Building surveying and mapping system based on unmanned aerial vehicle remote sensing technology

Abstract

The invention discloses a building surveying and mapping system based on an unmanned aerial vehicle remote sensing technology, and belongs to the technical field of building surveying and mapping. Comprising the following modules: a multi-sensor data acquisition module and an unmanned aerial vehicle data acquisition platform based on an integrated RGB camera, a laser radar and a multi-spectral sensor; the preliminary image processing module is used for generating standardized image data for scene analysis; the intelligent scene recognition module outputs a scene classification result and preliminary complexity grade evaluation; the self-adaptive decision module is used for generating an initial image processing parameter combination according to a scene recognition result; the dynamic image processing module is used for executing regional image registration, correction and seamless splicing based on the initial parameter combination to generate a building orthoimage; and the real-time quality evaluation module monitors and evaluates geometric accuracy and visual quality in the image processing process in real time.

Description

Technical Field

[0001] This application relates to the field of architectural surveying technology, and more specifically, to an architectural surveying system based on UAV remote sensing technology. Background Technology

[0002] With the rapid pace of urban construction, the importance of architectural surveying is becoming increasingly prominent. Traditional surveying mainly relies on manual on-site surveys or data acquisition using single sensors, which is not only time-consuming and labor-intensive but also greatly affected by environmental conditions, making it difficult to meet modern needs in terms of accuracy and efficiency. Unmanned aerial vehicle (UAV) remote sensing technology, due to its flexibility, wide coverage, and ability to quickly acquire high-resolution data, is gradually being widely applied in the field of architectural surveying.

[0003] However, most current UAV mapping systems rely on a single sensor, failing to fully reflect the structural details of buildings and the complexity of the surrounding environment, resulting in limited mapping results. Furthermore, image processing often uses fixed parameters, lacking strategies to adjust based on different mapping scenarios, leading to less than ideal processing results. In addition, the lack of real-time quality monitoring during the mapping process makes it difficult to guarantee the accuracy and completeness of the data, affecting subsequent application effectiveness.

[0004] In summary, how to achieve dynamic and intelligent processing of UAV remote sensing data based on multi-sensor fusion, improve the accuracy and efficiency of building surveying, and ensure real-time evaluation and feedback of surveying data quality has become an urgent technical problem to be solved. Summary of the Invention

[0005] To overcome a series of shortcomings in existing technologies, the purpose of this application is to provide a building surveying system based on UAV remote sensing technology, comprising the following modules: Multi-sensor data acquisition module, based on a UAV data acquisition platform integrating RGB camera, LiDAR and multispectral sensor; A preliminary image processing module is used to generate standardized image data for scene analysis; The intelligent scene recognition module outputs scene classification results and a preliminary complexity level assessment; The adaptive decision-making module generates an initial combination of image processing parameters based on the scene recognition results. The dynamic image processing module performs regional image registration, correction, and seamless stitching based on the initial parameter combination to generate building orthophotos; The real-time quality assessment module monitors and evaluates the geometric accuracy and visual quality during image processing in real time.

[0006] Furthermore, the RGB camera employs a 20-megapixel APS-C format sensor, equipped with a 24-70mm variable focal length lens, a shutter speed range of 1 / 2000 second to 1 second, and an ISO sensitivity range of 100-6400. It also integrates a three-axis electronic image stabilization system to reduce the impact of vibrations during flight. The lidar sensor uses a 905nm wavelength pulsed laser, with a scanning frequency of 100,000 points per second, a vertical field of view of 30 degrees, and a horizontal field of view of 270 degrees. It can operate stably within an ambient temperature range of -10℃ to +50℃. The range is 5-150 meters, and it has a built-in inertial measurement unit for attitude correction, recording the precise position and attitude information at the moment of each laser pulse emission in real time. The multispectral sensor covers the visible spectrum, near-infrared spectrum, and short-wave infrared spectrum, specifically including five independent channels: blue light band (450-520 nm), green light band (520-600 nm), red light band (630-690 nm), near-infrared band (760-900 nm), and short-wave infrared band (1550-1750 nm). Each channel is equipped with a 12-bit ADC converter, and the dynamic range reaches 72 dB.

[0007] Furthermore, the intelligent scene recognition module includes the following components: The feature extraction and fusion unit is responsible for performing deep feature extraction and multi-dimensional feature fusion on RGB images, LiDAR point clouds and multispectral data to generate a unified multi-dimensional feature vector. Building type identification unit, used to automatically identify building types and output classification labels; The environmental condition assessment unit is used to assess the impact of the current surveying and mapping environment on the quality of data acquisition. The preliminary complexity calculation unit performs a preliminary complexity assessment based on the identified building types and environmental conditions, providing a reference for parameter selection; The scene semantic annotation unit is used to generate comprehensive scene description tags that include building type, environmental conditions, and initial complexity level for use by the adaptive decision-making module.

[0008] Furthermore, the feature extraction and fusion unit adopts a deep convolutional neural network architecture, which includes 5 convolutional layers and 3 fully connected layers. Each convolutional layer uses a 3×3 convolutional kernel with a stride of 1 and a same padding method. The RGB image feature extraction network outputs a 512-dimensional feature vector, the LiDAR point cloud feature extraction adopts a PointNet network architecture to output a 256-dimensional feature vector, and the multispectral data extracts a 128-dimensional feature vector through a one-dimensional convolutional network. The three feature vectors are weighted and fused through an attention mechanism to finally generate a unified 896-dimensional multidimensional feature vector.

[0009] Furthermore, the preliminary complexity calculation unit performs quantitative evaluation based on three indicators: building outline complexity, surface texture complexity, and height variation complexity. Specifically, building outline complexity is calculated by the perimeter-to-area ratio after edge detection, with a value range of 1-20; surface texture complexity is calculated using the contrast and entropy values ​​of the gray-level co-occurrence matrix, normalized to a range of 0-10; and height variation complexity is calculated based on the elevation standard deviation of the lidar point cloud, with a range of 0-15.

[0010] Furthermore, the adaptive decision-making module includes the following components: The scene mapping unit maps the output of the intelligent scene recognition module to the candidate parameter set, and initially filters the parameter combinations that meet the processing requirements. The parameter optimization unit is used to perform weight evaluation and performance prediction on the mapped parameter set, determine the priority order of the initial parameter combination, and predict the processing efficiency. The confidence assessment unit estimates the uncertainty of the parameter optimization results, calculates the execution reliability index of the selected parameter combination, and automatically selects alternative parameter combinations when the confidence level is lower than the threshold. The parameter management unit sends the final selected and confidence-verified parameter combination to the dynamic image processing module, receives feedback information from the real-time quality assessment module to dynamically adjust the parameters, maintains the parameter adjustment history to support regional reprocessing decisions, and records all parameter configurations and update logs.

[0011] Furthermore, the scene mapping unit establishes a parameter library containing 36 predefined parameter sets. Each parameter set is optimized for a specific combination of building type and complexity level. The parameter sets include the feature point number threshold, matching distance threshold, control point density and stitching overlap in image registration, and the control point density and stitching overlap in geometric correction. The mapping process establishes an association matrix between scene features and parameter combinations and uses a nearest neighbor search algorithm to quickly locate the most suitable parameter combination.

[0012] Furthermore, the dynamic image processing module includes the following components: The region division unit performs preliminary region division based on the spatial distribution characteristics of the image, and dynamically adjusts the region boundaries based on the actual complexity feedback during the processing. The image registration unit precisely aligns multiple images within the same sub-region based on the current combination of image processing parameters. The geometric correction unit performs perspective and distortion correction on the registered images based on the sensor's exterior orientation elements and digital elevation model, ensuring the geometric consistency of images of the same area in the world coordinate system. The precise complexity assessment unit performs a precise reassessment of the spatial complexity of each region based on the high-quality image data after registration and correction. The regional reprocessing unit, when real-time quality assessment finds that the quality of a specific region is substandard, adjusts the historical and complexity assessment results based on parameters and performs reprocessing or local optimization on that region.

[0013] Furthermore, the precise complexity evaluation unit uses a convolutional neural network to extract multi-scale features of the building. The network includes a ResNet-50 backbone network and a feature pyramid network, with an input image size of 512×512 pixels. The complexity evaluation is quantified from three dimensions: geometric complexity, texture complexity, and occlusion complexity of the building. Each dimension outputs a score from 0 to 10, and the final complexity level is the weighted average of the scores of the three dimensions, with weights of 0.4, 0.3, and 0.3, respectively.

[0014] Furthermore, the real-time quality assessment module employs a multi-index comprehensive assessment method, including four aspects: geometric accuracy assessment, image sharpness assessment, color consistency assessment, and stitching gap assessment. Specifically: geometric accuracy is assessed through statistical evaluation of ground control point residuals, requiring a planar accuracy better than ±0.8 meters; image sharpness is calculated using the Laplacian operator to determine the gradient amplitude, with a threshold set to 30; color consistency is assessed through the difference in mean and standard deviation of overlapping areas of adjacent images, with difference thresholds of 10 and 5, respectively; and stitching gaps are identified through edge detection and morphological operations, with a gap width threshold set to 3 pixels.

[0015] Compared with the prior art, this application has the following beneficial effects: This application combines multi-sensor data from UAVs with deep learning scene recognition, adaptive parameter decision-making, and real-time quality assessment. It can dynamically select processing parameters according to different building environments and complete image registration, correction, and seamless stitching by region, achieving high-precision and high-efficiency building surveying. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the structure of a building surveying system based on UAV remote sensing technology disclosed in an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some embodiments of this invention, but not all embodiments.

[0018] 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.

[0019] The embodiments and directional terms described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0020] like Figure 1 As shown, a building surveying system based on UAV remote sensing technology includes the following modules: The multi-sensor data acquisition module, based on a UAV data acquisition platform integrating RGB camera, LiDAR and multispectral sensor, enables synchronous acquisition and initial preprocessing of multi-source heterogeneous data; The preliminary image processing module performs basic registration and preliminary correction on the preprocessed multi-source data to generate standardized image data for scene analysis. The intelligent scene recognition module automatically identifies building types and environmental conditions based on standardized image data, and outputs scene classification results and preliminary complexity level assessment. The adaptive decision-making module generates an initial combination of image processing parameters based on the scene recognition results. The dynamic image processing module performs regional image registration, correction, and seamless stitching based on the initial parameter combination to generate building orthophotos; The real-time quality assessment module monitors and evaluates the geometric accuracy and visual quality during image processing in real time. It provides timely feedback on processing results through an anomaly detection mechanism and triggers dynamic parameter adjustment or region reprocessing mechanisms when quality indicators deviate from the expected threshold.

[0021] This system enhances the richness and accuracy of building surveying data through synchronous acquisition of data from multiple sensors. Preliminary image processing standardizes multi-source data, laying the foundation for subsequent analysis. The intelligent scene recognition module automatically distinguishes between building types and their environments, assisting in determining processing strategies. The adaptive decision-making and dynamic image processing modules achieve precise regional correction and stitching, ensuring the integrity and accuracy of building orthophotos. A real-time quality assessment mechanism effectively monitors accuracy and visual quality during processing, promptly adjusting parameters or reprocessing abnormal areas.

[0022] Furthermore, the RGB camera employs a 20-megapixel APS-C format sensor, equipped with a 24-70mm variable focal length lens, a shutter speed range of 1 / 2000 second to 1 second, and an ISO sensitivity range of 100-6400. It also integrates a three-axis electronic image stabilization system to reduce the impact of vibrations during flight. The lidar sensor uses a 905nm wavelength pulsed laser, with a scanning frequency of 100,000 points per second, a vertical field of view of 30 degrees, and a horizontal field of view of 270 degrees. It can operate stably within an ambient temperature range of -10℃ to +50℃. The range is 5-150 meters, and it has a built-in inertial measurement unit for attitude correction, recording the precise position and attitude information at the moment of each laser pulse emission in real time. The multispectral sensor covers the visible spectrum, near-infrared spectrum, and short-wave infrared spectrum, specifically including five independent channels: blue light band (450-520 nm), green light band (520-600 nm), red light band (630-690 nm), near-infrared band (760-900 nm), and short-wave infrared band (1550-1750 nm). Each channel is equipped with a 12-bit ADC converter, and the dynamic range reaches 72 dB.

[0023] Furthermore, the preliminary image processing module extracts 500-2000 feature points from each image, establishes a feature point correspondence matrix, and calculates affine transformation parameters. During the basic registration process, the feature point matching threshold is set to 0.8, the distance ratio test threshold is set to 0.7, and the number of iterations is set to 1000. The accuracy of the preliminary correction is controlled within 0.5 pixels to ensure high consistency of multi-source data in spatial location.

[0024] Furthermore, the intelligent scene recognition module includes the following components: The feature extraction and fusion unit is responsible for performing deep feature extraction and multi-dimensional feature fusion on RGB images, LiDAR point clouds and multispectral data to generate a unified multi-dimensional feature vector. Building type identification unit, used to automatically identify building types and output classification labels; The environmental condition assessment unit is used to assess the impact of the current surveying and mapping environment on the quality of data acquisition. The preliminary complexity calculation unit performs a preliminary complexity assessment based on the identified building types and environmental conditions, providing a reference for parameter selection; The scene semantic annotation unit is used to generate comprehensive scene description tags that include building type, environmental conditions, and initial complexity level for use by the adaptive decision-making module.

[0025] As described above, the intelligent scene recognition module improves the comprehensiveness and accuracy of scene perception by fusing multi-source data to extract a unified feature vector. Building type recognition and environmental condition assessment enable rapid identification of surveying targets and external interference factors, helping to improve the targeting of subsequent processing. The complexity calculation unit performs a preliminary assessment of task difficulty based on multi-factor information, providing a basis for setting image processing parameters. The scene semantic annotation unit generates structured labels, providing key support for adaptive processing, and overall improving the automation level and operational efficiency of the surveying system.

[0026] Furthermore, the feature extraction and fusion unit adopts a deep convolutional neural network architecture, which includes 5 convolutional layers and 3 fully connected layers. Each convolutional layer uses a 3×3 convolutional kernel with a stride of 1 and a same padding method. The RGB image feature extraction network outputs a 512-dimensional feature vector, the LiDAR point cloud feature extraction adopts a PointNet network architecture to output a 256-dimensional feature vector, and the multispectral data extracts a 128-dimensional feature vector through a one-dimensional convolutional network. The three feature vectors are weighted and fused through an attention mechanism to finally generate a unified 896-dimensional multidimensional feature vector.

[0027] Furthermore, the building type identification unit establishes a building classification system covering residential buildings, commercial buildings, industrial buildings, educational buildings, medical buildings, sports buildings, transportation buildings, and cultural buildings, with each main category further subdivided into 3-5 subcategories; the building is classified based on a support vector machine classifier, with a radial basis function as the kernel function, a regularization parameter C set to 10.0, and a kernel parameter gamma set to 0.001.

[0028] Furthermore, the environmental condition assessment unit comprehensively evaluates four dimensions: illumination conditions, weather conditions, atmospheric visibility, and wind speed. Illumination conditions are assessed based on image brightness histogram statistics, calculating average grayscale values ​​and contrast indices. Weather conditions are determined through atmospheric transmittance analysis of multispectral data. Atmospheric visibility is calculated using Koschmead's law to determine the line-of-sight distance. Wind speed is estimated using data from UAV flight attitude sensors. The four assessment dimensions are assigned weights of 0.3, 0.25, 0.25, and 0.2, respectively, with a comprehensive score range of 0-100 points. A higher score indicates more favorable environmental conditions for data collection.

[0029] Furthermore, the preliminary complexity calculation unit performs quantitative evaluation based on three indicators: building outline complexity, surface texture complexity, and height variation complexity. Specifically, building outline complexity is calculated by the perimeter-to-area ratio after edge detection, with a value range of 1-20; surface texture complexity is calculated using the contrast and entropy values ​​of the gray-level co-occurrence matrix, normalized to a range of 0-10; and height variation complexity is calculated based on the elevation standard deviation of the lidar point cloud, with a range of 0-15.

[0030] Furthermore, the adaptive decision-making module includes the following components: The scene mapping unit maps the output of the intelligent scene recognition module to the candidate parameter set, and initially filters the parameter combinations that meet the processing requirements. The parameter optimization unit is used to perform weight evaluation and performance prediction on the mapped parameter set, determine the priority order of the initial parameter combination, and predict the processing efficiency. The confidence assessment unit estimates the uncertainty of the parameter optimization results, calculates the execution reliability index of the selected parameter combination, and automatically selects alternative parameter combinations when the confidence level is lower than the threshold. The parameter management unit sends the final selected and confidence-verified parameter combination to the dynamic image processing module, receives feedback information from the real-time quality assessment module to dynamically adjust the parameters, maintains the parameter adjustment history to support regional reprocessing decisions, and records all parameter configurations and update logs.

[0031] As described above, the adaptive decision-making module achieves precise parameter selection and preliminary optimization by matching scene recognition results with a parameter library, thereby improving the targeting of image processing. Weight evaluation and performance prediction help to rationally rank parameter combinations and improve processing efficiency. Confidence assessment ensures that the selected parameters have sufficient reliability, reducing processing risks caused by inappropriate parameters. Dynamic parameter management, combined with real-time quality feedback, supports flexible parameter adjustment and historical tracking, enhancing the system's adaptability and stability, and providing a basis for subsequent regional reprocessing.

[0032] Furthermore, the scene mapping unit establishes a parameter library containing 36 predefined parameter sets. Each parameter set is optimized for a specific combination of building type and complexity level. The parameter sets include the feature point number threshold, matching distance threshold, control point density and stitching overlap in image registration, and the control point density and stitching overlap in geometric correction. The mapping process establishes an association matrix between scene features and parameter combinations and uses a nearest neighbor search algorithm to quickly locate the most suitable parameter combination.

[0033] Furthermore, the parameter optimization unit employs a multi-objective optimization algorithm to evaluate candidate parameter combinations. The optimization objectives include three aspects: processing accuracy, processing speed, and resource consumption. Specifically, processing accuracy is estimated using the root mean square error from historical data statistics, with a target value controlled within 0.3 pixels; processing speed is calculated based on algorithm complexity and hardware performance, with a target of processing time not exceeding 15 minutes per square kilometer; resource consumption is assessed by evaluating CPU and memory usage, with a target of CPU usage below 80% and memory usage below 70%. The three objectives are combined using a weighted summation method to synthesize a comprehensive score, with weights of 0.5, 0.3, and 0.2, respectively.

[0034] Furthermore, the confidence assessment unit calculates confidence based on three factors: historical execution success rate, current scenario matching degree, and parameter combination complexity. Specifically, the historical execution success rate has a weight of 0.4, calculated by statistically analyzing the number of successful processing operations in similar scenarios; the current scenario matching degree has a weight of 0.35, calculated based on the cosine similarity of the scenario feature vectors; and the parameter combination complexity has a weight of 0.25, evaluated based on the number of parameters and their value range. The confidence threshold is set to 0.75, and alternative parameter combinations are automatically selected when the value is below this threshold.

[0035] Furthermore, the dynamic image processing module includes the following components: The region division unit performs preliminary region division based on the spatial distribution characteristics of the image, and dynamically adjusts the region boundaries based on the actual complexity feedback during the processing. The image registration unit precisely aligns multiple images within the same sub-region based on the current combination of image processing parameters. The geometric correction unit performs perspective and distortion correction on the registered images based on the sensor's exterior orientation elements and digital elevation model, ensuring the geometric consistency of images of the same area in the world coordinate system. The precise complexity assessment unit performs a precise reassessment of the spatial complexity of each region based on the high-quality image data after registration and correction. The regional reprocessing unit, when real-time quality assessment finds that the quality of a specific region is substandard, adjusts the historical and complexity assessment results based on parameters and performs reprocessing or local optimization on that region.

[0036] As described above, the dynamic image processing module dynamically divides image regions and optimizes processing boundaries using complexity feedback, improving the targeted nature of zoning processing. The image registration unit ensures precise alignment between multiple images of the same region, enhancing data fusion performance. The geometric correction unit corrects distortion and perspective aberration, ensuring the consistency and accuracy of images in the world coordinate system. The complexity assessment unit, based on high-quality data, meticulously evaluates the spatial characteristics of each region, supporting subsequent optimization. The regional reprocessing mechanism targets areas with quality anomalies, combining historical parameters and complexity information to implement targeted reprocessing, improving overall image quality and mapping accuracy.

[0037] Furthermore, the region division unit adopts an adaptive block-based strategy based on quadtree decomposition. Initially, the survey area is divided into square grids with a side length of 500 meters. Then, it is further subdivided according to the building density and complexity within each grid. Specifically, building density is calculated by the number of building outlines per unit area, with a threshold set at 20 buildings per square kilometer. Complexity is assessed based on the standard deviation of building height and the irregularity of the outline, with thresholds set at 15 meters and 0.6, respectively. When the building density or complexity within a grid exceeds the threshold, the grid is subdivided into 4 sub-grids, with the minimum block size limited to 100 meters × 100 meters, ensuring a balance between processing efficiency and accuracy.

[0038] Furthermore, the image registration unit employs a layered registration strategy to achieve precise alignment for multiple images within the same sub-region. The specific steps are as follows: First, based on the initial position and attitude information provided by GPS and IMU data, the approximate geometric transformation relationship between images is calculated; then, Harris corner points and SIFT feature points are extracted, feature point matching relationships are established, and precise affine transformation parameters are solved. During the registration process, the feature point extraction threshold is set to 0.01, the maximum number of feature points is 1500, the number of matching point pairs is required to be no less than 50 pairs, and the registration accuracy is controlled within 0.2 pixels.

[0039] Furthermore, the geometric correction unit performs rigorous geometric correction using exterior orientation elements provided by the UAV-borne GPS and attitude parameters provided by the airborne IMU. The correction process employs a digital elevation model to eliminate the influence of terrain undulations and uses a bilinear interpolation resampling method to maintain the radiometric characteristics of the original image.

[0040] Furthermore, the precise complexity evaluation unit uses a convolutional neural network to extract multi-scale features of the building. The network includes a ResNet-50 backbone network and a feature pyramid network, with an input image size of 512×512 pixels. The complexity evaluation is quantified from three dimensions: geometric complexity, texture complexity, and occlusion complexity of the building. Each dimension outputs a score from 0 to 10, and the final complexity level is the weighted average of the scores of the three dimensions, with weights of 0.4, 0.3, and 0.3, respectively.

[0041] Furthermore, the real-time quality assessment module employs a multi-index comprehensive assessment method, including four aspects: geometric accuracy assessment, image sharpness assessment, color consistency assessment, and stitching gap assessment. Specifically: geometric accuracy is assessed through statistical evaluation of ground control point residuals, requiring a planar accuracy better than ±0.8 meters; image sharpness is calculated using the Laplacian operator to determine the gradient amplitude, with a threshold set to 30; color consistency is assessed through the difference in mean and standard deviation of overlapping areas of adjacent images, with difference thresholds of 10 and 5, respectively; and stitching gaps are identified through edge detection and morphological operations, with a gap width threshold set to 3 pixels.

[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A building surveying system based on UAV remote sensing technology, characterized in that, Includes the following modules: Multi-sensor data acquisition module, based on a UAV data acquisition platform integrating RGB camera, LiDAR and multispectral sensor; A preliminary image processing module is used to generate standardized image data for scene analysis; The intelligent scene recognition module outputs scene classification results and a preliminary complexity level assessment; The adaptive decision-making module generates an initial combination of image processing parameters based on the scene recognition results. The dynamic image processing module performs regional image registration, correction, and seamless stitching based on the initial parameter combination to generate building orthophotos; The real-time quality assessment module monitors and evaluates the geometric accuracy and visual quality during image processing in real time.

2. The building surveying system based on UAV remote sensing technology according to claim 1, characterized in that, The RGB camera uses a 20-megapixel APS-C format sensor, equipped with a 24-70mm variable focal length lens, a shutter speed range of 1 / 2000 second to 1 second, and an ISO sensitivity range of 100-6400. It also integrates a three-axis electronic image stabilization system to reduce the impact of vibrations during flight. The lidar sensor uses a 905nm wavelength pulsed laser, with a scanning frequency of 100,000 points per second, a vertical field of view of 30 degrees, and a horizontal field of view of 270 degrees. It can operate stably within an ambient temperature range of -10℃ to +50℃, and its ranging range is [not specified]. The range is 5-150 meters, and it has an inertial measurement unit for attitude correction, recording the precise position and attitude information at the moment of each laser pulse emission in real time. The multispectral sensor covers the visible spectrum, near-infrared spectrum, and short-wave infrared spectrum, specifically including five independent channels: blue light band (450-520 nm), green light band (520-600 nm), red light band (630-690 nm), near-infrared band (760-900 nm), and short-wave infrared band (1550-1750 nm). Each channel is equipped with a 12-bit ADC converter, and the dynamic range reaches 72 dB.

3. The building surveying system based on UAV remote sensing technology according to claim 1, characterized in that, The intelligent scene recognition module includes the following components: The feature extraction and fusion unit is responsible for performing deep feature extraction and multi-dimensional feature fusion on RGB images, LiDAR point clouds and multispectral data to generate a unified multi-dimensional feature vector. Building type identification unit, used to automatically identify building types and output classification labels; The environmental condition assessment unit is used to assess the impact of the current surveying and mapping environment on the quality of data acquisition. The preliminary complexity calculation unit performs a preliminary complexity assessment based on the identified building types and environmental conditions, providing a reference for parameter selection; The scene semantic annotation unit is used to generate comprehensive scene description tags that include building type, environmental conditions, and initial complexity level for use by the adaptive decision-making module.

4. The building surveying system based on UAV remote sensing technology according to claim 3, characterized in that, The feature extraction and fusion unit adopts a deep convolutional neural network architecture, which includes 5 convolutional layers and 3 fully connected layers. Each convolutional layer uses a 3×3 convolutional kernel with a stride of 1 and a same padding method. The RGB image feature extraction network outputs a 512-dimensional feature vector, the LiDAR point cloud feature extraction adopts the PointNet network architecture to output a 256-dimensional feature vector, and the multispectral data extracts a 128-dimensional feature vector through a one-dimensional convolutional network. The three feature vectors are weighted and fused through an attention mechanism to finally generate a unified 896-dimensional multidimensional feature vector.

5. A building surveying system based on UAV remote sensing technology according to claim 3, characterized in that, The preliminary complexity calculation unit performs quantitative evaluation based on three indicators: building outline complexity, surface texture complexity, and height variation complexity. Specifically, building outline complexity is calculated by the perimeter-to-area ratio after edge detection, with a value range of 1-20; surface texture complexity is calculated using the contrast and entropy values ​​of the gray-level co-occurrence matrix, normalized to a range of 0-10; and height variation complexity is calculated based on the elevation standard deviation of the lidar point cloud, with a range of 0-15.

6. The building surveying system based on UAV remote sensing technology according to claim 1, characterized in that, The adaptive decision-making module includes the following components: The scene mapping unit maps the output of the intelligent scene recognition module to the candidate parameter set, and initially filters the parameter combinations that meet the processing requirements. The parameter optimization unit is used to perform weight evaluation and performance prediction on the mapped parameter set, determine the priority order of the initial parameter combination, and predict the processing efficiency. The confidence assessment unit estimates the uncertainty of the parameter optimization results, calculates the execution reliability index of the selected parameter combination, and automatically selects alternative parameter combinations when the confidence level is lower than the threshold. The parameter management unit sends the final selected and confidence-verified parameter combination to the dynamic image processing module, receives feedback information from the real-time quality assessment module to dynamically adjust the parameters, maintains the parameter adjustment history to support regional reprocessing decisions, and records all parameter configurations and update logs.

7. A building surveying system based on UAV remote sensing technology according to claim 6, characterized in that, The scene mapping unit establishes a parameter library containing 36 predefined parameter sets. Each parameter set is optimized for a specific combination of building type and complexity level. The parameter sets include the feature point number threshold, matching distance threshold, control point density and stitching overlap in image registration, and the control point density and stitching overlap in geometric correction. The mapping process establishes an association matrix between scene features and parameter combinations and uses a nearest neighbor search algorithm to quickly locate the most suitable parameter combination.

8. A building surveying system based on UAV remote sensing technology according to claim 1, characterized in that, The dynamic image processing module includes the following components: The region division unit performs preliminary region division based on the spatial distribution characteristics of the image, and dynamically adjusts the region boundaries based on the actual complexity feedback during the processing. The image registration unit precisely aligns multiple images within the same sub-region based on the current combination of image processing parameters. The geometric correction unit performs perspective and distortion correction on the registered images based on the sensor's exterior orientation elements and digital elevation model, ensuring the geometric consistency of images of the same area in the world coordinate system. The precise complexity assessment unit performs a precise reassessment of the spatial complexity of each region based on the high-quality image data after registration and correction. The regional reprocessing unit, when real-time quality assessment finds that the quality of a specific region is substandard, adjusts the historical and complexity assessment results based on parameters and performs reprocessing or local optimization on that region.

9. A building surveying system based on UAV remote sensing technology according to claim 8, characterized in that, The precise complexity evaluation unit uses a convolutional neural network to extract multi-scale features of buildings. The network includes a ResNet-50 backbone network and a feature pyramid network. The input image size is 512×512 pixels. The complexity evaluation is quantified from three dimensions: geometric complexity, texture complexity, and occlusion complexity. Each dimension outputs a score from 0 to 10. The final complexity level is the weighted average of the scores of the three dimensions, with weights of 0.4, 0.3, and 0.3, respectively.

10. A building surveying system based on UAV remote sensing technology according to claim 8, characterized in that, The real-time quality assessment module employs a multi-index comprehensive assessment method, including four aspects: geometric accuracy assessment, image sharpness assessment, color consistency assessment, and stitching gap assessment. Specifically: geometric accuracy is assessed through statistical evaluation of ground control point residuals, requiring a planar accuracy better than ±0.8 meters; image sharpness is calculated using the Laplacian operator to determine gradient amplitude, with a threshold set to 30; color consistency is assessed through the difference in mean and standard deviation of overlapping areas of adjacent images, with difference thresholds of 10 and 5, respectively; and stitching gaps are identified through edge detection and morphological operations, with a gap width threshold set to 3 pixels.

Images