A method and system for detecting hidden fire points based on UAV thermal imaging

By separating the low-frequency temperature base component and the high-frequency radiation differential component, and combining the second-order curvature characteristics and radiation power changes, geometric correction is performed using a digital surface model. This solves the problem of identifying and locating hidden fire points in complex surface environments, achieving higher detection accuracy and stability.

CN123118749APending Publication Date: 2026-11-17SHUHUI QIANKUN TECH CO LTD
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611580467.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-13
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively identify and locate hidden fire points in complex terrain environments. Conventional infrared image detection methods are prone to false alarms or missed detections and have difficulty distinguishing between targets that continuously emit heat and those that accumulate heat.

Method used

By separating the low-frequency temperature base component and the high-frequency radiation differential component, weak thermal anomalies are identified using second-order curvature features and neighborhood comparison information. Geometric correction is then performed by combining radiation power changes and digital surface models to extract the three-dimensional geographic coordinates of hidden smoldering points.

Benefits of technology

It improves the accuracy of identifying and locating hidden fire points, reduces false alarms and missed detections, and enhances the detection capability in complex surface environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118749A_ABST
    Figure CN123118749A_ABST
Patent Text Reader

Abstract

This invention relates to the field of infrared image pattern recognition technology, and discloses a method and system for detecting hidden fire points based on UAV thermal imaging. The method includes: acquiring a geometrically registered thermal image sequence and generating an enhanced thermal image; calculating the second-order partial derivatives of pixels to construct the Hessian matrix and solve for the thermal seepage anisotropy index, and combining this with the contrast activation response to screen candidate fire point masks; spatially integrating the net radiative heat flux density within the mask to solve for the instantaneous total radiative power, and confirming hidden smoldering fire points based on the temporal derivatives of the endogenous heat release power and the total radiative power; and calculating the three-dimensional geographic coordinates of the intersection point between the fire point centroid ray vector and the digital surface model. This invention characterizes the thermal resistance physical properties as anisotropic morphological features of the image, and, in conjunction with a radiation temporal discrimination mechanism, eliminates heat storage interference, effectively improving the signal-to-noise ratio for hidden fire point identification.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of infrared image pattern recognition technology, and in particular relates to a method and system for detecting hidden fire points based on UAV thermal imaging. Background Technology

[0002] Using drones equipped with thermal imaging devices to collect surface thermal radiation images and identify abnormally high-temperature areas from the images is a common method in fire patrols. In areas with complex surface cover, such as forests and grasslands, hidden smoldering fires may be located in thick layers of dead leaves or soil. The heat released by the fire is affected by thermal resistance when passing through the cover layer and diffuses to the surrounding area during conduction. The temperature rise after reaching the surface is usually weak. Because the pores and microcracks inside the cover medium are not evenly distributed, the heat transferred to the surface often appears in infrared images as unevenly distributed thermal textures extending along the direction of microcracks. In contrast, exposed rocks, dry sand, and other ground features that do not generate heat themselves can also form local high-temperature areas after being exposed to sunlight, but their temperature distribution in images is usually more uniform and appears as clumps.

[0003] Conventional infrared image detection methods typically process image pixels as two-dimensional intensity data, relying primarily on absolute temperature thresholds to segment targets and using Gaussian smoothing filters to reduce noise. These methods mainly judge based on pixel radiance. When lowering the segmentation threshold to detect weak surface temperature rises, large areas of exposed rock and other high-temperature features heated by sunlight are easily misidentified as fire sources, resulting in numerous false alarms. Conversely, when raising the segmentation threshold to reduce such high-temperature interference, faint thermal textures formed by hidden fire points beneath the overburden layer are easily drowned out by noise, leading to missed detections. Furthermore, conventional smoothing filters may weaken the fine thermal texture features distributed along media fissures. Relying solely on a single frame of infrared image makes it difficult to distinguish between active fire sources continuously releasing heat and those exposed to sunlight. For heat-storing ground features that gradually cool down after heating, in addition to the difficulty for inspection drones to maintain absolute stability of the physical line of sight solely through mechanical gimbals under sudden wind disturbances in the field, existing fire point detection and location methods also have deficiencies at the data processing level. For example, Chinese invention patent with authorization announcement number CN111982291B discloses a fire point location method, device and system based on drones. It segments overheated fire points in thermal images based on preset temperature thresholds and completes geographic coordinate conversion by combining drone attitude and digital elevation model. However, this scheme takes the absolute exceedance of high temperature as a preset premise for the establishment of fire points, which is difficult to cope with the slight temperature rise after the thermal resistance of the forest cover layer decays and the anisotropic crack texture. Under the interference of solar heat storage, it is very easy to miss detection or false alarm.

[0004] Therefore, the technical problem to be solved by this invention is how to identify weak thermal anomalies by utilizing the second-order curvature features of the image and neighborhood contrast information at different scales during the infrared image detection process, and how to distinguish between continuously exothermic targets and heat storage interference by combining the change of radiation power over time, while using digital surface models to perform geometric correction on the fire point location, thereby improving the identification and three-dimensional positioning capabilities of hidden fire points in complex surface cover environments. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A method for detecting hidden fire points based on UAV thermal imaging, comprising the following steps:

[0006] Step S1: Obtain the geometric registration thermal image sequence after multi-source sensor registration; separate the low-frequency temperature substrate component and the high-frequency radiation differential component, adjust the gain coefficient according to the low-frequency local variance, and superimpose to generate an enhanced thermal image;

[0007] Step S2: Calculate the two-dimensional second-order partial derivatives of the pixels in the enhanced thermal image to construct the Hessian matrix, decompose to obtain the maximum and minimum absolute eigenvalues ​​and calculate the thermal anisotropy index; use the concentric ring difference operator to calculate the contrast activation response, and generate candidate fire point connected domain masks when the contrast activation response is positive and the thermal anisotropy index is not less than the preset cutoff threshold.

[0008] Step S3: Calculate the net radiative heat flux density of the pixels within the connected domain mask of the candidate fire point and perform spatial integration to solve for the instantaneous total radiative power; extract the temporal derivative of the instantaneous total radiative power; confirm the hidden smoldering fire point when the endogenous heat release power inverted based on the instantaneous total radiative power is not less than the preset critical threshold and the temporal derivative is not less than the preset continuous heat release threshold; filter out the heat storage false target when the temporal derivative is less than the preset decay threshold.

[0009] Step S4: Based on the airborne smooth attitude matrix and the intrinsic parameter matrix of the thermal imaging sensor, the centroid pixel coordinates of the concealed smoldering fire point are converted into spatial ray unit vectors. The intersection coordinates of the spatial ray unit vectors and the digital surface model of the ground are calculated, and the three-dimensional geographic coordinates of the ground are output.

[0010] Preferably, in step S1, the geometric registration thermal image sequence obtained after multi-source sensor registration is superimposed to generate an enhanced thermal image, including: step S11, uniformly accessing the hardware time reference source, using the line exposure trigger signal of the thermal imaging sensor as a reference, performing cubic Hermite polynomial interpolation on the airborne positioning coordinates and three-axis attitude angles, combining the intrinsic parameter matrix of the thermal imaging sensor and the rigid body mounting lever arm vector, mapping the discrete infrared field of view to the tangent plane of the reference geodetic coordinate system, and generating a geometric registration thermal image sequence; step S12, using anisotropic guided edge-preserving filtering based on structural tensor to separate the low-frequency temperature base component and the high-frequency radiation differential component; statistically analyzing the local temperature standard deviation of the low-frequency temperature base component in a 15-pixel × 15-pixel neighborhood, using the sum of the local temperature standard deviation and a preset constant 0.5 as the divisor, and the scaling constant 2.4 as the dividend, and adding 1.0 to the quotient to obtain the adaptive gain coefficient; using the adaptive gain coefficient to amplify the high-frequency radiation differential component and superimposing it onto the low-frequency temperature base component.

[0011] Preferably, step S2 uses a concentric ring differential operator to calculate the contrast activation response, including: step S21, constructing a three-layer concentric ring multi-kernel differential operator with radii of 2 pixels, 5 pixels and 9 pixels respectively; calculating the contrast step response between the average radiance of the inner core region and the concentric ring mask regions of the outer layers, and taking the minimum contrast step response at the three-layer scale as the contrast activation response of the central pixel; when the contrast activation response is greater than zero and the thermal permeation anisotropy index is not less than a preset truncation threshold, confirming that the corresponding pixel has active permeation texture features, preserving local hot spots and outputting candidate fire point connected domain masks and corresponding initial confidence scores; suppressing large-area homogeneous heated regions with non-monotonic convex contrast step responses at each scale.

[0012] Preferably, before solving for the instantaneous total radiated power, step S3 further includes gradient symmetry compensation for the candidate fire point connected domain mask: Step S31, calculate the eight-directional radial temperature gradient magnitude of the candidate fire point connected domain mask centered on the centroid; calculate the variance and mean of the radial temperature gradient magnitude in each direction, use the sum of the mean and a set constant 0.1 as the divisor, and the variance as the dividend to calculate the gradient dispersion ratio; subtract the gradient dispersion ratio from 1.0 to obtain the gradient symmetry compensation coefficient; when the gradient symmetry compensation coefficient is less than 0.35, determine that the corresponding connected domain is a unipolar divergent asymmetric pseudo-hot spot, and reduce the initial confidence level; remove candidate connected domains with an initial confidence level lower than a preset confidence threshold.

[0013] Preferably, in step S3, the net radiative heat flux density of the pixels within the candidate fire point connected region mask is calculated and spatially integrated to solve for the instantaneous total radiative power, including: step S32, based on the Stefan-Boltzmann law, using the Earth's surface thermal emissivity of 0.94 and the Boltzmann constant of 5.67 × 10⁻⁶. -8 W·m-2 ·K -4 The net radiative heat flux density is calculated by dividing the fourth power of the candidate connected region pixel temperature by the fourth power of the ambient background temperature. The ground coverage area of ​​a single pixel is calculated based on the flight altitude and the instantaneous field of view of the sensor, and the instantaneous total radiative power is obtained by summing the data across the entire region. The endogenous heat release power is calculated based on the smoldering radiative efficiency ratio of 0.15 to 0.22, and combined with the standard lower flammability value of 1.82 × 10⁻⁶ for combustible materials. 7 The mass loss rate per unit time is calculated using J / kg; the preset critical threshold is 15.0W, corresponding to a minimum mass loss rate of 8.24 × 10⁻⁶. -7 kg / s, with a preset continuous heat release threshold of -0.015W / s and a preset decay threshold of -0.05W / s.

[0014] Preferably, step S3 further includes preliminary screening of candidate fire point connected domain masks based on spatial adaptive dual threshold criteria: Step S33, statistically analyze the average background temperature of ground objects and the standard deviation of neighborhood temperatures within the sliding window, and set the absolute cutoff threshold for open flames to 373.15K; add the average background temperature of ground objects to the product of the dynamic factor and the standard deviation of neighborhood temperatures to form the hidden smoldering floating threshold, where the dynamic factor takes a value of 1.85 to 2.30; only candidate connected domains with local peak temperatures between the hidden smoldering floating threshold and the absolute cutoff threshold for open flames are input into the subsequent instantaneous total radiant power spatial integral and temporal derivative discrimination process.

[0015] Preferably, step S4 converts the centroid pixel coordinates of the concealed smoldering fire point into a spatial ray unit vector based on the airborne smooth attitude matrix and the thermal imaging sensor intrinsic parameter matrix, including: step S41, extracting the instantaneous measured angular velocities of yaw angle, pitch angle, and roll angle, and setting the single-step maximum angular velocity threshold to 15.0° / s; when the instantaneous measured angular velocity of any axis exceeds the single-step maximum angular velocity threshold, replacing the current sampled value with the single-step maximum angular velocity threshold, and calculating the smooth attitude matrix; converting the centroid pixel coordinates into a ray vector in the camera normalized coordinate system based on the thermal imaging sensor intrinsic parameter matrix; projecting the ray vector onto the geocentric coordinate system and normalizing it through the pod-mounted rotation matrix and smooth attitude matrix to generate a spatial ray unit vector.

[0016] Preferably, step S4, which calculates the coordinates of the intersection point between the spatial ray unit vector and the digital surface model, includes: step S42, projecting a probe ray along the spatial ray unit vector using the geodetic coordinates of the camera's optical center, traversing the elevation grid patches of the digital surface model using a hierarchical bounding box, and constructing an elevation intersection residual equation; using a numerical iteration method to solve for the ray forward step distance parameter, and substituting it into the ray equation to obtain the geodetic coordinates of the intersection point; when multiple intersection points are generated in the undulating canyon terrain, using the forward distance priority queue arbitration criterion, uniquely locking the first intersection point with the smallest spatial Euclidean distance along the line of sight forward direction, and eliminating false intersection points that penetrate the preceding ridge terrain.

[0017] Preferably, after step S4, the method further includes dynamic tracking and situation analysis of fire points: Step S43, input the calculated three-dimensional geographic coordinates of the earth into the interactive multi-model Kalman filter algorithm to smooth the temporal trajectory and remove jump noise; statistically analyze the three-dimensional centroid drift rate and geometric convex hull area expansion rate of the connected fire point group per unit time, fuse the local wind direction and speed vectors with the elevation, slope and aspect data extracted from the digital surface model of the earth's surface, and fit the fire spread trend matrix inside the earth's surface cover; when the cruise field of view is obscured by dense smoke, resulting in the loss of less than 15 consecutive observations, the virtual coordinates of the fire points are extrapolated based on the previous inertial filtering to maintain the continuity of the tracking trajectory; when the observation is restored, the fire points are re-matched and anchored, and an early warning signal containing the fire spread level and the recommended isolation zone cut-off position is output.

[0018] A fire concealed fire point detection system based on UAV thermal imaging is used to realize a fire concealed fire point detection method based on UAV thermal imaging, including a heat map sequence enhancement module, a seepage feature extraction module, a temporal radiation discrimination module, and a three-dimensional spatial positioning module.

[0019] The thermal image sequence enhancement module is used to acquire the geometrically registered thermal image sequence after multi-source sensor registration; it separates the low-frequency temperature substrate component and the high-frequency radiation differential component, adjusts the gain coefficient according to the low-frequency local variance, and superimposes them to generate an enhanced thermal image;

[0020] The seepage feature extraction module, connected to the heat map sequence enhancement module, is used to calculate the two-dimensional second-order partial derivatives of pixels in the enhanced heat image to construct the Hessian matrix, decompose to obtain the maximum and minimum absolute eigenvalues ​​and calculate the thermal seepage anisotropy index; the concentric ring difference operator is used to calculate the contrast activation response, and when the contrast activation response is positive and the thermal seepage anisotropy index is not less than the preset cutoff threshold, candidate fire point connected component masks are generated.

[0021] The temporal radiation identification module, connected to the seepage feature extraction module, is used to calculate the net radiative heat flux density of pixels within the candidate fire point connected domain mask and perform spatial integration to solve for the instantaneous total radiative power; extract the temporal derivative of the instantaneous total radiative power; confirm the hidden smoldering fire point when the inverted endogenous heat release power is not less than the preset critical threshold and the temporal derivative is not less than the preset continuous heat release threshold; and filter out the thermal storage false targets when the temporal derivative is less than the preset decay threshold.

[0022] The three-dimensional spatial positioning module, connected to the temporal radiometric discrimination module, is used to convert the centroid pixel coordinates of the concealed smoldering fire point into a spatial ray unit vector based on the airborne smooth attitude matrix and the intrinsic parameter matrix of the thermal imaging sensor. It then calculates the intersection coordinates of the spatial ray unit vector and the digital surface model of the earth, and outputs the three-dimensional geographic coordinates of the earth.

[0023] Compared with existing technologies, the present invention provides a method and system for detecting hidden fire points based on UAV thermal imaging, which has the following advantages:

[0024] 1. In the detection of hidden fire points using UAV thermal imaging, by separating the low-frequency temperature base component and the high-frequency radiation differential component, and combining the thermal infiltration anisotropy index and contrast activation response to identify local hot spots, the directional thermal texture formed along the pores of the medium by hidden smoldering fire points can be better preserved, while eliminating interference from heated ground objects with relatively uniform temperature distribution. This reduces two types of problems that occur when relying solely on the level of radiation brightness temperature for judgment: weak fire points are easily missed, and high-temperature backgrounds such as bare rock are easily misjudged as fire points, making the identification of hidden fire points in complex surface environments more reliable.

[0025] 2. By analyzing the instantaneous total radiant power of the candidate fire point's connected domain and its change over time, the differences in the temporal variation patterns of continuous heat release and natural heat storage can be used to further distinguish targets. Underground smoldering fire points continuously release heat, while exposed rocks and other heat-storing objects heated by sunlight usually gradually dissipate heat and cool down. Therefore, this can reduce the continuous false alarms caused by the inability to distinguish between the two types of targets based solely on the instantaneous temperature of a single frame, and reduce the interference of sunlight-heated objects and artificial highly reflective debris on fire point determination.

[0026] 3. By smoothing changes in airborne attitude and combining spatial ray unit vectors with digital surface models of the ground to determine the spatial location of the fire point, the influence of instantaneous line-of-sight swaying caused by airflow disturbances on the positioning results can be suppressed, while correcting the projection position deviation caused by complex undulating terrain. This allows for the stable conversion of the fire point location in the thermal image into three-dimensional geographic coordinates, improving the stability and accuracy of fire point spatial positioning under complex terrain cruising conditions in the field. Attached Figure Description

[0027] Figure 1This is a flowchart of the multi-stage detection and three-dimensional spatial positioning of concealed fire points according to the present invention;

[0028] Figure 2 This is the candidate connected component multi-level feature screening and false target removal graph of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0030] A method for detecting hidden fire points based on UAV thermal imaging includes the following steps:

[0031] Step S1: Obtain the geometric registration thermal image sequence after multi-source sensor registration; separate the low-frequency temperature substrate component and the high-frequency radiation differential component, adjust the gain coefficient according to the low-frequency local variance, and superimpose to generate an enhanced thermal image;

[0032] Step S2: Calculate the two-dimensional second-order partial derivatives of the pixels in the enhanced thermal image to construct the Hessian matrix, decompose to obtain the maximum and minimum absolute eigenvalues ​​and calculate the thermal anisotropy index; use the concentric ring difference operator to calculate the contrast activation response, and generate candidate fire point connected domain masks when the contrast activation response is positive and the thermal anisotropy index is not less than the preset cutoff threshold.

[0033] Step S3: Calculate the net radiative heat flux density of the pixels within the connected domain mask of the candidate fire point and perform spatial integration to solve for the instantaneous total radiative power; extract the temporal derivative of the instantaneous total radiative power; confirm the hidden smoldering fire point when the endogenous heat release power inverted based on the instantaneous total radiative power is not less than the preset critical threshold and the temporal derivative is not less than the preset continuous heat release threshold; filter out the heat storage false target when the temporal derivative is less than the preset decay threshold.

[0034] Step S4: Based on the airborne smooth attitude matrix and the intrinsic parameter matrix of the thermal imaging sensor, the centroid pixel coordinates of the concealed smoldering fire point are converted into spatial ray unit vectors. The intersection coordinates of the spatial ray unit vectors and the digital surface model of the ground are calculated, and the three-dimensional geographic coordinates of the ground are output.

[0035] Preferably, in step S1, the geometric registration thermal image sequence obtained after multi-source sensor registration is superimposed to generate an enhanced thermal image, including: step S11, uniformly accessing the hardware time reference source, using the line exposure trigger signal of the thermal imaging sensor as a reference, performing cubic Hermite polynomial interpolation on the airborne positioning coordinates and three-axis attitude angles, combining the intrinsic parameter matrix of the thermal imaging sensor and the rigid body mounting lever arm vector, mapping the discrete infrared field of view to the tangent plane of the reference geodetic coordinate system, and generating a geometric registration thermal image sequence; step S12, using anisotropic guided edge-preserving filtering based on structural tensor to separate the low-frequency temperature base component and the high-frequency radiation differential component; statistically analyzing the local temperature standard deviation of the low-frequency temperature base component in a 15-pixel × 15-pixel neighborhood, using the sum of the local temperature standard deviation and a preset constant 0.5 as the divisor, and the scaling constant 2.4 as the dividend, and adding 1.0 to the quotient to obtain the adaptive gain coefficient; using the adaptive gain coefficient to amplify the high-frequency radiation differential component and superimposing it onto the low-frequency temperature base component.

[0036] Preferably, step S2 uses a concentric ring differential operator to calculate the contrast activation response, including: step S21, constructing a three-layer concentric ring multi-kernel differential operator with radii of 2 pixels, 5 pixels and 9 pixels respectively; calculating the contrast step response between the average radiance of the inner core region and the concentric ring mask regions of the outer layers, and taking the minimum contrast step response at the three-layer scale as the contrast activation response of the central pixel; when the contrast activation response is greater than zero and the thermal permeation anisotropy index is not less than a preset truncation threshold, confirming that the corresponding pixel has active permeation texture features, preserving local hot spots and outputting candidate fire point connected domain masks and corresponding initial confidence scores; suppressing large-area homogeneous heated regions with non-monotonic convex contrast step responses at each scale.

[0037] Preferably, before solving for the instantaneous total radiated power, step S3 further includes gradient symmetry compensation for the candidate fire point connected domain mask: Step S31, calculate the eight-directional radial temperature gradient magnitude of the candidate fire point connected domain mask centered on the centroid; calculate the variance and mean of the radial temperature gradient magnitude in each direction, use the sum of the mean and a set constant 0.1 as the divisor, and the variance as the dividend to calculate the gradient dispersion ratio; subtract the gradient dispersion ratio from 1.0 to obtain the gradient symmetry compensation coefficient; when the gradient symmetry compensation coefficient is less than 0.35, determine that the corresponding connected domain is a unipolar divergent asymmetric pseudo-hot spot, and reduce the initial confidence level; remove candidate connected domains with an initial confidence level lower than a preset confidence threshold.

[0038] Preferably, in step S3, the net radiative heat flux density of the pixels within the candidate fire point connected region mask is calculated and spatially integrated to solve for the instantaneous total radiative power, including: step S32, based on the Stefan-Boltzmann law, using the Earth's surface thermal emissivity of 0.94 and the Boltzmann constant of 5.67 × 10⁻⁶. -8 W·m⁻²·K⁻4 The net radiative heat flux density is calculated by dividing the fourth power of the candidate connected region pixel temperature by the fourth power of the ambient background temperature. The ground coverage area of ​​a single pixel is calculated based on the flight altitude and the instantaneous field of view of the sensor, and the instantaneous total radiative power is obtained by summing the data across the entire region. The endogenous heat release power is calculated based on the smoldering radiative efficiency ratio of 0.15 to 0.22, and combined with the standard lower flammability value of 1.82 × 10⁻⁶ for combustible materials. 7 The mass loss rate per unit time is calculated using J / kg; the preset critical threshold is 15.0W, corresponding to a minimum mass loss rate of 8.24 × 10⁻⁶. -7 kg / s, with a preset continuous heat release threshold of -0.015W / s and a preset decay threshold of -0.05W / s.

[0039] Preferably, step S3 further includes preliminary screening of candidate fire point connected domain masks based on spatial adaptive dual threshold criteria: Step S33, statistically analyze the average background temperature of ground objects and the standard deviation of neighborhood temperatures within the sliding window, and set the absolute cutoff threshold for open flames to 373.15K; add the average background temperature of ground objects to the product of the dynamic factor and the standard deviation of neighborhood temperatures to form the hidden smoldering floating threshold, where the dynamic factor takes a value of 1.85 to 2.30; only candidate connected domains with local peak temperatures between the hidden smoldering floating threshold and the absolute cutoff threshold for open flames are input into the subsequent instantaneous total radiant power spatial integral and temporal derivative discrimination process.

[0040] Preferably, step S4 converts the centroid pixel coordinates of the concealed smoldering fire point into a spatial ray unit vector based on the airborne smooth attitude matrix and the thermal imaging sensor intrinsic parameter matrix, including: step S41, extracting the instantaneous measured angular velocities of yaw angle, pitch angle, and roll angle, and setting the single-step maximum angular velocity threshold to 15.0° / s; when the instantaneous measured angular velocity of any axis exceeds the single-step maximum angular velocity threshold, replacing the current sampled value with the single-step maximum angular velocity threshold, and calculating the smooth attitude matrix; converting the centroid pixel coordinates into a ray vector in the camera normalized coordinate system based on the thermal imaging sensor intrinsic parameter matrix; projecting the ray vector onto the geocentric coordinate system and normalizing it through the pod-mounted rotation matrix and smooth attitude matrix to generate a spatial ray unit vector.

[0041] Preferably, step S4, which calculates the coordinates of the intersection point between the spatial ray unit vector and the digital surface model, includes: step S42, projecting a probe ray along the spatial ray unit vector using the geodetic coordinates of the camera's optical center, traversing the elevation grid patches of the digital surface model using a hierarchical bounding box, and constructing an elevation intersection residual equation; using a numerical iteration method to solve for the ray forward step distance parameter, and substituting it into the ray equation to obtain the geodetic coordinates of the intersection point; when multiple intersection points are generated in the undulating canyon terrain, using the forward distance priority queue arbitration criterion, uniquely locking the first intersection point with the smallest spatial Euclidean distance along the line of sight forward direction, and eliminating false intersection points that penetrate the preceding ridge terrain.

[0042] Preferably, after step S4, the method further includes dynamic tracking and situation analysis of fire points: Step S43, input the calculated three-dimensional geographic coordinates of the earth into the interactive multi-model Kalman filter algorithm to smooth the temporal trajectory and remove jump noise; statistically analyze the three-dimensional centroid drift rate and geometric convex hull area expansion rate of the connected fire point group per unit time, fuse the local wind direction and speed vectors with the elevation, slope and aspect data extracted from the digital surface model of the earth's surface, and fit the fire spread trend matrix inside the earth's surface cover; when the cruise field of view is obscured by dense smoke, resulting in the loss of less than 15 consecutive observations, the virtual coordinates of the fire points are extrapolated based on the previous inertial filtering to maintain the continuity of the tracking trajectory; when the observation is restored, the fire points are re-matched and anchored, and an early warning signal containing the fire spread level and the recommended isolation zone cut-off position is output.

[0043] A fire concealed fire point detection system based on UAV thermal imaging is used to realize a fire concealed fire point detection method based on UAV thermal imaging, including a heat map sequence enhancement module, a seepage feature extraction module, a temporal radiation discrimination module, and a three-dimensional spatial positioning module.

[0044] The thermal image sequence enhancement module is used to acquire the geometrically registered thermal image sequence after multi-source sensor registration; it separates the low-frequency temperature substrate component and the high-frequency radiation differential component, adjusts the gain coefficient according to the low-frequency local variance, and superimposes them to generate an enhanced thermal image;

[0045] The seepage feature extraction module, connected to the heat map sequence enhancement module, is used to calculate the two-dimensional second-order partial derivatives of pixels in the enhanced heat image to construct the Hessian matrix, decompose to obtain the maximum and minimum absolute eigenvalues ​​and calculate the thermal seepage anisotropy index; the concentric ring difference operator is used to calculate the contrast activation response, and when the contrast activation response is positive and the thermal seepage anisotropy index is not less than the preset cutoff threshold, candidate fire point connected component masks are generated.

[0046] The temporal radiation identification module, connected to the seepage feature extraction module, is used to calculate the net radiative heat flux density of pixels within the candidate fire point connected domain mask and perform spatial integration to solve for the instantaneous total radiative power; extract the temporal derivative of the instantaneous total radiative power; confirm the hidden smoldering fire point when the inverted endogenous heat release power is not less than the preset critical threshold and the temporal derivative is not less than the preset continuous heat release threshold; and filter out the thermal storage false targets when the temporal derivative is less than the preset decay threshold.

[0047] The three-dimensional spatial positioning module, connected to the temporal radiometric discrimination module, is used to convert the centroid pixel coordinates of the concealed smoldering fire point into a spatial ray unit vector based on the airborne smooth attitude matrix and the intrinsic parameter matrix of the thermal imaging sensor. It then calculates the intersection coordinates of the spatial ray unit vector and the digital surface model of the earth, and outputs the three-dimensional geographic coordinates of the earth.

[0048] Example 1: The current UAV patrol system is used for surface smoldering inspection in large-area mixed coniferous and broadleaf forest areas. The surface is covered with a layer of dry fallen leaves and humus with a thickness of 5cm to 30cm. There is also dense smoke obscuring the surface due to sudden changes in wind direction and localized high-temperature interference from exposed granite heated by sunlight. The UAV is equipped with a long-wave uncooled infrared thermal imager, a dual-light pod gyro-stabilized platform, a high-precision GNSS positioning module, and an inertial measurement unit. The infrared thermal imaging sensor has a spectral response band of 8μm to 14μm, an effective resolution of 640×512 pixels, a lens focal length of 19mm, a pixel physical size of 17μm, and the principal point coordinates of the pixel center are... To eliminate spatiotemporal asynchronous errors caused by the different integration periods of airborne pose data and thermal infrared detectors under turbulent airflow conditions, the system integrates the airborne GNSS positioning module, inertial measurement unit, and infrared thermal imaging sensor into a nanosecond-level hardware time reference source. Using the line exposure trigger signal of the infrared thermal imaging sensor as a unified reference, the system extracts the instantaneous longitude, latitude, and geodetic height output from the airborne GNSS, as well as the pitch, roll, and yaw angles output from the inertial measurement unit. A high-order interpolation using a cubic Hermitian polynomial is performed between the two sets of discrete GNSS data sampling points. Combined with the pre-calibrated intrinsic parameter matrix of the infrared thermal imager and the vector of the airborne rigid body mounting lever arm, the discrete infrared field of view is mapped to the tangent plane of the reference geodetic coordinate system, constructing a geometric registration thermal map sequence with spatial pose markers. Specifically, during the field of view mapping, the system first extracts the airborne positioning data... The three-dimensional geodetic coordinates of the antenna phase center are used to transform the eccentric lever arm vector between the positioning antenna and the optical center of the infrared camera in the pre-calibrated rigid body coordinate system into the local station center tangent plane coordinate system using the airborne instantaneous attitude rotation matrix obtained by interpolation. This vector is then superimposed onto the positioning antenna coordinates to complete the translational eccentricity compensation of the true spatial position of the camera's optical center. Subsequently, the focal length and principal point parameters in the intrinsic parameter matrix are used to convert the coordinates of each pixel in the thermal image into the camera's normalized line of sight. Two three-dimensional coordinate system rotations are then performed sequentially through the pod installation rotation matrix and the airborne attitude matrix to transform the coordinates into the line of sight under the reference geodetic coordinate system tangent plane. Finally, the coordinates are projected along the line of sight onto the set reference tangent plane elevation reference, and the normalized grid coordinates of each pixel on the ground reference tangent plane are calculated, achieving accurate geometric registration of the thermal image sequence in time and space references.

[0049] Hidden smoldering ignition points buried deep beneath the humus layer, after thermal radiation is attenuated by the physical thermal resistance of the porous medium and dissipated by lateral heat conduction, typically result in an effective radiant temperature rise to the surface that exceeds the surrounding ambient temperature by 1.5K to 12K. These ignition points are easily obscured by the gently changing terrain. The thermal image sequence enhancement module uses geometric registration to register the current thermal image within the thermal image sequence. As input, an anisotropic-guided edge-preserving filter based on the structure tensor is used to perform frequency domain splitting, separating the low-frequency temperature substrate component that reflects the background of a large area of ​​terrain. And the high-frequency radiation differential component reflecting local thermal anomaly characteristics In specific implementation, with Using itself as a guide image, the system first calculates the first-order spatial temperature gradient of the image along two orthogonal axes using first-order difference operators in the horizontal and vertical directions, respectively. This constructs a symmetric structure tensor matrix composed of the squares of the horizontal and vertical gradients, and the product of the horizontal and vertical gradients. Subsequently, a Gaussian smoothing kernel is used to perform local neighborhood weighted averaging on each component of the structure tensor matrix to suppress false edges caused by high-frequency random noise. Finally, analytical eigenvalue decomposition is performed on the smoothed structure tensor matrix to obtain the maximum eigenvalue representing the local edge intensity and energy extension direction, and its corresponding normal eigenvector. The image is processed by analyzing the tangential feature vector along the edge direction. Local coherence is calculated based on the difference between the maximum and minimum eigenvalues. When coherence is high, the pixel is determined to be located at a land boundary or geological fissure. The diffusion weight perpendicular to the edge normal is reduced, while the diffusion weight parallel to the edge tangential direction is increased. Isotropic uniform diffusion smoothing is applied to flat areas with near-zero coherence. Spatial convolution filtering is performed on the guiding image using a local kernel function with adaptive diffusion weights to smooth out subtle noise and preserve step edge contours, outputting a low-frequency temperature base component reflecting the continuous distribution of the large-scale terrain background. Finally, the image is... The brightness temperature values ​​of each pixel are subtracted point by point from the corresponding low-frequency temperature base component values, and the difference is used as the output of the high-frequency radiation differential component. The system calculates the low-frequency local variance of the low-frequency temperature base component within a 15-pixel × 15-pixel local sliding neighborhood, and takes the square root of this variance as the local temperature standard deviation. Its dimension is Kelvin; subsequently, an adaptive local variance mapping gain operator is constructed. In this context, constant 2.4 is a dimensionless gain scaling constant, and constant 0.5 is a Kelvin-dimensional zero-smoothing constant. This operator adjusts the amplification weights of the high-frequency radiation differential components, and the resulting superimposed components generate an enhanced thermal image. Under this parameter mapping, the local temperature rise in a low-variance homogeneous background achieves a dynamic enhancement of up to 5.8 times, while the gain weight of high-variance bare rock and steep slope edges automatically decays to around 1, suppressing strong radiation noise interference from the slowly varying temperature field in the background. The constant 0.5K corresponds to the typical system noise equivalent temperature difference limit of an uncooled infrared thermal imager under normal operating conditions. Setting the denominator smoothing constant to 0.5K ensures that the gain operator has a stable mathematical lower bound when the neighborhood temperature standard deviation approaches zero, preventing the sensor quantization background noise from being amplified infinitely and causing numerical divergence. The selection of the dimensionless scaling constant 2.4 is based on the image height. The matching determination of the dynamic range of the frequency edge and the display bit depth shows that when the infrared image bit depth is 14 bits, the high-frequency differential amplitude of weak temperature rise pixels without enhancement is usually less than 12 gray levels. If the scaling constant is less than 1.5, the maximum magnification of weak temperature rise in a homogeneous background is less than 4 times, and the target is still buried in the surface thermal noise. If the scaling constant is greater than 3.2, the magnification in a homogeneous background exceeds 7.4 times, and weak hot spots in the thermal infrared image produce highlight overflow artifacts. In the sample comparison test on a large area of ​​fallen leaf surface, when the scaling constant is set to 2.4, the theoretical maximum gain in a low variance homogeneous background remains at 5.8 times.

[0050] Two-dimensional intensity threshold screening typically presets the thermal radiation source as a symmetrical Gaussian hot spot. This can easily generate false alarms for lumpy absorbers such as exposed granite and dry sand under strong sunlight. The seepage feature extraction module utilizes the physical characteristics of combustion gases escaping along microcracks in porous media to enhance thermal images. For candidate pixels, the second-order partial derivatives in the local two-dimensional space are calculated to construct the pixel-level Hessian matrix. Then, eigenvalue decomposition is performed on the Hessian matrix to obtain the largest absolute eigenvalue representing the curvature of the local image texture. with the smallest absolute eigenvalue Combined with the prevention of zero constant Calculation of thermal anisotropy index ,in, The value is a dimensionless smoothing constant. Simultaneously, the system constructs three concentric ring-shaped multi-kernel differential operators with radii of 2 pixels, 5 pixels, and 9 pixels, respectively. Based on the radiometric calibration relationship of the thermal imaging sensor, the radiance corresponding to each pixel is obtained, and the average radiance of the inner core region is calculated. With the Average radiance of the concentric ring shielding area around the outer layer The difference between the two is taken as the contrast step response at each scale, and the minimum contrast step response at all three scales is taken as the contrast activation response of the central pixel. The system imposes joint constraints on the contrast activation response and the thermal percolation anisotropy index, while satisfying... and At that time, it was confirmed that the corresponding pixel possessed active percolation texture features and it was included in the candidate fire point connected region mask. Initial confidence level The contrast activation response is determined by normalization mapping based on the positive margin of the zero decision boundary and the positive margin of the thermal anisotropy index relative to the cutoff threshold of 3.5. For any margin close to the corresponding decision boundary, the initial confidence is reduced; when both margins increase simultaneously, the initial confidence is increased accordingly. The contrast step responses obtained at three scales (2 pixels, 5 pixels, and 9 pixels) are examined in scale order. When the response at the intermediate scale is higher than the responses at the adjacent inner and outer scales, forming a local bulge, the three... When the layer response cannot maintain a monotonic relationship with scale, it is judged as a non-monotonic bulge in the contrast step response at each scale, and the corresponding large-area homogeneous heated region is suppressed. For isotropic blocky high-temperature pixels with nearly uniform two-dimensional curvature, such as solar thermal storage granite, the minimum absolute eigenvalue and the maximum absolute eigenvalue are on the same order of magnitude, and the thermal permeation anisotropy index is lower than the cutoff threshold of 3.5. These pixels are then removed during the morphological feature screening stage. Regarding the setting of the cutoff threshold of 3.5 and the prevention zero constant of 0.05, porous media... When combustion gases escape along microcracks in a medium, the hot gas flow is blocked by the porous skeleton solid phase of the sidewalls in the normal direction of the main crack, forming a relatively steep transverse temperature gradient, while it spreads gently along the gas seepage path in the tangential direction of the crack. According to Darcy's law and the thermal diffusion equation in porous media seepage mechanics, the ratio of effective thermal conductivity in the main tangential direction to that in the normal direction is usually between 3.6 and 6.2. After mapping to the two-dimensional intensity curvature space of the image, the absolute value of the principal eigenvalue of its Hessian matrix in the normal direction is more than 3.6 times that of the secondary eigenvalue in the tangential direction. The truncation threshold is set as follows: The threshold is set at 3.5. If the threshold is below 3, the local anisotropy caused by thermal stress microcracks at the edge of the sunlit rock may exceed the criterion and generate a false alarm. If the threshold is above 4.5, the weak seepage texture after transverse smoothing by porous media at a depth of 30cm may be missed. The zero constant for prevention is set at 0.05. Based on the natural noise variance of the Hessian matrix eigenvalues ​​in the flat background area, its value is slightly higher than the average noise amplitude of the secondary eigenvalues ​​in the flat background, which is 0.038. This is used to avoid singular divergence of the denominator when the eigenvalues ​​are close to zero in a homogeneous background.

[0051] To prevent residual heat from the front-line extinguishing and environmental temperature drift from interfering with the identification of candidate areas, the system statistically analyzes the average temperature within the sliding window of the ground feature background field. with temperature standard deviation Both are in units of K, and the absolute cutoff threshold for open flame is set. The value is 373.15K, combined with a dynamic factor ranging from 1.85 to 2.30. Constructing a hidden smoldering floating threshold Only the local peak temperature is between and Candidate regions between them are preserved and imported into the subsequent dynamic discrimination process; subsequently, the system extracts the candidate fire point connected domain mask with thermal radiation centroid. octagonal radial temperature gradient modulus centered Calculate the variance of gradient magnitude in each direction. with the mean Through calculation The gradient symmetry compensation coefficient was obtained, where the constant 0.1 is a zero-smoothing constant with the same dimension as the mean gradient modulus. Thermal infiltration anisotropy feature extraction and eight-axis radial gradient symmetry compensation were performed at two different spatial scales. The thermal infiltration anisotropy index was used to calculate the second-order partial derivative in a 3×3 pixel local neighborhood to extract the ridge-like profile curvature of the pore edges emerging from surface microfractures. Eight-axis radial temperature gradient compensation was performed at a connected domain outer envelope scale containing dozens of pixels, with the sampling ray penetrating from the thermal radiation centroid to the connected domain. At the outer boundary, the macroscopic temperature envelope field formed by the diffusion of deep heat sources in all directions is statistically analyzed. Underground fire sources inside isotropic media exhibit a centrally divergent heat conduction ellipse with uniform gradient distribution in all directions. The measured gradient symmetry compensation coefficient is stable above 0.75. The diffuse area of ​​smoke blown in one direction or the yin-yang surface of rocks heated on one side by sunlight exhibits a unipolar gradient cliff, and the calculated value drops to below 0.35. Based on this, the system performs initial confidence decay on asymmetric pseudo-hot spots and removes interference patches with initial confidence values ​​lower than the preset confidence threshold.

[0052] The temporal radiation discrimination module, based on the Stefan-Boltzmann law, calls upon the thermal emissivity of the land cover. With Boltzmann constant W·m⁻²·K⁻ 4 Calculate the difference between the fourth power of the radiation temperature of each pixel within the candidate connected region and the fourth power of the ambient background temperature, and solve for the net radiative heat flux density of the pixel. ,in, For candidate pixel radiative brightness temperature, Using the ambient background temperature, and combining the drone's real-time cruise relative altitude with the instantaneous field of view of the thermal imager, the physical coverage projection area of ​​a single pixel on the ground is calculated. For connected components The instantaneous total radiated power is obtained by summing up all pixels in the region. Based on the theoretical radiation efficiency ratio of the smoldering reaction of combustible organic matter on the forest surface. Inversion to determine the endogenous heat release power of the underground heating core The standard lower calorific value of red pine needles and peat humus was then substituted into the values. J / kg Calculation of combustible material mass loss rate per unit time The smoldering heat release within the porous media cover layer includes not only heat conduction through the solid-phase framework but also upward convective seepage of pyrolysis gases generated by the combustion reaction, driven by buoyancy and thermal pressure, along micro-cracks in the porous soil framework and gaps between fallen leaves. The high-temperature seeping gases carry the heat of reaction out to the surface micro-regions, forming convective and radiative heat transfer windows at the openings of local cracks. The aforementioned radiation efficiency ratio of 0.15 to 0.22 is the equivalent apparent radiative heat transfer coefficient obtained after establishing quasi-steady-state micro-crack thermal convection channels, comprehensively considering the equivalent thermal resistance attenuation factor of the cover layer micropores and the local escape heat flux. It is used to represent the ratio between the net escape radiation energy flow from the surface after equivalent attenuation through the crack channels and the total chemical heat release energy flow from the deep reaction core. Within a continuous 30-second observation window, for an underground fire source in an active smoldering state, the calculated endogenous heat release power is 498.78 W, exceeding the critical threshold of 15 W for maintaining stable combustion, corresponding to a mass loss rate of... kg / s, and the time-series derivative of the instantaneous total radiated power The measured average is W / s, greater than the sustained heat release threshold W / s; Facing a bare granite dummy target that strongly absorbs heat during the day, during a 40-second observation period with solar radiation and gusts of wind disturbance, its instantaneous total radiant power monotonically decreases from 280.5 W to 276.2 W, and the calculated time derivative is: W / s, below the preset attenuation threshold W / s, the system filters out the false target of the thermal storage. The numerical setting method of the continuous heat release threshold -0.015W / s and the preset attenuation threshold -0.05W / s is as follows: Under the environmental convection interference of outdoor wind speeds of 2m / s to 6.5m / s, the natural cooling process of naturally exposed granite after the reduction of sunlight follows Newton's law of cooling. The heat dissipation rate of its measured instantaneous radiant power within a 40s time window is distributed between -0.062W / s and -0.125W / s; When an active, self-sustaining smoldering fire source continuously releases heat, and is affected by transient convective cooling caused by cold air gusts at the surface, the rate of decrease in thermal radiation maintained by deep high-temperature gas supply does not exceed -0.012 W / s. Under weak wind conditions, it usually exhibits positive heat release growth. Therefore, the continuous heat release threshold is set to -0.015 W / s, and the decay threshold is set to -0.05 W / s, in order to distinguish between hidden smoldering fire points that continuously release chemical heat and naturally cooled thermal storage features.

[0053] After confirming the effective concealment of the fire point, the 3D spatial positioning module extracts the pixel coordinates of the thermal radiation centroid of the connected domain of the fire point. The ray vector is converted into a ray vector in the camera's normalized coordinate system using the camera intrinsic parameter matrix. In response to sudden airflow disturbances during cruise, the system extracts the instantaneous measured angular velocities of yaw, pitch, and roll angles, and sets a single-step maximum angular velocity threshold. For samples exceeding the threshold value of 15° / s, amplitude limiting filtering is applied to calculate the airborne smooth attitude matrix. Combined with the installation of rotating matrix in the pod The ray vector was transformed to the ENU geocentric coordinate system of the Northeast Sky and normalized to generate a spatial ray unit vector. The single-step maximum angular velocity threshold of 15° / s is used in conjunction with the reverse mechanical isolation bandwidth of the airborne dual-optical pod gyro stabilization platform. Because the gyro stabilization platform actively isolates and filters out over 90% of the high-frequency airflow vibrations of the aircraft, the actual physical angular velocity of the infrared camera's optical axis relative to the ground remains within 10° / s during cruise mapping. High-frequency jump components exceeding 15° / s are treated as quantization noise and pulse impact anomalies in the inertial measurement unit under transient gust disturbances. The attitude smoothing filter module isolates abnormal outlier impacts through amplitude limiting and truncation, while internally maintaining an error integration delay queue with a discrete sampling period. After the transient pulse ends, the angle difference lost due to peak clipping is compensated back into the integration loop in subsequent stable sampling frames with an asymptotic damping step size, using the geodetic coordinates of the UAV camera's optical center. Projecting probe rays along spatial unit vectors, using hierarchical bounding boxes to accelerate traversal of elevation grid patches in the Digital Surface Model (DSM), including local gentle slope elevations. For a depth of 330m, construct the intersection residual equation between the space ray and the ground elevation. ,in The forward step distance parameter along the ray is used; the ray depth scalar is calculated by convergent solution using the Newton-Raphson numerical iteration method. The distance is 169.73m. Substituting this into the ray equation, the horizontal displacement in the due east direction is 120.02m and the displacement in the due south direction is 1.07m. The output three-dimensional geographic coordinates of the fire point are 118.453587°E, 32.124550°N, and an elevation of 330m. In the undulating deep mountain canyon terrain, when multiple grid intersection points are generated due to the obstruction of mountain ridges, the system activates the forward distance priority queue arbitration criterion, and uniquely locks the first intersection point with the smallest spatial Euclidean distance along the direction of the line of sight, eliminating false intersection points formed by the ray penetrating the near mountain and projecting to the far back slope.

[0054] The calculated 3D geographic coordinates of the fire points are input into an interactive multi-model Kalman filter algorithm to smooth the temporal positioning trajectory and remove single-frame jump noise. The 3D centroid drift rate and geometric convex hull area expansion rate of the connected fire point group are statistically analyzed per unit time, and the wind direction and speed vectors transmitted from local meteorological sensors are fused together. And elevation, slope, and aspect data extracted from digital surface models At the current fire location, elevation, slope, and aspect data The topographic elevation gradient vector composed of eastward and northward elevation changes Indicates; the diffusion driving vector formed by wind field and topography according to Calculate, where, and To establish a diffusion sensitivity constant related to forest stand type, the system uses the three-dimensional centroid drift rate as the translational state variable of the fire swarm and the geometric convex hull area expansion rate as the expansion state variable of the fire swarm. It also considers the translational state variable, expansion state variable, and diffusion driving vector at each observation time. Arranged chronologically, a matrix of fire spread trends within the surface cover was formed. During the calibration process, the wind direction sensitivity constant and the terrain slope sensitivity constant were determined by the surface combustible moisture content and combustible load. Under baseline conditions of 2.5 kg / m² to 4 kg / m² litter cover and 12% to 18% moisture content in coniferous forests, controlled ignition tests in a micro-wind tunnel were conducted to determine the horizontal spread velocity of the underground fire front at different wind speeds. The slope of the spread velocity relative to wind speed was used to determine the wind direction sensitivity constant as 0.035. When the combustible load increased to above 5 kg / m², this constant was adjusted to 0.045. The diffusion increment along the slope was measured in a slope variation test chamber under zero wind speed conditions. The partial derivative of the spread velocity with respect to the slope tangent is used to determine the terrain slope sensitivity constant as 0.012 m / s. When the compaction density of the leaf litter layer on the slope increases, leading to a decrease in air permeability, this constant is adjusted to 0.008 m / s. When the UAV passes through a dense smoke plume downwind in the cruise field of view, causing transient infrared signal attenuation and obstruction for less than 15 consecutive frames, the system switches to the pre-processing inertial filtering state, extrapolating the virtual coordinates of the fire point to maintain the continuity of the tracking trajectory. After observation recovery, the fire point is re-matched and anchored using a multi-scale template. The system determines the fire spread level based on the comparison results of the spread velocity, three-dimensional centroid drift rate, and geometric convex hull area expansion rate at consecutive moments in the spread trend matrix with the preset classification boundary on site, and then... The indicated main spread direction and the location of the fire front within the corresponding simulation time determine the recommended firebreak cut-off location, and the final output is an early warning dataset containing the fire spread level, the center geodetic coordinates, and the recommended firebreak cut-off location.

[0055] Example 2: The current test system is deployed in a field test site with the surface characteristics of a mixed coniferous and broadleaf forest. The surface cover includes dried red pine needles, dead branches, and peat humus, with a stepped distribution of cover thickness between 5cm and 35cm. A controlled underground smoldering heat source is deployed in the test site, and large areas of exposed granite, dry gravel piles, and metal fragments exposed to direct sunlight are selected as heat absorption comparison targets. The test platform uses a multi-rotor aircraft equipped with a long-wave uncooled infrared thermal imager with a spectral response band of 8μm to 14μm, a dual-beam gyro-stabilized pod, a high-precision differential GNSS positioning module, and an inertial measurement unit; the effective physical resolution of the thermal imager is 6. The lens has a 40×512 pixel resolution, a pixel pitch of 17μm, and an effective focal length of 19mm. The airborne hardware is uniformly connected to a hardware time reference source with a timing accuracy better than 10ns. Using the line exposure trigger signal of the infrared thermal imaging sensor as a reference, cubic Hermite polynomial interpolation is performed on the airborne position coordinates and three-axis attitude angles. A limiting and smoothing constraint is applied to the airframe attitude angular velocity under airflow disturbance conditions. Combined with the pre-calibrated intrinsic parameter matrix of the thermal imaging sensor and the rigid body mounting lever arm vector, the discrete infrared field of view is mapped to the tangent plane of the reference geodetic coordinate system, generating a geometrically registered thermal map sequence with external orientation parameter markings. During the test, the average ambient background temperature... The temperature is 296.2 K, and the standard deviation of the neighborhood temperature is [missing information]. At 1.6K, the thermal imager's field of view is superimposed with natural gust wind disturbances ranging from 2.0 m / s to 6.5 m / s and intermittent dense smoke obscuring the field of view.

[0056] When enhancing the geometrically registered heatmap sequence, an anisotropic-guided edge-preserving filter based on the structure tensor is used to separate the low-frequency temperature floor component and the high-frequency radiation differential component. The low-frequency local variance is statistically analyzed within a 15-pixel × 15-pixel sliding neighborhood, and the local temperature standard deviation is obtained from its square root. Gain adjustment uses a gain scaling constant of 2.4 and a zero-smoothing constant of 0.5K. In test conditions with coverage thicknesses ranging from 5cm to 30cm, the high-frequency radiation differential component actually obtains a gain weight of 2.0 to 2.7 times. Subsequently, the two-dimensional second-order partial derivatives of each pixel in the enhanced thermal image are calculated and the Hessian matrix is ​​constructed. The thermal infiltration anisotropy index is obtained using a zero-smoothing constant of 0.05. The concentric ring differential operator still uses a three-layer scale of 2 pixels, 5 pixels, and 9 pixels. Based on the radiometric calibration relationship of the thermal imaging sensor, it obtains the average radiance of the inner core region and the average radiance of each outer concentric ring region, calculates the contrast step response, and forms the contrast activation response. When the contrast activation response is positive and... Candidate pixels with a value not less than 3.5 are retained. After the candidate fire point connected components enter subsequent processing, the eight-directional radial temperature gradient modulus is calculated with the centroid as the center. The gradient symmetry compensation coefficient is obtained by using the variance and mean of the gradient modulus in each direction and a predetermined smoothing constant of 0.1. After initial screening using a spatially adaptive dual-threshold method, the net radiative heat flux density is calculated and spatially integrated. The endogenous heat release power is then inverted using a smoldering radiative efficiency ratio of 0.18. For each test condition, the time-series derivative of the instantaneous total radiated power was extracted within a 40-second observation window. The endogenous heat release power is compared with a preset critical threshold of 15W, and the time derivative is compared with a preset continuous heat release threshold of -0.015W / s and a preset decay threshold of -0.05W / s. After confirming the hidden smoldering fire point, the three-dimensional spatial positioning module combines the airborne smooth attitude matrix, the pod installation rotation matrix and the thermal imaging sensor intrinsic parameter matrix to generate a spatial ray unit vector. The elevation intersection is solved by the Digital Surface Model (DSM) to obtain the three-dimensional geographic coordinates of the fire point.

[0057] In shallow active smoldering conditions with a 5cm cover thickness, the surface temperature rise was 11.2K, and the local temperature standard deviation was [not specified]. The K value is 1.65K; after enhancing the thermal image to include seepage feature extraction, The value is 5.34, obtained by processing an eight-directional radial temperature gradient. The value is 0.86. After calculating the radiated power of the connected components of the candidate fire points, It is +0.038W / s. The W value was 586.2 W. Both values ​​fell within the criteria for identifying a concealed smoldering fire point, and the system confirmed the target as such. Subsequently, the system completed the intersection positioning of the spatial ray and the DSM (Distributed Sunlight Detection System). The absolute error of the three-dimensional positioning was 0.35 m. After the cover thickness increased to 15 cm, the surface temperature rose by 4.8 K. It is 1.18K, and after the same processing, it is obtained 4.82 0.81 +0.024W / s The value was 312.4W, and the system also confirmed the hidden smoldering point, with an absolute error of 0.42m in three-dimensional positioning.

[0058] When the cover thickness increases to 30cm, the effective surface temperature rises to 1.8K. The value is 0.92K; after enhancement and morphological feature extraction, Still 3.76, The value is 0.72, and the radiation power is being processed in time series. The value is +0.008 W / s, obtained from the inversion. The intensity was 128.5W. Based on this, the system identified a hidden smoldering ignition point. The absolute error of the three-dimensional positioning was 0.68m. When the cover thickness was further increased to 35cm, the surface temperature rise was 1.4K. It is 0.85K. It dropped to 3.52. It is 0.68; affected by wind field convective disturbance, It is -0.002W / s. The value is 84.1W. Existing criteria still confirm that the target is a hidden smoldering fire point. The absolute error of three-dimensional positioning is 0.94m. Test data under different cover thicknesses show that as the cover layer thickens, the surface temperature rise, the anisotropy index of thermal infiltration, and the endogenous heat release power all change accordingly with thermal resistance dissipation. The 35cm working condition is close to the weak thermal detection boundary under this thermal infrared detection condition.

[0059] To compare the impact of edge protection reinforcement and seepage feature extraction on the detection results, edge protection filtering and seepage feature extraction were also removed under the same heat source conditions of 15cm cover thickness and 4.8K surface temperature rise. Still 1.18K, but Not extracted It is 0.79. It is +0.021W / s. The value is 308.6W. Since the front end did not form a screening result for the anisotropic characteristics of thermal seepage, the detection result is recorded as a missed detection and not triggered. Subsequent spatial positioning was not triggered, and the three-dimensional positioning result is unable to locate.

[0060] When testing the bare granite rock with solar thermal storage, its apparent brightness temperature reached 338.2 K. The value is 1.94K, obtained from image morphology calculation. It is 1.42. The value is 0.26; within the same observation window, the instantaneous total radiative power decreases with convective dissipation. The value is -0.108 W / s. The endogenous heat release power term is recorded as having no chemical thermal self-consistency. Under the condition of canceling the comparison of combustion value time derivative discrimination, the target generates false alarms and outputs three-dimensional positioning results with an absolute error of 1.15 m. The negative time series change in this set of data corresponds to the cooling process of the solar thermal storage body. After retaining the time series derivative discrimination, it can be used to distinguish it from the change difference between it and the continuously releasing heat target.

[0061] As a comparative test, when verifying the ray tracing correction of the digital surface model, smoldering fire points with a 15cm cover thickness were selected and a gently sloping, undulating terrain was set. This condition... It is 1.18K. It is 4.8. It is 0.81. It is +0.023W / s. The fire point detection result was still confirmed as a hidden smoldering fire point at 315W. After canceling the DSM ray tracing correction, the elevation undulation of the gentle mountain slope and the perturbation of the aircraft's attitude angular velocity directly entered the spatial projection error, and the absolute error of three-dimensional positioning increased to 18.62m.

[0062] When comparing parameter boundaries, The cutoff threshold was lowered to 2, and the edge of the solar thermal storage granite was used as the test target. It is 2.12K. It is 2.45. It is 0.31. The value was -0.092 W / s, and the endogenous heat release power term was recorded as having no chemical-thermal self-consistency. Because 2.45 exceeded the lowered cutoff threshold, the detection results showed false alarms at the edge, with an absolute error of 1.28 m in 3D positioning. After raising the cutoff threshold to 5.5, the deep, weakly smoldering target with a 30 cm cover thickness was further processed. For 0.92K, 3.76 0.72 +0.008W / s It is 128.5W; due to If the result is below the adjusted cutoff threshold, the detection result becomes a false positive, which will be eliminated or missed, and subsequent location cannot be determined.

[0063] In another set of parameter comparisons, the dynamic factor The value was lowered to 1.5, and a high-temperature dry sand pile disturbed by wind was selected as the background target. It is 1.78K. It is 1.88. It is 0.38. The value was -0.065 W / s, and the endogenous heat release power term was recorded as having no chemical-thermal self-consistency. After the dynamic factor was reduced, the dry sand pile entered the floating threshold screening range, and the detection result was a false alarm in the background. The absolute error of three-dimensional positioning was 2.45 m.

[0064] In the above tests, under normal smoldering conditions, the following processes were performed sequentially: heat map sequence enhancement, joint screening of anisotropic characteristics of thermal infiltration and contrast activation response, gradient symmetry compensation, radiation power time sequence discrimination, and DSM intersection positioning. All intermediate data and the final judgment were generated continuously along the same processing chain. Under the control conditions, after changing the front-end feature extraction, time sequence discrimination, three-dimensional terrain correction, and parameter values, missed detection, thermal storage false alarm, positioning error amplification, and background false alarm appeared in the corresponding processing stages.

[0065] Example 3: The current UAV thermal imaging fire concealed fire detection system is deployed on an airborne multi-core embedded computing platform for patrolling concealed fires in forests where red pine and broad-leaved tree species are interspersed and the ground surface is accompanied by exposed gravel and a thick humus layer. The airborne multi-core embedded computing platform includes a graphics processing unit for image pipeline flow and a central processing unit responsible for spatial motion calculation and task scheduling. The two interact through a high-speed peripheral interconnect standard bus. When the airborne long-wave uncooled infrared thermal imager passes through areas with different canopy cover densities during the cruise, the background radiation substrate undergoes time-varying drift depending on the ground cover type. The thermal image sequence enhancement module extracts the low-frequency temperature substrate component. A spatiotemporally adaptive sliding statistical window is established, with a spatial analysis window of 64 pixels × 64 pixels. A single-order exponentially weighted moving average filter with a coefficient of 0.08 is used to update the local background field mean temperature on the time axis. with temperature standard deviation This allows the low-frequency temperature baseline component to track the gradual thermal radiation from the forest canopy to the bare soil and grassland. Simultaneously, the low-frequency local variance is statistically analyzed within a 15×15 pixel neighborhood, and the local temperature standard deviation is obtained from the square root of this variance. Substituting into the adaptive local variance mapping gain operator Adjusting the differential component of high-frequency radiation Local weights are superimposed to enhance the output thermal image. .

[0066] The seepage feature extraction module receives enhanced thermal images. Then, construct the standard deviation. A 5×5 pixel discrete Gaussian partial derivative convolution kernel group is used to perform convolution in the horizontal, vertical, and diagonal directions to obtain the horizontal second-order partial derivatives of the candidate pixels. Vertical second-order partial derivatives with cross partial derivatives Based on this, a pixel-level Hessian matrix is ​​established and its characteristic equation is solved. The maximum absolute eigenvalue is then calculated using the trace and determinant of the Hessian matrix. with the smallest absolute eigenvalue The two eigenvalues ​​satisfy: The one with the larger modulus value is denoted as . The smaller of the modulus values ​​is denoted as Combined with the prevention of zero constant Calculate the thermal anisotropy index: For isotropically heated granite and homogeneous dry soil, the second-order partial derivatives in the horizontal and vertical directions are close, and the cross derivatives approach zero. The calculated... The value is in the range of 1 to 1.8; the seepage thermal texture extending along the pores of the humic medium has a large curvature in the normal direction and is relatively gentle in the tangential direction. Its principal eigenvalue magnitude is more than 4 times that of the secondary eigenvalue. This result is jointly determined with the contrast activation response output by three concentric ring multi-kernel differential operators with radii of 2 pixels, 5 pixels, and 9 pixels. and At the same time, retain the corresponding pixel and output the connected component mask of the candidate fire point. .

[0067] The temporal radiation discrimination module uses the centroids of candidate connected components. Centered on a central point, eight radial rays are sampled outward at 45-degree intervals. Temperature sequences along these rays are obtained through linear interpolation, and the first derivative magnitude is calculated. The variance and mean of the gradient magnitudes in each direction are statistically analyzed to obtain the gradient symmetry compensation coefficient. The smoothing constant of 0.1K / pixel has the same dimensions as the mean gradient magnitude. Its value is determined based on the small pseudo-gradient calibration formed by the background noise of the infrared sensor. The calculation result is set to a zero lower bound. When the variance is greater than the sum of the mean and the constant, the output is locked to zero. The judgment threshold of 0.35 is determined based on the discrete distribution of asymmetric pseudo-hot spots and real fire spots: the eight-directional gradient discrete variance of the rock surface illuminated by unidirectional side light reaches more than 0.7 times the mean, and the resulting gradient symmetry compensation coefficient is distributed in the range of 0.15 to 0.31; when the underground fire spot propagates outward, the gradient magnitude in each direction is relatively balanced, and the variance is usually less than 0.5 times the mean, and the resulting compensation coefficient is distributed in the range of 0.68 to 0.89.

[0068] For those that pass the gradient symmetry check and whose peak temperature falls within the range determined by the dynamic factor The candidate connected regions of the spatially adaptive dual-threshold initial screening interval are selected using the surface thermal emissivity. With Boltzmann constant Calculate the net radiative heat flux density and perform spatial integration, then combine it with the smoldering radiative efficiency ratio. The endogenous heat release power is obtained. Within a 40-second observation window, the instantaneous total radiated power is sampled at 1-second intervals, and the time-series derivative is obtained using least-squares linear fitting. ;when and At that time, the corresponding candidate connected region is confirmed as a hidden smoldering point, and the centroid pixel coordinates are transmitted to the three-dimensional spatial positioning module.

[0069] The three-dimensional spatial positioning module extracts the maximum angular velocity threshold in a single step. Airborne smooth attitude matrix after amplitude limiting and filtering Rotary matrix installed with pod The camera's normalized ray vector is converted into a spatial ray unit vector in the geodetic station's coordinate system. The Digital Surface Model (DSM) employs a two-level hierarchical bounding box acceleration structure, based on the geodetic coordinates of the camera's optical center. Calculate the distance to the first intersection point with the outer bounding box of the target terrain region along the unit vector of the spatial ray. Distance from the point of launch ,by As the initial step depth scalar for the Newton-Raphson numerical iteration, for the grid cell where the current projection point of the ray is located, the elevation of the cell and its four adjacent grid nodes is extracted, and the local continuous elevation surface is reconstructed by bilinear surface interpolation. At the horizontal projection coordinates of the current step point, the interpolated elevation is calculated according to the bilinear weights, and the partial derivatives of the interpolation basis function with respect to the horizontal coordinates are obtained to obtain the continuous elevation partial derivatives in the eastward and northward directions, which are used as local terrain gradient components to participate in the calculation of residual derivatives.

[0070] The elevation intersection residual function is: Its first derivative is determined by subtracting the inner product of the horizontal ray component and the local terrain elevation gradient from the ray's celestial component. The partial derivative of the local terrain slope is obtained in real time by the DSM grid center difference. The single-step iteration update amount is limited to... Within 0.01m; when the absolute value of the step difference between two consecutive iterations is less than 0.02m or the absolute value of the residual is less than 0.02m, the iteration is terminated and a convergence depth scalar is output. When the line of sight passes through a continuous undulating canyon and generates multiple grid intersection solutions, each intersection solution is pushed into a positive distance priority queue. The first intersection point with the smallest depth scalar is selected along the positive propagation direction of the ray, and the three-dimensional geographic coordinates of the fire point are calculated accordingly.

[0071] After inputting the three-dimensional geographic coordinates of the fire point into the interactive multi-model Kalman filter algorithm, two kinematic sub-models, one for constant speed and the other for coordinated turning, are used for temporal trajectory smoothing. When the cruise field of view is obscured by dense smoke, the confidence of the target in continuous image frames decreases, or the infrared signal is interrupted, the measurement update residual injection is stopped and the Kalman gain is set to zero. The state vector is extrapolated based on the previous state transition matrix to maintain the continuity of the fire point spatial trajectory. During the measurement update interruption, a preset process noise covariance matrix is ​​accumulated in each extrapolation cycle, so that the state covariance increases with the extrapolation time. The position process noise variance of the constant speed kinematic sub-model is 0.05 m² / s², and the velocity process noise variance is 0.02 m² / s³. When measurements are continuously lost, the position and velocity components of the covariance matrix are accumulated cycle by cycle according to the corresponding process noise intensity.

[0072] When the smoke disperses with the local wind field and less than 15 frames are lost in continuous observation, the current extrapolated spatial coordinates are back-projected onto the focal plane of the thermal image. Within a local search area of ​​32 pixels × 32 pixels around the predicted pixel location, the bidirectional normalized cross-correlation matching coefficient is calculated by calling the multi-scale connected domain morphology template of the fire point locked before occlusion. When the maximum matching coefficient reaches 0.82 and the ratio of peak value to subpeak value is greater than 1.35, the fire point is re-anchored, and the three-dimensional geographic coordinates of the earth are recalculated using the current rematched pixels. The coordinates are then input into the Kalman filter to complete the covariance reset and state update.

[0073] After completing the state update, the 3D centroid drift rate and geometric convex hull area expansion rate of the connected fire point group are statistically analyzed per unit time, while local wind direction and speed vectors are also fused. In addition, elevation, slope, and aspect data extracted from the digital surface model of the land surface, and elevation, slope, and aspect data of the fire location in the form of topographic elevation gradient vectors. The diffusion vector is represented by the following formula: The system organizes the three-dimensional centroid drift rate, geometric convex hull area expansion rate, and diffusion vector into a diffusion trend matrix according to continuous time intervals, and determines the fire diffusion level based on the matrix. At the same time, it deduces the position of the fire front along the main diffusion direction corresponding to the diffusion vector, determines the coordinates of the recommended firebreak cutting position, and outputs an early warning signal containing the fire diffusion level, the spread velocity vector, and the recommended firebreak cutting position.

[0074] Example 4: This example combines Figures 1 to 2 This paper describes a method and system for detecting hidden fire points based on UAV thermal imaging, such as... Figure 1As shown, in step S1, the geometrically registered thermal image sequence after multi-source sensor registration is obtained, the low-frequency temperature substrate component and the high-frequency radiation differential component are separated, the gain coefficient is adjusted according to the low-frequency local variance, and the enhanced thermal image is generated by superposition. Step S2 is completed. The two-dimensional second-order partial derivatives of the pixels in the enhanced thermal image are calculated to construct the Hessian matrix. The maximum and minimum absolute eigenvalues ​​are obtained by decomposition, and the thermal permeation anisotropy index is calculated. The contrast activation response is calculated using the concentric ring difference operator. When the contrast activation response is positive and the thermal permeation anisotropy index is not less than the preset truncation threshold, a candidate fire point connected component mask is generated. Step S3 is completed. The candidate fire point connected component is calculated. The net radiative heat flux density of the pixels within the mask is spatially integrated to solve for the instantaneous total radiative power. The temporal derivative of the instantaneous total radiative power is extracted. When the endogenous heat release power derived from the instantaneous total radiative power is not less than a preset critical threshold and the temporal derivative is not less than a preset continuous heat release threshold, the hidden smoldering fire point is identified. When the temporal derivative is less than a preset decay threshold, the heat storage false target is filtered out. Step S4 is completed. Based on the airborne smooth attitude matrix and the intrinsic parameter matrix of the thermal imaging sensor, the centroid pixel coordinates of the hidden smoldering fire point are converted into a spatial ray unit vector. The intersection coordinates of the spatial ray unit vector and the digital surface model of the ground are calculated, and the three-dimensional geographic coordinates of the earth are output.

[0075] like Figure 2 As shown, the input is an enhanced thermal image. The system determines whether the contrast activation response is positive and whether the thermal anisotropy index is not less than a preset cutoff threshold. If the result is negative, large-area homogeneous heating regions are suppressed. If the result is positive, candidate fire point connected component masks are generated. For each generated candidate fire point connected component mask, the system determines whether its local peak temperature is between the smoldering floating threshold and the open flame absolute cutoff threshold. If the result is negative, the mask is removed. If the result is positive, the system determines whether the corresponding connected component is unipolar. Divergent asymmetric pseudo-hotspots; if the corresponding connected region is determined to be a unipolar divergent asymmetric pseudo-hotspot, then the initial confidence level is reduced and the pseudo-hotspot is removed. If the corresponding connected region is determined to be a non-unipolar divergent asymmetric pseudo-hotspot, then the instantaneous total radiation power is calculated. After calculating the instantaneous total radiation power, it is determined whether the endogenous heat release power is not less than the preset critical threshold and whether the time derivative is not less than the preset continuous heat release threshold. If the determination result is yes, then the hidden smoldering point is confirmed. When the time derivative is less than the preset attenuation threshold, the thermal storage pseudo-target is filtered out.

[0076] Example 5: Before deploying the detection system to the forest patrol site, a reference surface containing fallen leaf humus and sun-exposed bare rocks was selected in the takeoff area to collect thermal radiation data. The airborne computing platform received continuous reference frames collected by the infrared thermal imager and statistically analyzed the average background temperature of ground features within the window. 296.2K, temperature standard deviation The temperature was 1.6K, and the peak surface temperature of the bare rock heated by sunlight was measured. The dynamic factor is calculated by dividing the difference between the peak surface temperature and the average background temperature of ground features by the temperature standard deviation. The value is 299.56 K. The value is in the range of 1.85 to 2.30. Substituting this dynamic factor into the parameter relationship, we obtain... Based on this, the floating threshold for concealed smoldering was determined. This prevents false hotspots caused by local differences in sunlight from entering the subsequent feature extraction process.

[0077] During continuous drone patrols, the ground cover alternates between dense canopies, sparse shrubs, and bare soil. The heat map sequence enhancement module continuously calculates the temperature skewness coefficient and kurtosis coefficient of non-high-temperature background pixels in the current field of view within a 64×64 pixel spatiotemporal sliding window, and adjusts the dynamic factor online accordingly. When the drone moves from the canopy area into sparse bare ground and the background temperature skewness coefficient increases, the proportion of locally heated bare soil in the current field of view increases accordingly. The system increases the dynamic factor to the range of 2.20 to 2.30 in steps of 0.05, raising the smoldering floating threshold to suppress background noise from sun-exposed bare rocks. After the drone re-enters a large area of ​​homogeneous dense forest, the temperature distribution returns to a Gaussian normal and the skewness coefficient decreases. The dynamic factor is then adjusted to the range of 1.85 to 1.95 to respond to the weak temperature rise signal attenuated by the deep humus layer. The dynamic factor is thus continuously adjusted according to the changes in background temperature distribution during flight.

[0078] When there are mechanical mounting tolerances in the airborne thermal imaging sensor, inertial measurement unit, and positioning module, spatial geometric mapping mismatch will occur. Before takeoff, the system establishes a calibration field containing four high-contrast ground thermal targets with geodetic coordinates. The thermal imager collects the pixel coordinates of the centroid of thermal radiation of each target under different heading and pitch attitudes. The intrinsic parameter matrix of the thermal imaging sensor and the vector of the airborne rigid body mounting lever arm are calculated by minimizing the residual of spatial multi-line intersection. At the same time, the pod installation rotation matrix is ​​obtained. The angle correction value is installed, and the calibration result is back-projected onto the digital surface model (DSM) for verification. When the geometric residual of the planar position is less than 0.08m, the corresponding parameter matrix is ​​fixed to correct the effect of aerial attitude sway and terrain undulation on the unit vector of spatial light. The resulting projection bias is used in subsequent ray stepping numerical iterations to obtain the elevation intersection depth scalar. And the three-dimensional geographic coordinates of the fire point.

[0079] Example 6: During the inspection, localized plume diffusion can cause asymmetric dissipation of infrared radiation. The temporal radiation discrimination module extracts the eight-directional radial temperature gradient modulus of the centroid of the candidate fire point connected domain mask and calculates the gradient symmetry compensation coefficient. ,when At that time, it was treated as an asymmetric pseudo-hotspot with unipolar divergence, and the initial confidence level was adjusted. Perform linear decay, with decay weights taken as follows: The ratio to 0.35 is based on actual measurements. , Substitute and adjust the confidence level as follows: The corrected confidence level of 0.52 is not lower than the preset confidence threshold of 0.4, so the candidate connected component is retained. The linear decay relationship is proportionally corrected to the initial confidence level according to the degree of asymmetric distortion. When the confidence level is below 0.35 but above 0.20, the candidate region may correspond to the local gradient distortion formed after the real fire point is blocked by the unidirectional airflow plume. The continuous attenuation method is used to retain the conditions for further identification. The pre-set confidence threshold of 0.4 is calibrated by the field multi-source smoke interference test: the corrected confidence level of the real fire point after unilateral diffuse attenuation by the plume is distributed in the range of 0.42 to 0.58. The corrected confidence levels of the rock surface under unilateral sunlight and artificial metal reflective false targets are both below 0.32.

[0080] The time derivative of the instantaneous total radiated power Falling within the hysteresis transition range between the preset attenuation threshold of -0.05 W / s and the sustained heat release threshold of -0.015 W / s, the system extended the staring observation time window from 40 seconds to 60 seconds. The instantaneous total radiant power of the thermal storage rock sample continuously decreased from 210.5 W to 207.2 W. The calculated value after the extended observation period was... The obtained time derivative is -0.055W / s, which is lower than the preset decay threshold of -0.05W / s. The system outputs a filtering rejection command to remove the thermal storage rock sample from the candidate fire points.

[0081] The abrupt changes in the complex wind field caused the airborne measured angular velocity to continuously reach the single-step maximum angular velocity threshold. At that time, the attitude smoothing filter module calculates the airborne smoothed attitude matrix according to the limited angular velocity. After the measured angular velocity falls below the threshold, a damping difference compensation term with a decay step size of discrete sampling period is introduced to compensate for the phase lag of attitude calculation accumulated during multi-frame clipping.

[0082] In the dynamic tracking and filtering phase, the surface covered by red pine needles is used as the corresponding working condition, and the wind direction sensitivity constant is adopted according to the surface combustible material load. and terrain slope sensitivity constant Local wind speed The elevation, slope, and aspect data extracted from the digital surface model correspond to the slope modulus. Substituting this into the diffusion model yields the scalar value of the propagation velocity: Based on a 300-second timescale, the displacement of the fire front is... This displacement is used to determine the position of the fire front within the extrapolated timescale, and its spatial direction is determined by the two-dimensional vector synthesis of the wind field and the terrain.

[0083] In the calculation of the spread trend, the eastward and northward unit components of the wind direction on the horizontal tangent plane are first extracted and multiplied by the wind speed scalar and the wind direction sensitivity constant, respectively, to obtain the two-dimensional velocity vector components driven by the wind field. Then, the eastward and northward slope partial derivatives of the elevation grid of the fire point location along the slope rise direction are extracted to form the two-dimensional slope vector corresponding to the elevation slope and aspect data. This vector is then multiplied by the terrain slope sensitivity constant to obtain the terrain-driven two-dimensional velocity vector components. The two components are added together on the eastward and northward directions of the horizontal tangent plane to obtain the spread vector. The fire spread azimuth angle is determined based on the eastward and northward components of this vector. The three-dimensional centroid drift rate, the geometric convex hull area expansion rate, and this spread vector are written into the spread trend matrix at continuous time intervals. The fire front geodetic mapping position at the corresponding extrapolation time is obtained by extrapolating 42.54m along the spread vector direction. Based on this, the system determines the fire spread level and the recommended isolation zone cutoff position, and generates an early warning command set by generating the geodetic coordinates of the fire point center and sending it to the ground terminal.

[0084] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A method for detecting hidden fire points based on UAV thermal imaging, characterized in that, Includes the following steps: Step S1: Obtain the geometric registration heatmap sequence after multi-source sensor registration; separate... The low-frequency temperature base component and the high-frequency radiation differential component are superimposed to generate an enhanced thermal image by adjusting the gain coefficient based on the low-frequency local variance. Step S2: Calculate the two-dimensional second-order partial derivatives of the pixels in the enhanced thermal image to construct the Hessian matrix, decompose to obtain the maximum and minimum absolute eigenvalues, and calculate the thermal infiltration anisotropy index. The contrast activation response is calculated using a concentric ring difference operator. When the contrast activation response is positive and the thermal infiltration anisotropy index is not less than a preset cutoff threshold, a candidate fire point connected domain mask is generated. Step S3: Calculate the net radiative heat flux density of the pixels within the candidate fire point connected domain mask and perform spatial integration to solve for the instantaneous total radiative power. Extract the temporal derivative of the instantaneous total radiation power. When the endogenous heat release power derived from the instantaneous total radiation power is not less than a preset critical threshold and the temporal derivative is not less than a preset continuous heat release threshold, the hidden smoldering point is identified. When the temporal derivative is less than a preset decay threshold, the thermal storage false target is filtered out. Step S4: Based on the airborne smooth attitude matrix and the intrinsic parameter matrix of the thermal imaging sensor, the centroid pixel coordinates of the concealed smoldering fire point are converted into spatial ray unit vectors. The intersection coordinates of the spatial ray unit vectors and the digital surface model of the ground are calculated, and the three-dimensional geographic coordinates of the ground are output.

2. The method for detecting concealed fire points based on UAV thermal imaging according to claim 1, characterized in that, Step S1 involves acquiring the geometric registration thermal image sequence after multi-source sensor registration and overlaying it to generate an enhanced thermal image. This includes: Step S11, unifying the access to the hardware time reference source, using the line exposure trigger signal of the thermal imaging sensor as a reference, performing cubic Hermite polynomial interpolation on the airborne positioning coordinates and three-axis attitude angles, combining the intrinsic parameter matrix of the thermal imaging sensor and the rigid body mounting lever arm vector, mapping the discrete infrared field of view to the tangent plane of the reference geodetic coordinate system, and generating the geometric registration thermal image sequence; Step S12, using anisotropic guided edge-preserving filtering based on structural tensors to separate the low-frequency temperature substrate component and the high-frequency radiation differential component; statistically analyzing the local temperature standard deviation of the low-frequency temperature substrate component within a 15-pixel × 15-pixel neighborhood, using the sum of the local temperature standard deviation and a preset constant of 0.5 as the divisor, and the scaling constant of 2.4 as the dividend, and adding 1.0 to the quotient to obtain the adaptive gain coefficient; using the adaptive gain coefficient to amplify the high-frequency radiation differential component and overlaying it onto the low-frequency temperature substrate component.

3. The method for detecting concealed fire points based on UAV thermal imaging according to claim 1, characterized in that, Step S2 uses a concentric ring differential operator to calculate the contrast activation response, including: Step S21, constructing a three-layer concentric ring multi-kernel differential operator with radii of 2 pixels, 5 pixels and 9 pixels respectively; calculating the contrast step response between the average radiance of the inner core region and the concentric ring mask regions of the outer layers, and taking the minimum contrast step response at the three-layer scale as the contrast activation response of the central pixel; when the contrast activation response is greater than zero and the thermal permeation anisotropy index is not less than the preset truncation threshold, confirming that the corresponding pixel has active permeation texture features, preserving local hot spots and outputting candidate fire point connected domain masks and corresponding initial confidence scores; suppressing large-area homogeneous heated regions with non-monotonic convex contrast step responses at each scale.

4. The method for detecting concealed fire points based on UAV thermal imaging according to claim 3, characterized in that, Before solving for the instantaneous total radiation power, step S3 also includes gradient symmetry compensation for the candidate fire point connected domain mask: step S31, calculate the eight-directional radial temperature gradient modulus of the candidate fire point connected domain mask centered on the centroid. The variance and mean of the radial temperature gradient magnitude in each direction are statistically analyzed. The sum of the mean and the set constant 0.1 is used as the divisor, and the variance is used as the dividend to calculate the gradient dispersion ratio. The gradient symmetry compensation coefficient is obtained by subtracting the gradient dispersion ratio from 1.

0. When the gradient symmetry compensation coefficient is less than 0.35, the corresponding connected region is determined to be a unipolar divergent asymmetric pseudo-hot spot, and the initial confidence level is reduced. Candidate connected components whose initial confidence level is lower than the preset confidence threshold are removed.

5. The method for detecting concealed fire points based on UAV thermal imaging according to claim 1, characterized in that, Step S3 calculates the net radiative heat flux density of pixels within the candidate fire point connected region mask and performs spatial integration to solve for the instantaneous total radiative power, including: Step S32, based on the Stefan-Boltzmann law, using the Earth's surface thermal emissivity of 0.94 and the Boltzmann constant of 5.67 × 10⁻⁶. -8 W·m -2 ·K -4 The net radiative heat flux density is calculated by dividing the fourth power of the candidate connected region pixel temperature by the fourth power of the ambient background temperature. The ground coverage area of ​​a single pixel is calculated based on the flight altitude and the instantaneous field of view of the sensor, and the instantaneous total radiative power is obtained by summing the data across the entire region. The endogenous heat release power is calculated based on the smoldering radiative efficiency ratio of 0.15 to 0.22, and combined with the standard lower flammability value of 1.82 × 10⁻⁶ for combustible materials. 7 The mass loss rate per unit time is calculated using J / kg; the preset critical threshold is 15.0W, corresponding to a minimum mass loss rate of 8.24 × 10⁻⁶. -7 kg / s, with a preset continuous heat release threshold of -0.015W / s and a preset decay threshold of -0.05W / s.

6. The method for detecting concealed fire points based on UAV thermal imaging according to claim 1, characterized in that, Step S3 also includes preliminary screening of candidate fire point connected component masks based on spatial adaptive dual threshold criteria: Step S33, statistically analyze the average background temperature of ground objects and the standard deviation of neighborhood temperatures within the sliding window, and set the absolute cutoff threshold for open flames to 373.15K; add the average background temperature of ground objects to the product of the dynamic factor and the standard deviation of neighborhood temperatures to form the hidden smoldering floating threshold, where the dynamic factor ranges from 1.85 to 2.30; only candidate connected components with local peak temperatures between the hidden smoldering floating threshold and the absolute cutoff threshold for open flames are input into the subsequent instantaneous total radiant power spatial integral and temporal derivative discrimination process.

7. The method for detecting concealed fire points based on UAV thermal imaging according to claim 1, characterized in that, In step S4, the centroid pixel coordinates of the concealed smoldering fire point are converted into a spatial ray unit vector based on the airborne smooth attitude matrix and the intrinsic parameter matrix of the thermal imaging sensor. This includes: Step S41, extracting the instantaneous measured angular velocities of yaw, pitch, and roll angles, and setting the single-step maximum angular velocity threshold to 15.0° / s; when the instantaneous measured angular velocity in any axis exceeds the single-step maximum angular velocity threshold, the current sampled value is replaced with the single-step maximum angular velocity threshold, and the smooth attitude matrix is ​​calculated; the centroid pixel coordinates are converted into ray vectors in the camera normalized coordinate system based on the intrinsic parameter matrix of the thermal imaging sensor; the ray vectors are projected onto the geocentric coordinate system and normalized by installing the rotation matrix and smooth attitude matrix on the pod, generating a spatial ray unit vector.

8. The method for detecting concealed fire points based on UAV thermal imaging according to claim 1, characterized in that, Step S4 involves calculating the coordinates of the intersection point between the spatial ray unit vector and the digital surface model, including: Step S42, projecting a probe ray along the spatial ray unit vector using the geodetic coordinates of the camera's optical center, traversing the elevation grid patches of the digital surface model using a hierarchical bounding box, and constructing an elevation intersection residual equation; using a numerical iteration method to solve for the ray forward step distance parameter, and substituting it into the ray equation to obtain the geodetic coordinates of the intersection point; when multiple intersection points are generated in undulating canyon terrain, using the forward distance priority queue arbitration criterion, uniquely locking the first intersection point with the smallest spatial Euclidean distance along the line of sight forward direction, and eliminating false intersection points that penetrate the preceding ridge terrain.

9. A method for detecting concealed fire points based on UAV thermal imaging according to claim 1, characterized in that, Following step S4, the process also includes dynamic tracking and situation analysis of fire points: In step S43, the calculated three-dimensional geographic coordinates of the earth are input into the interactive multi-model Kalman filter algorithm to smooth the temporal trajectory and remove jump noise; the three-dimensional centroid drift rate and geometric convex hull area expansion rate of the connected fire point group are statistically analyzed per unit time, and the local wind direction and speed vectors are fused with the elevation, slope and aspect data extracted from the digital surface model to fit the fire spread trend matrix within the surface cover; when the cruise field of view is obscured by dense smoke, resulting in the loss of less than 15 consecutive observations, the virtual coordinates of the fire points are extrapolated based on the previous inertial filtering to maintain the continuity of the tracking trajectory. Once observations are restored, the fire point is re-matched and anchored, and an early warning signal containing the fire spread level and recommended firebreak location is output.

10. A fire concealment detection system based on UAV thermal imaging, used to implement the fire concealment detection method based on UAV thermal imaging as described in claim 1, characterized in that, It includes a heat map sequence enhancement module, a seepage feature extraction module, a temporal radiation discrimination module, and a three-dimensional spatial positioning module; The thermal image sequence enhancement module is used to acquire the geometrically registered thermal image sequence after multi-source sensor registration; it separates the low-frequency temperature substrate component and the high-frequency radiation differential component, adjusts the gain coefficient according to the low-frequency local variance, and superimposes them to generate an enhanced thermal image; The seepage feature extraction module, connected to the heat map sequence enhancement module, is used to calculate the two-dimensional second-order partial derivatives of pixels in the enhanced heat image to construct the Hessian matrix, decompose to obtain the maximum and minimum absolute eigenvalues, and calculate the thermal seepage anisotropy index. The contrast activation response is calculated using a concentric ring difference operator. When the contrast activation response is positive and the thermal infiltration anisotropy index is not less than a preset cutoff threshold, a candidate fire point connected domain mask is generated. The temporal radiation identification module, connected to the seepage feature extraction module, is used to calculate the net radiative heat flux density of pixels within the candidate fire point connected domain mask and perform spatial integration to solve for the instantaneous total radiative power; extract the temporal derivative of the instantaneous total radiative power; confirm the hidden smoldering fire point when the inverted endogenous heat release power is not less than the preset critical threshold and the temporal derivative is not less than the preset continuous heat release threshold; and filter out the thermal storage false targets when the temporal derivative is less than the preset decay threshold. The three-dimensional spatial positioning module, connected to the temporal radiometric discrimination module, is used to convert the centroid pixel coordinates of the concealed smoldering fire point into a spatial ray unit vector based on the airborne smooth attitude matrix and the intrinsic parameter matrix of the thermal imaging sensor. It then calculates the intersection coordinates of the spatial ray unit vector and the digital surface model of the earth, and outputs the three-dimensional geographic coordinates of the earth.

Citation Information

Patent Citations

  • A method, apparatus and system for fire point localization based on unmanned aerial vehicles (UAVs)

    CN111982291B