Multi-source information feedforward regulation positioning method for complex restricted environment unmanned aerial vehicle system

Abstract

The application provides a multi-source information feedforward regulation positioning method for an unmanned aerial vehicle system in a complex restricted environment, which comprises the following steps: firstly, time synchronization and coordinate unification are performed on laser point cloud, image data and inertial measurement data of multi-source observation data, and distortion removal preprocessing is performed; then, multi-dimensional quality indexes reflecting the geometric structure characteristics of the environment, the stability of feature matching and the change of sensor signals are extracted; through normalization and fusion modeling of various multi-dimensional quality indexes, continuous observation reliability is obtained; then, according to the observation reliability, the noise covariance of various observations is nonlinearly regulated, and the regulated covariance is introduced into an extended Kalman filter or a nonlinear optimization framework for state updating; and finally, the pose estimation result and the unmanned aerial vehicle system health index are outputted; the application can improve the precision and robustness of multi-source fusion positioning in a complex environment, reduce the adverse effects caused by false matching and degenerative scenes, and provide reliable pose estimation and system health evaluation for autonomous flight of the unmanned aerial vehicle.

Description

Technical Field

[0001] This invention relates to the field of UAV positioning and multi-sensor fusion perception technology, and in particular to a multi-source information feedforward adjustment positioning method for UAV systems in complex and confined environments. It relates to a multi-source information fusion positioning method based on observation quality modeling and feedforward covariance adjustment, which is designed to meet the positioning needs of UAVs in confined space environments. Background Technology

[0002] With the increasing application of unmanned aerial vehicles (UAVs) in confined space scenarios such as indoor inspection, pipeline inspection, and underground space exploration, high-precision and robust positioning and navigation technologies have become a key foundation for their autonomous flight. In confined space environments, there are often unavailable GNSS signals, narrow and complex spatial structures, unstable lighting conditions, and interference factors such as dust and smoke, making it difficult for UAVs to obtain stable and reliable pose information relying on a single sensor. LiDAR is prone to observation degradation in environments with repetitive structures or degraded geometric features; visual sensors are susceptible to weak textures, lighting changes, and motion blur, leading to unstable feature extraction; and while inertial measurement units (IMUs) have the advantage of high-frequency output, they suffer from unavoidable cumulative drift. Therefore, multi-sensor fusion positioning methods that integrate information from multiple sources such as LiDAR, visual sensors, and IMUs, utilizing the complementary characteristics of various sensors to achieve continuous and stable pose estimation in confined space environments, have become the main technical means for UAV positioning.

[0003] However, existing multi-sensor fusion localization methods still have the following shortcomings in UAV applications in confined spaces: First, most methods assume that sensor observation noise is a fixed parameter, lacking the ability to perceive and quantify the dynamic changes in observation quality with the environment in real time. In situations such as dust, weak texture, or near-wall flight disturbances, it is difficult to accurately reflect the reliability of observations. Second, due to the lack of effective assessment of the credibility of observation data, the system is unable to dynamically adjust the fusion weights according to the different sensor states. When a sensor experiences performance degradation, it may still be assigned a high weight, thereby affecting the overall positioning accuracy or even causing the system to diverge. Third, some methods rely on posterior residuals to adjust the observation weights, which is a post-correction mechanism. In the event of sudden interference or rapid environmental changes, there is a response lag, making it difficult to suppress the impact of low-quality observations on the system in a timely manner. In addition, existing positioning systems usually only output pose results, lacking real-time quantitative characterization of the reliability of positioning results, which is not conducive to the safe flight and decision-making control of UAVs in complex and confined spaces. Summary of the Invention

[0004] This invention proposes a multi-source information feedforward adjustment positioning method for UAV systems in complex and constrained environments. This method enables UAVs to achieve positioning and navigation functions in environments where GNSS is unavailable, structures are complex, and perception is degraded. It can effectively improve the accuracy and robustness of multi-source fusion positioning in complex environments, reduce the adverse effects of mismatch and degradation scenarios, and provide reliable pose estimation and system health assessment for autonomous flight of UAVs.

[0005] The present invention adopts the following technical solution.

[0006] A multi-source information feedforward adjustment positioning method for unmanned aerial vehicle (UAV) systems in complex and confined environments is proposed. The UAV system is equipped with a positioning system for collecting multi-source observation data. The positioning method constructs an observation reliability model by incorporating the quality assessment results of the multi-source observation data into the fusion process and dynamically adjusts the observation noise covariance using a feedforward approach. First, the laser point cloud, image data, and inertial measurement data from the multi-source observations are synchronized in time and unified in coordinates, and preprocessed with distortion correction. Then, based on the screening of observation validity, multi-dimensional quality indicators reflecting environmental geometric characteristics, feature matching stability, and sensor signal changes are extracted. By normalizing and fusing various multi-dimensional quality indicators, a continuous observation reliability is obtained to quantify the reliability of the observation data. Furthermore, the noise covariance of various observations is nonlinearly adjusted according to the observation reliability, reducing the weight of low-reliability observations and increasing the weight of high-reliability observations. The adjusted covariance is then introduced into an extended Kalman filter or nonlinear optimization framework for state updates to suppress the influence of abnormal observation data on the positioning results. Finally, the pose estimation results and the health index of the UAV flight system are output. This invention can effectively improve the accuracy and robustness of multi-source fusion positioning in complex environments, reduce the adverse effects of mismatch and degradation scenarios, and provide reliable pose estimation and system health assessment for UAV autonomous flight.

[0007] The positioning system is a multi-sensor fusion positioning system, including a lidar, a visual sensor, an inertial measurement unit, and a fusion computing module. The fusion computing module is used to jointly process multi-source observation data and output the pose estimation results of the UAV.

[0008] The positioning system includes a lidar, a visual sensor, and an inertial measurement unit, used to realize the positioning and navigation functions of UAVs in environments where GNSS is unavailable, structures are complex, and perception is degraded. The multi-source information feedforward adjustment positioning method based on observation quality modeling is described, but the present invention is not limited to the specific sensor combination forms mentioned above.

[0009] The fusion computing module is deployed on the airborne computing platform. The lidar, camera and IMU output data at frequencies of 10~20Hz, 15~30Hz and 100~400Hz respectively. The system uses the IMU time as a unified reference to perform time alignment on the multi-source data.

[0010] The positioning method takes multi-source observation information as input, and after completing the validity screening of observation data, constructs an observation quality assessment model based on multi-dimensional observation features, performs real-time quality quantification of observation data from each sensor, and adjusts the degree of participation of observation information according to the observation quality before fusion calculation. At the same time, it suppresses or removes low-confidence observations to achieve dynamic optimization of the fusion weight of multi-source information; including the following steps.

[0011] Step 1: Acquisition and Preprocessing of Multi-Source Observation Data. This involves acquiring lidar point cloud data, image data, and inertial measurement data. The data undergoes time synchronization, coordinate system transformation, and motion distortion correction to construct multi-source observation data under a unified time and spatial coordinate system. Specifically, the lidar point cloud and image data are time-synchronized. The point cloud is compensated using timestamp-based linear interpolation, and the image data is aligned using nearest neighbor matching or interpolation, with the time error controlled within 5ms. Subsequently, based on a pre-calibrated extrinsic parameter matrix, the data from each sensor are unified to the UAV's body coordinate system. Motion distortion correction is then performed on the lidar point cloud using IMU integration results to obtain multi-source observation data under a unified time and spatial reference.

[0012] Step Two: Initial Screening of Observation Validity. Based on preset geometric constraints and matching consistency conditions, the validity of multi-source observation data is judged. Point cloud data that does not meet the planar constraints, mismatched image features, and abnormal inertial data are removed to obtain a valid observation set for the quality assessment stage. Specifically, in the observation validity screening stage, outliers that are too far away or have insufficient neighbor points are removed from the point cloud data, and outliers are removed by plane fitting error; features are extracted from visual data using FAST or ORB methods, and geometric consistency is screened using RANSAC, retaining only matching results with an inlier ratio greater than 0.3; at the same time, statistical detection is performed on IMU data within a short time window, and when the acceleration change exceeds a set threshold, it is judged as outlier data and removed.

[0013] Step 3: For the effective observation set, extract multi-dimensional quality indicators characterizing the structural features and statistical properties of the observation data. These quality indicators include parameters reflecting the distribution characteristics of the environmental geometry, the stability of observation matching, and the characteristics of sensor signal changes, used to characterize the comprehensive quality of the current observation data in terms of spatial structure, matching accuracy, and signal stability. Specifically: In the observation quality assessment stage, multi-dimensional quality indicators are constructed for different sensors. For lidar, a local neighborhood covariance matrix is ​​constructed and eigenvalue decomposition is performed to assess whether there is degradation in the environmental geometry, while the mean of the point cloud matching residual characterizes the matching stability. For visual observation, the richness of image texture and matching reliability are assessed by calculating the feature matching in-point ratio, the average reprojection error, and the image information entropy, where the reprojection error is controlled within the range of 1 to 3 pixels. For IMU, the stability of the inertial signal is assessed by calculating the acceleration variance within a sliding window.

[0014] Step 4: Construct an observation reliability model based on the multidimensional quality indicators. Perform unified modeling and normalization on quality indicators from different sources to obtain continuous reliability values ​​that characterize the reliability of observation data, thereby realizing the transformation of observation quality from discrete judgment to continuous quantification. That is, normalize various quality indicators and construct an observation reliability model through weighted fusion to obtain continuous reliability values ​​in the range of 0 to 1. The weights of each sensor can be set according to experience or offline calibration.

[0015] Step 5: Based on the observation confidence level, the observation noise covariance is dynamically adjusted before the observation data enters the fusion calculation. This increases the noise covariance corresponding to low confidence observations and decreases the noise covariance corresponding to high confidence observations, thereby achieving feedforward allocation and pre-tuning of the weights of multi-source observation information. Specifically, before the observations enter the fusion calculation, the observation noise covariance is adjusted based on the observation confidence level. Specifically, when the confidence level is greater than 0.8, the covariance is reduced to enhance the observation constraint effect; when the confidence level is between 0.4 and 0.8, the original covariance is maintained; when the confidence level is less than 0.4, the covariance is increased to reduce its impact; when the confidence level is less than 0.2, the observation is directly removed, thereby achieving dynamic allocation of observation weights. This adjustment process is completed before state estimation.

[0016] Step Six: Adaptive Fusion State Update. In this step, the adjusted covariance matrix is ​​introduced into the state estimation model, and the state is updated using extended Kalman filtering or nonlinear optimization methods. During the update process, the influence of low-quality observations on the system is automatically suppressed based on the feedforward adjusted weights, thereby obtaining stable pose estimation results.

[0017] During the fusion computing phase, extended Kalman filtering or sliding window nonlinear optimization methods can be used for state updates. The IMU is used for state prediction, and laser and visual observations are used for updates. The adjusted covariance is used as a weight during the optimization process to adaptively suppress the influence of low-quality observations on the system and obtain stable pose estimation results.

[0018] Step 7: Output the positioning results and system health. After completing the state update, output the pose estimation results at the current moment. At the same time, construct a system health index based on the credibility of various observations to characterize the reliability of the current positioning results. Quantitatively evaluate the reliability of the positioning results to provide a reference for the upper-level decision-making system.

[0019] The method for constructing the lidar observation quality index in step three is as follows:

[0020] Construct the point cloud covariance matrix in the local neighborhood of each laser point:

[0021]

[0022] in, coordinates of neighboring points The mean coordinates of the neighborhood points. For the number of domain points;

[0023] The covariance matrix is ​​decomposed into eigenvalues ​​to obtain the eigenvalues. And construct a set of degradation indicators:

[0024]

[0025] Used to characterize the geometric observability of the current environment.

[0026] The quality indicators for laser observation also include the matching stability index, which is constructed as follows:

[0027] Calculate the mean of the point cloud matching residuals:

[0028]

[0029] in, For point-to-plane residuals, The number of valid matching points;

[0030] In step three, the structural indicators reflecting the observability of the environmental geometry are... By combining the residual index, which characterizes the stability of the matching, a comprehensive quality evaluation metric for lidar observations is constructed to provide a unified quantitative representation of lidar observation quality. The relevant formula is as follows:

[0031]

[0032] in, and These are the weighting coefficients.

[0033] The visual observation quality indicators mentioned in step three include feature matching stability and image texture information. These are used to construct a matching reliability index based on the feature matching results. The construction method is as follows:

[0034] Calculate the proportion of points within a feature point:

[0035]

[0036] in: The number of feature points required to satisfy geometric consistency constraints; This represents the total number of matched feature points; The ratio of interior points is used to characterize the reliability of feature matching.

[0037] Calculate the average reprojection error:

[0038]

[0039] in: For the first The observed pixel coordinates of each feature point; These are the predicted pixel coordinates estimated from the current state. Represents the Euclidean norm; This represents the mean of visual reprojection error;

[0040] In step three, the feature matching reliability index is... Combined with the observation accuracy index constructed based on reprojection error, a comprehensive quality evaluation metric for visual observation is constructed to provide a unified quantitative representation of visual observation quality. The formula is as follows:

[0041]

[0042] in, and These are the weighting coefficients.

[0043] In step three, the mass index of the inertial measurement unit is constructed by analyzing the acceleration changes within a statistical time window, and its calculation method is as follows:

[0044]

[0045] in: For a moment acceleration measurement value The average acceleration within the time window; The length of the time window; This is an index of acceleration variance;

[0046] In step three, the stability index constructed based on signal fluctuations is homogenized, and a comprehensive inertial observation quality evaluation metric is constructed to provide a unified quantitative characterization of the inertial measurement observation quality. The formula is as follows:

[0047]

[0048] in: These are the weighting coefficients.

[0049] In step four, the observation reliability model is constructed by weighted fusion of multidimensional quality indicators to characterize the reliability of observation data from different sources. Through modeling and continuous quantification of observation reliability, based on the multidimensional observation quality indicators, the reliability of observation data from different sources is uniformly modeled and continuously quantified. Specifically, the observation quality evaluation quantities corresponding to lidar, visual sensors, and inertial measurement units are normalized, and an observation reliability function is constructed to map the multidimensional quality indicators to continuous reliability values ​​at a unified scale. The formula is:

[0050]

[0051] in, These are weighting coefficients used to reflect the relative importance of different sensor observations in the current fusion process;

[0052] A nonlinear mapping function is introduced to transform the observation quality evaluation metric, and an extended form of the observation reliability model is constructed to enhance the distinguishing ability across different quality ranges. The formula is as follows:

[0053]

[0054] in, This is a monotonic mapping function used to enhance the response difference between high-quality and low-quality observations, thereby increasing the confidence level of the observations. As a unified quantitative indicator of the reliability of observation data, it is used to drive the feedforward adaptive adjustment of observation noise covariance, thereby completing the pre-allocation of observation weights before the observations enter the fusion calculation.

[0055] The multidimensional quality indicators include geometric structure indicators of laser observation, feature matching indicators of visual observation, and stability indicators of inertial measurement. After normalizing various indicators, a unified model is performed to obtain a continuous observation reliability value.

[0056] The observation reliability model is expressed as a weighted function of various quality indicators.

[0057] In step five, the feedforward adjustment mechanism specifically involves: using a nonlinear function that includes a critical threshold, adjustment sensitivity, and a maximum expansion penalty coefficient to map the observation confidence to the dynamic observation noise covariance in real time; when the observation confidence is lower than the critical threshold, the current observation noise covariance is exponentially expanded to a preset upper limit through the nonlinear function, thereby a priori reducing the Kalman gain of the corresponding sensor observation in the joint update before state estimation.

[0058] The feedforward adjustment mechanism is implemented based on observation quality, specifically: before the observation data enters the fusion calculation, it adjusts the data according to the observation reliability. The degree of participation of observation information is adjusted by feedforward adjustment, namely: constructing an adjustment function for the observation noise covariance based on the observation reliability, and adaptively adjusting the covariance corresponding to different observations, as shown in the formula:

[0059]

[0060] in, The original observation noise covariance matrix, The adjusted covariance matrix, This refers to the current observation confidence level output in step four. This is a preset system health threshold. This is the maximum expansion penalty coefficient, used to constrain the theoretical upper limit of covariance amplification and ensure the numerical stability of the filter; To adjust the sensitivity parameter and control the confidence level when it crosses a critical threshold Smooth, gradual rate of change; This is a higher-order compensation term for specific hardware characteristics.

[0061] In step seven, a system health index is constructed based on observation reliability to assess the reliability of the positioning results. Specifically, the reliability of observation data from different sources is comprehensively characterized based on the observation quality evaluation quantity or observation reliability corresponding to various sensor observations, resulting in the UAV system health index. This system health index reflects the overall operating status of the current positioning system, with a value range of 0 to 1. It characterizes the reliability of the current positioning results, provides a basis for UAV autonomous flight decisions, and can be used for upper-level decision-making, anomaly detection, or safety control, thereby achieving quantitative assessment and dynamic monitoring of the reliability of positioning results in complex environments.

[0062] The positioning system is integrated into a quadcopter UAV platform, which includes a frame, power system, flight control system, onboard computing module, lidar, visual sensor, and inertial measurement unit. The UAV achieves attitude control and spatial movement through its rotor power system. The four rotor power systems provide lift and attitude control torque to the UAV, enabling it to perform vertical takeoff and landing, hovering, low-speed flight, and in-situ turning, thus adapting to low-speed maneuvering and close-range obstacle avoidance requirements within confined spaces. The positioning system performs multi-source sensor fusion positioning to achieve attitude estimation in GNSS-unavailable environments.

[0063] The lidar is mounted on top of or above the front of the UAV, tilted forward at 10°-15° to provide a large horizontal field of view and forward environmental perception range, while minimizing obstruction of the point cloud field of view by the arms, propellers, and fuselage structure. The lidar's mounting direction maintains a fixed external parameter relationship with the UAV's coordinate system, ensuring that its coordinate system can be calibrated and transformed to the UAV's body coordinate system.

[0064] The camera of the vision sensor is mounted at the front of the drone body, with its optical axis pointing towards the main flight direction of the drone. It is used to acquire images of the forward environment. The fixed transformation relationship between the camera coordinate system and the drone body coordinate system is determined through extrinsic parameter calibration.

[0065] The inertial measurement unit (IMU) is built into the flight controller and is installed near the center of the UAV body or integrated with the flight controller module. It is used to provide high-frequency angular velocity and acceleration information. The IMU coordinate system is transformed to the unified UAV body coordinate system through external parameters or the body installation relationship.

[0066] The complex and confined environments referred to mainly refer to the environments faced by UAVs when performing inspection, detection, or positioning and navigation tasks in non-open spaces such as indoors, utility tunnels, underground spaces, tunnels, factories, narrow passages, and near-wall areas. These environments typically have the following characteristics:

[0067] The environment is usually a closed structure, GNSS signals are unavailable or severely attenuated, and it is impossible to obtain continuous and stable pose information by relying on external satellites or base stations for positioning;

[0068] The narrow space structure presents challenges such as close-to-wall flight, limited turning radius, and close proximity of the aircraft to obstacles. The airflow near the wall can easily affect inertial measurements and obstruct vision.

[0069] Complex or repetitive environmental structures, such as corridors, pipes, tunnels, and walls, can easily cause geometric degradation or unstable matching of lidar.

[0070] Unstable lighting conditions, including weak light, reflections, and abrupt changes in brightness, can easily affect visual feature extraction.

[0071] This invention discloses a multi-source information feedforward adjustment positioning method for UAVs in confined spaces based on observation quality modeling, relating to the fields of UAV positioning and multi-sensor fusion perception technology. This method is implemented based on a positioning system composed of LiDAR, visual sensors, inertial measurement units, and a fusion computing module. Its core lies in introducing the quality assessment results of multi-source observation data into the fusion process, and achieving feedforward dynamic adjustment of the observation noise covariance by constructing an observation reliability model. The method first synchronizes the time and coordinates of LiDAR point clouds, image data, and inertial measurement data, and performs preprocessing such as distortion correction. Based on the screening of observation effectiveness, multi-dimensional quality indicators reflecting environmental geometric characteristics, feature matching stability, and sensor signal changes are extracted. By normalizing and fusing various indicators, a continuous observation reliability is obtained, used to quantify the reliability of the observation data. Furthermore, the noise covariance of various observations is nonlinearly adjusted according to the observation reliability, reducing the weight of low-reliability observations and increasing the weight of high-reliability observations. The adjusted covariance is then introduced into an extended Kalman filter or nonlinear optimization framework for state updates, thereby suppressing the influence of abnormal observations on the positioning results. The final output includes pose estimation results and system health indicators. This invention can effectively improve the accuracy and robustness of multi-source fusion positioning in complex environments, reduce the adverse effects of mismatches and degradation scenarios, and provide reliable pose estimation and system health assessment for autonomous flight of UAVs.

[0072] The purpose of this invention is to propose a multi-sensor fusion positioning method based on observation quality modeling and feedforward adjustment for UAV positioning in confined spaces. This method addresses the problems in existing technologies, such as large errors and susceptibility to interference from single sensors due to factors like GNSS unavailability, environmental degradation, insufficient visual information, and flight disturbances in confined spaces, as well as the difficulty in quantifying observation quality, lack of dynamic adjustment capability for fusion weights, and delayed response to low-quality observations in multi-sensor fusion positioning methods. This invention enables real-time evaluation and feedforward adjustment of the quality of multi-source observations in confined spaces, and compared to the fixed covariance method, it can effectively reduce positioning errors and improve system robustness. Attached Figure Description

[0073] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0074] Appendix Figure 1 This is a schematic diagram of the system flow of the positioning method in an embodiment of the present invention;

[0075] Appendix Figure 2 This is a schematic diagram of the system principle architecture of the positioning method in an embodiment of the present invention;

[0076] Appendix Figure 3 This is a schematic diagram of a drone in an embodiment of the present invention. Detailed Implementation

[0077] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. This embodiment takes a UAV multi-sensor fusion positioning system in a confined space environment as an example. The system includes a lidar, a visual sensor, and an inertial measurement unit, and is used to realize the positioning and navigation function of the UAV in environments where GNSS is unavailable, the structure is complex, and the perception is degraded. The multi-source information feedforward adjustment positioning method based on observation quality modeling is described, but the present invention is not limited to the specific sensor combination form mentioned above.

[0078] As shown in the figure, a multi-source information feedforward adjustment positioning method for a UAV system in a complex and confined environment is presented. The UAV system is equipped with a positioning system for collecting multi-source observation data. The positioning method constructs an observation reliability model by incorporating the quality assessment results of the multi-source observation data into the fusion process and dynamically adjusts the observation noise covariance using a feedforward approach. First, the laser point cloud, image data, and inertial measurement data from the multi-source observations are synchronized in time and unified in coordinates, and preprocessed with distortion correction. Then, based on the screening of observation validity, multi-dimensional quality indicators reflecting environmental geometric characteristics, feature matching stability, and sensor signal changes are extracted. By normalizing and fusing various multi-dimensional quality indicators, a continuous observation reliability is obtained to quantify the reliability of the observation data. Furthermore, the noise covariance of various observations is nonlinearly adjusted according to the observation reliability, reducing the weight of low-reliability observations and increasing the weight of high-reliability observations. The adjusted covariance is then introduced into an extended Kalman filter or nonlinear optimization framework for state updates, suppressing the influence of abnormal observation data on the positioning results. Finally, the pose estimation results and the health index of the UAV flight system are output. This invention can effectively improve the accuracy and robustness of multi-source fusion positioning in complex environments, reduce the adverse effects of mismatch and degradation scenarios, and provide reliable pose estimation and system health assessment for UAV autonomous flight.

[0079] The positioning system is a multi-sensor fusion positioning system, including a lidar, a visual sensor, an inertial measurement unit, and a fusion computing module. The fusion computing module is used to jointly process multi-source observation data and output the pose estimation results of the UAV.

[0080] The positioning system includes a lidar, a visual sensor, and an inertial measurement unit, used to realize the positioning and navigation functions of UAVs in environments where GNSS is unavailable, structures are complex, and perception is degraded. The multi-source information feedforward adjustment positioning method based on observation quality modeling is described, but the present invention is not limited to the specific sensor combination forms mentioned above.

[0081] The fusion computing module is deployed on the airborne computing platform. The lidar, camera and IMU output data at frequencies of 10~20Hz, 15~30Hz and 100~400Hz respectively. The system uses the IMU time as a unified reference to perform time alignment on the multi-source data.

[0082] The positioning method takes multi-source observation information as input, and after completing the validity screening of observation data, constructs an observation quality assessment model based on multi-dimensional observation features, performs real-time quality quantification of observation data from each sensor, and adjusts the degree of participation of observation information according to the observation quality before fusion calculation. At the same time, it suppresses or removes low-confidence observations to achieve dynamic optimization of the fusion weight of multi-source information; including the following steps.

[0083] Step 1: Acquisition and Preprocessing of Multi-Source Observation Data. This involves acquiring lidar point cloud data, image data, and inertial measurement data. The data undergoes time synchronization, coordinate system transformation, and motion distortion correction to construct multi-source observation data under a unified time and spatial coordinate system. Specifically, the lidar point cloud and image data are time-synchronized. The point cloud is compensated using timestamp-based linear interpolation, and the image data is aligned using nearest neighbor matching or interpolation, with the time error controlled within 5ms. Subsequently, based on a pre-calibrated extrinsic parameter matrix, the data from each sensor are unified to the UAV's body coordinate system. Motion distortion correction is then performed on the lidar point cloud using IMU integration results to obtain multi-source observation data under a unified time and spatial reference.

[0084] Step Two: Initial Screening of Observation Validity. Based on preset geometric constraints and matching consistency conditions, the validity of multi-source observation data is judged. Point cloud data that does not meet the planar constraints, mismatched image features, and abnormal inertial data are removed to obtain a valid observation set for the quality assessment stage. Specifically, in the observation validity screening stage, outliers that are too far away or have insufficient neighbor points are removed from the point cloud data, and outliers are removed by plane fitting error; features are extracted from visual data using FAST or ORB methods, and geometric consistency is screened using RANSAC, retaining only matching results with an inlier ratio greater than 0.3; at the same time, statistical detection is performed on IMU data within a short time window, and when the acceleration change exceeds a set threshold, it is judged as outlier data and removed.

[0085] Step 3: For the effective observation set, extract multi-dimensional quality indicators characterizing the structural features and statistical properties of the observation data. These quality indicators include parameters reflecting the distribution characteristics of the environmental geometry, the stability of observation matching, and the characteristics of sensor signal changes, used to characterize the comprehensive quality of the current observation data in terms of spatial structure, matching accuracy, and signal stability. Specifically: In the observation quality assessment stage, multi-dimensional quality indicators are constructed for different sensors. For lidar, a local neighborhood covariance matrix is ​​constructed and eigenvalue decomposition is performed to assess whether there is degradation in the environmental geometry, while the mean of the point cloud matching residual characterizes the matching stability. For visual observation, the richness of image texture and matching reliability are assessed by calculating the feature matching in-point ratio, the average reprojection error, and the image information entropy, where the reprojection error is controlled within the range of 1 to 3 pixels. For IMU, the stability of the inertial signal is assessed by calculating the acceleration variance within a sliding window.

[0086] Step 4: Construct an observation reliability model based on the multidimensional quality indicators. Perform unified modeling and normalization on quality indicators from different sources to obtain continuous reliability values ​​that characterize the reliability of observation data, thereby realizing the transformation of observation quality from discrete judgment to continuous quantification. That is, normalize various quality indicators and construct an observation reliability model through weighted fusion to obtain continuous reliability values ​​in the range of 0 to 1. The weights of each sensor can be set according to experience or offline calibration.

[0087] Step 5: Based on the observation confidence level, the observation noise covariance is dynamically adjusted before the observation data enters the fusion calculation. This increases the noise covariance corresponding to low confidence observations and decreases the noise covariance corresponding to high confidence observations, thereby achieving feedforward allocation and pre-tuning of the weights of multi-source observation information. Specifically, before the observations enter the fusion calculation, the observation noise covariance is adjusted based on the observation confidence level. Specifically, when the confidence level is greater than 0.8, the covariance is reduced to enhance the observation constraint effect; when the confidence level is between 0.4 and 0.8, the original covariance is maintained; when the confidence level is less than 0.4, the covariance is increased to reduce its impact; when the confidence level is less than 0.2, the observation is directly removed, thereby achieving dynamic allocation of observation weights. This adjustment process is completed before state estimation.

[0088] Step Six: Adaptive Fusion State Update. In this step, the adjusted covariance matrix is ​​introduced into the state estimation model, and the state is updated using extended Kalman filtering or nonlinear optimization methods. During the update process, the influence of low-quality observations on the system is automatically suppressed based on the feedforward adjusted weights, thereby obtaining stable pose estimation results.

[0089] During the fusion computing phase, extended Kalman filtering or sliding window nonlinear optimization methods can be used for state updates. The IMU is used for state prediction, and laser and visual observations are used for updates. The adjusted covariance is used as a weight during the optimization process to adaptively suppress the influence of low-quality observations on the system and obtain stable pose estimation results.

[0090] Step 7: Output the positioning results and system health. After completing the state update, output the pose estimation results at the current moment. At the same time, construct a system health index based on the credibility of various observations to characterize the reliability of the current positioning results. Quantitatively evaluate the reliability of the positioning results to provide a reference for the upper-level decision-making system.

[0091] The method for constructing the lidar observation quality index in step three is as follows:

[0092] Construct the point cloud covariance matrix in the local neighborhood of each laser point:

[0093]

[0094] in, coordinates of neighboring points The mean coordinates of the neighborhood points. For the number of domain points;

[0095] The covariance matrix is ​​decomposed into eigenvalues ​​to obtain the eigenvalues. And construct a set of degradation indicators:

[0096]

[0097] Used to characterize the geometric observability of the current environment.

[0098] The quality indicators for laser observation also include the matching stability index, which is constructed as follows:

[0099] Calculate the mean of the point cloud matching residuals:

[0100]

[0101] in, For point-to-plane residuals, The number of valid matching points;

[0102] In step three, the structural indicators reflecting the observability of the environmental geometry are... By combining the residual index, which characterizes the stability of the matching, a comprehensive quality evaluation metric for lidar observations is constructed to provide a unified quantitative representation of lidar observation quality. The relevant formula is as follows:

[0103]

[0104] in, and These are the weighting coefficients.

[0105] The visual observation quality indicators mentioned in step three include feature matching stability and image texture information. These are used to construct a matching reliability index based on the feature matching results. The construction method is as follows:

[0106] Calculate the proportion of points within a feature point:

[0107]

[0108] in: The number of feature points required to satisfy geometric consistency constraints; This represents the total number of matched feature points; The ratio of interior points is used to characterize the reliability of feature matching.

[0109] Calculate the average reprojection error:

[0110]

[0111] in: For the first The observed pixel coordinates of each feature point; These are the predicted pixel coordinates estimated from the current state. Represents the Euclidean norm; This represents the mean of visual reprojection error;

[0112] In step three, the feature matching reliability index is... Combined with the observation accuracy index constructed based on reprojection error, a comprehensive quality evaluation metric for visual observation is constructed to provide a unified quantitative representation of visual observation quality. The formula is as follows:

[0113]

[0114] in, and These are the weighting coefficients.

[0115] In step three, the mass index of the inertial measurement unit is constructed by analyzing the acceleration changes within a statistical time window, and its calculation method is as follows:

[0116]

[0117] in: For a moment acceleration measurement value The average acceleration within the time window; The length of the time window; The acceleration variance index;

[0118] In step three, the stability index constructed based on signal fluctuations is homogenized, and a comprehensive inertial observation quality evaluation metric is constructed to provide a unified quantitative characterization of the inertial measurement observation quality. The formula is as follows:

[0119]

[0120] in: These are the weighting coefficients.

[0121] In step four, the observation reliability model is constructed by weighted fusion of multidimensional quality indicators to characterize the reliability of observation data from different sources. Through modeling and continuous quantification of observation reliability, based on the multidimensional observation quality indicators, the reliability of observation data from different sources is uniformly modeled and continuously quantified. Specifically, the observation quality evaluation quantities corresponding to lidar, visual sensors, and inertial measurement units are normalized, and an observation reliability function is constructed to map the multidimensional quality indicators to continuous reliability values ​​at a unified scale. The formula is:

[0122]

[0123] in, These are weighting coefficients used to reflect the relative importance of different sensor observations in the current fusion process;

[0124] A nonlinear mapping function is introduced to transform the observation quality evaluation metric, and an extended form of the observation reliability model is constructed to enhance the distinguishing ability across different quality ranges. The formula is as follows:

[0125]

[0126] in, This is a monotonic mapping function used to enhance the response difference between high-quality and low-quality observations, thereby increasing the confidence level of the observations. As a unified quantitative indicator of the reliability of observation data, it is used to drive the feedforward adaptive adjustment of observation noise covariance, thereby completing the pre-allocation of observation weights before the observations enter the fusion calculation.

[0127] The multidimensional quality indicators include geometric structure indicators of laser observation, feature matching indicators of visual observation, and stability indicators of inertial measurement. After normalizing various indicators, a unified model is performed to obtain a continuous observation reliability value.

[0128] The observation reliability model is expressed as a weighted function of various quality indicators.

[0129] In step five, the feedforward adjustment mechanism specifically involves: using a nonlinear function that includes a critical threshold, adjustment sensitivity, and a maximum expansion penalty coefficient to map the observation confidence to the dynamic observation noise covariance in real time; when the observation confidence is lower than the critical threshold, the current observation noise covariance is exponentially expanded to a preset upper limit through the nonlinear function, thereby a priori reducing the Kalman gain of the corresponding sensor observation in the joint update before state estimation.

[0130] The feedforward adjustment mechanism is implemented based on observation quality, specifically: before the observation data enters the fusion calculation, it adjusts the data according to the observation reliability. The degree of participation of observation information is adjusted by feedforward adjustment, namely: constructing an adjustment function for the observation noise covariance based on the observation reliability, and adaptively adjusting the covariance corresponding to different observations, as shown in the formula:

[0131]

[0132] in, The original observation noise covariance matrix, The adjusted covariance matrix, This refers to the current observation confidence level output in step four. This is a preset system health threshold. This is the maximum expansion penalty coefficient, used to constrain the theoretical upper limit of covariance amplification and ensure the numerical stability of the filter; To adjust the sensitivity parameter and control the confidence level when it crosses a critical threshold Smooth, gradual rate of change; This is a higher-order compensation term for specific hardware characteristics.

[0133] In step seven, a system health index is constructed based on observation reliability to assess the reliability of the positioning results. Specifically, the reliability of observation data from different sources is comprehensively characterized based on the observation quality evaluation quantity or observation reliability corresponding to various sensor observations, resulting in the UAV system health index. This system health index reflects the overall operating status of the current positioning system, with a value range of 0 to 1. It characterizes the reliability of the current positioning results, provides a basis for UAV autonomous flight decisions, and can be used for upper-level decision-making, anomaly detection, or safety control, thereby achieving quantitative assessment and dynamic monitoring of the reliability of positioning results in complex environments.

[0134] The positioning system is integrated into a quadcopter UAV platform, which includes a frame, power system, flight control system, onboard computing module, lidar, visual sensor, and inertial measurement unit. The UAV achieves attitude control and spatial movement through its rotor power system. The four rotor power systems provide lift and attitude control torque to the UAV, enabling it to perform vertical takeoff and landing, hovering, low-speed flight, and in-situ turning, thus adapting to low-speed maneuvering and close-range obstacle avoidance requirements within confined spaces. The positioning system performs multi-source sensor fusion positioning to achieve attitude estimation in GNSS-unavailable environments.

[0135] The lidar is mounted on top of or above the front of the UAV, tilted forward at 10°-15° to provide a large horizontal field of view and forward environmental perception range, while minimizing obstruction of the point cloud field of view by the arms, propellers, and fuselage structure. The lidar's mounting direction maintains a fixed external parameter relationship with the UAV's coordinate system, ensuring that its coordinate system can be calibrated and transformed to the UAV's body coordinate system.

[0136] The camera of the vision sensor is mounted at the front of the drone body, with its optical axis pointing towards the main flight direction of the drone. It is used to acquire images of the forward environment. The fixed transformation relationship between the camera coordinate system and the drone body coordinate system is determined through extrinsic parameter calibration.

[0137] The inertial measurement unit (IMU) is built into the flight controller and is installed near the center of the UAV body or integrated with the flight controller module. It is used to provide high-frequency angular velocity and acceleration information. The IMU coordinate system is transformed to the unified UAV body coordinate system through external parameters or the body installation relationship.

[0138] The complex and confined environments referred to mainly refer to the environments faced by UAVs when performing inspection, detection, or positioning and navigation tasks in non-open spaces such as indoors, utility tunnels, underground spaces, tunnels, factories, narrow passages, and near-wall areas. These environments typically have the following characteristics:

[0139] The environment is usually a closed structure, GNSS signals are unavailable or severely attenuated, and it is impossible to obtain continuous and stable pose information by relying on external satellites or base stations for positioning;

[0140] The narrow space structure presents challenges such as close-to-wall flight, limited turning radius, and close proximity of the aircraft to obstacles. The airflow near the wall can easily affect inertial measurements and obstruct vision.

[0141] Complex or repetitive environmental structures, such as corridors, pipes, tunnels, and walls, can easily cause geometric degradation or unstable matching of lidar.

[0142] Unstable lighting conditions, including weak light, reflections, and abrupt changes in brightness, can easily affect visual feature extraction.

Claims

1. A multi-source information feedforward adjustment positioning method for unmanned aerial vehicle (UAV) systems in complex and confined environments, characterized in that: The UAV system is equipped with a positioning system for collecting multi-source observation data. The positioning method constructs an observation reliability model by incorporating the quality assessment results of the multi-source observation data into the fusion process and dynamically adjusting the observation noise covariance using a feedforward approach. First, the laser point cloud, image data, and inertial measurement data from the multi-source observation data are synchronized in time and unified in coordinates, and distortion removal preprocessing is performed. Then, multi-dimensional quality indicators that reflect environmental geometric characteristics, feature matching stability, and sensor signal changes are extracted. By normalizing and fusing various multi-dimensional quality indicators, continuous observation reliability is obtained. Then, based on the observation reliability, the noise covariance of various observations is nonlinearly adjusted, and the adjusted covariance is introduced into the extended Kalman filter or nonlinear optimization framework for state update to suppress the influence of abnormal observation data on the positioning results. Output pose estimation results and unmanned aerial vehicle system health indicators.

2. The multi-source information feedforward adjustment positioning method for a complex and confined environment unmanned aerial vehicle system according to claim 1, characterized in that: The positioning system is a multi-sensor fusion positioning system, including a lidar, a visual sensor, an inertial measurement unit, and a fusion computing module. The fusion computing module is used to jointly process multi-source observation data and output the pose estimation results of the UAV.

3. The multi-source information feedforward adjustment positioning method for unmanned aerial vehicle systems in complex and confined environments according to claim 2, characterized in that: The positioning method takes multi-source observation information as input, and after completing the screening of the validity of the observation data, it constructs an observation quality assessment model based on multi-dimensional observation features, performs real-time quality quantification of the observation data of each sensor, and adjusts the degree of participation of the observation information according to the observation quality before the fusion calculation. At the same time, it suppresses or removes low-confidence observations to achieve dynamic optimization of the fusion weight of multi-source information. Includes the following steps; Step 1: Acquisition and Preprocessing of Multi-Source Observation Data. This involves acquiring lidar point cloud data, image data, and inertial measurement data. The data undergoes time synchronization, coordinate system transformation, and motion distortion correction to construct multi-source observation data under a unified time and spatial coordinate system. Specifically, the lidar point cloud and image data are time-synchronized. The point cloud is compensated using timestamp-based linear interpolation, and the image data is aligned using nearest neighbor matching or interpolation. Subsequently, based on a pre-calibrated extrinsic parameter matrix, the data from each sensor are unified to the UAV's body coordinate system. Motion distortion correction is then performed on the lidar point cloud using IMU integration results to obtain multi-source observation data under a unified time and spatial reference. Step 2: Initial Screening of Observation Validity. Based on preset geometric constraints and matching consistency conditions, the validity of multi-source observation data is judged. Point cloud data that does not meet the planar constraints, mismatched image features, and abnormal inertial data are removed to obtain a valid observation set for the quality assessment stage. Specifically, in the observation validity screening stage, outliers that are too far away or have insufficient neighbor points are removed from the point cloud data, and outliers are removed by plane fitting error; features are extracted from visual data using FAST or ORB methods, and geometric consistency is screened using RANSAC; at the same time, statistical detection is performed on IMU data within a short time window, and when the acceleration change exceeds a set threshold, it is judged as abnormal data and removed. Step 3: For the effective observation set, extract multi-dimensional quality indicators that characterize the structural features and statistical properties of the observation data. These quality indicators include parameters that reflect the distribution characteristics of the environmental geometric structure, the stability of observation matching, and the characteristics of sensor signal changes. They are used to characterize the comprehensive quality of the current observation data in terms of spatial structure, matching accuracy, and signal stability. Specifically, in the observation quality assessment stage, multi-dimensional quality indicators are constructed for different sensors. For lidar, a local neighborhood covariance matrix is ​​constructed and eigenvalue decomposition is performed to assess whether there is degradation in the environmental geometric structure. At the same time, the mean of the point cloud matching residuals is combined to characterize the matching stability. For visual observation, the richness of image texture and matching reliability are evaluated by calculating the feature matching inlier ratio, average reprojection error, and image information entropy; for IMU, the stability of inertial signal is evaluated by calculating the acceleration variance within a sliding window. Step 4: Construct an observation reliability model based on the multidimensional quality indicators, and perform unified modeling and normalization on the quality indicators from different sources to obtain a continuous reliability value that characterizes the reliability of the observation data, so as to realize the transformation of observation quality from discrete judgment to continuous quantification. Step 5: Based on the observation confidence level, dynamically adjust the observation noise covariance before the observation data enters the fusion calculation, so that the noise covariance corresponding to low confidence observation increases and the noise covariance corresponding to high confidence observation decreases, thereby realizing the feedforward allocation and pre-tuning of the weights of multi-source observation information. Step Six: Adaptive Fusion State Update. In this step, the adjusted covariance matrix is ​​introduced into the state estimation model, and the state is updated using extended Kalman filtering or nonlinear optimization methods. During the update process, the influence of low-quality observations on the system is automatically suppressed based on the feedforward adjusted weights. Step 7: Output the positioning results and system health. After completing the state update, output the pose estimation results at the current moment. At the same time, construct a system health index based on the confidence of various observations to characterize the reliability of the current positioning results.

4. The multi-source information feedforward adjustment positioning method for a complex and confined environment UAV system according to claim 3, characterized in that: The method for constructing the lidar observation quality index in step three is as follows: Construct the point cloud covariance matrix in the local neighborhood of each laser point: in, coordinates of neighboring points The mean coordinates of the neighborhood points. For the number of domain points; The covariance matrix is ​​decomposed into eigenvalues ​​to obtain the eigenvalues. And construct a set of degradation indicators: Used to characterize the geometric observability of the current environment. The quality indicators for laser observation also include the matching stability index, which is constructed as follows: Calculate the mean of the point cloud matching residuals: in, For point-to-plane residuals, The number of valid matching points; In step three, the structural indicators reflecting the observability of the environmental geometry are... By combining the residual index, which characterizes the stability of the matching, a comprehensive quality evaluation metric for lidar observations is constructed to provide a unified quantitative representation of lidar observation quality. The relevant formula is as follows: in, and These are the weighting coefficients.

5. The multi-source information feedforward adjustment positioning method for a complex and confined environment UAV system according to claim 4, characterized in that: The visual observation quality indicators mentioned in step three include feature matching stability and image texture information. These are used to construct a matching reliability index based on the feature matching results. The construction method is as follows: Calculate the proportion of points within a feature point: in: The number of feature points required to satisfy geometric consistency constraints; This represents the total number of matched feature points; The ratio of interior points is used to characterize the reliability of feature matching. Calculate the average reprojection error: in: For the first The observed pixel coordinates of each feature point; These are the predicted pixel coordinates estimated from the current state. Represents the Euclidean norm; This represents the mean of visual reprojection error; In step three, the feature matching reliability index is... Combined with the observation accuracy index constructed based on reprojection error, a comprehensive quality evaluation metric for visual observation is constructed to provide a unified quantitative representation of visual observation quality. The formula is as follows: in, and These are the weighting coefficients.

6. The multi-source information feedforward adjustment positioning method for a complex and confined environment unmanned aerial vehicle system according to claim 5, characterized in that: In step three, the mass index of the inertial measurement unit is constructed by analyzing the acceleration changes within a statistical time window, and its calculation method is as follows: in: For a moment acceleration measurement value The average acceleration within the time window; The length of the time window; The acceleration variance index; In step three, the stability index constructed based on signal fluctuations is homogenized, and a comprehensive inertial observation quality evaluation metric is constructed to provide a unified quantitative characterization of the inertial measurement observation quality. The formula is as follows: in: These are the weighting coefficients.

7. The multi-source information feedforward adjustment positioning method for a complex and confined environment unmanned aerial vehicle system according to claim 6, characterized in that: In step four, the observation reliability model is constructed by weighted fusion of multidimensional quality indicators to characterize the reliability of observation data from different sources. Through modeling and continuous quantification of observation reliability, based on the multidimensional observation quality indicators, the reliability of observation data from different sources is uniformly modeled and continuously quantified. Specifically, the observation quality evaluation quantities corresponding to lidar, visual sensors, and inertial measurement units are normalized, and an observation reliability function is constructed to map the multidimensional quality indicators to continuous reliability values ​​at a unified scale. The formula is: in, These are weighting coefficients used to reflect the relative importance of different sensor observations in the current fusion process; A nonlinear mapping function is introduced to transform the observation quality evaluation metric, and an extended form of the observation reliability model is constructed to enhance the distinguishing ability across different quality ranges. The formula is as follows: in, This is a monotonic mapping function used to enhance the response difference between high-quality and low-quality observations, thereby increasing the confidence level of the observations. As a unified quantitative indicator of the reliability of observation data, it is used to drive the feedforward adaptive adjustment of observation noise covariance, thereby completing the pre-allocation of observation weights before the observations enter the fusion calculation. The multidimensional quality indicators include geometric structure indicators of laser observation, feature matching indicators of visual observation, and stability indicators of inertial measurement. After normalizing various indicators, a unified model is performed to obtain a continuous observation reliability value. The observation reliability model is expressed as a weighted function of various quality indicators.

8. The multi-source information feedforward adjustment positioning method for a complex and confined environment unmanned aerial vehicle system according to claim 6, characterized in that: In step five, the feedforward adjustment mechanism specifically involves: using a nonlinear function that includes a critical threshold, adjustment sensitivity, and a maximum expansion penalty coefficient to map the observation confidence to the dynamic observation noise covariance in real time; when the observation confidence is lower than the critical threshold, the current observation noise covariance is exponentially expanded to a preset upper limit through the nonlinear function, thereby a priori reducing the Kalman gain of the corresponding sensor observation in the joint update before state estimation. The feedforward adjustment mechanism is implemented based on observation quality, specifically: before the observation data enters the fusion calculation, it adjusts the data according to the observation reliability. The degree of participation of observation information is adjusted by feedforward adjustment, namely: constructing an adjustment function for the observation noise covariance based on the observation reliability, and adaptively adjusting the covariance corresponding to different observations, as shown in the formula: in, The original observation noise covariance matrix, The adjusted covariance matrix, This refers to the current observation confidence level output in step four. This is a preset system health threshold. This is the maximum expansion penalty coefficient, used to constrain the theoretical upper limit of covariance amplification and ensure the numerical stability of the filter; To adjust the sensitivity parameter and control the confidence level when it crosses a critical threshold Smooth, gradual rate of change; This is a higher-order compensation term for specific hardware characteristics.

9. The multi-source information feedforward adjustment positioning method for a complex and confined environment unmanned aerial vehicle system according to claim 6, characterized in that: In step seven, a system health index is constructed based on observation reliability to evaluate the reliability of the positioning results. Specifically, the reliability of observation data from different sources is comprehensively characterized based on the observation quality evaluation quantity or observation reliability corresponding to various sensor observations to obtain the UAV system health index. The system health index reflects the overall operating status of the current positioning system, and its value ranges from 0 to 1. It is used to characterize the reliability of the current positioning results and provides a basis for the UAV's autonomous flight decision.

10. The multi-source information feedforward adjustment positioning method for a complex and confined environment unmanned aerial vehicle system according to claim 2, characterized in that: The positioning system is integrated into a rotary-wing UAV platform, which includes a frame, power system, flight control system, onboard computing module, lidar, visual sensor, and inertial measurement unit. The UAV achieves attitude control and spatial movement through the rotor power system, which provides lift and attitude control torque, enabling the UAV to perform vertical takeoff and landing, hovering, low-speed flight, and in-situ turning, thus adapting to low-speed maneuvering and close-range obstacle avoidance requirements in confined spaces. The positioning system performs multi-source sensor fusion positioning to achieve attitude estimation in GNSS-unavailable environments. The lidar is mounted on top of or above the front of the UAV, tilted forward at a 10°-15° angle. The lidar's mounting direction maintains a fixed external parameter relationship with the UAV's coordinate system, ensuring its coordinate system can be calibrated and transformed to the UAV's body coordinate system. The camera of the vision sensor is mounted at the front of the drone body, with its optical axis pointing towards the main flight direction of the drone. It is used to acquire images of the forward environment. The fixed transformation relationship between the camera coordinate system and the drone body coordinate system is determined through extrinsic parameter calibration. The inertial measurement unit (IMU) is built into the flight controller and is installed near the center of the UAV body or integrated with the flight controller module. It is used to provide high-frequency angular velocity and acceleration information. The IMU coordinate system is transformed to the unified UAV body coordinate system through external parameters or the body installation relationship.

Images