at different distances. The edge computing and fire detection module is deployed on the front-end edge computing device and has the lightweight deep learning model built in. It is used to receive the image with regional identification code output by the image acquisition and preprocessing module, call the resolution branch corresponding to the regional identification code to perform real-time fire identification, output the detection result containing the suspected fire area and confidence level, and perform spatiotemporal consistency judgment on the detection results of multiple consecutive frames to confirm the valid fire event. The positioning calculation module is used to calculate the actual geographical coordinates of the fire point through a geometric model based on the known height of the pole, camera intrinsic parameters, and the location of the valid fire event confirmed by the edge calculation and fire detection module in the image. The encrypted communication module is used to transmit fire information packets containing time, geographical location, and image data to the local server via a local area network in an encrypted manner. The local server and secondary verification module are used to receive the fire information packet sent by the encrypted communication module, perform secondary verification using a different algorithm than that on the edge computing device, and decide whether to trigger an alarm based on the verification result.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention effectively addresses the scale difference between near and far targets in large-scale monitoring scenarios through partitioned preprocessing and an adaptive multi-resolution detection mechanism. While maintaining the lightweight real-time detection capabilities of edge devices, it significantly improves the identification accuracy of small-sized and distant fires. Furthermore, it filters out transient interference through spatiotemporal consistency judgment and achieves accurate geographic coordinates and area estimation of fire points by combining a monocular geometric model. Finally, through secondary verification using heterogeneous algorithms, the false alarm rate is systematically reduced, ultimately constructing an automated fire early warning system with accurate detection, reliable positioning, and timely response. Attached Figure Description

[0013] Figure 1 This is a flowchart of the precise fire detection and positioning method based on fixed monocular vision on a pole, as proposed in Embodiment 2 of the present invention. Figure 2 This is a diagram of the lightweight deep learning model architecture proposed in Embodiment 3 of the present invention; Figure 3 This is a flowchart illustrating the steps of the two-dimensional space-channel interaction mechanism proposed in Embodiment 3 of the present invention. Figure 4 This is a flowchart illustrating the steps for calculating the actual geographical coordinates of the fire point as proposed in Embodiment 4 of the present invention. Detailed Implementation

[0014] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0015] Example 1 This embodiment proposes a precise fire detection and positioning system based on fixed monocular vision for upright poles, including: The image acquisition and preprocessing module is used to continuously acquire images of the monitoring area through a camera fixed to the top of the pole, divide the acquired monitoring images into regions according to their distance from the camera, and attach region identification codes to regions at different distances. The edge computing and fire detection module is deployed on the front-end edge computing device. It has a built-in lightweight deep learning model to receive images with regional identification codes output by the image acquisition and preprocessing module, call the resolution branch corresponding to the regional identification code to perform real-time fire identification, output detection results including suspected fire areas and confidence levels, and perform spatiotemporal consistency judgment on the detection results of multiple consecutive frames to confirm valid fire events. The positioning calculation module is used to calculate the actual geographical coordinates of the fire point in the image based on the known height of the pole, camera intrinsic parameters, and the location of valid fire events confirmed by the edge calculation and fire detection module through a geometric model. The encrypted communication module is used to transmit fire information packets containing time, geographical location, and image data to the local server via a local area network in an encrypted manner. The local server and secondary verification module are used to receive fire information packets sent by the encrypted communication module, perform secondary verification using a different algorithm than that on the edge computing device, and decide whether to trigger an alarm based on the verification result.

[0016] The image acquisition and preprocessing module, the edge computing and fire detection module, and the positioning computing module are deployed as edge-side components within the camera. The camera is a 1920×1080 RGB camera, employing a 1 / 2.8″ low-light sensor with a pixel size of 2.9 µm, a lens focal length of f=6 mm, a horizontal field of view of 73°-77°, and a vertical field of view of 10°-45°, ensuring a ground projection coverage width of 46-183 m for a distance of 30-120 m. The camera also includes a built-in processor and network communication module.

[0017] In a specific implementation case, the specific equipment parameters of the above system are as follows: Edge-end features: Custom 1920×1080 RGB camera with IMX335 sensor, 6mm lens, IP66 housing; Processor: Rockchip RK3566, quad-core A55 1.8 GHz, built-in 1 TOPS NPU; Network communication module: Gigabit RJ45 + 802.11ac Wi-Fi; Power supply module: 24 V / 60 W solar power + 20 Ah lithium battery, providing 7 days of battery life in cloudy or rainy weather. The local server has an Intel i5-13400 CPU, 16 GB of RAM, and comes pre-installed with Ubuntu 22.04, Docker Compose, InfluxDB 2.7, and Grafana 10.

[0018] In specific deployment: the pole is a 20 m hot-dip galvanized steel pipe, the foundation is a 1.2 m × 1.2 m × 1.5 m concrete, the top of the camera is tilted downwards at 12° when it is installed, the main point is aligned with the monitoring center, the calibration tool is a total station, and the lens distortion is calibrated using a checkerboard pattern by Zhang Zhengyou.

[0019] Example 2 This embodiment further develops a fire detection and positioning method adapted to the fixed-pole monocular vision fire accuracy detection and positioning system proposed in Embodiment 1, referring to... Figure 1 As shown, it includes: By using a camera fixed to the top of a pole, images of the monitored area are continuously acquired, resulting in a sequence of multiple consecutive monitoring images. The monitored images are preprocessed according to their distance from the camera, and region identification codes are added to regions at different distances. By using a preset distance threshold, the images are logically divided into different depth levels, and each region is labeled with a different identification code. This guides the subsequent detection model to adaptively call detection branches of different complexities for targets at different distances, thereby optimizing the allocation of computing resources. By utilizing a lightweight deep learning model deployed on edge computing devices, real-time fire detection is performed on acquired images. The output includes detection results containing suspected fire areas and confidence levels. The lightweight deep learning model calls different resolution branches to identify and predict the monitored images based on the region identifier code. The model can intelligently select different resolution branches for prediction based on the "region identifier code" assigned in the preprocessing stage. For example, a high-resolution branch is called for distant small target areas to capture details, while a standard branch is called for nearby large target areas to improve speed. Spatiotemporal consistency of detection results in multiple consecutive frames is judged. If the set conditions are met, it is confirmed as a valid fire event. By analyzing the existence of suspected fires in multiple consecutive frames of images, it is determined whether they are real continuous fires or transient interference. Based on the known height of the pole, camera intrinsic parameters, and the location of the effective fire event in the image, the actual geographical coordinates of the fire point are calculated through a geometric model. Using the known camera installation height, intrinsic parameters, and pixel coordinates of the fire in the image, the actual ground projection coordinates of the fire point relative to the pole are calculated through a perspective projection geometric model. This method transforms visual detection results into specific geographical location information usable by rescue forces, completing the process from perception to location. Fire information packets containing time, geographic location, and image data are encrypted and transmitted to a local server via a local area network (LAN). Structured fire information obtained from front-end processing is also packaged and transmitted via an encrypted LAN. Encryption ensures data security during transmission, preventing tampering or leakage. Transmission to the local server provides a foundation for subsequent, more complex verification, analysis, and decision-making. On the local server, a different algorithm than that used on the edge computing device is employed to perform secondary verification of the received fire information. Based on the verification results, the server determines whether to trigger an alarm. Specifically, the local server uses background subtraction and an HSV flame color model to perform secondary detection on the image data in the received fire information packet. If verification passes, it is marked as confirmed and an audible and visual alarm is triggered. If verification fails, it is marked as a false alarm, logged, and no alarm is triggered. This heterogeneous verification mechanism avoids the inherent biases or misjudgments of edge models in certain scenarios. The final alarm triggering is carefully decided by the server based on the dual verification results, thereby maximizing the accuracy of the alarm and improving the overall system reliability.

[0020] Specifically, the monitored images are preprocessed according to their distance from the camera, and region identification codes are added to regions at different distances. Set a zone distance threshold, add a far zone identifier code to the area in the image that exceeds the zone distance threshold, and add a near zone identifier code to the area in the image that is below the zone distance threshold. For example, in the specific implementation case of Example 1, it is necessary to realize fire monitoring within a ground area of ​​30-120 m. Set 90 m as the zone distance threshold, divide the 30-90 m range into the near zone, add a near zone identifier code to this area in the image, and divide the 90-120 m range into the far zone, add a far zone identifier code to this area in the image. The partition distance threshold can be set freely according to the actual situation during implementation. Generally, for areas with complex environments, such as forests with a lot of obstruction, the partition distance threshold is set small in order to enable high-resolution fire monitoring and identification over a larger area. For areas with simple environments, such as grasslands with less obstruction, the partition distance threshold is set large in order to reduce the computing load on identification and improve the identification speed.

[0021] This embodiment further proposes that the spatiotemporal consistency of the detection results of multiple consecutive frames be judged, and if the set conditions are met, the specific steps for confirming it as a valid fire event include: When a suspected fire area appears in a certain frame of an image, the frame and at least two consecutive frames acquired after it are cached as an image sequence for spatiotemporal consistency judgment. If all adjacent images in the spatiotemporal consistency judgment image sequence satisfy the condition that the overlap of the suspected fire area is higher than the overlap threshold and the confidence level of each frame image is higher than the confidence threshold, then it is judged as a valid fire event.

[0022] This system leverages the prior physical knowledge that fire events should possess temporal and spatial continuity. Specifically, when a suspected target is detected in a frame, the system does not immediately trigger an alarm. Instead, it initiates a verification window based on a short time sequence, requiring at least three consecutive frames. During verification, the overlap of the bounding boxes of suspected targets between adjacent frames is compared, using methods such as Intersection over Union (IoU), to ensure the continuous and reasonable movement of the target's position and to eliminate interference from momentarily appearing objects of similar color. Simultaneously, by checking that the detection confidence level of each frame remains above a high threshold, the stability of the observed features is ensured, eliminating confidence level fluctuations caused by light and shadow flickering. Only target sequences that simultaneously satisfy spatial continuity and confidence level stability are determined to be real, continuously burning, and valid fire events, thereby greatly improving the reliability of the front-end sensing results. By introducing a low-computational-cost spatiotemporal consistency verification mechanism, the system is provided with a powerful false alarm filtering capability without significantly increasing the burden on edge devices. It effectively filters out transient, isolated false detections caused by birds, vehicle lights, reflections, etc., basing alarm decisions on continuous and stable observational evidence, thereby significantly reducing the system's false alarm rate. Simultaneously, this mechanism integrates detection results from discrete image-level outputs into continuous events with a time dimension, providing a more reliable data foundation that better reflects the actual fire development patterns for subsequent precise positioning and coordinated response, enhancing the practicality and reliability of the entire early warning system.

[0023] Example 3 This embodiment, based on Embodiment 2, provides a detailed optimization design for the lightweight deep learning model, referring to... Figure 2 As shown, the lightweight deep learning model is trained based on the YOLOv8 architecture. The lightweight deep learning model sets up a P2 high-resolution detection branch, a P3 detection branch corresponding to traditional 8× downsampling, a P4 detection branch corresponding to traditional 16× downsampling, and a P5 detection branch corresponding to traditional 32× downsampling in the neck network. For the image region with the remote area identification code, the P2 high-resolution detection branch, the P4 detection branch, and the P5 detection branch are used for prediction and recognition. For image regions with near-field identification codes, P3, P4, and P5 detection branches are used for prediction and recognition.

[0024] In long-distance fire detection tasks, the small scale of the target is a core factor leading to missed detections. In YOLOv8s, its feature pyramid mainly utilizes P3 (8× downsampling), P4 (16×), and P5 (32×) layers for prediction. For long-distance fires commonly found in the ExDark dataset, after 8x downsampling by the P3 layer, the remaining pixels on the feature map are extremely small. This drastic spatial information compression makes it easy to lose weak, small target features. To address this issue, a P2 high-resolution detection branch is constructed in the Neck part. This branch constructs a high-resolution feature map that retains more fine-grained texture by upsampling the P3 layer features by 2× and concatenating them with the shallow P2 layer features (4× downsampling) output from the backbone network. However, introducing the P2 branch is a double-edged sword. While it restores spatial details, it also introduces a large amount of high-frequency background noise that is not filtered by the deep network. If the original P2 features are detected directly, this noise is easily misdetected as a fire. After the P2 features are fused with the upsampled P3 features, the resulting features are... Figure 1 On the one hand, this approach can provide more spatial details for small targets; on the other hand, it can easily introduce shot noise, false edge responses, and irrelevant high-frequency textures from low-light scenes into the detection process, thereby increasing the risk of false detection. Based on this, a two-dimensional space-channel interaction module is introduced after the high-resolution fusion features. This module aims to further recalibrate the shallow high-resolution features without introducing complex operators such as self-attention, suppress background responses irrelevant to the target, and enhance the structural information more crucial for the localization and recognition of small targets. This achieves refinement in the spatial domain, filtering out background interference while preserving high-resolution details. Reference Figure 3 As shown, the implementation process of the two-dimensional space-channel interaction module is as follows: The constructed high-resolution feature map that retains more fine-grained texture is denoted as feature map X. , B represents the tensor, C represents the number of channels, H represents the number of pixels or feature points in the vertical direction of the feature map, and W represents the number of pixels or feature points in the horizontal direction of the feature map. The feature map X is divided into two groups of sub-features according to the channel dimension. and , , ; For two sets of sub-features and ,use right Forward update is performed, and its expression is: ; Then use the updated Reverse calibration Its expression is: ; in, For the updated sub-features , For the updated sub-features , This represents element-wise multiplication. This represents the Sigmoid activation function. and Responsible for generating gated scaling items, and Provides supplementary offset information. , , and All are composed of the same BasicConv2d. , , and All four transformation functions maintain the spatial and channel dimensions of the features invariant, meaning their mapping relationships strictly follow... Each branch contains Conv2d and BatchNorm2d. Meanwhile, the module uses the Sigmoid function to generate a gated scaling factor, replacing the more aggressive exponential scaling transformation, thereby improving numerical stability during the training phase while maintaining nonlinear expressive power. Unlike unidirectional modulation, this cascaded interactive strategy of updating one set of features first and then reversing the calibration of the other set can reduce the asymmetry of information flow, allowing the two sets of features to gradually complete joint optimization during complementary modeling. For high-resolution features in remote fire detection, this process helps to reduce the interference of random noise and invalid background textures on subsequent detection tasks while preserving fine-grained texture information. Finally, the two sets of updated sub-features are... and The feature maps are re-stitched along the channel dimension to obtain the calibrated and updated feature maps. .

[0025] The aforementioned two-dimensional space-channel interaction mechanism does not simply re-convolve the original feature map, but rather explicitly recalibrates the high-resolution fused features through channel splitting, gated update, reverse calibration, and re-fusion. Its effectiveness is mainly reflected in two aspects: First, channel splitting allows the network to learn more complementary representations in the two sub-feature spaces, reducing redundant responses in shallow features; second, bidirectional gated update allows one set of features to be reweighted and supplemented with offset information under the conditional constraints of another set of features, thereby enhancing the target region response and suppressing background interference. For typical scenarios such as long-distance fire detection—characterized by weak target signals, strong background noise, rich shallow details, but insufficient semantics—it can more effectively improve the discriminative power of high-resolution features.

[0026] Example 4 This embodiment, based on Embodiment 2 or Embodiment 3, further proposes a specific calculation method for calculating the actual geographical coordinates of a fire point using a geometric model, based on the known height of the pole, camera intrinsic parameters, and the location of the effective fire event in the image. (Refer to...) Figure 4 As shown, this includes: the pixel size of the acquisition camera. and focal length f; Establish a ground projection coordinate system with the center of the bottom of the pole as the origin O, and the X-axis pointing east, the Y-axis pointing north, and the Z-axis pointing upward; Let the perpendicular distance from the optical center C of the camera to the origin O be... ; Record the latitude and longitude of point C. , Here are the longitude coordinates of point C. Here are the latitude coordinates of point C; The center pixel of the suspected fire area of ​​an effective fire event Convert to ground projection coordinates , The X-axis coordinate of the center of the suspected fire area in the ground projection coordinate system for a valid fire event. The Y-axis coordinate of the center of the suspected fire area of ​​an effective fire event in the ground projection coordinate system; The conversion formula is: ; ; ; Combining the latitude and longitude of point C The ENU→LLA algorithm is used to... Convert to fire point latitude and longitude ( , ), The longitude coordinates of the center of the suspected fire area for an effective fire incident. Latitude coordinates of the center of the suspected fire area of ​​a valid fire incident; The same method is used to calculate the latitude and longitude of the four corner points of the suspected fire area of ​​the effective fire event, and the suspected fire area of ​​the effective fire event is calculated based on the latitude and longitude of the four corner points.

[0027] Using known camera intrinsic parameters and installation height, the pixel coordinates of the fire point in the image can be calculated to coordinates in a local ground coordinate system with the base of the pole as the origin through perspective projection. This step essentially utilizes the principle of similar triangles to eliminate geometric distortion introduced by camera imaging, projecting the two-dimensional image points back onto the actual three-dimensional ground. Subsequently, using the standard geodetic coordinate transformation algorithm ENU→LLA, the local coordinates are combined with the known latitude and longitude of the camera's optical center to calculate the absolute geographic coordinates of the fire point. Furthermore, this method further estimates the approximate surface projection area of ​​the fire by calculating the latitude and longitude of the four corner points of the suspected fire area, providing richer information about the fire situation than a single center point. By employing the aforementioned methods for calculating the location and area of ​​a fire, the system outputs not just abstract image alarms, but precise geographic information with direct action guidance value. Its core benefit lies in achieving low-cost, high-precision monocular visual geolocation. Using only a fixed monocular camera and known installation parameters, without relying on expensive equipment such as binoculars or lidar, the system can calculate the latitude and longitude coordinates of the fire point and even estimate the burned area, greatly enhancing the practical value of the system output. This provides the fire command center with precise location and preliminary scale assessment of the fire, enabling rescue forces to quickly and accurately locate the fire scene, formulate firefighting routes, significantly shorten emergency response time, and improve the efficiency and scientific rigor of the entire early warning and response process.

[0028] In summary, the advantages of this invention are as follows: By employing partitioned preprocessing and an adaptive multi-resolution detection mechanism, it effectively addresses the scale difference between near and far targets in large-scale monitoring scenarios. While maintaining the lightweight real-time detection capabilities of edge devices, it significantly improves the identification accuracy of small-sized and distant fires. Furthermore, it filters out transient interference through spatiotemporal consistency judgment and achieves accurate geographic coordinates and area estimation of fire points by combining a monocular geometric model. Finally, through secondary verification using heterogeneous algorithms, the false alarm rate is systematically reduced, ultimately constructing an automated fire early warning system that is accurate in detection, reliable in positioning, and timely in response.

[0029] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for precise fire detection and positioning using a fixed monocular vision system for upright poles, characterized in that, include: The camera, fixed to the top of the pole, continuously captures images of the monitored area, obtains a sequence of multiple consecutive monitoring images, preprocesses the monitoring images according to their distance from the camera, and adds area identification codes to areas at different distances. Using a lightweight deep learning model deployed on an edge computing device, real-time fire detection is performed on the collected images, and the detection results, including suspected fire areas and confidence levels, are output. The lightweight deep learning model calls different resolution branches to identify and predict the monitored images based on the area identification code. The spatiotemporal consistency of the detection results in multiple consecutive frames is judged, and if the set conditions are met, it is confirmed as a valid fire event. Based on the known height of the pole, camera intrinsic parameters, and the location of the effective fire event in the image, the actual geographical coordinates of the fire point are calculated using a geometric model. Fire information packets containing time, geographical location, and image data are encrypted and transmitted to the local server via a local area network. On the local server, a different algorithm than that used on the edge computing device is used to perform secondary verification of the received fire information, and a decision is made on whether to trigger an alarm based on the verification result.

2. The method for accurate fire detection and positioning using fixed monocular vision on a pole, as described in claim 1, is characterized in that... The process of preprocessing the monitored images according to their distance from the camera, and attaching region identification codes to regions at different distances, specifically involves: Set a partition distance threshold, add a far zone identifier code to the area of ​​the image that exceeds the partition distance threshold, and add a near zone identifier code to the area of ​​the image that is below the partition distance threshold.

3. The method for accurate fire detection and positioning using fixed monocular vision on a pole, as described in claim 2, is characterized in that... The lightweight deep learning model is trained based on the YOLOv8 architecture. The lightweight deep learning model sets up a P2 high-resolution detection branch, a P3 detection branch corresponding to traditional 8× downsampling, a P4 detection branch corresponding to traditional 16× downsampling, and a P5 detection branch corresponding to traditional 32× downsampling in the neck network. For the image region with the remote area identification code, the P2 high-resolution detection branch, the P4 detection branch, and the P5 detection branch are used for prediction and recognition. For image regions with near-field identification codes, P3, P4, and P5 detection branches are used for prediction and recognition.

4. The method for accurate fire detection and positioning using fixed monocular vision on a pole, as described in claim 3, is characterized in that... The feature map construction process of the P2 high-resolution detection branch is as follows: The traditional 8× downsampled P3 layer features output by the backbone network are upsampled by 2× and then concatenated with the traditional 4× downsampled P2 layer shallow features output by the backbone network to construct a high-resolution feature map that retains more fine-grained texture.

5. The method for accurate fire detection and positioning using fixed monocular vision on a pole, as described in claim 4, is characterized in that... The feature map construction process of the P2 high-resolution detection branch also includes calibrating the stitched high-resolution feature map through a two-dimensional space-channel interactive calibration mechanism, which specifically includes: The constructed high-resolution feature map that retains more fine-grained texture is denoted as feature map X. , B represents the tensor, C represents the number of channels, H represents the number of pixels or feature points in the vertical direction of the feature map, and W represents the number of pixels or feature points in the horizontal direction of the feature map. The feature map X is divided into two groups of sub-features according to the channel dimension. and , , ; For two sets of sub-features and ,use right Forward update is performed, and its expression is: ； Then use the updated Reverse calibration Its expression is: ； in, For the updated sub-features , For the updated sub-features , This represents element-wise multiplication. This represents the Sigmoid activation function. and Responsible for generating gated scaling items, and Provides supplementary offset information. , , and All are composed of the same BasicConv2d; Finally, the two sets of updated sub-features are... and The feature maps are re-stitched along the channel dimension to obtain the calibrated and updated feature maps. .

6. The method for accurate fire detection and positioning using fixed monocular vision on a pole, as described in claim 5, is characterized in that... The step of performing a spatiotemporal consistency judgment on the detection results of multiple consecutive frames, and confirming a valid fire event if the set conditions are met, specifically includes: When a suspected fire area appears in a certain frame of an image, the frame and at least two consecutive frames acquired after it are cached as an image sequence for spatiotemporal consistency judgment. If all adjacent images in the spatiotemporal consistency judgment image sequence satisfy the condition that the overlap of the suspected fire area is higher than the overlap threshold and the confidence level of each frame image is higher than the confidence threshold, then it is judged as a valid fire event.

7. The method for accurate fire detection and positioning using fixed monocular vision on a pole, as described in claim 6, is characterized in that... The calculation of the actual geographical coordinates of the fire point using a geometric model, based on the known height of the pole, camera intrinsic parameters, and the location of the effective fire event in the image, specifically includes: Capture camera pixel size and focal length f; Establish a ground projection coordinate system with the center of the bottom of the pole as the origin O, and the X-axis pointing east, the Y-axis pointing north, and the Z-axis pointing upward; Let the perpendicular distance from the optical center C of the camera to the origin O be... ; Record the latitude and longitude of point C. , Here are the longitude coordinates of point C. Here are the latitude coordinates of point C; The center pixel of the suspected fire area of ​​an effective fire event Convert to ground projection coordinates , The X-axis coordinate of the center of the suspected fire area in the ground projection coordinate system for a valid fire event. The Y-axis coordinate of the center of the suspected fire area of ​​an effective fire event in the ground projection coordinate system; Combining the latitude and longitude of point C The ENU→LLA algorithm is used to... Convert to fire point latitude and longitude ( , ), The longitude coordinates of the center of the suspected fire area for an effective fire incident. Latitude coordinates of the center of the suspected fire area of ​​a valid fire incident; The same method is used to calculate the latitude and longitude of the four corner points of the suspected fire area of ​​the effective fire event, and the suspected fire area of ​​the effective fire event is calculated based on the latitude and longitude of the four corner points.

8. The method for accurate fire detection and positioning using fixed monocular vision on a pole, as described in claim 7, is characterized in that... The process of performing secondary verification of the received fire information on the local server side using an algorithm different from that on the edge computing device, and determining whether to trigger an alarm based on the verification result, specifically includes: The background difference + HSV flame color model is used to perform secondary detection on the image data in the received fire information packet. If the verification is successful, it is marked as confirmed and an audible and visual alarm is triggered. If the verification fails, it is marked as a false alarm, the log is saved, and no alarm is triggered.

9. A fire detection and positioning system based on fixed monocular vision for poles, used to implement the fire detection and positioning method based on fixed monocular vision for poles as described in any one of claims 1-8, characterized in that, include: The image acquisition and preprocessing module is used to continuously acquire images of the monitoring area through a camera fixed to the top of the pole, divide the acquired monitoring images into regions according to their distance from the camera, and attach region identification codes to regions at different distances. The edge computing and fire detection module is deployed on the front-end edge computing device and has the lightweight deep learning model built in. It is used to receive the image with regional identification code output by the image acquisition and preprocessing module, call the resolution branch corresponding to the regional identification code to perform real-time fire identification, output the detection result containing the suspected fire area and confidence level, and perform spatiotemporal consistency judgment on the detection results of multiple consecutive frames to confirm the valid fire event. The positioning calculation module is used to calculate the actual geographical coordinates of the fire point through a geometric model based on the known height of the pole, camera intrinsic parameters, and the location of the valid fire event confirmed by the edge calculation and fire detection module in the image. The encrypted communication module is used to transmit fire information packets containing time, geographical location, and image data to the local server via a local area network in an encrypted manner. The local server and secondary verification module are used to receive the fire information packet sent by the encrypted communication module, perform secondary verification using a different algorithm than that on the edge computing device, and decide whether to trigger an alarm based on the verification result.

10. The precise fire detection and positioning system with fixed monocular vision for poles as described in claim 9, characterized in that, The camera fixed to the top of the pole is a 1920×1080 RGB camera. The camera uses a 1 / 2.8″ low-light sensor with a pixel size of 2.9 µm, a lens focal length of f=6 mm, a horizontal field of view of 73°-77°, and a vertical field of view of 10°-45°, ensuring a ground projection coverage width of 46-183 m for a distance of 30-120 m.

Images