A road collapse hidden danger unmanned aerial vehicle exploration early warning system and method

Abstract

This invention discloses a UAV-based detection and early warning system and method for potential road collapse hazards, relating to the field of road infrastructure safety monitoring and intelligent early warning technology. It includes a UAV image acquisition module that acquires infrared image data of the road area and associates the infrared image data with acquisition time information and spatial location information to organize it into time-series infrared image data. In this invention, infrared image data with a binding relationship between acquisition time information and spatial location information is acquired by the UAV, and continuous time-series infrared image data is constructed based on this. This achieves unified organization of infrared information in the time and spatial dimensions during road inspection, providing a reliable data foundation for subsequent stability analysis. It effectively avoids the problem of missing time correlation in traditional single-frame analysis. By introducing a task variable modeling image correction mechanism, the infrared images can maintain stable and consistent radiation characteristics even under complex inspection conditions.

Description

Technical Field

[0001] This invention relates to the field of road infrastructure safety monitoring and intelligent early warning technology, and in particular to a UAV-based early warning system and method for detecting potential road collapse hazards. Background Technology

[0002] Road collapse, as a sudden and hidden infrastructure safety risk, often evolves gradually from factors such as underground cavities, pipeline damage, and soil erosion. In its early stages, it lacks obvious visual features at the surface level, and traditional methods relying on manual inspections or localized point detection are insufficient to detect potential hazards in a timely manner. With the continuous expansion of urban road scale and the increasing complexity of underground space structures, how to conduct continuous, wide-area, and timely monitoring of road collapse hazards without interfering with traffic operations or damaging the road surface structure has become an urgent technical problem to be solved in the field of road safety management.

[0003] Current methods for detecting road subsidence hazards primarily rely on ground-based radar detection, manual drilling sampling, and fixed-point sensor monitoring. Ground-based radar, limited by detection efficiency and coverage, struggles to support large-scale, high-frequency inspections; manual drilling is destructive and costly, making sustained implementation difficult; fixed-point monitoring only reflects local conditions and cannot comprehensively assess road safety. Existing solutions are mostly based on single-frame or short-time-series analysis, making them susceptible to environmental changes, flight attitude, and sensor noise, resulting in insufficient stability. Furthermore, current technologies lack sustained constraints on anomalies over time and spatial consistency analysis of cross-modal data. Risk assessment often relies on single indicators or thresholds, failing to accurately reflect the true risk of road subsidence hazards. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a UAV-based early warning system and method for detecting and preventing road collapse hazards.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a road collapse hazard detection and early warning system and method using unmanned aerial vehicles (UAVs), comprising: a UAV image acquisition module, which acquires infrared image data of a road area and associates the infrared image data with acquisition time information and spatial location information to organize and form time-series infrared image data; a task variable modeling and image correction module, which extracts brightness distribution, thermal radiation stability, and spatial noise characteristics from the time-series infrared image data, constructs an initial task variable parameter set by combining UAV flight status information and environmental condition information, performs self-updating processing on the initial task variable parameter set to form a dynamic task variable parameter set, and performs radiometric consistency correction, spatial noise suppression, and time-series stabilization processing on the time-series infrared image data based on the dynamic task variable parameter set to form corrected infrared image data; and an infrared anomaly confidence generation module. The system comprises the following modules: a module for generating anomaly confidence image data and anomaly uncertainty image data based on the spatial distribution characteristics and variation amplitude of thermal features in calibrated infrared image data; a module for temporal consistency analysis that constructs the anomalous structural features of thermal anomaly regions based on calibrated infrared image data, anomaly confidence image data, and anomaly uncertainty image data, and forms candidate area image data for collapse hazards based on the degree of preservation of the anomalous structural features in the time dimension; a module for cross-modal image consistency constraints that maps the underground anomaly image data corresponding to the candidate areas for collapse hazards to the infrared image spatial coordinate system, forming image consistency constraint result data; and a module for risk assessment and early warning generation that generates road collapse hazard risk image data and early warning information data based on the candidate area image data for collapse hazards, anomaly confidence image data, and image consistency constraint result data.

[0006] As a further description of the above technical solution: The UAV image acquisition module uses the UAV to collect infrared image data of the road area frame by frame during the inspection process. When collecting each frame of infrared image data, the corresponding acquisition time information and spatial location information are written synchronously. The acquisition time information and spatial location information are bound to the corresponding infrared image data at the frame level. The bound infrared image data is subjected to integrity checks, and image frames with missing acquisition time information or missing spatial location information are removed. The infrared image data that passes the integrity check are sorted according to the acquisition time information. The sorted infrared image data is organized into a continuous image sequence according to the acquisition time order. The continuous image sequence, together with the corresponding acquisition time information and spatial location information, is organized to form time-series infrared image data.

[0007] As a further description of the above technical solution: The task variable modeling image correction module reads infrared image data frame by frame from time-series infrared image data, maintaining the binding relationship between the acquisition time information and spatial location information and the corresponding infrared image data. For each frame of infrared image data, it calculates the brightness distribution statistics, thermal radiation stability statistics, and spatial noise statistics, and organizes the calculation results into image statistical features. It reads the UAV flight status information and environmental condition information corresponding to each frame of infrared image data, and organizes the reading results into an initial task variable parameter set. Based on the image statistical features, it constructs statistical deviation data corresponding to the target's stable statistical state, performs noise-sensitive suppression processing on the statistical deviation data, and introduces noise-sensitive suppression during the processing. The following steps are taken: First, noise suppression bias data is generated by controlling the noise suppression bias coefficients. Then, task variable parameters are updated based on this data, with a self-updating convergence coefficient introduced during the update process, forming a dynamic task variable parameter set. Next, radiometric consistency correction is performed on this dynamic task variable parameter set to generate radiometrically corrected image data. Then, spatial noise suppression is performed on the radiometrically corrected image data based on the dynamic task variable parameter set to generate denoised image data. Finally, temporal stabilization is performed on the denoised image data based on the dynamic task variable parameter set to generate the current frame correction result. The correction results for each frame are then organized according to the acquisition time information, maintaining the binding relationship between spatial location information and the corresponding correction results, to form corrected infrared image data.

[0008] As a further description of the above technical solution: The infrared anomaly confidence generation module reads infrared image data frame by frame from the calibrated infrared image data, maintaining the binding relationship between the acquisition time information and spatial location information and the corresponding infrared image data. For each frame of calibrated infrared image data, it calculates brightness distribution statistics, thermal radiation stability statistics, and spatial noise statistics, and organizes the calculation results into image statistical features. It extracts thermal feature values ​​from the calibrated infrared image data corresponding to the same spatial location at adjacent acquisition time points, and calculates thermal feature variation amplitude data based on these values. It constructs spatial distribution data of thermal features based on the thermal feature values ​​in each frame of calibrated infrared image data, and calculates spatial consistency data based on this spatial distribution data. Finally, it correlates the thermal feature variation amplitude data with the spatial distribution data. Consistent data are jointly organized to form candidate anomaly intensity data. Based on the candidate anomaly intensity data and combined with thermal radiation stability statistics, a confidence assignment process is performed to generate anomaly confidence image data that spatially corresponds one-to-one with the corrected infrared image data. Based on spatial noise statistics and combined with thermal characteristic change amplitude data, an uncertainty assignment process is performed to generate anomaly uncertainty image data that spatially corresponds one-to-one with the corrected infrared image data. The anomaly confidence image data and anomaly uncertainty image data are organized according to the acquisition time information sequence, while maintaining the binding relationship between spatial location information and corresponding image frames, to form anomaly confidence image data and anomaly uncertainty image data that are consistent with the time sequence of the corrected infrared image data.

[0009] As a further description of the above technical solution: The temporal consistency analysis module reads infrared image data frame by frame from the calibrated infrared image data, maintaining the binding relationship between the acquisition time information and spatial location information and the corresponding infrared image data; it also reads anomalous confidence image data frame by frame from the anomalous confidence image data, maintaining the binding relationship between the acquisition time information and spatial location information and the corresponding anomalous confidence image data; and it reads anomalous uncertainty image data frame by frame from the anomalous uncertainty image data, maintaining the binding relationship between the acquisition time information and spatial location information and the corresponding anomalous uncertainty image data. Within each frame of anomalous confidence image data, it locates high-confidence spatial locations and extracts spatially consistent thermal anomaly regions from the corresponding frame of calibrated infrared image data, forming candidate anomalous region data. It then performs region connectivity organization processing on each frame of candidate anomalous region data, forming a candidate anomalous region set data. Finally, it extracts spatial morphological description information and boundary structure description information from each candidate anomalous region in the candidate anomalous region set data, and organizes the extraction results into heterogeneous... The process involves several steps: First, establishing consistent structural features. Then, calculating the degree of structural preservation for anomalous structural features at spatially consistent locations between adjacent acquisition time points, forming a structural preservation sequence data. Next, performing memory-based recursive aggregation on the structural preservation sequence data, introducing a structural preservation memory coefficient to form overall structural preservation result data. Finally, performing uncertainty suppression processing on the overall structural preservation result data based on anomalous uncertainty image data, introducing an adaptive uncertainty suppression coefficient to form suppressed structural preservation result data. Based on the suppressed structural preservation result data, performing consistency judgment processing to eliminate spatial locations where structural preservation results do not meet stability requirements, retaining stable anomalous location data that meet stability requirements. Then, performing region connectivity organization processing on the stable anomalous location data to form a set of stable anomalous regions. Finally, mapping the set of stable anomalous regions to image data of candidate areas for collapse hazards, and organizing them according to the acquisition time information to form image data of candidate areas for collapse hazards.

[0010] As a further description of the above technical solution: The cross-modal image consistency constraint module reads the candidate image data of the collapse hazard candidate area frame by frame from the collapse hazard candidate area image data, and maintains the binding relationship between the acquisition time information and spatial location information and the corresponding collapse hazard candidate area image data. The module also reads the underground anomaly image data corresponding to the collapse hazard candidate area image data, and maintains the binding relationship between the spatial location information and the corresponding underground anomaly image data. Spatial registration processing is performed on the underground anomaly image data, and the underground anomaly image data is mapped to the infrared image spatial coordinate system to form mapped underground anomaly image data. Spatial overlap calculation processing is performed on the collapse hazard candidate area image data and the mapped underground anomaly image data to form anomaly spatial location consistency data. Anomaly structure distribution alignment processing is performed on the collapse hazard candidate area image data and the mapped underground anomaly image data to form anomaly structure distribution consistency data. The anomaly spatial location consistency data and the anomaly structure distribution consistency data are jointly organized to form image consistency constraint result data. The image consistency constraint result data is bound to the corresponding acquisition time information and spatial location information, and organized according to the acquisition time information order to form image consistency constraint result data.

[0011] As a further description of the above technical solution: The risk assessment and early warning generation module reads candidate collapse hazard image data frame by frame from the candidate collapse hazard image data, maintaining the binding relationship between the acquisition time information and spatial location information and the corresponding candidate collapse hazard image data. The risk assessment and early warning generation module also reads abnormal credibility image data frame by frame from the abnormal credibility image data, maintaining the binding relationship between the acquisition time information and spatial location information and the corresponding abnormal credibility image data. Furthermore, the risk assessment and early warning generation module reads the image consistency constraint result data corresponding to the candidate collapse hazard image data, maintaining the binding relationship between the acquisition time information and spatial location information and the corresponding image consistency constraint result data. It locates the spatial range of the candidate area in each frame of candidate collapse hazard image data and extracts spatially consistent abnormal credibility area data from the corresponding frame of abnormal credibility image data. Finally, it performs regional-level statistical analysis on the abnormal credibility area data to form a regional credibility index corresponding to the candidate area. The process involves: reading image consistency constraint result data corresponding to the spatial range of candidate regions and organizing this data into candidate region consistency indices; performing weight adjustment on the region credibility indices based on these indices to form weighted region credibility indices; extracting abnormal structure preservation characteristics within the spatial range of candidate regions and organizing the results into candidate region structure preservation indices; jointly organizing the weighted region credibility indices and candidate region structure preservation indices to form candidate region risk intensity data; performing risk intensity mapping on the candidate region risk intensity data and writing the mapping results to the corresponding spatial location in the infrared image to form road collapse hazard risk image data; performing risk level classification on the candidate region risk intensity data to form candidate region risk level data; and binding the candidate region risk level data with spatial location information and organizing it according to the acquisition time sequence to form early warning information data.

[0012] As a further description of the above technical solution: A method for detecting and warning of potential road collapse hazards using drones, comprising the following steps: During inspections, drones are used to collect infrared image data of road areas frame by frame. Simultaneously, corresponding acquisition time and spatial location information are written into the data. The acquisition time and spatial location information are then bound to the corresponding infrared image data at the frame level, and infrared image data lacking acquisition time or spatial location information is discarded. The bound infrared image data is sorted according to the acquisition time information to form continuous time-series infrared image data. Brightness distribution statistics, thermal radiation stability statistics, and spatial noise statistics are calculated frame by frame for the time-series infrared image data to form image statistical features. The drone flight data corresponding to the infrared image data is then read. Status and environmental condition information are organized into an initial set of task variable parameters. Statistical deviation data is constructed based on image statistical features, and noise-sensitive suppression processing is performed on the statistical deviation data, introducing a noise-sensitive suppression coefficient to form noise-suppressed deviation data. The task variable parameters are updated based on the noise-suppressed deviation data, introducing a self-updating convergence coefficient to form a dynamic set of task variable parameters. Radiometric consistency correction, spatial noise suppression, and temporal stabilization processing are then performed on the temporal infrared image data based on the dynamic set of task variable parameters to form corrected infrared image data. Based on the spatial distribution characteristics and variation amplitude of thermal features in the corrected infrared image data, an anomaly confidence image is generated. This process involves using data and image data with anomalies and uncertainties. High-confidence spatial locations are identified within the anomaly confidence image data, and spatially consistent thermal anomaly regions are extracted from the calibrated infrared image data to form candidate anomaly region data. Spatial morphological and boundary structure descriptions are extracted from these candidate anomaly region data to form anomaly structure features. Temporal consistency analysis is performed based on the retention of these features between adjacent acquisition time points, introducing a structure-preserving memory coefficient. Uncertainty suppression is then performed using the anomaly uncertainty image data to form stable anomaly region data. This stable anomaly region data is organized into candidate collapse hazard region image data. Corresponding underground anomaly image data is read and mapped to the infrared image spatial coordinate system. Spatial location consistency analysis and anomaly structure distribution consistency analysis are performed between the candidate collapse hazard image data and the mapped underground anomaly image data to form image consistency constraint result data. Based on the anomaly confidence image data, image consistency constraint result data, and candidate collapse hazard region image data, risk intensity data corresponding to the candidate regions is generated. This risk intensity data is mapped to the infrared image space to form road collapse hazard risk image data. Early warning information data associated with the risk level is generated based on this risk intensity data.

[0013] The present invention has the following beneficial effects: 1. In this invention, infrared image data with a binding relationship between acquisition time information and spatial location information is first acquired by a UAV. On this basis, continuous temporal infrared image data is constructed, realizing the unified organization of infrared information in the time and spatial dimensions during road inspection. This provides a reliable data foundation for subsequent stability analysis and effectively avoids the problem of missing time correlation in traditional single-frame analysis. By introducing a task variable modeling image correction mechanism, the task variable parameter set is dynamically updated based on image statistical features, flight status information, and environmental condition information. Radiometric consistency correction, spatial noise suppression, and temporal stabilization processing are performed on the infrared image, so that the infrared image can still maintain stable and consistent radiometric characteristics under complex inspection conditions, significantly reducing the impact of environmental changes, attitude fluctuations, and noise interference on the detection results.

[0014] 2. In this invention, during the anomaly analysis stage, thermal characteristic change amplitude data and spatial consistency data are constructed simultaneously. Combined with thermal radiation stability statistics and spatial noise statistics, anomaly credibility image data and anomaly uncertainty image data are generated. This distinguishes stable anomalies from random disturbances at the pixel level, improving the reliability of infrared anomaly identification results. By performing cross-time structure preservation analysis on the anomaly structural features and introducing a structure preservation memory coefficient for recursive aggregation, the preservation characteristics of the anomaly structure in the time series are fully characterized. This effectively identifies long-term stable anomalies caused by underground structural problems, avoiding the misjudgment of instantaneous thermal anomalies as potential collapse hazards. Furthermore, by incorporating anomaly uncertainty information, an uncertainty adaptive suppression coefficient is introduced to suppress the structure preservation results, effectively controlling the impact of high uncertainty regions on the final judgment result, thereby improving the robustness of stable anomaly judgment.

[0015] 3. In this invention, a cross-modal image consistency constraint mechanism is used to perform consistency analysis on the candidate areas of potential collapse hazards obtained by infrared inspection and the underground anomaly image data at the spatial location and structural distribution levels. This strengthens the correspondence between surface thermal anomalies and underground anomalies, significantly improving the overall credibility of the collapse hazard judgment. In the risk judgment and early warning stage, the anomaly credibility, structural preservation characteristics, and cross-modal consistency results are comprehensively utilized to perform regional-level risk intensity calculation and risk level classification on the candidate areas. This enables the generated road collapse hazard risk images and early warning information to more accurately reflect the actual risk level and enhances the system's adaptability to complex road scenarios. Attached Figure Description

[0016] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Reference Figure 1-2 This invention provides an embodiment of a road collapse hazard detection and early warning system and method using unmanned aerial vehicles (UAVs), comprising: a UAV image acquisition module, which acquires infrared image data of a road area and associates the infrared image data with acquisition time information and spatial location information to organize it into time-series infrared image data; a task variable modeling and image correction module, which extracts brightness distribution, thermal radiation stability, and spatial noise features from the time-series infrared image data, constructs an initial task variable parameter set by combining UAV flight status information and environmental condition information, performs self-updating processing on the initial task variable parameter set to form a dynamic task variable parameter set, and performs radiometric consistency correction, spatial noise suppression, and time-series stabilization processing on the time-series infrared image data based on the dynamic task variable parameter set to form corrected infrared image data; and an infrared anomaly confidence generation module. Based on the spatial distribution characteristics and variation amplitude of thermal features in the corrected infrared image data, anomaly confidence image data and anomaly uncertainty image data are generated. The temporal consistency analysis module constructs the anomalous structural features of the thermal anomaly region based on the corrected infrared image data, anomaly confidence image data, and anomaly uncertainty image data, and forms candidate area image data for collapse hazards based on the degree of preservation of the anomalous structural features in the time dimension. The cross-modal image consistency constraint module maps the underground anomaly image data corresponding to the candidate areas for collapse hazards to the infrared image spatial coordinate system to form image consistency constraint result data. The risk judgment and early warning generation module generates road collapse hazard risk image data and early warning information data based on the candidate area image data for collapse hazards, anomaly confidence image data, and image consistency constraint result data.

[0019] In this embodiment, the UAV covers the road area according to a preset flight path during the inspection process, and collects infrared image data of the road area frame by frame at fixed time intervals during the flight.

[0020] While acquiring each frame of infrared image data, the UAV simultaneously records the acquisition time information corresponding to that frame of infrared image data, and simultaneously records the spatial location information corresponding to that frame of infrared image data. The spatial location information is consistent with the current flight position of the UAV.

[0021] The system performs frame-level binding processing on the acquisition time information, spatial location information and corresponding infrared image data, so that each frame of infrared image data corresponds to unique acquisition time information and spatial location information.

[0022] After completing frame-level binding, the system performs an integrity check on all bound infrared image data, determining frame by frame whether the infrared image data simultaneously contains acquisition time information and spatial location information.

[0023] For infrared image data that lacks acquisition time information or spatial location information, the system will remove the corresponding image frame from the subsequent processing flow, and only retain the infrared image data that passes the integrity check.

[0024] The system sorts the infrared image data that has passed the integrity check according to the corresponding acquisition time information, so that the infrared image data forms a continuous and orderly arrangement in the time dimension.

[0025] After sorting by time, the system organizes the sorted infrared image data into a continuous image sequence according to the acquisition time order.

[0026] The system organizes the continuous image sequence with the acquisition time information and spatial location information corresponding to each frame of infrared image data in a unified manner, so that the time dimension and spatial location information are consistent at the sequence level, thereby forming time-series infrared image data.

[0027] In this embodiment, the task variable modeling image correction module reads infrared image data frame by frame from the temporal infrared image data in the order of acquisition time, and maintains the binding relationship between the acquisition time information and spatial location information and the corresponding infrared image data without change during the reading process.

[0028] For each frame of infrared image data read, the system calculates the brightness distribution statistics within the image space to quantify the overall distribution of pixel brightness in the infrared image; at the same time, it calculates the thermal radiation stability statistics to characterize the thermal radiation change state of the infrared image at the current acquisition time point; and it calculates the spatial noise statistics to characterize the distribution characteristics of random noise in the infrared image in the spatial dimension.

[0029] The system organizes brightness distribution statistics, thermal radiation stability statistics, and spatial noise statistics in a unified manner to form image statistical features corresponding to the current frame of infrared image data.

[0030] While acquiring image statistical features, the system reads the UAV flight status information and environmental condition information corresponding to the infrared image data of the current frame, and organizes the UAV flight status information and environmental condition information to form an initial task variable parameter set corresponding to the current frame.

[0031] The system constructs statistical deviation data corresponding to the target's stable statistical state based on image statistical features, so that the statistical deviation data characterizes the degree of deviation between the current frame infrared image data and the stable state.

[0032] For statistical deviation data, the system performs noise sensitivity suppression processing. During the processing, a noise sensitivity suppression coefficient is introduced to suppress fluctuations caused by spatial noise in the statistical deviation data, resulting in noise-suppressed deviation data.

[0033] The system updates the task variable parameters in the initial task variable parameter set based on noise suppression deviation data, and introduces a self-updating convergence coefficient during the update process to control the update magnitude of the task variable parameters, thus forming a dynamic task variable parameter set corresponding to the current frame.

[0034] After obtaining the set of dynamic task variable parameters, the system performs radiometric consistency correction processing on the current frame infrared image data based on the set of dynamic task variable parameters, corrects the radiometric differences in the infrared image caused by changes in acquisition conditions, and forms radiometrically corrected image data.

[0035] Subsequently, the system performs spatial noise suppression processing on the radiometrically corrected image data based on the dynamic task variable parameter set, suppressing the residual spatial noise in the image to form denoised image data.

[0036] After completing spatial noise suppression, the system performs temporal stabilization processing on the denoised image data based on the dynamic task variable parameter set, so that the current frame image data is consistent with the adjacent frames in the time dimension, forming the current frame correction result.

[0037] The system organizes the current frame correction results corresponding to each acquisition time point in the order of acquisition time information, and maintains the binding relationship between spatial location information and corresponding correction results without change during the organization process, thereby forming corrected infrared image data.

[0038] In this embodiment, the task variable modeling image correction module reads infrared image data frame by frame from the time-series infrared image data according to the order of the acquisition time information identifier. During the reading process, it synchronously reads the acquisition time information and spatial location information bound to the infrared image data of that frame, and keeps the binding relationship between the acquisition time information and spatial location information and the corresponding infrared image data unchanged.

[0039] For the currently read single-frame infrared image data, all valid pixels are traversed within the image pixel space. The brightness value of each pixel is statistically summarized, and a brightness distribution statistic is calculated based on the pixel brightness distribution. This statistic is used to characterize the overall brightness distribution of the infrared image frame. The formula for calculating the brightness distribution statistic is as follows: , This is a luminance distribution statistic used to characterize the overall luminance level of the current frame of the infrared image. For the first The brightness value of each effective pixel. This represents the number of valid pixels in the current frame of the infrared image.

[0040] After calculating the brightness distribution statistics, based on the thermal radiation values ​​of each pixel in the current frame of infrared image data, the changes in thermal radiation between adjacent pixels are statistically analyzed, and the concentration of thermal radiation fluctuations within the same frame is quantified to form thermal radiation stability statistics. The formula for calculating thermal radiation stability statistics is as follows: , This is a statistic on thermal radiation stability, used to characterize the concentration of thermal radiation variations within the same frame. For the first The thermal radiation value of each pixel In order to be with the first Thermal radiation values ​​of adjacent pixels in a pixel space. This represents the number of pixel pairs that participate in the adjacent pixel pair statistics.

[0041] Subsequently, multiple spatial locations were selected in the current frame of infrared image data to statistically analyze the local fluctuations in pixel brightness and thermal radiation values. Random fluctuation components were summarized and analyzed to calculate spatial noise statistics, which characterize the spatial distribution of noise in the current frame of infrared image. The formula for calculating spatial noise statistics is as follows: , This is a spatial noise statistic used to characterize the noise distribution intensity in the current frame of the infrared image. For the first The intensity of random fluctuations of pixels within a spatial region The number of spatial regions is determined.

[0042] The brightness distribution statistics, thermal radiation stability statistics, and spatial noise statistics are organized in a unified manner to form image statistical features that correspond one-to-one with the current frame of infrared image data.

[0043] After completing the image statistical feature construction, the UAV flight status information bound to the current frame infrared image data is read, and the flight attitude, flight altitude and motion state description information of the UAV at the current acquisition time are extracted.

[0044] Simultaneously, environmental condition information bound to the current frame of infrared image data is read, and the content reflecting changes in external acquisition conditions in the environmental condition information is organized.

[0045] The drone's flight status information and environmental condition information are jointly organized to form an initial set of task variable parameters corresponding to the current frame of infrared image data.

[0046] Based on the image statistical features corresponding to the current frame, a comparison is made with the stable statistical state of the target. The deviations of these image statistical features in brightness distribution, thermal radiation stability, and spatial noise dimensions are calculated to form statistical deviation data. The formula for calculating statistical deviation data is as follows: , This is statistical deviation data, used to characterize the degree of deviation of the current frame's statistical state from the target's stable statistical state. As a reference value for the target stable brightness distribution, The target is a stable reference value for thermal radiation stability. The target stable spatial noise statistical reference value.

[0047] For statistical bias data, noise sensitivity suppression processing is performed. During this process, a noise sensitivity suppression coefficient is introduced to weaken the random disturbance component reflected by spatial noise statistics in the statistical bias data, while retaining the stability bias component, thus obtaining noise-suppressed bias data. The formula for calculating noise-suppressed bias data is as follows: , This is noise suppression bias data. This is the original statistical deviation data. This is the noise sensitivity suppression coefficient.

[0048] Read the spatial noise statistics from the image statistical features corresponding to the current frame of infrared image data, and simultaneously read the spatial noise statistics corresponding to adjacent acquisition time points, keeping the acquisition time information in the same order. When reading, locate adjacent acquisition time points according to the acquisition time information, and write the spatial noise statistics of the two frames into the same calculation cache. The spatial noise statistics of the current frame are compared with those of adjacent acquisition time points to calculate the difference, thus obtaining the temporal fluctuation amplitude of the spatial noise. The difference calculation is performed based on the numerical correspondence between the spatial noise statistics of the two frames, using absolute difference calculation. The difference result is recorded as the temporal fluctuation result corresponding to the current frame. The formula for calculating the temporal fluctuation amplitude of spatial noise is as follows: , This represents the amplitude of spatial noise fluctuations over time. This represents the spatial noise statistics at the current data collection time point. This represents the spatial noise statistics for the adjacent previous data collection time point.

[0049] The spatial range corresponding to the current frame infrared image data is divided into multiple spatial regions, and the spatial noise statistics are calculated for each spatial region. The spatial region is divided into grids according to the image spatial coordinates so that each spatial region covers a continuous set of pixels. The random fluctuation components of the pixel values ​​in each spatial region are statistically summarized to obtain the spatial noise statistics corresponding to that spatial region. Discreteness calculations are performed on the spatial noise statistics corresponding to each spatial region to obtain the spatial noise dispersion in the spatial dimension. The dispersion calculation is performed by performing distribution dispersion statistics on the numerical set of spatial noise statistics for each spatial region, and the dispersion statistics results are recorded as the spatial dispersion results corresponding to the current frame; the formula for calculating the spatial dispersion of spatial noise is as follows: , The degree of dispersion of spatial noise in spatial dimensions, For the first Spatial noise statistics for each spatial region This represents the average value of spatial noise statistics for all spatial regions.

[0050] The spatial noise intensity level of the current frame is formed by jointly normalizing the fluctuation amplitude of spatial noise in the temporal dimension and the dispersion in the spatial dimension. The joint normalization process performs range normalization on both the temporal fluctuation result and the spatial dispersion result, and then performs joint organization on the normalized results to obtain a single intensity scalar. The formula for calculating the spatial noise intensity level is as follows: , The spatial noise intensity level is used to drive the generation of the noise sensitivity suppression coefficient. The weights are a combination of time and space dimensions. This is the normalization function.

[0051] Monotonic mapping is performed based on the spatial noise intensity level to obtain the noise sensitivity suppression coefficient corresponding to the spatial noise intensity level. The monotonic mapping process generates the corresponding coefficient value according to the numerical value of the spatial noise intensity level, and keeps the coefficient value from decreasing as the spatial noise intensity level increases. During the system initialization phase, multiple frames of corrected infrared image data generated during historical inspections are read, and the spatial noise statistics corresponding to each frame are summarized. Perform distribution statistics on the aggregated spatial noise statistics to extract the maximum and minimum stable ranges of spatial noise statistics in historical data; The noise suppression intensity corresponding to the maximum and minimum stable value ranges of the spatial noise statistics is taken as the maximum and minimum allowable values ​​of the noise sensitivity suppression coefficient. The maximum and minimum allowable values ​​are written into the range constraint rules of the noise sensitivity suppression coefficient, serving as the upper and lower bounds of the noise sensitivity suppression coefficient value; The noise sensitivity suppression coefficient is subjected to value range constraint processing to form a noise sensitivity suppression coefficient for statistical bias data processing. The value range constraint processing truncates the coefficient value according to the upper and lower bounds, and the constrained coefficient is written into the noise sensitivity suppression process of the current frame. The formula for calculating the noise sensitivity suppression coefficient is as follows: , This is the noise sensitivity suppression coefficient. It is a monotonic mapping function that satisfies When it increases No reduction.

[0052] Using noise suppression deviation data as the basis for updating, each task variable parameter in the initial task variable parameter set is adjusted item by item. During the task variable parameter update process, a self-updating convergence coefficient is introduced to constrain the magnitude of each parameter adjustment, so that the change process of the task variable parameters remains continuous. After the parameter update is completed, a dynamic task variable parameter set corresponding to the current frame infrared image data is formed.

[0053] Read the noise suppression deviation data corresponding to the current frame and the noise suppression deviation data corresponding to the previous acquisition time point, keeping the acquisition time information in the same order. When reading, locate the previous acquisition time point according to the acquisition time information and write the noise suppression deviation data of the two frames into the same calculation cache. The noise suppression deviation data of the current frame is compared with the noise suppression deviation data of the previous acquisition time point to calculate the deviation change magnitude. The difference calculation is performed according to the numerical correspondence between the two frames of noise suppression deviation data, and the difference result is recorded as the deviation change result corresponding to the current frame. The formula for calculating the deviation change magnitude is: , This represents the magnitude of the noise suppression bias over time. This is the noise suppression deviation data at the current acquisition time point. This is the noise suppression deviation data from the previous acquisition time point.

[0054] Read the thermal radiation stability statistics corresponding to the current frame and the thermal radiation stability statistics corresponding to the previous acquisition time point. When reading, locate the corresponding frame according to the acquisition time information and write the thermal radiation stability statistics of the two frames into the same calculation cache. The difference between the thermal radiation stability statistics of the current frame and the thermal radiation stability statistics of the previous acquisition time point is calculated to obtain the magnitude of the change in thermal radiation stability. The difference calculation is performed by performing an absolute difference calculation on the thermal radiation stability statistics of the two frames, and the difference result is recorded as the stability change result corresponding to the current frame; the formula for calculating the magnitude of the change in thermal radiation stability is as follows: , This represents the magnitude of change in thermal radiation stability over time. This is the statistical measure of thermal radiation stability at the current data collection time. This is the thermal radiation stability statistic for the previous data collection point.

[0055] The deviation variation amplitude and the thermal radiation stability variation amplitude are jointly normalized to form an updated instability index. The joint normalization process performs range normalization on both the deviation and stability variation results, and then performs joint organization on the normalized results to obtain a single instability scalar. The updated instability index calculation formula is as follows: , To update the instability index, This represents the combined weight of deviation changes and stability changes.

[0056] Read the initial task variable parameter set corresponding to the current frame and the dynamic task variable parameter set corresponding to the previous acquisition time point. During reading, keep the field order of the parameter items consistent and write the two sets of parameter sets into the same calculation cache. Perform item-by-item difference calculations on the initial and dynamic task variable parameter sets to obtain the parameter update magnitude level. The item-by-item difference calculation performs numerical and absolute value processing on a one-to-one correspondence between parameter items, and then summarizes the difference results of each parameter item to form a single magnitude scalar. The formula for calculating the parameter update magnitude level is: , Update the amplitude level of the parameter. For the current data collection time point The values ​​that can be taken from each task variable parameter. The parameter value corresponds to the previous data collection time point. This represents the number of task variable parameters.

[0057] The instability index and the parameter update magnitude are jointly mapped to form a self-updating convergence coefficient. The joint mapping process generates the corresponding coefficient values ​​according to the constraint that the coefficient values ​​do not increase when the instability index or the parameter update magnitude increases. During the system initialization phase, the system reads the set of dynamic task variable parameters generated during the historical inspection process, and simultaneously reads the noise suppression deviation data and thermal radiation stability statistics of the corresponding frames. Joint statistics were performed on the parameter update magnitude, noise suppression deviation change magnitude, and thermal radiation stability change magnitude in historical data. We selected time periods with small parameter update magnitudes, small noise suppression deviation changes, and small thermal radiation stability changes, as well as time periods with large parameter update magnitudes, large noise suppression deviation changes, or large thermal radiation stability changes. The mapping results of the update intensity of the task variable parameters within the above time period are used as the maximum and minimum allowable values ​​of the self-update convergence coefficient; The maximum and minimum allowed values ​​are written into the value range constraint rules of the self-updating convergence coefficient, serving as the upper and lower bounds of the self-updating convergence coefficient value; The self-updating convergence coefficients are subjected to value range constraints to form self-updating convergence coefficients used in the task variable parameter update process. The value range constraints truncate the coefficient values ​​according to the upper and lower bounds, and the constrained coefficients are written into the task variable parameter update process of the current frame. The formula for calculating the self-updating convergence coefficients is as follows: , For self-updating convergence coefficients, It is a joint mapping function that satisfies the condition that the coefficient values ​​do not increase when the instability or parameter update magnitude increases.

[0058] Based on a dynamic set of task variable parameters, radiometric consistency correction is performed on the current frame of infrared image data.

[0059] In this process, based on the correction parameters reflected in the dynamic task variable parameter set, the radiometric response of each pixel in the infrared image is adjusted to ensure that the current frame of the infrared image is consistent with that of adjacent frames in terms of radiometric properties, thus forming radiometrically corrected image data. Task variable parameter update formula: , For the updated task variable parameters, The values ​​are assigned to the corresponding parameters in the initial task variable parameter set. For self-updating convergence coefficients.

[0060] After obtaining the radiometrically corrected image data, spatial noise suppression processing is performed on the radiometrically corrected image data based on the dynamic task variable parameter set. High-noise regions reflected by spatial noise statistics are identified within the image space, and pixel fluctuations in these regions are suppressed. Stable regions retain their original characteristics, forming denoised image data. Based on the dynamic task variable parameter set, temporal stabilization processing is performed on the denoised image data. In this process, the denoised image data of the current frame is aligned with the denoised image data of adjacent acquisition time points, and the positions of pixels with abrupt changes in the time dimension are adjusted to keep the current frame image smooth and continuous in the time series, forming the current frame correction result.

[0061] The current frame correction results corresponding to each acquisition time point are arranged in the order of the acquisition time information identifier.

[0062] During the arrangement process, the binding relationship between each frame's correction result and its corresponding spatial location information remains unchanged, ultimately forming complete corrected infrared image data.

[0063] In this embodiment, the infrared anomaly confidence generation module receives the corrected infrared image data and reads the infrared image data frame by frame according to the time sequence corresponding to the acquisition time information, while keeping the binding relationship between the acquisition time information, spatial location information and the corresponding infrared image data unchanged.

[0064] After reading each frame of infrared image data, the brightness distribution statistics are first calculated within the image space of that frame. During the calculation process, the brightness values ​​at each spatial location in that frame are summarized and statistically analyzed, and the statistical results are written into the brightness distribution statistics record corresponding to that frame.

[0065] After completing the calculation of the brightness distribution statistics, the thermal radiation stability statistics are calculated within the image space of the infrared image data of that frame. During the calculation, adjacent frames of the infrared image data of that frame are aligned and read in time sequence, and the thermal characteristic values ​​of adjacent frames at the same spatial location are statistically analyzed. The statistical results of the differences are written into the thermal radiation stability statistics record corresponding to that frame.

[0066] After completing the calculation of thermal radiation stability statistics, the spatial noise statistics are calculated within the image space of the infrared image data of that frame. During the calculation, the local neighborhood of each spatial location in the frame is scanned, and the degree of fluctuation of thermal characteristic values ​​in the local neighborhood is statistically analyzed. The fluctuation statistics results are written into the spatial noise statistics record corresponding to that frame.

[0067] The brightness distribution statistics, thermal radiation stability statistics, and spatial noise statistics are organized at the frame level so that each frame of corrected infrared image data corresponds to a set of image statistical features, and the binding relationship between this set of image statistical features and the acquisition time information and spatial location information remains unchanged.

[0068] After forming the image statistical features, thermal feature value extraction is performed on the corrected infrared image data corresponding to the same spatial location at adjacent acquisition time points. During the extraction process, spatial location information is used as a consistency constraint. Thermal feature values ​​are read at the same spatial location in two adjacent frames, and the reading results are written into the thermal feature value sequence record.

[0069] The thermal characteristic change amplitude data is calculated based on the thermal characteristic value sequence record. During the calculation process, the difference calculation is performed on the thermal characteristic values ​​of adjacent acquisition time points, and the difference calculation result is written into the thermal characteristic change amplitude data of the corresponding spatial location. At the same time, the binding relationship between the thermal characteristic change amplitude data and the acquisition time information and spatial location information remains unchanged.

[0070] In each frame of corrected infrared image data, thermal feature spatial distribution data is constructed based on thermal feature values. During the construction process, the image spatial grid is used as the organization method. The thermal feature values ​​at each spatial location of the frame are written into the corresponding grid positions according to the spatial location information, forming thermal feature spatial distribution data that corresponds one-to-one with the frame.

[0071] Spatial consistency data is calculated based on the spatial distribution data of thermal features. During the calculation process, the local neighborhood of each spatial location is traversed, the thermal feature values ​​of the spatial location and its local neighborhood are read, and the degree of aggregation of thermal feature values ​​in the local neighborhood is measured for consistency. The consistency measurement results are written into the spatial consistency data of the corresponding spatial location, while keeping the binding relationship between the spatial consistency data and the acquisition time information and spatial location information unchanged.

[0072] Thermal feature variation amplitude data and spatial consistency data are jointly organized to form candidate anomaly intensity data. During the joint organization process, spatial location information is used as the key to write thermal feature variation amplitude data and spatial consistency data corresponding to the same spatial location into the same record structure. The record structure is then expanded according to the spatial location of the image to form a spatial distribution representation of the candidate anomaly intensity data.

[0073] Based on candidate anomaly intensity data and combined with thermal radiation stability statistics, a confidence assignment process is performed. During the assignment process, candidate anomaly intensity data is read at each spatial location, and thermal radiation stability statistics of the corresponding frame at that spatial location are read simultaneously. The two are jointly calculated at the same spatial location to obtain the confidence assignment result for that spatial location. The confidence assignment result is then written into the anomaly confidence image data at the corresponding spatial location, so that the anomaly confidence image data and the corrected infrared image data spatially correspond one-to-one.

[0074] Uncertainty assignment processing is performed based on spatial noise statistics and thermal feature variation amplitude data. During the assignment process, thermal feature variation amplitude data is read at each spatial location, and spatial noise statistics of the corresponding frame at that spatial location are read simultaneously. The two are jointly calculated at the same spatial location to obtain the uncertainty assignment result for that spatial location. The uncertainty assignment result is then written into the abnormal uncertainty image data at the corresponding spatial location, so that the abnormal uncertainty image data and the corrected infrared image data spatially correspond one-to-one.

[0075] The abnormal confidence image data and abnormal uncertainty image data are organized at the frame level according to the acquisition time information. During the organization process, the binding relationship between the spatial location information and the corresponding image frame is kept unchanged, so that the abnormal confidence image data and abnormal uncertainty image data are consistent with the corrected infrared image data in time sequence.

[0076] In this embodiment, the timing consistency analysis module reads infrared image data frame by frame from the calibrated infrared image data according to the order of the acquisition time information identifiers. During the reading process, it synchronously reads the acquisition time information and spatial location information bound to the infrared image data of that frame. During the reading process, the binding relationship between the acquisition time information and spatial location information and the corresponding infrared image data remains unchanged.

[0077] After completing the reading of the corrected infrared image data, the abnormal confidence image data is read frame by frame from the abnormal confidence image data according to the same acquisition time information order, and the binding relationship between the acquisition time information and spatial location information and the corresponding abnormal confidence image data is kept unchanged during the reading process.

[0078] Following the same acquisition time information sequence as described above, abnormal uncertainty image data is read frame by frame from the abnormal uncertainty image data, and the binding relationship between the acquisition time information, spatial location information and the corresponding abnormal uncertainty image data is kept unchanged during the reading process.

[0079] After reading the abnormal credibility image data corresponding to the same acquisition time point, all spatial locations within the image space range of the abnormal credibility image data of that frame are traversed one by one, and spatial locations that meet the predetermined credibility screening conditions are located based on the numerical values ​​recorded in the abnormal credibility image data. The located spatial locations are then written into the high credibility spatial location set of the current frame.

[0080] For each high-confidence spatial location that has been located, in the corrected infrared image data at the corresponding acquisition time point, according to the consistency constraint of spatial location information, the infrared image data of that spatial location and its neighborhood range are read, and the infrared image data within that spatial range are organized into a thermal anomaly region to obtain candidate anomaly region data corresponding to the current frame.

[0081] After the candidate anomaly region data is generated, region connectivity organization processing is performed on all candidate anomaly region data in the current frame. In the region connectivity organization processing, candidate anomaly regions that are adjacent to each other or have overlapping relationships are merged according to spatial adjacency, and the merged results are organized to form a set of candidate anomaly region data.

[0082] For each candidate anomaly region in the candidate anomaly region set data, spatial morphological description information is extracted within the spatial range of the candidate anomaly region. The characteristics of the candidate anomaly region in terms of spatial scale, regional outline and area distribution are quantitatively recorded. At the same time, the boundary structure description information of the candidate anomaly region is extracted. The continuity of the boundary and the changes in the boundary shape of the candidate anomaly region are recorded. The spatial morphological description information and the boundary structure description information are jointly organized to form anomaly structure features that correspond one-to-one with the candidate anomaly region.

[0083] After constructing the abnormal structural features, the structural preservation degree of abnormal structural features with consistent spatial locations at adjacent acquisition time points is calculated according to the acquisition time information sequence. During the structural preservation degree calculation, the differences in spatial morphological description information and boundary structure description information of corresponding abnormal structural features in two adjacent frames are quantified, forming a structural preservation sequence data reflecting the changes of abnormal structures over time. The formula for quantifying the differences in abnormal structural features is as follows: , Collection time point Below, spatial location The structural difference quantity corresponding to the abnormal structural features. Collection time point Below, spatial location Spatial morphological description information, Collection time point Below, spatial location Boundary structure description information, The previous data collection time point adjacent to the current data collection time point. Structural difference measurement is used to quantify differences in descriptive information. The formula for calculating structure-preserving sequence data is as follows: , Collection time point Below, spatial location The structure preserves the values ​​of the sequence data. : Anomalous structural feature differences : Normalization function, used to map differences to a uniform scale.

[0084] After obtaining the structure-preserving sequence data, a memory-based recursive aggregation process is performed. During this process, the structure-preserving sequence data is read sequentially according to the acquisition time information. A structure-preserving memory coefficient is introduced during aggregation to recursively influence the structure-preserving results at historical time points, thus forming the overall structure-preserving result data. The formula for calculating the overall structure-preserving result is as follows: , Collection time point The overall structure below maintains the result data. The overall structure of the data from the previous data collection point is preserved. The current structure preserves the sequence data. : Structure retention memory coefficient.

[0085] Read the structure-preserving sequence data in the order of acquisition time information, locate the structure-preserving sequence data value corresponding to the current acquisition time point, and locate the structure-preserving sequence data value corresponding to the previous acquisition time point to maintain consistency of spatial location information; The absolute difference between the current acquisition time point and the previous acquisition time point is calculated to obtain the structural stability change range relative to the previous acquisition time point. This difference calculation is performed separately at each spatial location. The formula for calculating the structural stability change range is as follows: , Spatial location At the time of collection The structure maintains the same range of change as the previous data collection point. , The values ​​of the structure-preserving sequence data corresponding to adjacent acquisition time points.

[0086] Read the acquisition time information corresponding to the current acquisition time point and the previous acquisition time point, calculate the time interval between the two acquisition time points, and use the time interval to perform time normalization processing on the structural change amplitude to obtain the change amplitude per unit time scale; the formula for calculating the time-normalized structural change amplitude is as follows: , The structure maintains its range of change over a unit time scale. The time interval between the current data collection point and the previous data collection point.

[0087] At the same spatial location, continue reading the structure-preserved sequence data values ​​corresponding to the previous acquisition time point. Calculate the change amplitude between the current acquisition time point and the previous acquisition time point, and between the previous acquisition time point and the acquisition time point before that. Perform a joint comparison of the two change amplitudes to obtain the stability of the change at the current acquisition time point. The formula for calculating the stability of the change is: , Collection time point Below, spatial location The degree of stability of the change, : The normalized variation of the current time period : The normalized change in the previous time period.

[0088] At the same spatial location, the structure-preserved sequence data values ​​at the current acquisition time point and the structure-preserved sequence data values ​​at historical acquisition time points are statistically analyzed to obtain the historical value fluctuation range for that spatial location. Based on the historical value fluctuation range, the stability of the change is normalized. The historical fluctuation range normalization formula is as follows: , Based on the stability of changes after normalization of historical fluctuation range, , Spatial location The minimum and maximum values ​​of the stability of changes at historical data collection points.

[0089] The structure retention memory coefficient is generated based on the stability of the normalized change, so that the structure retention memory coefficient is larger when the stability of the change is high, and smaller when the stability of the change is low. The structure-preserving memory coefficients are subjected to range constraints, and values ​​outside the range are truncated to boundary values, forming the structure-preserving memory coefficients used for memory recursive aggregation processing. The formula for calculating the structure-preserving memory coefficients is as follows: , Spatial location The structure retains the memory coefficient. The stability of changes after normalization. Formula for the structural memory coefficient constraint: , The constrained structure retains the memory coefficient. , : The upper and lower bounds allowed for the structure-preserving memory coefficient.

[0090] After the overall structure preservation result data is generated, the abnormal uncertainty image data corresponding to the overall structure preservation result data is read synchronously. Based on the uncertainty level recorded in the abnormal uncertainty image data, uncertainty suppression processing is performed on the overall structure preservation result data. During the processing, an adaptive uncertainty suppression coefficient is introduced to weaken the structure preservation result corresponding to the spatial location with a high uncertainty level, thus forming the suppressed structure preservation result data.

[0091] Read the abnormal uncertainty image data in the order of acquisition time information, locate the abnormal uncertainty image data frame corresponding to the current acquisition time point, and keep the binding relationship between spatial location information and abnormal uncertainty image data frame unchanged; In the current abnormal uncertainty image data frame, all spatial locations are traversed according to spatial location information, and the uncertainty value corresponding to each spatial location is read to form a one-to-one correspondence between the spatial location and the uncertainty value of the current frame. Perform distribution statistics on all uncertainty values ​​in the current frame to obtain the central tendency and dispersion of the uncertainty values ​​in the current frame, and use the statistical results as a reference for the uncertainty scale of the current frame; At each spatial location, the uncertainty value of that location is normalized by the uncertainty scale reference of the current frame to obtain the relative uncertainty level of that spatial location; the formula for calculating the relative uncertainty level is: , Spatial location The relative level of uncertainty : The original uncertainty values ​​recorded in the abnormal uncertainty image data.

[0092] At each spatial location, the relative uncertainty level is compared with the uncertainty value at the corresponding spatial location at historical data collection time points. The fluctuation amplitude of the relative uncertainty level over time is calculated, and the fluctuation amplitude is normalized. The formula for calculating the time fluctuation amplitude of uncertainty is as follows: , Spatial location The uncertainty of time fluctuation range.

[0093] The relative uncertainty level and the normalized time fluctuation amplitude are jointly organized to generate the uncertainty adaptive suppression coefficient corresponding to the spatial location. The uncertainty adaptive suppression coefficient is larger when the relative uncertainty level is high or the time fluctuation amplitude is large, and smaller when the relative uncertainty level is low and the time fluctuation amplitude is small. The adaptive suppression coefficient for uncertainty is subjected to a range constraint, and values ​​outside the range are truncated to boundary values, forming the adaptive suppression coefficient for uncertainty suppression. The formula for calculating the adaptive suppression coefficient for uncertainty is as follows: , Spatial location Uncertainty adaptive suppression coefficient, The joint weight of relative uncertainty level and time fluctuation amplitude. Uncertainty adaptive suppression coefficient constraint formula: , : Uncertainty adaptive suppression coefficient after constraints , The upper and lower bounds of the adaptive suppression coefficient for uncertainty.

[0094] Subsequently, a consistency determination process is performed based on the suppressed structure preservation results data. The structure preservation results for each spatial location are evaluated one by one. Spatial locations whose structure preservation results do not meet the stability requirements are removed from subsequent processing, and only spatial locations whose structure preservation results meet the stability requirements are retained, forming stable anomaly location data. The formula for calculating the suppressed structure preservation results is as follows: , Structure preservation results after uncertainty suppression. Consistency determination formula: , Stability determination result identifier Threshold for determining structural stability.

[0095] After obtaining stable anomaly location data, region connectivity organization is performed on the data, aggregating stable anomaly locations according to spatial adjacency to form a set of stable anomaly regions. Rules for forming stable anomaly regions: , Collection time point The set of stable anomaly regions below, : A region connectivity organization operator based on spatial adjacency.

[0096] Finally, the stable anomaly region set data is mapped to the collapse hazard candidate region image data, and the collapse hazard candidate region image data is organized according to the acquisition time information. During the organization process, the binding relationship between spatial location information and corresponding image frames is kept unchanged, forming the output collapse hazard candidate region image data.

[0097] In this embodiment, after the cross-modal image consistency constraint module accesses the image data of the candidate area of ​​the collapse hazard and the underground anomaly image data, it first reads the image data of the candidate area of ​​the collapse hazard frame by frame according to the acquisition time information. When reading each frame, it synchronously reads the acquisition time information and spatial location information bound to that frame. During the reading process, the binding relationship between the acquisition time information, spatial location information and the corresponding image frame remains unchanged. The system reads underground anomaly image data corresponding to the current frame's candidate area image data for potential collapse hazards. During the reading process, the binding relationship between the spatial location information and the underground anomaly image data remains unchanged. The spatial location information of the underground anomaly image data and the current frame's candidate area image data for potential collapse hazards is written into the same registration cache, allowing the two types of image data to be retrieved point by point according to spatial location information within the same cache. The underground anomaly image data is acquired by underground detection equipment corresponding to the road area during inspection or detection, or is formed by reading from historical underground detection results. The underground anomaly image data has already undergone imaging processing before entering the system, and is provided to the cross-modal image consistency constraint module using spatial location information as an index.

[0098] Spatial registration processing is performed on the underground anomaly image data. In the spatial registration process, spatial location information is used as the constraint benchmark. The corresponding values ​​of the underground anomaly image data are read one by one in spatial location. At the same time, the spatial location coordinates of the current frame's candidate area image data of collapse hazard are read one by one in spatial location. The spatial location coordinates of the underground anomaly image data are transformed to the infrared image spatial coordinate system according to the spatial location information. The transformed spatial location and the values ​​of the underground anomaly image data are written into the mapping cache to form the mapped underground anomaly image data. Spatial overlap calculation processing is performed on the image data of candidate areas of collapse hazards and the mapped underground anomaly image data. In the spatial overlap calculation processing, the two types of image data are read point by point according to the spatial location information. The spatial location set corresponding to the candidate area is located in the image data of candidate areas of collapse hazards, and the set of underground anomaly spatial locations consistent with the spatial location set is located in the mapped underground anomaly image data. The intersection calculation and coverage ratio calculation are performed on the two sets at the spatial location information level. The intersection spatial location and coverage ratio are written into the consistency record to form anomaly spatial location consistency data. Anomaly structure distribution alignment processing is performed on the image data of candidate areas for collapse hazards and the mapped underground anomaly image data. In the anomaly structure distribution alignment processing, spatial location information is used as an index to extract the spatial distribution expression within the candidate area in the image data of candidate areas for collapse hazards according to the spatial range of the candidate area. In the mapped underground anomaly image data, the spatial distribution expression consistent with the spatial range of the candidate area is extracted. Then, the spatial location difference statistics and cluster consistency statistics are performed on the two types of spatial distribution expressions within the spatial range of the candidate area. The difference statistics results and cluster consistency statistics results are written into the structure distribution record to form anomaly structure distribution consistency data. The data on the consistency of abnormal spatial locations and the data on the consistency of abnormal structural distribution are jointly organized. During the joint organization process, the acquisition time information and spatial location information are used as the joint key. The values ​​of the data on the consistency of abnormal spatial locations and the values ​​of the data on the consistency of abnormal structural distribution corresponding to the same spatial location information under the same acquisition time information are written into the same joint record. The joint record is then expanded according to the spatial location information to form the image consistency constraint result data. The image consistency constraint result data is bound with the corresponding acquisition time information and spatial location information. In the binding process, the joint record of the image consistency constraint result data is read one by one, and the acquisition time information and spatial location information fields of each joint record are checked to see if they are consistent with the current frame buffer. If the check passes, it is written to the binding index; if it fails, the joint record is removed. The image consistency constraint result data after binding is organized in the order of acquisition time information. During the organization process, the joint records are sorted according to the acquisition time information, and a continuous frame sequence is formed according to the sorting result. The binding relationship between spatial location information and corresponding joint records is kept unchanged. The image consistency constraint result data organized in the order of acquisition time information is output.

[0099] In this embodiment, after the risk judgment and early warning generation module accesses the image data of the candidate area of ​​the collapse hazard, the image data of the abnormality confidence, and the image consistency constraint result data, it first reads the image data of the candidate area of ​​the collapse hazard frame by frame in the order of the acquisition time information. When reading each frame, it simultaneously reads the acquisition time information and spatial location information bound to that frame. During the reading process, the binding relationship between the acquisition time information, spatial location information, and the corresponding image frame remains unchanged.

[0100] The abnormal confidence image data is read frame by frame according to the acquisition time information consistent with the image data of the candidate area of ​​collapse hazard. The acquisition time information and spatial location information bound to the abnormal confidence image data of the frame are read synchronously. During the reading process, the binding relationship between the abnormal confidence image data and the acquisition time information and spatial location information is kept unchanged. The abnormal confidence image data under the same acquisition time information and the image data of the candidate area of ​​collapse hazard are written into the same frame-level buffer, so that the two types of image data can be read point by point according to the spatial location information.

[0101] Read the image consistency constraint result data corresponding to the image data of the candidate area of ​​collapse hazard in the current frame, and simultaneously read the acquisition time information and spatial location information bound to the image consistency constraint result data. During the reading process, keep the binding relationship between the image consistency constraint result data and the acquisition time information and spatial location information unchanged, and write the image consistency constraint result data into the same frame-level buffer so that the image consistency constraint result data can be read point by point according to the spatial location information and the image data of the candidate area of ​​collapse hazard.

[0102] In the current frame of the candidate area image data of the collapse hazard, all spatial locations are traversed according to the spatial location information to locate the set of spatial locations marked as candidate areas in the candidate area image data of the collapse hazard. Then, the set of spatial locations is processed by the region connectivity organization to aggregate the spatial locations of the candidate areas that are spatially adjacent or have a connection relationship, so as to obtain the set of candidate area spatial ranges of the current frame. Each set of candidate area spatial ranges is written into the candidate area index of the current frame.

[0103] For each candidate region spatial range in the candidate region index of the current frame, extract the abnormal confidence region data that is consistent with the spatial range of the candidate region in the abnormal confidence image data according to the spatial location information. During the extraction process, read the abnormal confidence value in the spatial range of the candidate region at each spatial location, and write the read abnormal confidence value into the region value set according to the spatial location information, so that the abnormal confidence region data and the candidate region spatial range are in one-to-one correspondence.

[0104] After generating abnormal credibility region data, regional statistical analysis is performed on the abnormal credibility region data. In the regional statistical analysis, counting statistics, summation statistics, mean statistics, extreme value statistics and discrete statistics are performed on the abnormal credibility value set within the candidate region spatial range. The statistical results are written into the regional credibility index corresponding to the candidate region. During the writing process, the binding relationship between the regional credibility index and the current candidate region spatial range remains unchanged.

[0105] In the image consistency constraint result data, extract the set of image consistency constraint result data values ​​that are consistent with the spatial range of the candidate region according to the spatial location information. During the extraction process, read the constraint values ​​within the spatial range of the candidate region at each spatial location, and write the read constraint values ​​into the region value set according to the spatial location information, so that the region value set and the candidate region spatial range are in one-to-one correspondence, and organize the region value set into the candidate region consistency index.

[0106] The regional credibility index is weighted based on the candidate regional consistency index. In the weighting process, regional statistical summarization is performed on the candidate regional consistency index to obtain the regional consistency summary result used for adjustment. The regional consistency summary result is then converted into a weighting adjustment factor. Subsequently, the regional credibility index is scaled and updated using the weighting adjustment factor to form the weighted regional credibility index. During the update process, the binding relationship between the spatial range of the candidate region and the weighted regional credibility index remains unchanged.

[0107] Anomalies in the structure preservation characteristics within the candidate region are extracted. During the extraction process, the corresponding spatial location set is located according to the spatial range of the candidate region, and the structure preservation result value corresponding to the current frame is located according to the acquisition time information. The structure preservation result value is read spatially within the candidate region and regional summary statistics are performed. The regional summary statistics results are written into the candidate region structure preservation index. During the writing process, the binding relationship between the candidate region structure preservation index and the candidate region spatial range remains unchanged.

[0108] The weighted regional credibility index and the candidate region structure preservation index are jointly organized. During the joint organization process, the two types of indicators are checked for consistency according to the spatial range of the candidate region. After the check is passed, the two types of indicators are written into the same candidate region record, and the candidate region record is fused and calculated to form candidate region risk intensity data. During the formation process, the binding relationship between the candidate region risk intensity data and the candidate region spatial range is kept unchanged.

[0109] Risk intensity mapping processing is performed on the candidate region risk intensity data. In the mapping processing, the spatial range of the candidate region is used as the writing mask. The set of spatial locations corresponding to the spatial range of the candidate region is located in the infrared image space. The risk intensity data of the candidate region is written to the image location corresponding to the set of spatial locations. Default values ​​are written to the spatial locations of non-candidate regions to form the current frame road collapse hazard risk image data. The binding relationship between the acquisition time information and spatial location information and the current frame road collapse hazard risk image data remains unchanged.

[0110] Risk level classification processing is performed on the candidate region risk intensity data. In the classification process, interval mapping is performed on the candidate region risk intensity data to map the risk intensity values ​​to discrete candidate region risk level data, and a binding relationship is established between the candidate region risk level data and the corresponding candidate region spatial range to form the current frame candidate region risk level data set.

[0111] The candidate area risk level data is bound to spatial location information. During the binding process, the spatial location information of the candidate area is written into the record field corresponding to the candidate area risk level data, and the collection time information is written into the same record. This ensures that each record contains collection time information, spatial location information, and risk level value. The records are then organized according to the order of collection time information to form early warning information data, while maintaining consistency between the early warning information data and the road collapse hazard risk image data in the order of collection time information.

[0112] Example 1: In this example, the UAV covers the target road area according to a predetermined inspection path and collects infrared image data of the road area frame by frame at fixed time intervals during the flight.

[0113] During the acquisition of each frame of infrared image data, the corresponding acquisition time information and spatial location information are synchronously written, and the acquisition time information and spatial location information are bound to the corresponding infrared image data at the frame level to form an infrared image data frame with time and spatial identifiers.

[0114] The system performs integrity checks on infrared image data that has completed frame-level binding, removes infrared image data frames that are missing acquisition time information or spatial location information, and sorts the infrared image data that has passed the integrity check according to the acquisition time information to form continuous time-series infrared image data.

[0115] The task variable modeling image correction module reads the temporal infrared image data frame by frame according to the acquisition time information, and keeps the binding relationship between the acquisition time information, spatial location information and the corresponding infrared image data unchanged.

[0116] For each frame of infrared image data, the brightness distribution statistics, thermal radiation stability statistics, and spatial noise statistics are calculated, and the three types of statistics are uniformly organized into image statistical features that correspond one-to-one with the infrared image data of that frame.

[0117] Simultaneously, the drone flight status information and environmental condition information bound to the infrared image data of that frame are read, and the two are combined to form an initial task variable parameter set.

[0118] The system constructs statistical deviation data based on image statistical features, performs noise-sensitive suppression processing on the statistical deviation data and introduces a noise-sensitive suppression coefficient to form noise-suppressed deviation data.

[0119] The system updates the initial task variable parameter set based on noise suppression deviation data, and introduces a self-updating convergence coefficient during the update process to form a dynamic task variable parameter set.

[0120] The system performs radiometric consistency correction, spatial noise suppression, and temporal stabilization on infrared image data sequentially based on a dynamic set of task variable parameters to form corrected infrared image data.

[0121] The infrared anomaly confidence generation module reads the corrected infrared image data frame by frame in the order of acquisition time, and keeps the binding relationship between acquisition time information and spatial location information unchanged.

[0122] The system extracts thermal feature values ​​from adjacent acquisition time points at the same spatial location and calculates the thermal feature change amplitude data; Simultaneously, thermal feature spatial distribution data is constructed in single-frame corrected infrared image data, and spatial consistency data is calculated based on the thermal feature spatial distribution data; The system combines thermal characteristic change amplitude data with spatial consistency data to form candidate anomaly intensity data, and combines thermal radiation stability statistics to generate anomaly confidence image data, while combining spatial noise statistics to generate anomaly uncertainty image data.

[0123] The temporal consistency analysis module is based on calibrated infrared image data, anomaly confidence image data, and anomaly uncertainty image data. It locates high-confidence spatial locations in the anomaly confidence image data and extracts the corresponding thermal anomaly regions in the calibrated infrared image data to form candidate anomaly region data.

[0124] The system performs regional connectivity organization processing on the candidate anomaly region data to form a set of candidate anomaly region data, and extracts spatial morphological description information and boundary structure description information for each candidate anomaly region to form anomaly structure features.

[0125] The system performs structure preservation calculations on abnormal structural features between adjacent acquisition time points, forming structure preservation sequence data, and introduces structure preservation memory coefficients to perform memory recursive aggregation processing to form overall structure preservation result data.

[0126] The system introduces an adaptive uncertainty suppression coefficient based on abnormal uncertainty image data, performs uncertainty suppression processing on the overall structure preservation result data, and obtains stable abnormal location data through consistency judgment. After regional connectivity organization, it forms candidate area image data for collapse hazards.

[0127] The cross-modal image consistency constraint module reads the underground anomaly image data corresponding to the image data of the candidate area of ​​the collapse hazard, and maps the underground anomaly image data to the infrared image spatial coordinate system.

[0128] The system performs consistency analysis of abnormal spatial location and consistency analysis of abnormal structural distribution based on the image data of candidate areas of potential collapse hazards and the mapped underground anomaly image data, and forms image consistency constraint result data.

[0129] The risk assessment and early warning generation module performs regional statistical analysis on the abnormal credibility image data, anomaly credibility image data, and image consistency constraint result data based on the candidate area image data of the collapse hazard, and generates regional credibility index based on the image consistency constraint result data. It then performs weight adjustment on the regional credibility index based on the image consistency constraint result data.

[0130] The system combines the abnormal structure preservation characteristics of candidate areas to form risk intensity data of candidate areas, and maps the risk intensity data to the infrared image space to generate risk image data of potential road collapse hazards, while also generating early warning information data associated with the risk level.

[0131] Example 2: In this example, the drone repeatedly inspects the same road area during multiple inspection periods, and collects infrared image data at fixed time intervals during each inspection period.

[0132] During the acquisition of each frame of infrared image data, the acquisition time information and spatial location information are synchronously written, and the two are bound to the infrared image data at the frame level.

[0133] The system performs integrity checks on the infrared image data acquired during different inspection periods, and sorts the infrared image data that passes the check according to the order of acquisition time information to form continuous time-series infrared image data for multiple time periods.

[0134] The task variable modeling image correction module performs frame-by-frame processing on the temporal infrared image data of each inspection period, calculates the brightness distribution statistics, thermal radiation stability statistics, and spatial noise statistics to form image statistical features, and constructs an initial set of task variable parameters by combining the UAV flight status information and environmental condition information.

[0135] The system introduces a noise sensitivity suppression coefficient and a self-updating convergence coefficient based on statistical deviation data, and performs self-updating processing on the task variable parameter set to form a dynamic task variable parameter set corresponding to each inspection period.

[0136] The system performs radiometric consistency correction, spatial noise suppression, and temporal stabilization on infrared image data based on a dynamic set of task variable parameters, forming multi-time-period corrected infrared image data.

[0137] The infrared anomaly confidence generation module calculates thermal characteristic change amplitude data and spatial consistency data in the multi-time period corrected infrared image data according to the acquisition time sequence, and generates anomaly confidence image data and anomaly uncertainty image data.

[0138] The system maintains a consistent binding relationship between the spatial location information and the acquisition time information of the image data with abnormal credibility and the image data with abnormal uncertainty in different inspection periods.

[0139] The temporal consistency analysis module performs structure preservation calculations on spatially consistent abnormal structural features within the data range across inspection periods, and introduces structure preservation memory coefficients to perform memory recursive aggregation processing over a longer time span.

[0140] The system incorporates anomaly uncertainty image data and introduces an adaptive uncertainty suppression coefficient to perform uncertainty suppression processing on the overall structure-preserving result data, and forms stable anomaly location data through consistency determination.

[0141] The system performs regional connectivity organization processing on stable abnormal location data to form image data of candidate areas of potential collapse hazards that exist stably across inspection periods.

[0142] The cross-modal image consistency constraint module performs spatial mapping and consistency analysis on image data of candidate areas for collapse hazards and underground anomaly image data to form image consistency constraint result data.

[0143] The risk assessment and early warning generation module calculates the risk intensity and classifies the risk level of candidate areas based on abnormal credibility image data, image consistency constraint result data, and image data of candidate areas for road collapse hazards, and generates risk image data of road collapse hazards and corresponding early warning information data.

[0144] Through time-series consistency analysis of multi-period data, this embodiment achieves the identification and risk labeling of long-term stable collapse hazard areas within the existing system.

[0145] A method for detecting and warning of potential road collapse hazards using unmanned aerial vehicles (UAVs) comprises the following steps: During inspection, the UAV collects infrared image data of the road area frame by frame, simultaneously writing corresponding acquisition time and spatial location information. The acquisition time and spatial location information are then bound to the corresponding infrared image data at the frame level, and infrared image data lacking acquisition time or spatial location information is discarded. The bound infrared image data is sorted according to the acquisition time information to form continuous temporal infrared image data. For each frame of the temporal infrared image data, brightness distribution statistics, thermal radiation stability statistics, and spatial noise statistics are calculated to form image statistical features. The process involves: reading UAV flight status and environmental condition information corresponding to infrared image data; organizing this information into an initial set of task variable parameters; constructing statistical deviation data based on image statistical features; performing noise-sensitive suppression processing on the statistical deviation data; introducing a noise-sensitive suppression coefficient to form noise-suppressed deviation data; updating the task variable parameters based on the noise-suppressed deviation data; introducing a self-updating convergence coefficient to form a dynamic set of task variable parameters; and performing radiometric consistency correction, spatial noise suppression, and temporal stabilization processing on the temporal infrared image data based on the dynamic set of task variable parameters to form corrected infrared image data; and finally, analyzing the spatial distribution characteristics and changes of thermal features in the corrected infrared image data. The amplitude is measured to generate anomaly confidence image data and anomaly uncertainty image data. High-confidence spatial locations are located in the anomaly confidence image data, and spatially consistent thermal anomaly regions are extracted from the calibrated infrared image data to form candidate anomaly region data. Spatial morphological description information and boundary structure description information are extracted from the candidate anomaly region data to form anomaly structure features. Temporal consistency analysis is performed based on the degree of preservation of these anomaly structure features between adjacent acquisition time points, introducing a structure-preserving memory coefficient. Uncertainty suppression is then performed in conjunction with the anomaly uncertainty image data to form stable anomaly region data. This stable anomaly region data is then organized into collapse hazard candidate region image data, and read... Underground anomaly image data corresponding to the image data of candidate areas for potential collapse hazards is obtained and mapped to the infrared image spatial coordinate system. Spatial location consistency analysis and anomaly structure distribution consistency analysis are performed on the image data of candidate areas for potential collapse hazards and the mapped image data of underground anomalies to form image consistency constraint result data. Based on the anomaly confidence image data, the image consistency constraint result data and the image data of candidate areas for potential collapse hazards, risk intensity data corresponding to the candidate areas is generated. The risk intensity data is mapped to the infrared image space to form road collapse hazard risk image data. Based on the risk intensity data, early warning information data associated with the risk level is generated.

[0146] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A road subsidence hazard detection and early warning system and method using unmanned aerial vehicles (UAVs), characterized in that: include: The UAV image acquisition module acquires infrared image data of the road area and associates the acquisition time information and spatial location information with the infrared image data to form time-series infrared image data. The task variable modeling image correction module extracts brightness distribution, thermal radiation stability, and spatial noise features from time-series infrared image data. It combines UAV flight status information and environmental condition information to construct an initial task variable parameter set. The initial task variable parameter set is self-updated to form a dynamic task variable parameter set. Based on the dynamic task variable parameter set, the module performs radiometric consistency correction, spatial noise suppression, and temporal stabilization processing on the time-series infrared image data to form corrected infrared image data. The infrared anomaly confidence generation module generates anomaly confidence image data and anomaly uncertainty image data based on the spatial distribution characteristics and variation amplitude of thermal features in the corrected infrared image data. The temporal consistency analysis module constructs abnormal structural features of thermal anomaly regions based on corrected infrared image data, anomaly confidence image data, and anomaly uncertainty image data, and forms candidate image data of collapse hazard areas based on the degree of preservation of abnormal structural features in the time dimension. The cross-modal image consistency constraint module maps the underground anomaly image data corresponding to the candidate area of ​​collapse hazard to the infrared image spatial coordinate system to form image consistency constraint result data; The risk assessment and early warning generation module generates road collapse hazard risk image data and early warning information data based on image data of candidate areas for collapse hazards, image data of anomaly confidence, and image consistency constraint results.

2. The road collapse hazard unmanned aerial vehicle (UAV) exploration and early warning system and method according to claim 1, characterized in that: The drone image acquisition module uses the drone to collect infrared image data of the road area frame by frame during the inspection process, and synchronously writes the corresponding acquisition time information and spatial location information when acquiring each frame of infrared image data; The time and spatial location information of the acquisition are bound to the corresponding infrared image data at the frame level; Perform an integrity check on the bound infrared image data and remove image frames that are missing acquisition time information or spatial location information. Sort the infrared image data that pass the integrity check according to the acquisition time information and organize the sorted infrared image data into a continuous image sequence according to the acquisition time order. A continuous sequence of images, along with corresponding acquisition time and spatial location information, is organized to form temporal infrared image data.

3. The road collapse hazard unmanned aerial vehicle (UAV) exploration and early warning system and method according to claim 1, characterized in that: The task variable modeling image correction module reads infrared image data frame by frame from the time-series infrared image data and maintains the binding relationship between the acquisition time information and spatial location information and the corresponding infrared image data; For each frame of infrared image data, calculate the brightness distribution statistics, thermal radiation stability statistics, and spatial noise statistics, and organize the calculation results into image statistical features; Read the UAV flight status information and environmental condition information corresponding to each frame of infrared image data, and organize the reading results into an initial task variable parameter set; Based on image statistical features, statistical deviation data corresponding to the target stable statistical state is constructed. Noise-sensitive suppression processing is performed on the statistical deviation data, and a noise-sensitive suppression coefficient is introduced during the processing to form noise-suppressed deviation data. The task variable parameters are updated based on noise suppression bias data, and a self-updating convergence coefficient is introduced during the update process to form a dynamic set of task variable parameters. Radiometric consistency correction is performed based on a dynamic task variable parameter set to generate radiometrically corrected image data. Spatial noise suppression processing is performed on radiometrically corrected image data based on a dynamic task variable parameter set to form denoised image data. The denoised image data is subjected to temporal stabilization processing based on the dynamic task variable parameter set to form the correction result of the current frame; The correction results of each frame are organized in the order of acquisition time information, and the binding relationship between spatial location information and corresponding correction results is maintained to form corrected infrared image data.

4. The road collapse hazard unmanned aerial vehicle (UAV) exploration and early warning system and method according to claim 1, characterized in that: The infrared anomaly confidence generation module reads infrared image data frame by frame from the calibrated infrared image data and maintains the binding relationship between the acquisition time information and spatial location information and the corresponding infrared image data. For each frame of corrected infrared image data, calculate the brightness distribution statistics, thermal radiation stability statistics, and spatial noise statistics, and organize the calculation results into image statistical features; Thermal feature values ​​are extracted from the corrected infrared image data of the same spatial location at adjacent acquisition time points, and the thermal feature change amplitude data is calculated based on the thermal feature values; In each frame of corrected infrared image data, thermal feature spatial distribution data is constructed based on thermal feature values, and spatial consistency data is calculated based on thermal feature spatial distribution data; By combining thermal characteristic variation amplitude data with spatial consistency data, candidate anomaly intensity data is formed. Based on candidate anomaly intensity data and combined with thermal radiation stability statistics, a confidence assignment process is performed to generate anomaly confidence image data that corresponds one-to-one with the corrected infrared image data space. Based on spatial noise statistics and combined with thermal characteristic variation amplitude data, uncertainty assignment processing is performed to generate abnormal uncertainty image data that spatially corresponds one-to-one with the corrected infrared image data. The abnormal confidence image data and abnormal uncertainty image data are organized in the order of acquisition time information, and the spatial location information is bound to the corresponding image frame to form abnormal confidence image data and abnormal uncertainty image data that are consistent with the time order of the corrected infrared image data.

5. The road collapse hazard detection and early warning system and method according to claim 1, characterized in that: The timing consistency analysis module reads infrared image data frame by frame from the calibrated infrared image data and maintains the binding relationship between the acquisition time information and spatial location information and the corresponding infrared image data. Read the abnormal credibility image data frame by frame from the abnormal credibility image data, and maintain the binding relationship between the acquisition time information and spatial location information and the corresponding abnormal credibility image data; Read the abnormal uncertainty image data frame by frame from the abnormal uncertainty image data, and maintain the binding relationship between the acquisition time information and spatial location information and the corresponding abnormal uncertainty image data; In each frame of anomaly confidence image data, high confidence spatial locations are located, and spatially consistent thermal anomaly regions are extracted from the corresponding frame of corrected infrared image data to form candidate anomaly region data. Perform region connectivity organization processing on each frame of candidate anomaly region data to form a set of candidate anomaly region data; For each candidate anomaly region in the candidate anomaly region set data, spatial morphological description information and boundary structure description information are extracted, and the extraction results are organized into anomaly structure features. Between adjacent acquisition time points, the degree of structure preservation is calculated for anomalous structural features at spatially consistent locations to form structure-preserved sequence data; Memory recursive aggregation is performed on the structure-preserving sequence data, and structure-preserving memory coefficients are introduced during the processing to form the overall structure-preserving result data; Uncertainty suppression processing is performed on the overall structure preservation result data based on abnormal uncertainty image data, and an adaptive uncertainty suppression coefficient is introduced during the processing to form suppressed structure preservation result data; Based on the suppressed structure preservation results data, a consistency judgment process is performed to remove spatial locations where the structure preservation results do not meet the stability requirements, and retain stable abnormal location data where the structure preservation results meet the stability requirements. Perform region connectivity organization processing on stable anomaly location data to form a set of stable anomaly region data; The data of stable abnormal regions is mapped to image data of candidate areas for collapse hazards, and then organized into image data of candidate areas for collapse hazards according to the order of collection time information.

6. The road collapse hazard detection and early warning system and method according to claim 1, characterized in that: The cross-modal image consistency constraint module reads the image data of the candidate areas of collapse hazards frame by frame from the image data of the candidate areas of collapse hazards, and maintains the binding relationship between the acquisition time information and spatial location information and the corresponding image data of the candidate areas of collapse hazards; The cross-modal image consistency constraint module reads the underground anomaly image data corresponding to the image data of the candidate area of ​​the collapse hazard, and maintains the binding relationship between the spatial location information and the corresponding underground anomaly image data; Spatial registration processing is performed on the underground anomaly image data, and the underground anomaly image data is mapped to the infrared image spatial coordinate system to form mapped underground anomaly image data; Spatial overlap calculation processing is performed on the image data of candidate areas of potential collapse hazards and the mapped underground anomaly image data to form consistent data of anomaly spatial location; Anomaly structure distribution alignment processing is performed on the image data of candidate areas of potential collapse hazards and the mapped underground anomaly image data to form consistent anomaly structure distribution data; The consistency data of abnormal spatial locations and the consistency data of abnormal structural distributions are jointly organized to form image consistency constraint result data; The image consistency constraint result data is bound to the corresponding acquisition time information and spatial location information, and then organized in the order of acquisition time information to form the image consistency constraint result data.

7. The road collapse hazard unmanned aerial vehicle (UAV) exploration and early warning system and method according to claim 1, characterized in that: The risk assessment and early warning generation module reads the image data of the candidate areas of collapse hazards frame by frame from the image data of the candidate areas of collapse hazards, and maintains the binding relationship between the acquisition time information and spatial location information and the corresponding image data of the candidate areas of collapse hazards; The risk assessment and early warning generation module reads abnormal credibility image data frame by frame from the abnormal credibility image data, and maintains the binding relationship between the acquisition time information and spatial location information and the corresponding abnormal credibility image data. The risk assessment and early warning generation module reads the image consistency constraint result data corresponding to the image data of the candidate area of ​​the collapse hazard, and maintains the binding relationship between the acquisition time information and spatial location information and the corresponding image consistency constraint result data; In each frame of candidate area image data for potential collapse, the spatial range of the candidate area is located, and spatially consistent abnormal confidence area data is extracted from the corresponding frame of abnormal confidence image data. Perform regional-level statistical analysis on data from regions with abnormal credibility to generate regional credibility indices for candidate regions; Read the image consistency constraint result data corresponding to the spatial range of the candidate region, and organize the image consistency constraint result data into candidate region consistency index; The regional credibility index is weighted and adjusted based on the candidate region consistency index to form a weighted regional credibility index. Extraction processing is performed on the abnormal structure preservation characteristics within the spatial range of the candidate region, and the extraction results are organized into candidate region structure preservation indices; The weighted regional credibility index and the candidate regional structure preservation index are jointly organized to form candidate regional risk intensity data. Risk intensity mapping processing is performed on the risk intensity data of the candidate areas, and the mapping results are written into the corresponding positions in the infrared image space to form risk image data of potential road collapse hazards. Risk level classification processing is performed on the candidate area risk intensity data to generate candidate area risk level data; The risk level data of candidate areas is bound with spatial location information, and the data is organized into early warning information data according to the order of collection time.

8. A method for applying to a road subsidence hazard unmanned aerial vehicle (UAV) exploration and early warning system as described in any one of claims 1-7, characterized in that: The steps are as follows: During the inspection, drones are used to collect infrared image data of the road area frame by frame. During the collection process, the corresponding collection time information and spatial location information are written simultaneously. The collection time information and spatial location information are bound to the corresponding infrared image data at the frame level, and infrared image data with missing collection time information or spatial location information are removed. The bound infrared image data is sorted according to the acquisition time information to form continuous time-series infrared image data. The brightness distribution statistics, thermal radiation stability statistics and spatial noise statistics are calculated frame by frame for the time-series infrared image data to form image statistical features. Read the UAV flight status information and environmental condition information corresponding to the infrared image data, organize them into an initial task variable parameter set, construct statistical deviation data based on image statistical features, perform noise-sensitive suppression processing on the statistical deviation data, introduce a noise-sensitive suppression coefficient, and form noise-suppressed deviation data. The task variable parameters are updated based on the noise suppression bias data. A self-updating convergence coefficient is introduced to form a dynamic task variable parameter set. The radiometric consistency correction, spatial noise suppression and temporal stabilization processing are performed on the time-series infrared image data based on the dynamic task variable parameter set to form corrected infrared image data. Based on the spatial distribution characteristics and variation amplitude of thermal features in the corrected infrared image data, abnormal confidence image data and abnormal uncertainty image data are generated. High confidence spatial locations are located in the abnormal confidence image data, and spatially consistent thermal anomaly regions are extracted from the corrected infrared image data to form candidate anomaly region data. Spatial morphological description information and boundary structure description information are extracted from candidate anomaly region data, and anomaly structure features are organized to form them. Temporal consistency analysis is performed based on the degree of preservation of the anomaly structure features between adjacent acquisition time points. A structure preservation memory coefficient is introduced, and uncertainty suppression is performed in combination with anomaly uncertainty image data to form stable anomaly region data. The stable anomaly area data is organized into image data of candidate areas for collapse hazards. The underground anomaly image data corresponding to the image data of candidate areas for collapse hazards is read and mapped to the infrared image spatial coordinate system. Based on the image data of candidate areas for collapse hazards and the mapped underground anomaly image data, spatial location consistency analysis and anomaly structure distribution consistency analysis are performed to form image consistency constraint result data. Based on the anomaly confidence image data, image consistency constraint result data and image data of candidate areas for collapse hazards, risk intensity data corresponding to the candidate areas are generated. Risk intensity data is mapped onto infrared image space to form risk image data of potential road collapse hazards. Based on the risk intensity data, early warning information data associated with the risk level is generated.

Images