Unmanned aerial vehicle intelligent surveying and mapping system carrying multi-mode sensor and method

Abstract

The invention discloses an unmanned aerial vehicle intelligent surveying and mapping system and method carrying a multi-modal sensor, and belongs to the technical field of unmanned aerial vehicles, and the method comprises the steps: extracting reference sensor data scene features, and generating an atomization subtask; dynamically matching and establishing a sensor execution combination according to the sensor capability portrait; the driving combination collects multi-modal data, and a collaborative data packet is generated through space-time alignment and consistency verification; evaluating a task value and settling resource consumption by combining energy consumption and time parameters, and dynamically adjusting a flight path; the unmanned aerial vehicle executes according to a path, and iteratively optimizes a sensor combination based on a real-time pose and a sensing state; and performing spatial recombination and attribute normalization on the collaborative data packet, and outputting a structured data packet. Real-time scene feature extraction and semantic understanding are adopted to generate a task, and a closed-loop control strategy of a sensor combination and a flight path is dynamically optimized in combination with task value evaluation and resource consumption settlement, so that efficient, intelligent and high-quality collaborative data acquisition of a target area can be realized.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and in particular to an intelligent mapping system and method for UAVs equipped with multimodal sensors. Background Technology

[0002] Intelligent mapping using unmanned aerial vehicles (UAVs) refers to the use of UAV platforms equipped with multimodal sensors such as visible light cameras, lidar, thermal imagers, and hyperspectral analyzers to collect data on the Earth's surface in an automated or intelligent manner, and then using computer vision and image processing technologies to generate geospatial information products such as maps and 3D models. This technology plays an increasingly important role in fields such as topographic mapping, disaster emergency response, and urban management, and its core lies in how to efficiently and accurately acquire and process multi-source sensing data.

[0003] Among related technologies, Chinese invention patent application CN120411837A discloses a system and method for real-time target detection and intelligent recognition based on UAV imagery. This includes: 1) A dual-modal sensor is used to achieve pixel-level data alignment via a spatiotemporal synchronization error control matrix; 2) A dynamic feature fusion module is used to capture the differential characterization of visible light and methane data and enhance concentration-spatial joint features through cross-modal weight learning, leakage characteristic perception, and adaptive fusion units; 3) A regional attention calculation module is constructed to generate sub-regional attention scores based on multi-dimensional data such as historical leakage records and pipeline density, and dynamically allocate computing resource priorities; 4) A model optimization module combines attention levels and real-time computing load to dynamically adjust channel resources through a baseline retention ratio and dynamic pruning strategies.

[0004] Regarding the aforementioned technologies, the inventors believe that they have technical defects in practical applications. They are limited to preset sensor configurations and static prior resource scheduling. Their sensor systems adopt a fixed coupling mode, lacking flexibility to cope with hardware failures or task changes. Furthermore, resource allocation relies excessively on historical data and cannot dynamically optimize the trajectory based on the information gain and data confidence generated in real time during the surveying process. This results in significant deficiencies in task autonomy, multimodal data quality closed-loop, and resource utilization in complex surveying scenarios. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an intelligent mapping system and method for unmanned aerial vehicles (UAVs) equipped with multimodal sensors. It employs real-time scene feature extraction and semantic understanding to generate tasks, and combines task value assessment and resource consumption settlement to dynamically optimize the sensor combination and flight path closed-loop control strategy, thereby achieving efficient, intelligent, and high-quality collaborative data acquisition of target areas.

[0006] The above objectives can be achieved through the following approach:

[0007] A method for intelligent mapping of unmanned aerial vehicles (UAVs) equipped with multimodal sensors includes: extracting scene features and performing semantic understanding on real-time perception data collected by reference sensors on the UAV platform to generate an atomized subtask set; acquiring a sensor capability profile library recording the performance indicators of each sensor; matching the capability requirements of each subtask in the atomized subtask set; dynamically constructing execution units for each subtask to generate sensor execution combinations; driving the sensor execution combinations to collect multimodal data from a target area; performing spatiotemporal alignment and consistency verification on the collected data to generate collaborative acquisition data packets with confidence labels; combining statistical parameters recording unit action energy consumption and time overhead to evaluate the task value and calculate resource consumption of the collaborative acquisition data packets with confidence labels to generate a dynamically adjusted flight path; controlling the UAV to execute according to the dynamically adjusted flight path; iterating the sensor execution combinations using real-time feedback of pose and perception status until a task completion threshold is reached; and performing spatial reorganization and attribute normalization processing on the collaborative acquisition data packets generated during task execution to generate a final structured data packet.

[0008] Optionally, generating the atomized subtask set includes: acquiring real-time video stream data collected by a reference sensor; using a visual feature recognition algorithm to perform target detection and state classification on the real-time video stream data to identify targets of interest and states in the scene; acquiring a rule base for defining task mappings; mapping the targets of interest and states to data acquisition requirements to form an initial subtask list; acquiring the current pose information and map prior information of the UAV; and performing spatial and logical relationship analysis on the initial subtask list to eliminate redundant tasks and establish dependencies between tasks, thereby generating the atomized subtask set.

[0009] Optionally, the generation of sensor execution combinations includes: obtaining data requirement descriptions for each subtask from the atomized subtask set; obtaining initial confidence levels reported by each sensor reflecting its own hardware health status, and querying the capability scores of each sensor in response to the data requirement descriptions from the sensor capability profile library; and, based on the data requirement descriptions, the capability scores, and the initial confidence levels, performing calculations using a multi-objective optimization algorithm to select sensor configurations that meet the task quality threshold and power consumption threshold, thereby generating sensor execution combinations.

[0010] Optionally, the method further includes: when a key sensor failure or data quality continuously falling below a quality threshold is detected during real-time fusion and verification, re-evaluating the sensor capability profile library, setting the capability score of the failed sensor to zero, and obtaining an updated capability profile library; based on the updated capability profile library, re-matching and calculating for the affected sub-tasks to generate alternative sensor execution combinations.

[0011] Optionally, generating a collaborative acquisition data package with confidence labels includes: controlling the sensors in the combination to collect data from the target area according to a unified spatiotemporal reference to obtain raw multi-source data; obtaining verification rules that define cross-modal feature consistency standards, performing real-time cross-validation on the raw multi-source data to generate cross-validation results; and assigning a comprehensive confidence value to the feature information in the raw multi-source data by weighted summation of the cross-validation results and the initial confidence, thereby generating a collaborative acquisition data package with confidence labels.

[0012] Optionally, generating the dynamically adjusted flight path includes: reading the job time budget and energy budget defined in the storage medium to obtain a resource virtual budget; calculating the resource overhead generated by executing the set of atomized subtasks according to the statistical parameters, and deducting it from the resource virtual budget to obtain a remaining virtual budget; analyzing the collaborative acquisition data packets, and quantifying them according to the number of newly discovered targets, the degree of certainty improvement in target status, and the increase in confidence label value to obtain information gain; and evaluating the acquisition value density of unexecuted areas using an online path planning algorithm based on the remaining virtual budget and the information gain, optimizing the global flight path, and generating the dynamically adjusted flight path.

[0013] Optionally, the method includes: marking high-density information regions and low-density information regions in a preset acquisition value density distribution map based on the information gain; increasing the sampling frequency and resolution level for the high-density information regions and decreasing the sampling frequency and resolution level for the low-density information regions to generate an optimized global acquisition accuracy strategy; and inputting the optimized global acquisition accuracy strategy as a heading constraint into the path planning algorithm to generate a dynamically adjusted flight path that satisfies the optimal solution for resource allocation.

[0014] Optionally, the step of iterating the sensor execution combination using real-time feedback of pose and perception state includes: acquiring real-time feedback of current pose information and perception state information during the execution of the dynamically adjusted flight path by the UAV; based on the current pose information and the perception state information, determining the execution conditions of the unexecuted subtasks in the atomized subtask set to obtain the current task execution state; when the current task execution state changes, re-matching capability requirements according to the atomized subtask set to generate an updated sensor execution combination; and adjusting the flight path and sensor operating parameters according to the updated sensor execution combination to generate an updated flight path.

[0015] Optionally, generating the final structured data packet includes: acquiring the collaborative acquisition data packet and pose information generated during task execution; spatially aligning and integrating the collaborative acquisition data packet based on the pose information to obtain a spatially consistent data set; and performing unified scale processing and consistency organization on the multi-source data attributes in the spatially consistent data set to generate the final structured data packet.

[0016] Based on the same inventive concept, this invention also provides an intelligent mapping system for unmanned aerial vehicles (UAVs) equipped with multimodal sensors, including a task atomization module for extracting scene features and semantic understanding from real-time perception data collected by reference sensors mounted on the UAV platform, generating an atomized subtask set; a dynamic loading module for acquiring a sensor capability profile library recording the performance indicators of each sensor, matching capability requirements for each subtask in the atomized subtask set, dynamically assembling execution units for each subtask, and generating sensor execution combinations; and a collaborative acquisition module for driving the sensor execution combinations to acquire multimodal data of the target area, and processing the acquired data. The system performs spatiotemporal alignment and consistency verification to generate collaborative acquisition data packets with confidence labels. A path adaptation module combines statistical parameters recording unit action energy consumption and time overhead to evaluate the task value and calculate resource consumption of the collaborative acquisition data packets with confidence labels, generating a dynamically adjusted flight path. A data reconstruction module controls the UAV to execute according to the dynamically adjusted flight path and iterates the sensor execution combination using real-time feedback of pose and perception status until the task completion threshold is reached. It then performs spatial reorganization and attribute normalization processing on the collaborative acquisition data packets generated during task execution to generate the final structured data packet.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] This invention enables scene understanding and task decomposition through real-time visual perception, and dynamically matches the optimal sensor execution combination for the decomposed sub-tasks. This changes the traditional surveying and mapping mode of fixed routes and all-time sensor operation, realizing the transformation from passive data collection to active task-driven, and improving the intelligence level and resource utilization efficiency of surveying and mapping tasks.

[0019] This invention introduces a real-time consistency verification and confidence assessment mechanism during data acquisition and possesses online fault self-healing capability for sensor failures. This design ensures the intrinsic quality and reliability of multimodal data from the data source, avoids task failures due to equipment malfunctions or environmental changes, and improves the accuracy and usability of the final generated data.

[0020] This invention establishes a framework for task value assessment based on information gain and path adaptive adjustment based on energy consumption. This framework enables UAVs to evaluate data collection returns in real time and autonomously optimize flight paths and data collection accuracy, prioritizing limited endurance resources to areas with the highest information value, thereby maximizing the overall benefits of the mapping task under the same resource constraints.

[0021] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating an intelligent mapping method for unmanned aerial vehicles (UAVs) equipped with multimodal sensors, according to an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of sensor combination selection based on multi-objective optimization according to an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of three-dimensional dynamic path planning according to an embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of the structure of an unmanned aerial vehicle (UAV) intelligent mapping system equipped with a multimodal sensor, according to an embodiment of the present invention. Detailed Implementation

[0027] 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.

[0028] Reference Figure 1One embodiment of the present invention proposes an intelligent mapping method for UAVs equipped with multimodal sensors. It adopts real-time scene feature extraction and semantic understanding to generate tasks, and combines task value assessment and resource consumption settlement to dynamically optimize the sensor combination and flight path closed-loop control strategy, which can achieve efficient, intelligent and high-quality collaborative data collection of target areas.

[0029] The method described in this embodiment specifically includes: S1. Extract scene features and perform semantic understanding on the real-time perception data collected by the reference sensors on the UAV platform to generate an atomized subtask set; Optionally, the generation of the atomized subtask set includes: The system acquires real-time video stream data collected by a benchmark sensor, and uses a visual feature recognition algorithm to perform target detection and state classification on the real-time video stream data, identifying the target of interest and its state in the scene. Obtain a rule base for defining task mappings, map the targets of interest and states to data collection requirements, and form an initial list of subtasks; The system acquires the current pose information of the UAV and prior map information, performs spatial and logical relationship analysis on the initial subtask list, eliminates redundant tasks and establishes dependencies between tasks, and generates an atomized subtask set.

[0030] Specifically, the drone platform acquires continuous real-time video stream data from its onboard base sensor, typically a forward-looking wide-angle visible light camera, such as 1920x1080 resolution video at 30 frames per second. This data stream is fed into a visual feature recognition algorithm running in the onboard computing unit. This algorithm is a deep learning-based convolutional neural network model, pre-trained, capable of real-time analysis of the input image. Its core function is to identify and locate specific targets in the scene and classify their states, thereby identifying targets of interest and their states within the scene. For example, in a disaster assessment scenario, the algorithm can identify targets of interest such as "collapsed buildings," "damaged roads," or "stuck vehicles" with a confidence threshold higher than 0.85, and output their positions, bounding boxes, and classification labels in the image coordinate system.

[0031] Access the locally stored rule base used to define task mappings. A rule is either a structured lookup table or a set of "IF-THEN" logical statements that associates the target of interest and its state output by the visual feature recognition algorithm with specific surveying data acquisition requirements. For example, a rule might be defined as follows: if a "collapsed building" is identified, the triggered data acquisition requirements are "acquiring 3D laser point clouds of the target area with centimeter-level accuracy" and "acquiring multi-angle oblique images with a resolution better than 5 megapixels," thus forming an initial list of subtasks containing target location, task type, and data specifications.

[0032] Acquire real-time pose information from the UAV navigation system, including latitude, longitude, altitude, and aircraft attitude, and combine it with pre-loaded map prior information, such as a digital elevation model (DEM). Initiate a spatial and logical relationship analysis program to process the initial subtask list. In spatial analysis, traverse the task pairs in the list, based on the following conditions... Determine if the task is redundant: , in, and Representing two independent initial subtasks, The geographical locations of the two task objectives are calculated based on pose information and prior map information. and The Euclidean distance between them It is a preset spatial proximity threshold, such as 15 meters. It measures the data collection requirements of the two tasks. and The similarity function has a range of values ​​between 0 and 1. This is a similarity threshold, such as 0.8. When the condition... When true, the two subtasks are merged to eliminate redundancy. Logically, dependencies are established based on task type; for example, the "region overview" task has higher priority than the "detailed modeling" task within its coverage area. Through redundancy removal and dependency construction, a final set of indivisible atomic subtasks with a clear execution order and dependencies is generated.

[0033] For example, taking the task of searching for collapsed buildings by a drone after an earthquake as an example, the system first captures a real-time video stream at a rate of 30 frames per second using a forward-looking wide-angle visible light camera, and then uses the YOLO series deep learning model in the onboard computing unit to analyze the video frames. In the video frame at second T, the algorithm identifies a "collapsed building" target with a confidence level of 0.92. According to the rule in the rule base that "if a collapsed building is identified, a three-dimensional laser point cloud with centimeter-level precision for the target area must be obtained," the initial subtask A is generated. Then, at second T+2, the algorithm identifies the other side of the same target at an adjacent location, generating the initial subtask B. At this time, the system starts redundancy removal processing to obtain the target geographic coordinates of subtask A. The target geographic coordinates of subtask B are (116.3000, 39.9000). The coordinates are (116.3001, 39.9001). The system first converts the latitude and longitude into projected plane coordinates and calculates the Euclidean distance between the two points. The ground distance corresponding to the longitude difference is 8.5 meters, and the ground distance corresponding to the latitude difference is 11.0 meters. The calculation process is as follows: Meters. Because the calculated result of 13.9 meters is less than the preset spatial proximity threshold. The meter reading met the first criterion. Subsequently, the system calculated the similarity. Since both tasks are triggered by "collapsed buildings," their data collection requirements are... They are completely identical, therefore the similarity function calculates a value of 1.0, which is greater than the preset similarity threshold. The second condition is met. Given the condition... If the result is true, the system determines that subtask B is redundant compared to A, and therefore merges them, outputting an atomic subtask containing explicit coordinates and the requirement of "high-precision point cloud acquisition". This method transforms real-time visual perception into a set of de-redundant atomic tasks, achieving logical decoupling and structuring of surveying instructions. It not only eliminates repetitive detection requirements but also provides conflict-free and accurate input for subsequent path planning, improving task planning efficiency in complex scenarios.

[0034] S2. Obtain a sensor capability profile library that records the performance indicators of each sensor, perform capability requirement matching for each subtask in the atomized subtask set, dynamically assemble execution units for each subtask, and generate sensor execution combinations. Optionally, the generating sensor performs a combination including: Obtain the data requirement description for each subtask from the set of atomized subtasks; Obtain the initial confidence level of each sensor's hardware health status reported by each sensor, and query the capability score of each sensor in response to the data requirement description from the sensor capability profile library. Based on the data requirement description, the capability score, and the initial confidence level, a multi-objective optimization algorithm is used to select sensor configurations that meet the task quality threshold and power consumption threshold, and a sensor execution combination is generated.

[0035] Specifically, the data requirement description for the current subtask to be executed is extracted from the set of atomized subtasks. This description is structured data that clarifies the type of data required, such as "3D point cloud" or "hyperspectral image," as well as key quality indicators, such as "point cloud density of not less than 100 points / square meter" or "spectral resolution of 5 nanometers."

[0036] The system requests initial confidence levels from all onboard sensors, reflecting their hardware health status. This initial confidence level is a normalized value between 0 and 1, generated in real-time by the sensor's built-in self-test program based on status parameters such as temperature, voltage stability, and optical component cleanliness; 1 represents perfect condition. Using the data requirement description as an index, the system queries a pre-built sensor capability profile library to obtain a capability score for each sensor for that task. This capability score, also a quantified value between 0 and 1, characterizes the degree of match between the sensor's physical performance and the task requirements, and is pre-determined through ground calibration experiments.

[0037] After obtaining the above input, a multi-objective optimization algorithm is launched. The core objective is to minimize the combined power consumption while satisfying the task quality threshold. Assume the UAV carries N sensors, and an N-dimensional binary decision vector X represents a sensor combination. This indicates selecting the i-th sensor, otherwise setting it to 0. The objective function and constraints for this optimization can be expressed as solving problem B under the premise of satisfying condition A. Condition A is the task quality threshold constraint: Problem B, namely, the goal of minimizing power consumption. , in, Representative sensor combination The overall task quality score is calculated using a non-linear aggregation function. The calculation shows that the function The input is a capability score for all selected sensors. Compared with the initial confidence level The product set of . The first one obtained from the sensor capability profile database Each sensor scores the capability of the current task. It is the first The initial confidence level reported by each sensor. This is the minimum acceptable quality standard set for the current subtask, such as 0.8. This represents the total power consumption of combination X. It is the first The rated operating power consumption of each sensor, in watts, is a static parameter obtained from the equipment specifications. The algorithm employs a genetic algorithm with simulated annealing or elite retention strategies, iteratively searching the feasible solution space to ultimately output a decision vector X that satisfies the quality constraints and minimizes power consumption. The set of sensors represented by this vector X is the optimal sensor execution combination, which is then locked and activated by the system. For example... Figure 2 As indicated by the pentagram in the middle, this point is the selected optimal execution combination X, which minimizes energy consumption while ensuring that the task quality meets the standards.

[0038] For example, for the high-precision point cloud acquisition atomization subtask, the system needs to select the optimal combination from the airborne sensor array. The quality threshold is set in the data requirements description. The confidence level is 0.85. The airborne sensors include a lidar and a tilting camera. The system first acquires status parameters: the lidar generates slight heat due to prolonged operation, and the initial confidence level reported by the self-test procedure is... The initial confidence level is 0.95; the tilted camera is in perfect condition. The rating is 1.0. A query of the capability profile database shows that the LiDAR capability score is 1.0 for point cloud acquisition tasks. The score is 0.90, while the ability of a tilting camera to generate point clouds via photogrammetry is rated as 0.90. The score is 0.60. The system uses a multi-objective optimization algorithm for calculation, first verifying whether the quality of the single-sensor combination meets the standard. If only LiDAR is selected, its comprehensive quality score is calculated through an aggregation function. Therefore, the reliability is calculated using the probability complement model, i.e. Substitute the values ​​to calculate .because This single-sensor solution meets the quality constraints. Next, the power consumption is calculated; the rated power consumption of the lidar... If the power consumption is 40 watts, then the total power consumption is 40 watts. If a combination of "LiDAR + tilt camera" is attempted, the quality score would be... Although the quality is higher, the total power consumption increases to 55 watts. According to the objective function... Under the premise of meeting the quality threshold of 0.85, the system ultimately selects the "only LiDAR on" option with the lowest power consumption as the optimal sensor combination and generates the corresponding binary decision vector. This method utilizes a multi-objective optimization algorithm to dynamically select sensor combinations, minimizing energy consumption while meeting task quality requirements. It also incorporates device health confidence scores to mitigate hardware vulnerabilities and achieve a balance between system performance, operating costs, and reliability.

[0039] Optionally, the method further includes: When a critical sensor failure or data quality consistently below the quality threshold is detected during real-time fusion and verification, the sensor capability profile library is re-evaluated, the capability score of the failed sensor is set to zero, and an updated capability profile library is obtained. Based on the updated capability profile library, the affected subtasks are re-matched and recalculated to generate alternative sensor execution combinations.

[0040] Specifically, the trigger condition for this process is that an anomaly is detected during the real-time data fusion and verification process by the collaborative acquisition module. The status of each critical sensor in the current optimal sensor execution combination is monitored in real time. A critical sensor is one whose failure would directly cause the current subtask to fail to reach its quality threshold. Anomaly detection is performed in two ways, and the replanning process is triggered when the following condition F is met: , in, This represents the hardware self-test status code obtained from the critical sensor numbered k. When it is FALSE, it indicates that the sensor has reported a fatal hardware error, such as overheating or communication interruption, which means that the critical sensor has failed. Let i be the consistency check score of the i-th collaboratively collected data packet. The quality threshold is dynamically set and is slightly lower than the quality threshold used during task planning. It is a counting function within a sliding time window, which counts the number of times the data quality score is below a threshold within a recent period, such as the last 10 collection cycles. This is a preset threshold for the number of consecutive alarms, such as 5. The latter part of this condition means that the data quality is consistently below the quality threshold, indicating that the sensor may be experiencing performance degradation that has not been detected by the self-test program.

[0041] Once trigger condition F is met, a failed or degraded sensor is identified, and the sensor capability profile library is immediately and dynamically updated. The unique identifier of the failed sensor is locked, and the values ​​of all corresponding capability scores are forcibly set to zero in the sensor capability profile library data structure loaded in memory. This generates an updated capability profile library that reflects the current true availability of the hardware.

[0042] Based on the updated capability profile library, matching calculations are re-performed for the affected subtasks. Affected subtasks include not only tasks currently in execution but interrupted by sensor failure, but also all tasks in the atomized subtask set that have not yet started. Using the data requirement descriptions of these tasks and the updated capability profile library as input, the multi-objective optimization algorithm is invoked again. Since the capability score of the failed sensor has been reset to zero, the algorithm's solution space changes, and it will search for new optimal solutions among the remaining available sensors. After calculation, one or more alternative optimal sensor execution combinations are generated, and based on the current flight position and resource budget, the most suitable one is selected for immediate execution, thereby achieving seamless task continuation or orderly degradation.

[0043] For example, during the drone's mission, a background monitoring program runs continuously to ensure robustness. During data collection, consistency checks are continuously performed on the data packets from the most recent 10 periods, resulting in a score sequence. There were 7 instances where the dynamic quality threshold was lower than 7. (Set to 0.70). The system calculates the counting function within the sliding time window. The result was 7. This count value is greater than the preset threshold for the number of consecutive alarms. (5 times), even if the hardware status code of the lidar is displayed at this time. The system still displays normally, and it determines that condition F has been triggered, indicating a latent degradation in sensor performance (such as interference from smoke). The system then initiates a replanning process. First, it locks the lidar's ID in the in-memory sensor capability profile library and forcibly sets its capability score for that task to 0. Then, the system re-invokes the optimization algorithm based on the updated profile library. At this point, the lidar fails to meet the quality requirements. The algorithm searches the remaining resources and finds that enabling a "tilted camera" in conjunction with a "low-speed, high-overlap flight mode" can improve its capability score to 0.88. The system executes a recalculation: based on a confidence level of 1.0 and a capability score of 0.88, the final quality score for the tilted camera solution is 0.88, which satisfies the threshold condition of being greater than 0.85. Therefore, the system automatically generates an alternative solution to switch to tilted camera operation, achieving online fault self-healing for the task. This method endows the system with online fault perception and self-healing replanning capabilities, automatically eliminating faulty nodes and matching alternative solutions when sensors fail. This ensures that missions can be seamlessly continued or orderly degraded in the event of hardware failure, enhancing the robustness and continuity of drone operations.

[0044] S3. Drive the sensor to perform combined multimodal data acquisition of the target area, and perform spatiotemporal alignment and consistency verification on the acquired data to generate a collaborative acquisition data packet with confidence label; Optionally, generating the collaborative acquisition data packet with confidence labels includes: The sensor assembly is controlled to collect data from the target area according to a unified spatiotemporal reference, thereby acquiring raw multi-source data. Obtain the validation rules that define the cross-modal feature consistency standard, perform real-time cross-validation on the original multi-source data, and generate cross-validation results; By performing a weighted summation of the cross-validation results and the initial confidence level, a comprehensive confidence level value is assigned to the feature information in the original multi-source data, generating a collaborative acquisition data package with a confidence level label.

[0045] Specifically, the system controls each sensor in the selected optimal sensor array to collect data on the target area based on a unified spatiotemporal reference. This spatiotemporal reference is provided by the high-precision Global Navigation Satellite System (GNSS) and Inertial Measurement Unit (INS) onboard the UAV. It assigns a precise timestamp (within microseconds) and centimeter-level spatial position and attitude information to each sensor's output data frame or data point, whether it's laser point cloud, visible light imagery, or hyperspectral data. The output of this operation includes raw, multi-source data with both temporal and spatial labels.

[0046] The system retrieves validation rules defining cross-modal feature consistency criteria from local storage. These rules are a set of algorithms built upon prior knowledge that the same object in the physical world should exhibit correlated features under different sensors. For example, a rule might define that the position of a 3D edge of a building detected in a LiDAR point cloud, after being projected onto a visible light image at the same time through coordinate transformation, should coincide with a high-gradient pixel region in the image within a 2-pixel error range. The validation rules are then run on the original multi-source data to perform real-time cross-validation, and the quantized cross-validation results are output.

[0047] The cross-validation results are weighted with the initial confidence scores of each sensor to assign a final comprehensive confidence score to the core feature information in the currently collected data. The formula is: , In this formula, This represents the overall confidence level. The result represents the cross-validation result and is a score that reflects the degree of agreement between multimodal data. It is the initial confidence level of the i-th sensor in the optimal sensor combination participating in this data acquisition, provided by the sensor self-test program. These are preset weighting coefficients, ranging from 0 to 1, used to balance the importance of external data consistency and the sensor's own health status in the final confidence assessment. When the task demands extremely high accuracy in data fusion... The value will tend to be large, such as 0.7. Is assigned to the first The weights of each sensor reflect their importance or dominance in the current task. The calculated weights... It is used as a metadata tag and packaged together with the original multi-source data that has been spatiotemporally aligned to form the final collaborative acquisition data package with confidence labels.

[0048] For example, after a UAV completes a set of data acquisitions using a tilting camera, the system needs to generate a collaborative acquisition data packet with confidence labels. The system first uses the spatiotemporal reference provided by GNSS / INS to precisely align the acquired images with the location information. Then, the system runs consistency verification rules, back-projecting feature points in the images onto a known 3D map using a photogrammetric algorithm, and calculates the reprojection error. The calculated normalized cross-validation result... The initial confidence level is 0.82. At this point, the confidence level of the cameras participating in the data acquisition is... The weight is 1.0, which is the weight of a single sensor operation. The system uses a weighted formula to calculate the overall confidence level, which is set to 1. Preset weighting coefficients The value is set to 0.6 to emphasize external validation results. The calculation process is as follows: First, the validation part is calculated. Then calculate the sensor state part. Finally, add the two parts together. The system uses 0.892 as a metadata tag and writes it into the image data header file, signifying the comprehensive reliability of the data packet in terms of geometric accuracy and device status. This method generates data packets with confidence labels through multi-sensor cross-validation, achieving quality quantification and spatiotemporal unification of source data. This solves the problem of difficulty in distinguishing data quality, reduces the cost of post-registration, and provides a data foundation for subsequent analysis.

[0049] S4. Combining the statistical parameters of energy consumption and time expenditure of the recorded unit action, perform task value assessment and resource consumption calculation on the collaborative acquisition data packet with confidence label, and generate a dynamically adjusted flight path. Optionally, generating the dynamically adjusted flight path includes: Read the job time budget and energy budget defined in the storage medium to obtain the virtual resource budget; The resource overhead generated by executing the set of atomized subtasks is calculated according to the statistical parameters and deducted from the virtual resource budget to obtain the remaining virtual budget; The collaboratively collected data packets are analyzed, and the information gain is obtained by quantifying the number of newly discovered targets, the degree of certainty improvement in the target status, and the increase in the confidence label. Based on the remaining virtual budget and the information gain, the collection value density of the unexecuted area is evaluated using an online path planning algorithm, and the global flight path is optimized to generate a dynamically adjusted flight path.

[0050] Specifically, at mission initiation, pre-set mission time and energy budgets are read from storage media, such as configuration files on the onboard computer. For example, the total flight time should not exceed 45 minutes, and the total energy consumption should not exceed 85% of the battery capacity. These two values ​​are loaded by the system as the initial virtual resource budget, serving as hard constraints for all decisions throughout the mission.

[0051] During the execution of a set of atomized subtasks by the UAV, resource costs are calculated based on preset statistical parameters after each subtask or flight segment is completed. These statistical parameters are a database containing various UAV flight maneuvers, such as hovering, level flight, and climb, as well as the energy consumption and time costs per unit time under different sensor operating modes. The actual cost of the executed tasks is deducted from the virtual resource budget, and the remaining virtual budget is updated and maintained in real time.

[0052] Deep analysis was performed on the newly generated collaboratively acquired data packets, using the information gain evaluation function. To calculate the value of this data collection: , In this formula, This represents the quantized information gain. , , These are preset weighting coefficients used to adjust the importance of different gain sources, and their sum is usually normalized to 1. This represents the number of newly discovered targets of interest that were not recorded in the prior map information through this data packet analysis, and is directly output by the target detection algorithm. This represents the increase in the certainty of the state of an identified target. For example, if the probability of damage to a building increases from 0.6 in the previous assessment to 0.95 after data collection, the increase is 0.35. This is the sum of the certainty improvements of all objectives. This represents the improvement in data quality reflected by the confidence label attached to the collaboratively collected data packet when repeatedly collecting data from the same area. For example, if the confidence score improves from 0.7 in the previous collection to 0.9 in the current collection, the improvement value is 0.2. This is the sum of the confidence increases for all overlapping regions.

[0053] The calculated information gain and remaining virtual budget are fed into an online path planning algorithm, such as a fast exploratory random tree algorithm based on information entropy. The core action of this algorithm is to evaluate the acquisition value density of all unexplored areas, i.e., the ratio of the potential information gain to the required resource cost of performing acquisition in these areas. The algorithm prioritizes planning flights to areas with high acquisition value density while ensuring that the total estimated cost of the new planned path does not exceed the remaining virtual budget. By solving this constrained optimization problem, an optimized, entirely new global flight path is generated—a dynamically adjusted flight path—and this path is immediately sent to the flight control system for execution.

[0054] For example, after completing data acquisition in the current area, the system needs to decide on the next flight action. First, the system reads the current remaining virtual budget, with a remaining flight time of 20 minutes. Next, the system analyzes the newly generated collaboratively acquired data packets to calculate the information gain. Three new stranded vehicles were discovered in the images collected during this phase of the data collection. The probability of certainty regarding the known extent of damage to collapsed buildings increased from 0.6 to 0.9, i.e. =, and the overall confidence level of the map data for this area increased by 0.2, that is The preset weighting coefficients are as follows: , , Substitute the values ​​into the formula to verify the calculation: After obtaining this quantified gain, the path planning algorithm based on information entropy searches the remaining unexplored areas of the map and finds that region X has a high expected data collection value density. The algorithm calculates that flying to and operating in region X will take 8 minutes and consume 40% of the remaining budget, while flying to another region Y will take 5 minutes but has extremely low expected gains. After comprehensively evaluating the benefit-cost ratio, the system generates a dynamically adjusted flight path that prioritizes region X and instructs the flight control system to execute it. This method constructs a dynamic path planning closed loop based on information gain and resource budget, enabling the UAV to autonomously adjust its flight path to prioritize data collection in high-value areas. This breaks the limitations of preset flight paths and maximizes the effective information acquisition of a single flight within a limited endurance time.

[0055] Optionally, the method includes: Based on the information gain, high-density information regions and low-density information regions are marked in the preset data acquisition value density distribution map; Increase the sampling frequency and resolution level for the high-density information region, and decrease the sampling frequency and resolution level for the low-density information region to generate an optimized global acquisition accuracy strategy. The optimized global acquisition accuracy strategy is used as a heading constraint and input into the path planning algorithm to generate a dynamically adjusted flight path that satisfies the optimal solution for resource allocation.

[0056] Specifically, based on the continuously calculated information gain, a value score is dynamically assigned to each grid cell on an internally maintained gridded map covering the entire operational area—a pre-defined data collection value density distribution map. For example, the value score of a 10m x 10m grid represents the information gain generated when a drone flies over that area. Two dynamic thresholds are set: a high-value threshold and a low-value threshold. When the value score of a grid consistently exceeds the high-value threshold, the area is marked as a high-information-density area; conversely, it is marked as a low-information-density area.

[0057] For grid areas marked as high-density information regions, the system automatically increases the sampling frequency and resolution for those areas. Increasing the sampling frequency means instructing the UAV to reduce its flight speed and increase the camera's image interval from, for example, 0.5 Hz to 2 Hz; increasing the resolution means instructing the UAV to reduce its flight altitude from the standard cruising altitude of 50 meters to 20 meters to obtain higher ground resolution images, or instructing the LiDAR to increase the scanning beam density. Conversely, for low-density information regions, the system adopts the opposite strategy, instructing the UAV to fly at a higher speed, a lower image frequency, or a higher flight altitude to save time and energy.

[0058] The optimized global acquisition accuracy strategy is input into the online path planning algorithm as a dynamic heading constraint. This heading constraint does not refer to insurmountable obstacles, but rather is an important component of the path cost function. When evaluating any candidate path, the algorithm calculates not only its length and flight time, but also the corresponding data acquisition time and energy consumption based on the accuracy strategy of the areas the path traverses. For example, the cost of a path segment traversing a high-information-density area will significantly increase due to low-speed, low-altitude flight and high-frequency sensing operations. The objective function of the path planning algorithm is to find a path that maximizes global information gain by performing the highest-precision acquisition in the highest-value areas while quickly traversing low-value areas, all while satisfying the remaining virtual budget. This path, which integrates "where to go" and "how to do it," is the final dynamically adjusted flight path that satisfies the optimal solution for resource allocation. Figure 3 As shown, when the drone flies over high-value dark areas, its trajectory drops significantly, reducing its flight altitude to obtain sub-meter high-resolution images; while when flying over low-value light areas, the drone automatically climbs to its cruising altitude to reduce energy consumption and expand its field of view coverage.

[0059] For example, after determining the macroscopic path to region X, the system further refines the specific flight strategy within that region. Based on the continuously updated data collection value density distribution map, the system identifies a "high-density information area" grid at the center of region X, whose value score consistently exceeds the high-value threshold. For this grid, the system generates a "high-precision data collection" strategy, instructing the UAV to reduce its flight altitude from 50 meters to 30 meters and its flight speed from 10 meters per second to 5 meters per second to increase sampling density. Correspondingly, the path planning algorithm incorporates these parameters when calculating the path cost across this grid. The grid has a side length of 100 meters; if traversed at a low speed, the time cost is... seconds, compared to high-speed passage For every 10 seconds added to the time cost, energy consumption also increases accordingly. The algorithm weighs these specific physical overheads against the expected high information gain, confirming that even with the increased cost, the strategy still significantly improves the overall task value without exceeding the total budget. This allows the algorithm to lock onto the 3D flight path containing the speed and altitude changes, achieving refined spatial allocation of resources. This method transforms regional value density into a differentiated acquisition accuracy strategy, enabling on-demand allocation of mapping resources in the spatial dimension. By acquiring high-value areas meticulously and quickly passing through low-value areas, it avoids resource waste caused by uniform sampling and optimizes the overall energy efficiency ratio.

[0060] S5. Control the UAV to execute according to the dynamically adjusted flight path, and use the real-time feedback of pose and perception status to iterate the sensor execution combination until the mission end threshold is reached. Perform spatial reorganization and attribute normalization processing on the collaborative acquisition data packets generated during the mission execution to generate the final structured data packets.

[0061] Optionally, the step of iterating the sensor's performance using real-time feedback of pose and perception state includes: The current pose information and perception status information are obtained in real time during the execution of the dynamically adjusted flight path by the UAV; Based on the current pose information and the perception state information, the execution conditions of the subtasks that have not yet been executed in the atomized subtask set are judged to obtain the current task execution state; When the current task execution state changes, the capability requirements are rematched based on the atomized subtask set to generate an updated sensor execution combination. Based on the updated sensor execution combination, the flight path and sensor operating parameters are adjusted to generate an updated flight path.

[0062] Specifically, real-time feedback from the flight control system and perception module during the execution of the dynamically adjusted flight path is obtained at high frequencies, such as 10 Hz. This feedback mainly includes two types of information: First, current pose information, including the drone's precise geographic coordinates, altitude, and fuselage attitude angles. Second, perception status information, which is an environmental summary obtained after preliminary processing of the raw data from the reference sensors, such as whether there are obstructions in the field of view and whether the current lighting conditions meet the imaging requirements.

[0063] Based on the acquired current pose and perception state information, the system iterates through all subtasks marked "not yet executed" in the atomized subtask set, and performs execution condition checks on each subtask. This check process is multi-dimensional. For example, for a detailed modeling subtask of a target point, its execution conditions might include: the current horizontal distance between the UAV and the target is less than 30 meters, the line of sight is unobstructed, and the target is located in the center of the camera's field of view. These check results are then aggregated to obtain a snapshot of the current task execution state, indicating which pending tasks have now entered the executable window.

[0064] The system continuously compares the task execution status from the previous moment to the current moment. When a change in the current task execution status is detected—for example, a new high-priority subtask becomes executable due to a drone reaching its designated location, or a task being prepared for execution becomes unexecutable due to a sudden weather event causing sunlight levels to fall below a threshold—a re-matching calculation is triggered. Based on the changed status of the set of atomized subtasks, the capability requirement matching process is restarted. Because the list of tasks to be executed or their priorities have changed, the input to the multi-objective optimization algorithm also changes, resulting in an updated optimal sensor execution combination that differs from the previous one.

[0065] Once the updated optimal sensor execution combination is generated, it's crucial to ensure that the UAV's subsequent actions match this new configuration. For example, if the new combination activates a lidar sensor, and the previous path didn't consider the optimal sweep angle for the lidar scan, the sensor operating parameters will be adjusted based on the new combination, and the flight path will be further optimized locally or globally to generate an updated flight path. This might mean adding a hovering maneuver to the existing flight path to complete a 360-degree scan, or adjusting the flight altitude to achieve the lidar's optimal operating distance. The updated flight path and sensor parameters will be immediately deployed for execution, completing a full dynamic iterative closed loop.

[0066] For example, when executing a refined path, the UAV establishes a real-time feedback control loop. The system reads the pose and perception status at a frequency of 10 Hz. When the UAV flies to a distance of 50 meters from the target building, the perception status information shows that the current ambient light intensity suddenly drops to 300 lumens, below the optimal operating threshold of 500 lumens for the visible light camera. The system immediately performs an execution condition judgment on the currently unexecuted "building facade texture acquisition" subtask, determining that the current lighting conditions cannot meet the original pure visible light acquisition requirements. This state change immediately triggers the sensor reconfiguration process. The system updates the input conditions, lowers the capability score of the visible light camera, and re-evaluates the resource pool. At this time, the system discovers that the onboard thermal infrared camera is unaffected by light and has a high capability score in detecting human thermal signals. The multi-objective optimization algorithm then outputs a new optimal combination: "Activate the thermal infrared camera to replace the visible light camera." Upon receiving this instruction, the system not only switches the sensors but also automatically fine-tunes the distance between the UAV and the building based on the field of view parameters of the thermal infrared camera, generates an updated flight path, and executes it immediately, ensuring the continuity of the mission in dynamic environments. This method utilizes high-frequency pose and sensor feedback to establish a real-time control loop, enabling immediate response to environmental changes such as occlusion and lighting, and dynamic strategy adjustments. This ensures that sensor configuration and flight attitude are always within the optimal operating window, guaranteeing a high success rate for data acquisition in dynamic environments.

[0067] Optionally, generating the final structured data packet includes: Acquire the collaborative acquisition data packet and pose information generated during task execution; Based on the pose information, the collaboratively acquired data packets are spatially aligned and integrated to obtain a spatially consistent data set; The multi-source data attributes in the spatially consistent data set are processed with a unified scale and consistent in order to generate the final structured data package.

[0068] Specifically, the process acquires all collaboratively acquired data packets with confidence labels generated throughout the entire mission, along with the pose information precisely corresponding to the timestamp of each data packet. This pose information is recorded by high-precision GNSS / INS and includes the 3D coordinates and attitude angles at each acquisition. Based on the pose information of each data packet, a coordinate transformation algorithm is used to uniformly transform all source data, including laser point clouds, visible light imagery, hyperspectral cubes, etc., from their respective sensor coordinate systems or body coordinate systems to a common global geographic coordinate system, such as WGS84. After this step, a spatially consistent dataset that is precisely registered and eliminates the effects of parallax is obtained.

[0069] Multi-source data attributes in spatially consistent datasets undergo unified scaling and consistency processing. For color attributes, images captured by different cameras under varying lighting conditions are adjusted to a unified color temperature and brightness standard using color correction algorithms. For geometric attributes, 3D point clouds generated from visible light images using a dense matching algorithm are fused with point clouds directly acquired by LiDAR, filling data gaps and improving overall density. For spectral attributes, spectral data from different bands are organized into a unified spectral library format.

[0070] Spatially aligned and attribute-normalized data, along with its metadata, including but not limited to the source sensor, acquisition time, original confidence label, and processing records for each data point, are packaged into a single, hierarchical data file or database. This file, the final structured data package, is internally organized; for example, all data is indexed by a geographic grid, with each grid containing all multimodal information for that region.

[0071] For example, after the mission concludes, the system needs to generate the final structured data packet. The system aggregates all data collected during the flight, including point cloud data acquired via lidar and image data acquired via cameras. First, utilizing the high-precision GNSS / INS pose information attached to each data packet, the system transforms all point cloud and image feature points to the WGS84 global geographic coordinate system using a coordinate transformation matrix, eliminating spatial parallax. Next, the system addresses attribute heterogeneity, mapping the RGB color information of the image data onto the corresponding lidar point cloud to generate a true-color point cloud, and then applying the calculated comprehensive confidence label. This quality attribute is attached to each data object. Finally, the system encapsulates this spatially aligned, attribute-fused, and quality-labeled data into a single HDF5 file according to the geographic grid index. This file has a clear internal structure; for example, the "Grid_01" node contains the color point cloud, original imagery, and corresponding metadata table for the region, forming a standardized mapping product that can be directly used for post-disaster 3D reconstruction. This method eliminates parallax and format differences between heterogeneous data by unifying the spatial benchmark and attribute dimensions of multi-source data. The output standardized structured data package supports rapid retrieval and direct application, shortening the cycle from data acquisition to downstream applications such as modeling and evaluation.

[0072] Based on the same inventive concept, such as Figure 4 As shown, the present invention also provides an intelligent mapping system for unmanned aerial vehicles equipped with multimodal sensors, comprising: The task atomization module is used to extract scene features and understand semantics from real-time perception data collected by the baseline sensors on the UAV platform, and generate a set of atomized subtasks. The dynamic loading module is used to acquire a sensor capability profile library that records the performance indicators of each sensor, match the capability requirements of each subtask in the atomized subtask set, dynamically assemble execution units for each subtask, and generate sensor execution combinations. The collaborative acquisition module is used to drive the sensors to perform combined multimodal data acquisition of the target area, and to perform spatiotemporal alignment and consistency verification on the acquired data to generate collaborative acquisition data packets with confidence labels. The path adaptation module is used to combine statistical parameters of energy consumption and time cost of recording unit actions to evaluate the task value and calculate resource consumption of the collaboratively acquired data packets with confidence labels, and generate a dynamically adjusted flight path. The data reconstruction module is used to control the UAV to execute according to the dynamically adjusted flight path, and to iterate the sensor execution combination using the real-time feedback of pose and perception status until the mission end threshold is reached. The module performs spatial reconstruction and attribute normalization processing on the collaborative acquisition data packets generated during the mission execution to generate the final structured data packets.

[0073] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method is applicable to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0074] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for intelligent mapping using an unmanned aerial vehicle (UAV) equipped with a multimodal sensor, characterized in that, The method includes: Scene features are extracted and semantic understanding is performed on the real-time perception data collected by the benchmark sensors on the UAV platform to generate an atomized set of subtasks; Obtain a sensor capability profile library that records the performance indicators of each sensor, match the capability requirements of each subtask in the atomized subtask set, dynamically assemble execution units for each subtask, and generate sensor execution combinations. The sensor is driven to perform multimodal data acquisition of the target area, and the acquired data is spatiotemporally aligned and verified for consistency to generate a collaborative acquisition data packet with confidence labels. By combining statistical parameters of energy consumption and time expenditure of the recorded unit action, the collaborative acquisition data packet with confidence label is evaluated for task value and resource consumption is calculated to generate a dynamically adjusted flight path. The drone is controlled to execute the dynamically adjusted flight path, and the sensor execution combination is iterated using the real-time feedback of pose and perception status until the mission end threshold is reached. The collaborative acquisition data packets generated during the mission execution are spatially reorganized and attribute normalized to generate the final structured data packets.

2. The intelligent mapping method for unmanned aerial vehicles equipped with multimodal sensors according to claim 1, characterized in that, The set of generated atomized subtasks includes: The system acquires real-time video stream data collected by a benchmark sensor, and uses a visual feature recognition algorithm to perform target detection and state classification on the real-time video stream data, identifying the target of interest and its state in the scene. Obtain a rule base for defining task mappings, map the targets of interest and states to data collection requirements, and form an initial list of subtasks; The system acquires the current pose information of the UAV and prior map information, performs spatial and logical relationship analysis on the initial subtask list, eliminates redundant tasks, establishes dependencies between tasks, and generates an atomized subtask set.

3. The intelligent mapping method for unmanned aerial vehicles equipped with multimodal sensors according to claim 1, characterized in that, The generated sensor execution combination includes: Obtain the data requirement description for each subtask from the set of atomized subtasks; Obtain the initial confidence level of each sensor's hardware health status reported by each sensor, and query the capability score of each sensor in response to the data requirement description from the sensor capability profile library. Based on the data requirement description, the capability score, and the initial confidence level, a multi-objective optimization algorithm is used to select sensor configurations that meet the task quality threshold and power consumption threshold, and a sensor execution combination is generated.

4. The intelligent mapping method for unmanned aerial vehicles equipped with multimodal sensors according to claim 1, characterized in that, The method further includes: When a critical sensor failure or data quality consistently below the quality threshold is detected during real-time fusion and verification, the sensor capability profile library is re-evaluated, the capability score of the failed sensor is set to zero, and an updated capability profile library is obtained. Based on the updated capability profile library, the affected subtasks are re-matched and recalculated to generate alternative sensor execution combinations.

5. The intelligent mapping method for unmanned aerial vehicles equipped with multimodal sensors according to claim 3, characterized in that, The generation of collaborative acquisition data packets with confidence labels includes: The sensor assembly is controlled to collect data from the target area according to a unified spatiotemporal reference, thereby acquiring raw multi-source data. Obtain the validation rules that define the cross-modal feature consistency standard, perform real-time cross-validation on the original multi-source data, and generate cross-validation results; By performing a weighted summation of the cross-validation results and the initial confidence level, a comprehensive confidence level value is assigned to the feature information in the original multi-source data, generating a collaborative acquisition data package with a confidence level label.

6. The intelligent mapping method for unmanned aerial vehicles equipped with multimodal sensors according to claim 1, characterized in that, The generated dynamically adjusted flight path includes: Read the job time budget and energy budget defined in the storage medium to obtain the virtual resource budget; The resource overhead generated by executing the set of atomized subtasks is calculated according to the statistical parameters and deducted from the virtual resource budget to obtain the remaining virtual budget; The collaboratively collected data packets are analyzed, and the information gain is obtained by quantifying the number of newly discovered targets, the degree of certainty improvement in the target status, and the increase in the confidence label. Based on the remaining virtual budget and the information gain, the collection value density of the unexecuted area is evaluated using an online path planning algorithm, and the global flight path is optimized to generate a dynamically adjusted flight path.

7. The intelligent mapping method for unmanned aerial vehicles equipped with multimodal sensors according to claim 6, characterized in that, The method includes: Based on the information gain, high-density information regions and low-density information regions are marked in the preset data acquisition value density distribution map; Increase the sampling frequency and resolution level for the high-density information region, and decrease the sampling frequency and resolution level for the low-density information region to generate an optimized global acquisition accuracy strategy. The optimized global acquisition accuracy strategy is used as a heading constraint and input into the path planning algorithm to generate a dynamically adjusted flight path that satisfies the optimal solution for resource allocation.

8. The intelligent mapping method for unmanned aerial vehicles equipped with multimodal sensors according to claim 1, characterized in that, The iterative combination of the sensor execution using real-time feedback of pose and perception state includes: Acquire the current pose information and perception status information fed back in real time during the execution of the dynamically adjusted flight path by the UAV; Based on the current pose information and the perception state information, the execution conditions of the subtasks that have not yet been executed in the atomized subtask set are judged to obtain the current task execution state; When the current task execution state changes, the capability requirements are rematched based on the atomized subtask set to generate an updated sensor execution combination. Based on the updated sensor execution combination, the flight path and sensor operating parameters are adjusted to generate an updated flight path.

9. The intelligent mapping method for unmanned aerial vehicles equipped with multimodal sensors according to claim 1, characterized in that, The generation of the final structured data packet includes: Acquire the collaborative acquisition data packet and pose information generated during task execution; Based on the pose information, the collaboratively acquired data packets are spatially aligned and integrated to obtain a spatially consistent data set; The multi-source data attributes in the spatially consistent data set are processed with a unified scale and consistent in order to generate the final structured data package.

10. A UAV intelligent mapping system equipped with multimodal sensors, characterized in that, The system includes: The task atomization module is used to extract scene features and understand semantics from real-time perception data collected by the baseline sensors on the UAV platform, and generate a set of atomized subtasks. The dynamic loading module is used to acquire a sensor capability profile library that records the performance indicators of each sensor, match the capability requirements of each subtask in the atomized subtask set, dynamically assemble execution units for each subtask, and generate sensor execution combinations. The collaborative acquisition module is used to drive the sensors to perform combined multimodal data acquisition of the target area, and to perform spatiotemporal alignment and consistency verification on the acquired data to generate collaborative acquisition data packets with confidence labels. The path adaptation module is used to combine statistical parameters of energy consumption and time cost of recording unit actions to evaluate the task value and calculate resource consumption of the collaboratively acquired data packets with confidence labels, and generate a dynamically adjusted flight path. The data reconstruction module is used to control the UAV to execute according to the dynamically adjusted flight path, and to iterate the sensor execution combination using the real-time feedback of pose and perception status until the mission end threshold is reached. The module performs spatial reconstruction and attribute normalization processing on the collaborative acquisition data packets generated during the mission execution to generate the final structured data packets.

Images