An unmanned aerial vehicle-based construction site hazard source identification system and method

Abstract

The application discloses a construction site hazard source identification system and method based on a UAV, belongs to the technical field of construction site hazard source identification, and acquires construction site image data by using a UAV and extracts initial target detection features; scene visibility fluctuation indexes and carrier attitude disturbance indexes are calculated based on image data and UAV attitude data, first target detection features are extracted according to the scene visibility fluctuation indexes; second target detection features are obtained according to the carrier attitude disturbance indexes; detection features are fused, and environment compensation is carried out in combination with the scene visibility fluctuation indexes and the carrier attitude disturbance indexes, a dynamic update judgment threshold is updated through threshold self-adaptive learning, a rule deviation degree is calculated according to hazard source identification features and a dynamic threshold set, and a hazard state is judged; the application can maintain high identification precision under extreme conditions such as insufficient illumination, rain and fog interference and unstable UAV attitude, and improves the stability and reliability of construction site hazard source identification.

Description

Technical Field

[0001] This invention relates to the field of hazard identification technology at construction sites, specifically to a construction site hazard identification system and method based on unmanned aerial vehicles (UAVs). Background Technology

[0002] Hazard identification at construction sites refers to the systematic search, analysis, and confirmation of all potential hazards that could cause personal injury, equipment damage, or environmental impact before and during construction activities. By identifying these hazards, construction companies can take effective preventative and control measures in advance, thereby reducing the likelihood of accidents and ensuring safe and orderly construction.

[0003] In existing technologies, hazard source rule bases typically rely on fixed thresholds (such as safe distances between personnel and machinery, color feature matching, etc.) for hazard assessment. However, in extreme environments such as insufficient lighting at night, strong winds causing personnel posture shifts, and rain blurring camera images, image quality and target detection results can deviate significantly. This leads to AI misjudging personnel positions, machinery boundaries, and protective equipment such as safety helmets due to jitter, shifts, or color misidentifications. Consequently, fixed threshold rules cannot accurately match actual risk conditions, resulting in missed or false alarms. This problem is particularly prominent when drones perform nighttime inspections because rapid changes in lighting and the susceptibility of drone posture to wind make it difficult for static threshold-based hazard source matching mechanisms to maintain reliability and stability. Summary of the Invention

[0004] The purpose of this invention is to provide a construction site hazard identification system and method based on unmanned aerial vehicles (UAVs) to address the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying hazardous sources at construction sites based on unmanned aerial vehicles (UAVs), comprising: By acquiring image data of the construction site through drones, and using target detection models to identify personnel, machinery and equipment and safety protection equipment, initial target detection features are obtained; Calculate the scene visibility fluctuation index and the carrier attitude disturbance index based on image data and UAV attitude data. Based on the scene visibility fluctuation index, the image is subjected to illumination enhancement, rain removal, or fog removal processing to generate an enhanced image, and the enhanced first target detection feature is extracted from the enhanced image. Based on the carrier attitude perturbation index, jitter compensation and attitude stabilization correction are performed on the image to generate a corrected image, and the second target detection feature after attitude correction is extracted from the corrected image. The initial target detection features, the enhanced first target detection features, and the pose-corrected second target detection features are fused to obtain fused target detection features. The environmental compensation calculation is performed by inputting the target detection features, scene visibility fluctuation index and carrier attitude disturbance index into the compensation model. Based on the calculation results, the personnel position, mechanical boundary and safety protection features are dynamically corrected to obtain the hazard source identification features after environmental compensation. Based on the hazard source identification features, threshold adaptive learning is performed to generate a dynamic threshold set that adjusts with environmental changes; Based on the hazard identification features and the dynamic threshold set, the rule deviation is calculated. When the rule deviation exceeds the set threshold, it is determined that there is a risk of missed or misjudged detection, and the hazard identification result of the construction site is output.

[0006] Preferably, the calculation of the scene visibility fluctuation index includes the following steps: Atmospheric scattering model inversion is performed based on image data to estimate the transmittance of the image in multiple local regions, and the deviation of the transmittance in the spatial distribution is calculated to obtain the transmittance perturbation value. Fourier frequency domain decomposition is performed on the image data to extract high-frequency energy components, and the sharpness fluctuation value is calculated based on the attenuation rate of high-frequency energy at adjacent time points. Based on the attitude data of the UAV, the pitch angle, roll angle, yaw angle and angular velocity information at the corresponding moment are obtained, and the coupling degree between the image blur degree and the attitude change is calculated to obtain the attitude disturbance value. The transmittance perturbation value, sharpness fluctuation value, and attitude perturbation value are fused and calculated according to a preset weight to generate the scene visibility fluctuation index.

[0007] Preferably, the calculation of the carrier attitude disturbance index includes the following steps: Optical flow field is extracted based on image data using optical flow estimation algorithm, and theoretical optical flow field is constructed based on UAV attitude angle changes. The residual signal energy between the actual optical flow field and the theoretical optical flow field is used as the optical flow residual perturbation value. Based on image data, a blind deblurring algorithm is used to estimate the image blur kernel, and the directional offset between the main direction of the blur kernel and the direction of change of the UAV attitude angle is calculated to obtain the directional perturbation value of the blur kernel. The attitude perturbation value, optical flow residual perturbation value, and fuzzy kernel direction perturbation value are fused and calculated according to a preset weight to obtain the carrier attitude perturbation index.

[0008] Preferably, the enhanced image is generated by performing illumination enhancement, rain removal, or fog removal on the image based on the scene visibility fluctuation index, including the following steps: The degree of illumination disturbance is determined based on the scene visibility fluctuation index, and an illumination enhancement model is constructed. The image brightness is reconstructed using an illumination enhancement algorithm based on illumination component estimation and reflection component constraint to obtain an illumination-enhanced image. When the transmittance perturbation value corresponding to the scene visibility fluctuation index exceeds the set transmittance threshold, the image transmittance is estimated and corrected using a dehazing algorithm based on atmospheric scattering inversion, and a dehazed image is generated. When the high-frequency energy attenuation rate reflected by the scene visibility fluctuation index exceeds the set attenuation threshold, the rain texture component is extracted using a rain texture separation-based deraining algorithm, and the rain texture component is separated from the original image to generate a derained image. The image with the highest visibility is selected from the enhanced image, the dehazed image, or the derained image as the enhanced image, and the enhanced image is input into the target detection model to extract the enhanced first target detection feature.

[0009] Preferably, the process of performing jitter compensation and attitude stabilization correction on the image based on the carrier attitude perturbation index to generate a corrected image includes the following steps: The attitude jitter level is determined based on the attitude disturbance index of the carrier. When the attitude disturbance index exceeds the set attitude compensation threshold, an attitude transformation matrix is ​​constructed based on the pitch angle, roll angle and yaw angle in the UAV attitude data. The image data is reversed by using the attitude transformation matrix to compensate for the imaging geometric offset caused by attitude change pixel by pixel to obtain a preliminary attitude correction image. The image gradient stability is calculated for the initial attitude correction image. Motion blur estimation is performed for regions with insufficient gradient stability, and the local regions are deblurred according to the estimated blur kernel to generate the correction image. The corrected image is input into the target detection model to extract the second target detection features after pose correction.

[0010] Preferably, the acquisition of the fused target detection features includes the following steps: Based on the initial target detection features, the enhanced first target detection features, and the pose-corrected second target detection features, feature alignment processing is performed on the detection position, detection confidence, and feature vector of the same target in different images to obtain an aligned feature set. Calculate the feature consistency score based on the similarity value between each feature vector in the aligned feature set and the reference feature vector; A weighted fusion operation is performed on the aligned feature set according to the feature consistency score as the weight, so as to obtain the fused target feature vector; The target detection bounding box position and detection confidence are updated based on the fused target feature vector to generate the fused target detection features.

[0011] Preferably, obtaining the environmentally compensated hazard source identification features includes the following steps: The fused target detection features, the scene visibility fluctuation index, and the carrier attitude disturbance index are input into the compensation model; Adjust the brightness and texture features in the fused target detection features based on the scene visibility fluctuation index; Based on the carrier attitude perturbation index, position offset correction is performed on the spatial position information in the fused target detection features; The results after illumination feature compensation and posture feature compensation are input into the dynamic correction rules of the compensation model to perform comprehensive correction on personnel position, safety protection features and mechanical boundaries, and output the hazard source identification features after environmental compensation.

[0012] Preferably, the threshold adaptive learning includes the following steps: Based on the environmentally compensated hazard source identification features, feature parameters are extracted to characterize personnel location features, mechanical boundary features, and safety protection features, and environmental disturbance weights are constructed according to the scene visibility fluctuation index and the carrier attitude disturbance index. Based on the environmental disturbance weights, a disturbance sensitivity analysis is performed on the preset thresholds to determine the adjustment direction and magnitude of the thresholds for each characteristic parameter under different environmental conditions. A threshold learning model is constructed based on the feature parameters and the threshold adjustment direction, and the threshold update amount is solved by minimizing the hazard source determination error. The threshold update amount is superimposed with the preset threshold to generate a threshold set that dynamically adjusts with changes in the environment.

[0013] Preferably, the calculation of the rule deviation includes the following steps: Based on the hazard source identification features, feature parameters for determining the hazardous state are extracted from personnel location features, mechanical boundary features, and safety protection features, and the feature parameters are paired with the corresponding thresholds in the dynamic threshold set; Calculate the feature offset for each set of feature parameters and the corresponding threshold; The offset weights are set according to the degree of danger of each feature offset, and the offset weights are weighted and accumulated to obtain an offset score for comprehensively characterizing the accuracy of rule matching. The offset score is compared with a set rule reliability threshold to generate a rule deviation score.

[0014] This invention also provides a construction site hazard identification system based on unmanned aerial vehicles (UAVs), comprising: Image acquisition module: Acquires image data of the construction site through drones, and uses target detection models to identify personnel, machinery and equipment and safety protection equipment to obtain initial target detection features; Calculation module: Calculates scene visibility fluctuation index and carrier attitude disturbance index based on image data and UAV attitude data; Image enhancement processing module: performs illumination enhancement, rain removal, or fog removal on the image based on the scene visibility fluctuation index to generate an enhanced image, and extracts the enhanced first target detection features from the enhanced image; Attitude compensation and correction module: Performs jitter compensation and attitude stabilization correction on the image according to the carrier attitude perturbation index, generates a corrected image, and extracts the second target detection features after attitude correction from the corrected image; Multi-source feature fusion module: performs feature fusion on the initial target detection features, the enhanced first target detection features, and the pose-corrected second target detection features to obtain fused target detection features; Environmental compensation calculation module: The module integrates target detection features, scene visibility fluctuation index and carrier attitude disturbance index into the compensation model to perform environmental compensation calculation. Based on the calculation results, it dynamically corrects personnel positions, mechanical boundaries and safety protection features to obtain the hazard source identification features after environmental compensation. Threshold adaptive learning module: Performs threshold adaptive learning based on the hazard source identification features to generate a dynamic threshold set that adjusts with changes in the environment; Hazard source rule deviation analysis module: Calculates the rule deviation degree based on the hazard source identification features and the dynamic threshold set. When the rule deviation degree exceeds the set threshold, it determines that there is a risk of missed or misjudged identification, and outputs the hazard source identification results at the construction site.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention achieves dual-dimensional quantitative perception of extreme environmental interference and UAV attitude changes at construction sites by constructing a scene visibility fluctuation index and a carrier attitude disturbance index, making image enhancement and attitude correction processes targeted and adaptive. Multiple processing methods, such as illumination enhancement, rain removal, fog removal, and jitter compensation, are all based on dynamic triggering and fine adjustment of quantitative environmental indicators, significantly improving image input quality and comprehensively enhancing the detection accuracy of personnel, machinery, and safety equipment under different environmental conditions. The method of this invention effectively solves the problems of missed detections and false detections that easily occur in existing technologies under conditions of low light, rain, fog, and jitter, enhancing the robustness and reliability of the hazard source identification process.

[0016] 2. This invention, based on the multi-source fusion of enhanced first target detection features, posture-corrected second target detection features, and initial target detection features, achieves comprehensive extraction of target spatial location, appearance features, and protection information. Through the combined action of dynamic threshold learning and rule deviation evaluation mechanisms, it can automatically adjust the judgment threshold and identify rule deviations according to environmental changes, thereby improving the intelligence level of hazard identification. By introducing an environmental compensation model, this invention can dynamically correct personnel position shifts, mechanical boundary deformation, and weakening of protection features, ensuring that the final hazard identification results maintain high stability and reliability in rapidly changing and complex construction environments, significantly improving overall identification capabilities compared to traditional methods. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of the method of the present invention.

[0019] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0021] Example 1, please refer to Figure 1 As shown in this embodiment, a method for identifying hazardous sources at construction sites based on unmanned aerial vehicles (UAVs) includes: By acquiring image data of the construction site using drones, and using target detection models to identify personnel, machinery, equipment, and safety protection equipment, initial target detection features are obtained.

[0022] In this invention, a drone is first used as a data acquisition carrier to obtain visual information about the construction site. Specifically, the drone sequentially takes aerial photographs of the construction area according to a preset flight path, collecting multi-angle image or video data, including top-down, side-view, and close-up views. To ensure complete data coverage, the drone can achieve full coverage data acquisition based on geofencing, waypoint planning, and altitude control, and record shooting pose information in real time during the acquisition process.

[0023] After acquisition, the image data is input into a preset target detection model for initial recognition. The target detection model can be a deep learning-based target recognition network used to automatically detect and locate key objects in construction site images, including personnel, machinery, safety equipment such as helmets, and other elements that may affect safety risk assessment. The model outputs initial target detection results containing information such as object category, location coordinates, and confidence level through processes such as convolutional feature extraction, candidate region generation, and target classification.

[0024] In this embodiment, the identified personnel locations, machinery boundaries, and safety equipment wearing status are all used as the foundational features for subsequent environmental compensation, threshold learning, and hazard identification. By combining wide-area data acquisition from UAVs with automated visual recognition, this invention enables the rapid acquisition of accurate and structured initial target detection features in large-scale construction scenarios, providing a data foundation for subsequent hazard identification.

[0025] The scene visibility fluctuation index and the carrier attitude disturbance index are calculated based on image data and UAV attitude data.

[0026] The methods for obtaining the scene visibility fluctuation index include: In this embodiment, an atmospheric scattering model is constructed from image data to estimate transmittance. The atmospheric scattering model is based on the optical imaging mechanism, and its core definition is that the actual brightness of an object's surface in the image is attenuated by scattering from suspended particles in the air before entering the imaging device. To estimate the degree of attenuation, the minimum brightness component is extracted as the dark channel value for each local region of the image, and this value is used to estimate the transmittance of that region. Subsequently, the transmittance perturbation is calculated based on the distribution of transmittance across all regions. Specifically, the difference between the transmittance of each region and the average transmittance is calculated, and the dispersion of all differences is used as the transmittance perturbation value. An increase in the transmittance perturbation value indicates a more drastic change in visibility caused by fog, dust, or raindrops.

[0027] In this embodiment, to measure the changes in image texture information caused by environmental degradation, a Fourier domain transformation is performed on the image, and high-frequency energy components are extracted. Fourier frequency domain decomposition converts the image brightness matrix into a frequency domain representation, and the image detail is quantified by calculating the energy distribution in high-frequency regions. To calculate the sharpness fluctuation value, the high-frequency energy is calculated for images acquired at different times, and the attenuation ratio of high-frequency energy at adjacent times is compared. The absolute value of the attenuation ratio is used as the sharpness fluctuation value. A larger sharpness fluctuation value indicates more significant texture degradation caused by low light, rain, fog, moisture, or dirt.

[0028] In this embodiment, to address image jitter caused by UAV flight, attitude disturbance values ​​are obtained through mathematical analysis of UAV attitude data. UAV attitude data includes pitch angle, roll angle, yaw angle, and corresponding angular velocity. A functional relationship is established between the rate of change of angular velocity and the degree of image blur, quantifying the coupling between the two. To calculate the degree of image blur, a convolution operation is performed on the image brightness gradient to obtain the gradient magnitude; a decrease in gradient magnitude indicates an increase in blur. Subsequently, a correlation analysis is performed between the blur degree change curve and the attitude angular velocity change curve, and the absolute value of the correlation coefficient is used as the attitude disturbance value. A higher attitude disturbance value indicates a more significant impact of imaging jitter on image sharpness.

[0029] A scene visibility fluctuation index is generated by establishing a weighted fusion model. The weighted fusion model uses transmittance perturbation, sharpness fluctuation, and pose perturbation as input parameters, and linearly combines these parameters by setting weight coefficients. The weight coefficients are set based on the sensitivity to environmental changes. The transmittance perturbation weight reflects the impact of airborne scattering on the image, the sharpness fluctuation weight reflects the impact of texture degradation on visibility, and the pose perturbation weight reflects the impact of jitter on image quality. The fusion result is the scene visibility fluctuation index; a higher value indicates a stronger impact of extreme environmental interference on image visibility.

[0030] The methods for obtaining the carrier attitude perturbation index include: In this embodiment, to characterize the local pixel motion caused by attitude changes in the image, an optical flow estimation algorithm is employed on the image data. The optical flow estimation algorithm obtains the motion vector of each pixel by solving for brightness consistency constraints and gradient smoothing constraints on images acquired at consecutive acquisition times, thus forming the actual optical flow field. Simultaneously, based on the change in the UAV's attitude angle between adjacent acquisition times, a theoretical optical flow field is calculated using the spatial rotation matrix. The theoretical optical flow field represents the motion of all pixels assumed to be caused only by attitude rotation. By performing pixel-by-pixel subtraction between the actual and theoretical optical flow fields, and calculating the total energy of the set of difference vectors, this total energy is defined as the optical flow residual perturbation value. A larger optical flow residual perturbation value indicates a more pronounced non-attitude-induced motion blur and drift effect in the image.

[0031] In this embodiment, to estimate the direction and degree of blurring caused by attitude changes during image capture, a blind deblurring algorithm is used to extract a blur kernel from the image data. The blur kernel is defined as a two-dimensional matrix structure describing the direction and degree of image blurring. The shape of the blur kernel is estimated by reconstructing the image sharpness target in reverse. The extracted blur kernel has a specific principal direction, which is obtained by performing eigenvalue decomposition on the blur kernel matrix. Subsequently, the attitude change direction is calculated based on the change direction of the UAV's attitude angle at adjacent acquisition times. The directional offset between the principal direction of the blur kernel and the attitude change direction is used as the blur kernel direction perturbation value. The larger the directional offset, the more significant the contribution of attitude change to image blurring, and the higher the degree of perturbation.

[0032] In this embodiment, to comprehensively reflect the degree of attitude disturbance caused by the combined effects of image gradient changes, optical flow residual changes, and blur kernel direction shifts, the attitude disturbance values, optical flow residual disturbance values, and blur kernel direction disturbance values ​​are fused. A weighted linear combination model is used for fusion, constructing a fusion calculation expression by assigning weight coefficients to each of the three disturbance values. The weight coefficients reflect the sensitivity of different disturbance values ​​to changes in image quality. The weighted sum is defined as the carrier attitude disturbance index. A larger carrier attitude disturbance index indicates more severe imaging jitter caused by attitude changes. This carrier attitude disturbance index is used for subsequent image correction processing and hazard identification.

[0033] The image is enhanced by illumination enhancement, rain removal, or fog removal based on the scene visibility fluctuation index to generate an enhanced image, and the enhanced first target detection features are extracted from the enhanced image.

[0034] In this embodiment, the presence of significant brightness attenuation in the image is determined based on the illumination perturbation value in the scene visibility fluctuation index. When the illumination perturbation value is higher than a preset illumination enhancement threshold, an illumination enhancement model based on illumination component estimation is used to enhance the image.

[0035] The construction of the illumination enhancement model includes the following steps: Perform a logarithmic transformation on the luminance component of each pixel in the image to represent the image as a product of the illumination component and the reflection component; The least squares method is used to solve the illumination component, making the illumination component change smoothly in space, and the local gradient of the reflection component is used as a constraint to ensure that the enhancement process does not destroy the original texture. Perform an inverse exponential transform on the obtained illumination components to restore the luminance range; The restored illumination and reflection components are recombined to generate an enhanced illumination image.

[0036] The brightness level and detail contrast of the illumination-enhanced image are higher than those of the original image, enabling subsequent target detection to still have good recognition capabilities in low-light environments.

[0037] In this embodiment, when the transmittance perturbation value in the scene visibility fluctuation index exceeds the set transmittance threshold, a dehazing algorithm based on an atmospheric scattering model is used to process the image, and the following steps are performed: The image is divided into multiple local regions, and the minimum brightness value of each region is extracted as the dark channel value. The initial transmittance of the region is estimated based on the dark channel value. Transmittance represents the proportion of light remaining after passing through the air medium. A smooth transmittance map is obtained by constraining the spatial continuity of transmittance based on the initial transmittance using soft matting. Based on the smoothed transmittance map and atmospheric light intensity, the true reflected light intensity of the object surface is calculated using the brightness inverse function to obtain the dehazed image.

[0038] Dehazed images can restore scattered and blurred edge details and significantly improve image contrast, enabling target detection models to maintain stable output in environments such as haze and dust.

[0039] In this embodiment, when the high-frequency energy attenuation rate in the scene visibility fluctuation index exceeds a set attenuation threshold, a rain deraining algorithm based on rain texture separation is performed on the image. This algorithm is constructed according to the following steps: The image is decomposed into high-frequency components containing texture and low-frequency components containing the main structure by performing multi-scale wavelet decomposition. Rain patterns with linear stripe structures are identified in high-frequency components, and the rain patterns are characterized by constructing a sparse rain pattern dictionary. The high-frequency components are decomposed using a sparse coding method, and the rain pattern component and the content texture component are obtained respectively. The rain streak component is removed from the high-frequency component and then reconstructed with the low-frequency component to obtain the rain-removed image.

[0040] De-raining images can significantly reduce texture degradation caused by raindrop occlusion, restoring the clarity of target contours.

[0041] After the enhanced illumination image, dehazed image, and derained image are generated, visibility is evaluated on these images. The visibility evaluation includes the following three quantitative indicators: Brightness uniformity: Brightness uniformity is calculated based on the distribution uniformity of the brightness histogram of the enhanced image; Contrast enhancement ratio: Calculated based on the change in local contrast of the image before and after enhancement; High-frequency energy recovery rate: Calculates the ratio of high-frequency energy in the enhanced image to that in the original image.

[0042] The visibility evaluation value is obtained by summing the three evaluation indicators according to their weights, and the image with the highest evaluation value is used as the enhanced image.

[0043] The enhanced image is input into the object detection model, which obtains the enhanced first object detection features through steps such as convolutional feature extraction, bounding box generation, and feature classification. Because the visibility of the enhanced image is significantly improved, the enhanced first object detection features have higher localization accuracy and object recognition reliability.

[0044] To verify the actual improvement effect of illumination enhancement, dehazing, and deraining processing on target detection in this embodiment, images were acquired under different environmental conditions, and the detection results were compared with those of unenhanced images. The experimental environment included: Low-light environments at night (light intensity below 30 lx); Hazy environment (visibility range of 5 m to 12 m); Raindrop interference environment (raindrop diameter 1 mm to 3 mm); A sunny, normal environment served as the control group.

[0045] Quantitative data on light enhancement effects include: In low-light environments: The brightness, visible detail, and detection accuracy of the illumination-enhanced images are significantly improved, demonstrating the practical effectiveness of the illumination-enhanced model.

[0046] Quantitative data on defogging effectiveness include: In a smoggy environment: Transmittance estimation and atmospheric scattering inversion significantly reduce the blurring caused by obstructions.

[0047] Quantitative data on rain removal effectiveness include: In raindrop interference environment: De-raining significantly improves the representation of edge and texture information.

[0048] The enhanced first target detection features generated from the three types of enhanced images are compared with the initial target detection features of the original images: Positioning error reduced by approximately 18%–35%; Target classification confidence improved by approximately 12%–28%; Small target detection capability improved by approximately 22%–41%; This indicates that augmentation processing driven by the scene visibility fluctuation index can significantly improve the input quality of the target detection model, thereby enhancing detection accuracy and stability.

[0049] The image is jitter compensated and attitude stabilization corrected according to the carrier attitude perturbation index to generate a corrected image, and the second target detection feature after attitude correction is extracted from the corrected image.

[0050] In this embodiment, to determine whether image attitude correction is required, the impact of UAV attitude changes on image quality is assessed using the carrier attitude disturbance index. When the carrier attitude disturbance index exceeds a set attitude compensation threshold, attitude correction processing is performed on the image. The attitude compensation threshold characterizes the acceptable range of attitude disturbances, and its value is obtained through extensive flight experiments.

[0051] To describe the impact of UAV attitude changes on image spatial position, pitch, roll, and yaw rotation matrices were constructed based on UAV attitude data. Each rotation matrix was constructed using a spatial rotation formula in a three-dimensional coordinate system. The three rotation matrices were then multiplied in the order of attitude changes to obtain the attitude transformation matrix. The attitude transformation matrix describes the overall transformation of the imaging plane caused by attitude disturbances.

[0052] In this embodiment, to compensate for the spatial offset caused by the pose change, inverse pose mapping is performed on the original image. Inverse pose mapping maps image pixels from the disturbed position back to their original imaging position before the pose change by performing matrix inversion on the pose transformation matrix. The implementation of inverse pose mapping includes the following steps: Extract the two-dimensional coordinates of each pixel in the image plane; Extend the two-dimensional coordinates to homogeneous coordinates and input the inverse of the attitude transformation matrix; The corrected position of each pixel is reconstructed based on the matrix multiplication result; Bilinear interpolation is used to calculate the grayscale or color value of the corrected pixel position.

[0053] A preliminary attitude correction image is obtained by inverse attitude mapping. The overall geometry of this image is consistent with the real scene, which significantly reduces the image shift caused by drone jitter.

[0054] Although inverse pose mapping can correct spatial offsets, it cannot completely eliminate local motion blur caused by rapid pose changes. To further improve image sharpness, image gradient stability is calculated on the initially pose-corrected image. Image gradient stability is calculated through the following steps: Gradient calculations are performed on the initial attitude correction image to obtain the horizontal and vertical gradients. Calculate the gradient magnitude and gradient direction for each pixel; Perform a difference operation on the gradient magnitude change at the same location between images at adjacent time points; If the change in gradient magnitude exceeds the gradient stability threshold, the region is identified as a motion blur region.

[0055] For regions identified as motion-blurred, a deblurring algorithm based on blur kernel inversion is used for restoration. The blur kernel is estimated by minimizing image reconstruction error and includes two parameters: blur direction and blur length. Based on the estimated blur direction and blur length, an inverse convolution operation is constructed to deblur the local region, restoring texture details. The corrected image is obtained after deblurring.

[0056] In this embodiment, the corrected image is input into the target detection model. A convolution operation is performed through a feature extraction layer to obtain a multi-scale feature map. Then, a candidate region generation algorithm extracts potential target regions, and a target classification algorithm determines the target category within each region. The final output feature set is defined as the pose-corrected second target detection feature.

[0057] Compared to uncorrected images, the edges, contours, and spatial positions of objects in corrected images are more stable, resulting in higher positioning accuracy, lower false detection rate, and more stable feature representation capabilities for the second target detection features after pose correction, providing a more reliable data foundation for subsequent hazard source identification.

[0058] By capturing the same scene under different drone flight attitudes, the geometric shift, blurriness, and detection accuracy of the images before and after correction are quantitatively compared. Experimental attitude conditions include: Slight shaking (angular velocity) ); Moderate jitter (angular velocity) ); Severe shaking (angular velocity) ).

[0059] Based on the inverse pose mapping and pose transformation matrix correction, the image offset is improved as follows: Reverse attitude mapping can effectively eliminate the overall offset caused by attitude changes.

[0060] The gradient stability of the initial attitude correction image is calculated and compared with the correction image generated by fuzzy kernel inversion: This proves that gradient stability detection combined with fuzzy kernel inversion can effectively recover details.

[0061] Compared with uncorrected images: target localization offset reduced by approximately 29%–58%; target bounding box stability improved by approximately 18%–39%; classification accuracy improved by approximately 11%–22%; and missed detections due to jitter were avoided by approximately 17%–33%. This demonstrates that the correction mechanism driven by the carrier attitude perturbation index can significantly improve the stability and reliability of target detection features.

[0062] The initial target detection features, the enhanced first target detection features, and the pose-corrected second target detection features are fused to obtain fused target detection features.

[0063] In this embodiment, to reduce the deviations in target localization and feature representation caused by different image processing stages, feature alignment processing is required for the initial target detection features, the enhanced first target detection features, and the pose-corrected second target detection features. The feature alignment process includes the following steps: Based on the detection box positions of the corresponding targets in the three types of target detection features, the target position reference system is unified using the spatial coordinate remapping method; The detection confidence scores of each detection result are normalized to ensure that the detection confidence scores are within the same range. The feature vectors of each detection result are dimensionally standardized to ensure that similarity comparisons can be made between feature vectors.

[0064] The three feature sets after alignment processing form the alignment feature set.

[0065] In this embodiment, to determine the matching degree between multiple detection results, a set of reference feature vectors is selected from the alignment feature set. The reference feature vector is defined as the one with the highest confidence among the three types of feature vectors. Subsequently, feature similarity is calculated between the other feature vectors and the reference feature vectors. Feature similarity is calculated as follows: Calculate the inner product of the reference feature vector and the target feature vector; Calculate the vector lengths of both vectors; The ratio of the inner product to the product of the vector length is used as the feature consistency score.

[0066] The higher the feature consistency score, the more consistent the directions of the two feature vectors in the feature space and the higher the degree of matching.

[0067] In this embodiment, multiple feature vectors are combined into a single representative feature vector using a weighted fusion method. The specific steps of the weighted fusion are as follows: A weight value is calculated based on the feature consistency score of each feature vector, and the weight value is proportional to the feature consistency score. Multiply each dimension value of each feature vector by its corresponding weight value; Perform a dimension-by-dimensional summation operation on all weighted eigenvectors; The summation result is normalized to form the fused target feature vector.

[0068] The fused target feature vector can comprehensively reflect the detection results from multiple sources, improving the stability and robustness of feature representation.

[0069] After the target feature vectors are fused, the target detection bounding box position and detection confidence are updated to ensure the consistency of target localization information. The update method is as follows: The target locations of the three types of detection results are weighted and averaged to obtain a weighted reference feature consistency score. The detection confidence scores of the three types of detection results are weighted and averaged to ensure that the weights are consistent with the feature consistency score. The updated target location and detection confidence are combined with the fused target feature vector to form the fused target detection feature.

[0070] Fusion of target detection features can effectively eliminate the differences caused by illumination enhancement, defogging and attitude correction, so that the target detection results remain stable and consistent under multiple environmental conditions.

[0071] The environmental compensation model is calculated by integrating target detection features, scene visibility fluctuation index and carrier attitude disturbance index. Based on the calculation results, the personnel position, mechanical boundary and safety protection features are dynamically corrected to obtain the hazard source identification features after environmental compensation.

[0072] In this embodiment, the target detection features, scene visibility fluctuation index, and carrier attitude perturbation index are jointly input into the compensation model. The compensation model describes the impact of environmental perturbations on detection accuracy by constructing a feature correction function. The steps for constructing the feature correction function include: An environmental disturbance quantification vector is established based on the scene visibility fluctuation index and the carrier attitude disturbance index to characterize the combined effects of light attenuation, fog obstruction and attitude jitter. The environmental disturbance quantization vector is correlated with the fused target detection features, and the correction function parameters are solved by minimizing the target detection error. The correction function is used to describe the offset relationship between environmental disturbances and personnel position characteristics, mechanical boundary characteristics and safety protection characteristics, so as to provide a mathematical basis for subsequent correction.

[0073] In this embodiment, to compensate for feature weakening caused by insufficient lighting and rain / fog obstruction, a lighting compensation weight is calculated based on the scene visibility fluctuation index. The lighting compensation weight is obtained through the following steps: The degree of light attenuation is calculated based on the light disturbance value and transmittance disturbance value in the scene visibility fluctuation index; Using the degree of light attenuation as an adjustment factor, brightness reconstruction is performed on the brightness features in the fused target detection features, so that the brightness features are restored to a level that matches the normal visibility conditions; Texture enhancement processing is performed on the texture features in the fused target detection features according to the texture attenuation ratio, making edge details and safety protection features (such as safety helmet texture) clearer.

[0074] The dynamically compensated lighting features and texture features form an ambient lighting correction feature set.

[0075] In this embodiment, to eliminate target position offset caused by attitude jitter, a position offset correction amount is determined based on the carrier attitude disturbance index. The position offset correction amount is determined according to the following steps: The impact of attitude changes on target position is estimated based on the attitude coupling perturbation value and optical flow residual perturbation value in the carrier attitude perturbation index; A position offset vector is constructed based on the relationship between the orientation change direction and the offset. The position offset vector is used to describe the direction and magnitude of pixel offset. The position offset vector is applied to the personnel position features and mechanical boundary features in the fused target detection features, so that the position features are restored to their spatial position before the attitude disturbance.

[0076] The corrected positional features have higher spatial accuracy.

[0077] In this embodiment, the features after illumination feature compensation and pose feature compensation are input into the dynamic correction rules of the compensation model. The dynamic correction rules perform differentiated correction on different categories of features based on the environmental disturbance quantization vector, specifically including: A position correction coefficient is applied to the personnel position characteristics to ensure that the actual spatial position of the personnel is consistent with the detection position; Texture enhancement coefficients are applied to safety protection features so that features such as safety helmets and reflective vests can still be accurately identified in extreme environments; Apply boundary stabilization coefficients to mechanical boundary features to prevent the mechanical shape from deforming due to vibration.

[0078] After dynamic correction, environmentally compensated hazard identification features are generated, which are used for subsequent hazard classification and risk assessment.

[0079] Based on the hazard source identification features, threshold adaptive learning is performed to generate a dynamic threshold set that adjusts with environmental changes.

[0080] In this embodiment, feature parameters describing the criteria for determining different hazard types are extracted based on the hazard source identification features after environmental compensation. These parameters include personnel location features, mechanical boundary features, and safety protection features. Subsequently, the scene visibility fluctuation index and the carrier attitude disturbance index are combined to construct an environmental disturbance weight. This environmental disturbance weight is used to quantify the impact of environmental changes on the stability of the feature parameters, and its construction method is as follows: The brightness disturbance, transmittance disturbance, and texture attenuation in the scene visibility fluctuation index are assigned weight coefficients to the lighting environment, air scattering environment, and texture quality environment, respectively. The spatial migration environment is assigned weight coefficients based on the attitude coupling perturbation value, optical flow residual perturbation value, and fuzzy kernel direction perturbation value in the carrier attitude perturbation index, respectively. The above weighting coefficients are combined proportionally to form environmental disturbance weights, which are used for subsequent threshold adjustments.

[0081] In this embodiment, to determine the adjustment direction and magnitude of each threshold under different environmental conditions, a perturbation sensitivity analysis is performed on the preset thresholds. The threshold perturbation sensitivity analysis is completed through the following steps: The environmental disturbance weights are correlated with the corresponding changes in feature parameters to calculate the threshold shift trend caused by environmental changes. Based on the threshold shift trend, it is determined whether the thresholds for different feature categories should be increased or decreased in the current environment. For example, the safety distance threshold should be increased or the color recognition threshold should be decreased in a low-light environment. The threshold adjustment range is calculated, and the adjustment range is proportional to the environmental disturbance weights to ensure that the threshold adjustment is more significant when the environmental influence is stronger.

[0082] The direction and magnitude of threshold adjustment can be obtained through perturbation sensitivity analysis.

[0083] In this embodiment, to achieve adaptive threshold adjustment, a threshold learning model is constructed based on the feature parameters and the threshold adjustment direction. The steps for constructing the threshold learning model are as follows: The feature parameters, threshold adjustment direction, and environmental disturbance weights are used as input variables. The objective function is the hazard identification error, which is calculated based on the number and severity of misclassifications. The threshold update amount is solved by minimizing the hazard source determination error, so that the adjusted threshold can best adapt to the identification needs under the current environmental conditions.

[0084] The threshold learning model ensures that the threshold update process is convergent and interpretable.

[0085] In this embodiment, the threshold update amount is superimposed on a preset threshold to form a dynamic threshold set. The dynamic threshold set represents the optimal judgment threshold under environmental changes, and its generation method is as follows: Perform addition on each preset threshold and its corresponding threshold update; apply range constraints to the updated thresholds to keep them within a valid range; combine all updated thresholds to form a dynamic threshold set to guide subsequent hazard source determination.

[0086] The dynamic threshold set can be adjusted in real time according to changes in lighting, rain and fog, and attitude disturbances, thereby enabling hazard source identification to maintain high stability and high accuracy in complex environments.

[0087] Based on the hazard identification features and the dynamic threshold set, the rule deviation is calculated. When the rule deviation exceeds the set threshold, it is determined that there is a risk of missed or misjudged detection, and the hazard identification result of the construction site is output.

[0088] In this embodiment, to assess whether the hazard identification results deviate from the dynamic threshold set, feature parameters for rule comparison are extracted item by item from the environmentally compensated hazard identification features. The feature parameters include, by category: Personnel location characteristics: coordinates of the personnel's center point, the closest distance between the personnel and the equipment, and the change in the personnel's direction of movement; Mechanical boundary characteristics: boundary coordinates of the bounding rectangle of the mechanical equipment, direction of mechanical movement, and expansion amount of mechanical telescopic components; Safety protection features: color response value of safety helmet, intensity of texture feature of reflective clothing, and form matching degree of safety protective equipment.

[0089] Subsequently, the aforementioned feature parameters are paired with corresponding thresholds in the dynamic threshold set. The dynamic threshold set includes dynamic safety distance thresholds, dynamic mechanical intrusion thresholds, and dynamic security protection identification thresholds. The pairing method is established according to the mapping relationship of "feature type and threshold type are completely consistent", forming feature-threshold pairs to prepare for offset calculation.

[0090] In this embodiment, to quantitatively describe the deviation of hazard identification features from the dynamic threshold set, a feature offset is calculated for each feature-threshold pair. For example: The linear feature offset calculation method is applicable to features such as personnel distance and machine boundary. The offset is defined as the difference between the actual value of the feature parameter and the value of the corresponding dynamic threshold. For example, if the closest distance between a person and a machine is 2.3 meters and the dynamic safety distance threshold is 3.0 meters, then the offset is -0.7 meters.

[0091] The non-linear feature offset calculation method is applicable to features such as lighting, texture, and safety protection. The offset is calculated based on the rate of change of the feature parameter. For example, the response curve of a safety helmet's color response value to changes in lighting is non-linear; therefore, a non-linear offset is constructed as follows: Calculate the instantaneous rate of change of the feature parameters; calculate the rate of change of the dynamic threshold with environmental disturbances; and use the absolute value of the difference between the two rates as the nonlinear feature offset.

[0092] Offset sign rule: A negative offset indicates that the threshold has been exceeded (increased security risk), while a positive offset indicates that it is still within the safe range of the threshold.

[0093] The above methods can be used to obtain multiple offsets in three dimensions: personnel, machinery, and safety protection.

[0094] In this embodiment, since different feature offsets represent different risk levels, it is necessary to construct offset weights. The offset weights are determined as follows: Personnel location has the highest weighting and is used to reflect the high risk when personnel approach hazardous equipment. The mechanical boundary has the second highest weight, used to reflect the potential collision risk caused by mechanical motion; The safety protection features have a lower weight and are used to reflect secondary risks caused by insufficient protection.

[0095] The offset score is constructed through the following steps: Each feature offset is multiplied by its corresponding offset weight; all weighted offsets are then added together to form an offset score; the higher the offset score, the greater the deviation from the standard rule. The offset score serves as a comprehensive indicator to represent the overall degree of deviation from the rule.

[0096] In this embodiment, to determine the reliability of rule matching in the current environment, the offset score is compared with a rule reliability threshold. The rule reliability threshold is defined as the maximum tolerable range of offset scores, and this threshold is obtained based on historical data statistics.

[0097] When the deviation score is less than the rule reliability threshold, the rule deviation is low, indicating that the hazard identification features are reliable. When the offset score is greater than or equal to the rule reliability threshold, the rule deviation is high, indicating that the hazard source identification may be off track, requiring an early warning or further correction.

[0098] Rule deviation can quantify rule matching error, ensuring that hazard identification remains reliable under conditions such as changes in lighting, rain and fog interference, and drone attitude disturbances.

[0099] Example 2, please refer to Figure 2 As shown in this embodiment, a construction site hazard identification system based on unmanned aerial vehicles (UAVs) includes: Image acquisition module: Acquires image data of the construction site through drones, and uses target detection models to identify personnel, machinery and equipment and safety protection equipment to obtain initial target detection features; Calculation module: Calculates scene visibility fluctuation index and carrier attitude disturbance index based on image data and UAV attitude data; Image enhancement processing module: performs illumination enhancement, rain removal, or fog removal on the image based on the scene visibility fluctuation index to generate an enhanced image, and extracts the enhanced first target detection features from the enhanced image; Attitude compensation and correction module: Performs jitter compensation and attitude stabilization correction on the image according to the carrier attitude perturbation index, generates a corrected image, and extracts the second target detection features after attitude correction from the corrected image; Multi-source feature fusion module: performs feature fusion on the initial target detection features, the enhanced first target detection features, and the pose-corrected second target detection features to obtain fused target detection features; Environmental compensation calculation module: The module integrates target detection features, scene visibility fluctuation index and carrier attitude disturbance index into the compensation model to perform environmental compensation calculation. Based on the calculation results, it dynamically corrects personnel positions, mechanical boundaries and safety protection features to obtain the hazard source identification features after environmental compensation. Threshold adaptive learning module: Performs threshold adaptive learning based on the hazard source identification features to generate a dynamic threshold set that adjusts with changes in the environment; Hazard source rule deviation analysis module: Calculates the rule deviation degree based on the hazard source identification features and the dynamic threshold set. When the rule deviation degree exceeds the set threshold, it determines that there is a risk of missed or misjudged identification, and outputs the hazard source identification results at the construction site.

[0100] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for identifying hazardous sources at construction sites based on unmanned aerial vehicles (UAVs), characterized in that: include: By acquiring image data of the construction site through drones, and using target detection models to identify personnel, machinery and equipment and safety protection equipment, initial target detection features are obtained; Calculate the scene visibility fluctuation index and the carrier attitude disturbance index based on image data and UAV attitude data. Based on the scene visibility fluctuation index, the image is subjected to illumination enhancement, rain removal, or fog removal processing to generate an enhanced image, and the enhanced first target detection feature is extracted from the enhanced image. Based on the carrier attitude perturbation index, jitter compensation and attitude stabilization correction are performed on the image to generate a corrected image, and the second target detection feature after attitude correction is extracted from the corrected image. The initial target detection features, the enhanced first target detection features, and the pose-corrected second target detection features are fused to obtain fused target detection features. The environmental compensation calculation is performed by inputting the target detection features, scene visibility fluctuation index and carrier attitude disturbance index into the compensation model. Based on the calculation results, the personnel position, mechanical boundary and safety protection features are dynamically corrected to obtain the hazard source identification features after environmental compensation. Based on the hazard source identification features, threshold adaptive learning is performed to generate a dynamic threshold set that adjusts with environmental changes; Based on the hazard identification features and the dynamic threshold set, the rule deviation is calculated. When the rule deviation exceeds the set threshold, it is determined that there is a risk of missed or misjudged detection, and the hazard identification result of the construction site is output.

2. The method for identifying hazardous sources at construction sites based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that: The calculation of the scene visibility fluctuation index includes the following steps: Atmospheric scattering model inversion is performed based on image data to estimate the transmittance of the image in multiple local regions, and the deviation of the transmittance in the spatial distribution is calculated to obtain the transmittance perturbation value. Fourier frequency domain decomposition is performed on the image data to extract high-frequency energy components, and the sharpness fluctuation value is calculated based on the attenuation rate of high-frequency energy at adjacent time points. Based on the attitude data of the UAV, the pitch angle, roll angle, yaw angle and angular velocity information at the corresponding moment are obtained, and the coupling degree between the image blur degree and the attitude change is calculated to obtain the attitude disturbance value. The transmittance perturbation value, sharpness fluctuation value, and attitude perturbation value are fused and calculated according to a preset weight to generate the scene visibility fluctuation index.

3. The method for identifying hazardous sources at construction sites based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that: The calculation of the carrier attitude disturbance index includes the following steps: Optical flow field is extracted based on image data using optical flow estimation algorithm, and theoretical optical flow field is constructed based on UAV attitude angle changes. The residual signal energy between the actual optical flow field and the theoretical optical flow field is used as the optical flow residual perturbation value. Based on image data, a blind deblurring algorithm is used to estimate the image blur kernel, and the directional offset between the main direction of the blur kernel and the direction of change of the UAV attitude angle is calculated to obtain the directional perturbation value of the blur kernel. The attitude perturbation value, optical flow residual perturbation value, and fuzzy kernel direction perturbation value are fused and calculated according to a preset weight to obtain the carrier attitude perturbation index.

4. The method for identifying hazardous sources at construction sites based on unmanned aerial vehicles (UAVs) according to claim 2, characterized in that: Based on the scene visibility fluctuation index, the image is enhanced by illumination, de-raining, or de-hazing to generate an enhanced image, including the following steps: The degree of illumination disturbance is determined based on the scene visibility fluctuation index, and an illumination enhancement model is constructed. The image brightness is reconstructed using an illumination enhancement algorithm based on illumination component estimation and reflection component constraint to obtain an illumination-enhanced image. When the transmittance perturbation value corresponding to the scene visibility fluctuation index exceeds the set transmittance threshold, the image transmittance is estimated and corrected using a dehazing algorithm based on atmospheric scattering inversion, and a dehazed image is generated. When the high-frequency energy attenuation rate reflected by the scene visibility fluctuation index exceeds the set attenuation threshold, the rain texture component is extracted using a rain texture separation-based deraining algorithm, and the rain texture component is separated from the original image to generate a derained image. The image with the highest visibility is selected from the enhanced image, the dehazed image, or the derained image as the enhanced image, and the enhanced image is input into the target detection model to extract the enhanced first target detection feature.

5. The method for identifying hazardous sources at construction sites based on unmanned aerial vehicles (UAVs) according to claim 3, characterized in that: Based on the carrier attitude disturbance index, jitter compensation and attitude stabilization correction are performed on the image to generate a corrected image, including the following steps: The attitude jitter level is determined based on the attitude disturbance index of the carrier. When the attitude disturbance index exceeds the set attitude compensation threshold, an attitude transformation matrix is ​​constructed based on the pitch angle, roll angle and yaw angle in the UAV attitude data. The image data is reversed by using the attitude transformation matrix to compensate for the imaging geometric offset caused by attitude change pixel by pixel to obtain a preliminary attitude correction image. The image gradient stability is calculated for the initial attitude correction image. Motion blur estimation is performed for regions with insufficient gradient stability, and the local regions are deblurred according to the estimated blur kernel to generate the correction image. The corrected image is input into the target detection model to extract the second target detection features after pose correction.

6. The method for identifying hazardous sources at construction sites based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that: The acquisition of the fused target detection features includes the following steps: Based on the initial target detection features, the enhanced first target detection features, and the pose-corrected second target detection features, feature alignment processing is performed on the detection position, detection confidence, and feature vector of the same target in different images to obtain an aligned feature set. Calculate the feature consistency score based on the similarity value between each feature vector in the aligned feature set and the reference feature vector; A weighted fusion operation is performed on the aligned feature set according to the feature consistency score as the weight, so as to obtain the fused target feature vector; The target detection bounding box position and detection confidence are updated based on the fused target feature vector to generate the fused target detection features.

7. The method for identifying hazardous sources at construction sites based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that: The obtained hazard source identification features after environmental compensation include the following steps: The fused target detection features, the scene visibility fluctuation index, and the carrier attitude disturbance index are input into the compensation model; Adjust the brightness and texture features in the fused target detection features based on the scene visibility fluctuation index; Based on the carrier attitude perturbation index, position offset correction is performed on the spatial position information in the fused target detection features; The results after illumination feature compensation and posture feature compensation are input into the dynamic correction rules of the compensation model to perform comprehensive correction on personnel position, safety protection features and mechanical boundaries, and output the hazard source identification features after environmental compensation.

8. The method for identifying hazardous sources at construction sites based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that: The threshold adaptive learning includes the following steps: Based on the environmentally compensated hazard source identification features, feature parameters are extracted to characterize personnel location features, mechanical boundary features, and safety protection features, and environmental disturbance weights are constructed according to the scene visibility fluctuation index and the carrier attitude disturbance index. Based on the environmental disturbance weights, a disturbance sensitivity analysis is performed on the preset thresholds to determine the adjustment direction and magnitude of the thresholds for each characteristic parameter under different environmental conditions. A threshold learning model is constructed based on the feature parameters and the threshold adjustment direction, and the threshold update amount is solved by minimizing the hazard source determination error. The threshold update amount is superimposed with the preset threshold to generate a threshold set that dynamically adjusts with changes in the environment.

9. The method for identifying hazardous sources at construction sites based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that: The calculation of the rule deviation includes the following steps: Based on the hazard source identification features, feature parameters for determining the hazardous state are extracted from personnel location features, mechanical boundary features, and safety protection features, and the feature parameters are paired with the corresponding thresholds in the dynamic threshold set; Calculate the feature offset for each set of feature parameters and the corresponding threshold; The offset weights are set according to the degree of danger of each feature offset, and the offset weights are weighted and accumulated to obtain an offset score for comprehensively characterizing the accuracy of rule matching. The offset score is compared with a set rule reliability threshold to generate a rule deviation score.

10. A construction site hazard identification system based on unmanned aerial vehicles (UAVs), used to implement the construction site hazard identification method based on UAVs as described in any one of claims 1-9, characterized in that: include: Image acquisition module: Acquires image data of the construction site through drones, and uses target detection models to identify personnel, machinery and equipment and safety protection equipment to obtain initial target detection features; Calculation module: Calculates scene visibility fluctuation index and carrier attitude disturbance index based on image data and UAV attitude data; Image enhancement processing module: performs illumination enhancement, rain removal, or fog removal on the image based on the scene visibility fluctuation index to generate an enhanced image, and extracts the enhanced first target detection features from the enhanced image; Attitude compensation and correction module: Performs jitter compensation and attitude stabilization correction on the image according to the carrier attitude perturbation index, generates a corrected image, and extracts the second target detection features after attitude correction from the corrected image; Multi-source feature fusion module: performs feature fusion on the initial target detection features, the enhanced first target detection features, and the pose-corrected second target detection features to obtain fused target detection features; Environmental compensation calculation module: The module integrates target detection features, scene visibility fluctuation index and carrier attitude disturbance index into the compensation model to perform environmental compensation calculation. Based on the calculation results, it dynamically corrects personnel positions, mechanical boundaries and safety protection features to obtain the hazard source identification features after environmental compensation. Threshold adaptive learning module: Performs threshold adaptive learning based on the hazard source identification features to generate a dynamic threshold set that adjusts with changes in the environment; Hazard source rule deviation analysis module: Calculates the rule deviation degree based on the hazard source identification features and the dynamic threshold set. When the rule deviation degree exceeds the set threshold, it determines that there is a risk of missed or misjudged identification, and outputs the hazard source identification results at the construction site.

Images