Unmanned aerial vehicle intelligent accurate positioning system and method based on multi-source information fusion

Abstract

The invention discloses an unmanned aerial vehicle intelligent precise positioning system and method based on multi-source information fusion, and belongs to the technical field of unmanned aerial vehicle navigation positioning, and the system comprises a multi-sensor data collection module which is used for obtaining an original physical signal required by positioning; the data preprocessing and synchronizing module is used for processing original physical signals and ensuring the consistency of multi-source data; the feature extraction and semantic recognition module is used for extracting environment features and industrial equipment semantic information from original data and providing geometric and semantic constraints for positioning; the multi-source information fusion positioning module is used for fusing multi-source data and constructing a factor graph containing sensor error terms and semantic constraint factors; the positioning calculation and state estimation module is used for calculating the positioning pose of the unmanned aerial vehicle based on the fused data; according to the unmanned aerial vehicle intelligent accurate positioning system and method based on multi-source information fusion provided by the invention, by fusing multi-source information, the positioning problem in a complex industrial environment is solved, and accurate identification and positioning of target equipment are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle navigation positioning, in particular to an intelligent precise positioning system and method for unmanned aerial vehicles based on multi-source information fusion. BACKGROUND

[0002] With the wide application of unmanned aerial vehicles in industrial inspection (such as power inspection, oil pipeline inspection), logistics distribution, surveying and mapping and other fields, the requirements for positioning accuracy and intelligence level are increasingly stringent. The traditional single sensor positioning scheme (such as relying only on GPS / RTK which is susceptible to blocked environment, relying only on vision which fails in complex lighting or textureless scene, relying only on IMU which has cumulative error) has obvious limitations and cannot meet the high-precision and high-reliability positioning requirements in complex scenes (such as inside industrial plants, between urban high-rise buildings, and complex terrains in the wild).

[0003] At the same time, there is an increasing demand for unmanned aerial vehicles to identify specific industrial equipment (such as valves, instruments, and power transmission tower components) in industrial scenarios and assist positioning based on device semantic information to achieve more intelligent inspection and operation planning. Therefore, there is an urgent need for an intelligent precise positioning system for unmanned aerial vehicles that integrates multi-source sensor information, has geometric positioning accuracy and semantic constraints, and can adapt to complex environments. SUMMARY

[0004] The purpose of the present application is to provide an intelligent precise positioning system and method for unmanned aerial vehicles based on multi-source information fusion to solve the problems in the background art.

[0005] To achieve the above purpose, the present application provides an intelligent precise positioning system for unmanned aerial vehicles based on multi-source information fusion, comprising: a multi-sensor data acquisition module for acquiring original physical signals required for positioning; a data preprocessing and synchronization module for processing the original physical signals to ensure the consistency of multi-source data; a feature extraction and semantic recognition module for extracting environmental features and industrial equipment semantic information from the original data to provide geometric and semantic constraints for positioning; a multi-source information fusion positioning module for fusing multi-source data by constructing a factor graph containing sensor error terms and semantic constraint factors; a positioning solution and state estimation module for calculating the positioning pose of the unmanned aerial vehicle based on the fused data.

[0006] The present application also provides an intelligent precise positioning method for unmanned aerial vehicles based on multi-source information fusion, comprising the following steps: S1, obtaining the original physical signals required for positioning, including visible light images and infrared image data collected by the binocular vision unit, Beidou / GNSS satellite signals received by the RTK unit, data collected by the ultrasonic sensor array unit, three-dimensional point cloud data collected by the LiDAR point cloud information acquisition unit, acceleration and angular velocity of the unmanned aerial vehicle measured by the IMU inertial measurement unit, and auxiliary height information collected by the barometer; S2, processing the original physical signals to ensure the consistency of multi-source data; S3, extracting environmental features and industrial equipment semantic information from the original data to provide geometric and semantic constraints for positioning; S4, fusing multi-source data by constructing a factor graph containing sensor error terms and semantic constraint factors; S5, calculating the unmanned aerial vehicle positioning pose based on the fused data.

[0007] Preferably, step S2 specifically includes: Time synchronization and registration, linear interpolation method is used to align sensor data of different sampling frequencies to a unified time reference, ensuring that each frame of data matches in time; Outlier rejection, abnormal data outside the reasonable range is detected and rejected by setting a threshold; the determination formula is: ; Wherein, is the sensor measurement value; is the sliding window mean; is the standard deviation; Missing value completion, forward difference method is used to complete the short-time missing sensor data; Spatial registration, all sensor data is converted to a geographic coordinate system: ; Wherein, is the conversion matrix; is the translation vector; Kalman filter is used to suppress IMU noise, smooth angular velocity and acceleration data, and the IMU state update equation is: ; Wherein, is the state transition matrix; is the Kalman gain; is the observation matrix; represents the optimal estimate value of the IMU state at the time; represents the optimal estimate value of the IMU state at the time.

[0008] Preferably, step S3 specifically includes: S31, through the dual-branch Swin-Transformer structure, feature extraction is performed on visible light images and infrared images respectively, as well as ultrasonic and LiDAR assisted feature extraction. Ultrasonic sensor data is used to generate equipment close-range features. , LiDAR auxiliary features extract line and surface features by performing curvature calculation on LiDAR point cloud data : ;in, It is a function for calculating the curvature of point cloud data; Represents a point in LiDAR point cloud data; is the eigenvalue of the point cloud neighborhood covariance matrix; S32, visible light features of each stage output of Swin-Transformer and infrared characteristics , achieve fusion through autonomous learning weights, and then integrate ultrasonic and LiDAR features into fusion features through channel splicing; S33. Construct top-down and bottom-up feature pyramids to deepen multi-scale feature fusion and output multi-scale device feature maps. ; S34: The corresponding feature map detection head predicts the device category and location, constructs semantic constraints, and finally outputs a fused feature vector containing geometric features, semantic features, and constraints, providing dual constraints for positioning: ;in, is the spatial geometric features obtained from the sensor; is the semantic feature vector; is the constraint vector.

[0009] Preferably, step S32 realizes feature fusion by constructing a three-stage autonomous learning strategy of feature interaction-attention learning-dynamic weighting, and the specific formula is: ; ; ; ; in, It is the intermediate feature vector obtained by summing the visible light and infrared features, and then processing them through global average pooling and a fully connected layer; is the fully connected layer; GAP is the global average pooling, is the adaptive weight; It is a fully connected layer related to visible light features; It is a fully connected layer related to infrared features; It is the feature obtained by fusion of visible light feature and infrared feature; Integrate ultrasonic and LiDAR features into fusion features through channel stitching: ,in, For the Ultrasound characteristics of the stage; For the LiDAR features of the stage, Represents a concatenation operation on the channel dimension.

[0010] Preferably, the feature pyramid in step S33 specifically includes: Top-down fusion: high-level semantic features are upsampled and then concatenated with shallow features: ; Bottom-up enhancement: shallow positioning features are downsampled and then concatenated with high-level features: ; in, In the top-down fusion process, Feature maps of layers; is the convolution operation; is the upsampling operation; For the The feature map of the layer is usually a high-level semantic feature map; for Convolution operation; In the bottom-up enhancement process, Feature maps of layers; is the downsampling operation; For the The feature map of the layer is usually a shallow positioning feature map.

[0011] Preferably, the factor graph model construction in step S4 specifically includes: S41. Define the system state vector ; in, is the three-dimensional position; is the quaternion attitude; Three-dimensional velocity; IMU gyro / accelerometer bias; set the initial state based on the LiDAR SLAM initial frame pose and IMU bias calibration results and the covariance matrix ; S42. Add factor types and build error models and their corresponding information matrices. Factor types include LiDARSLAM factors. INS factor , UWB factor and semantic constraint factors .

[0012] Preferably, step S42 specifically comprises: LiDAR SLAM factor The error model and information matrix of the LiDAR SLAM factor are respectively: ; ; wherein, is a spatial line model fitted by points ; is a spatial plane model fitted by points ; is the sum of errors of all line features and plane features in the same frame of point cloud; is the standard deviation of noise; is an identity matrix; INS factor The error model and information matrix of the INS factor are respectively: ; ; wherein, is the IMU raw measurement value; is the IMU pre-integration calculation, is the state at the last time, is the current control input; is the standard deviation of noise of the IMU angular velocity measurement; is the standard deviation of noise of the IMU acceleration measurement; UWB factor The error model and information matrix of the UWB factor are respectively: ; ; wherein, is the UWB measured distance; is the theoretically calculated distance of state prediction; is the position vector of the reference point; is a non-line-of-sight error compensation term; is an environment distinguishing parameter, if it is a non-line-of-sight environment, then =1, otherwise 0; is the standard deviation of noise of UWB ranging; semantic constraint factor The error model and information matrix of the semantic constraint factor are respectively: ; ; wherein, is the predicted size of the device; a priori standard size for the device; a noise standard deviation for the size constraint; a noise standard deviation for the spatial relationship constraint; a diagonal matrix.

[0013] Preferably, step S5 minimizes the factor graph error based on the error model and information matrix in the multi-source information fusion positioning module using the Levenberg-Marquardt algorithm: ; wherein, the unmanned aerial vehicle pose variable to be optimized; the information matrix defines the error term the weight in the Mahalanobis distance space, traversing each type of factor, a small deviation in the error term in the high confidence dimension is amplified, thereby dominating the optimization direction; the solution result as the unmanned aerial vehicle positioning pose result after fusing multi-source data.

[0014] Therefore, the unmanned aerial vehicle intelligent precise positioning system and method based on multi-source information fusion have the following beneficial effects: (1) The multi-source sensor data such as binocular vision, RTK, ultrasonic sensor array, IMU, barometer, etc. are fused, which is complementary to each other, effectively overcomes the shortcomings of single sensor in occlusion, complex lighting, no texture, etc. scene, and improves the robustness and reliability of positioning in complex environment; (2) Multi-modal features are extracted and adaptively fused, multi-scale feature utilization is deepened combined with bidirectional feature pyramid, industrial equipment semantic information is recognized and equipment semantic constraint is constructed, which meets the dual needs of equipment recognition and precise positioning in industrial inspection scene; (3) A factor graph containing multi-sensor error terms and semantic constraint factors is constructed, the weights of each error term are reasonably distributed through the information matrix, multi-source information is fully utilized, geometric positioning and semantic constraint are effectively fused, and finally high-precision and high-reliability unmanned aerial vehicle positioning pose is output, which meets the industrial precise positioning demand.

[0015] The technical solutions of the present application will be further described in detail below with the help of the drawings and examples. DRAWINGS

[0016] Figure 1 is the structure flow chart of the unmanned aerial vehicle intelligent precise positioning system based on multi-source information fusion of the present application. DETAILED DESCRIPTION

[0017] The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort shall fall within the scope of protection of the present invention.

[0018] Example 1

[0019] See also Figure 1 ,The intelligent and precise positioning method of UAV based on multi-source information fusion includes the following steps: S1. Acquire the original physical signals required for positioning, including visible light and infrared image data collected by the binocular vision unit, Beidou / GNSS satellite signals received by the RTK unit, data collected by the ultrasonic sensor array unit, 3D point cloud data collected by the LiDAR point cloud information acquisition unit, drone acceleration and angular velocity measured by the IMU inertial measurement unit, and auxiliary altitude information collected by the barometer; S2. Process the original physical signal to ensure the consistency of multi-source data. The specific implementation process is as follows: Time synchronization and registration, using linear interpolation to align sensor data with different sampling frequencies to a unified time base, ensuring that each frame of data matches in time; suppose sensor A is at time The measurement value is , sensor B at time and The measurement value is and ,like , then align to The sensor B measurement value is: .

[0020] Outlier elimination: By setting a threshold, abnormal data that exceeds a reasonable range (such as RTK jump value and ultrasonic multipath reflection error) is detected and eliminated; the judgment formula is: ; in, is the sensor measurement value; is the sliding window mean; is the standard deviation; Missing value completion, the forward difference method is used to complete the short-term missing sensor data; if the time If data is missing, use the previous moment and The measured value 、 Estimate: ; Spatial registration, converting all sensor data into a geographic coordinate system: ; wherein, is a transformation matrix; is a translation vector; The Kalman filter is used to suppress the IMU noise, smooth the angular velocity and acceleration data, and the IMU state update equation is: ; wherein, is a state transition matrix; is a Kalman gain; is an observation matrix; represents the optimal estimation value of the IMU state at the time; represents the optimal estimation value of the IMU state at the time.

[0021] S3, extracting environmental features and industrial equipment semantic information from raw data to provide geometric and semantic constraints for positioning; specifically including: S31, feature extraction is performed on visible light images (containing industrial equipment texture details) and infrared images (containing equipment thermal radiation features) respectively through a double-branch Swin-Transformer structure, and ultrasonic and LiDAR auxiliary feature extraction is performed, the ultrasonic sensor data is used to generate equipment near-distance features , and the LiDAR auxiliary feature is extracted by calculating the curvature of the LiDAR point cloud data to extract line and surface features : ; wherein, is a function for calculating the curvature of the point cloud data; represents a point in the LiDAR point cloud data; is the eigenvalue of the point cloud neighborhood covariance matrix; S32, the visible light features and infrared features output by each stage of the Swin-Transformer are fused by learning the weights independently, and the formula is: ; ; ; ; wherein, is an intermediate feature vector obtained by summing the visible light and infrared features, and then processing through global average pooling and a fully connected layer; is a fully connected layer; GAP is global average pooling, is a fully connected layer related to the visible light features; It is a fully connected layer related to infrared features; It is the feature obtained by fusion of visible light feature and infrared feature; The weight learning process is supervised by the loss function (such as the back propagation of positioning error, dynamic adjustment of weights to make the fusion features more suitable for positioning tasks). Based on the learned weights, the visible light and infrared features are linearly weighted fused to allow the fusion features to be dynamically adjusted with the scene (such as in daytime scenes). Approaching 0.7-0.9, in night / occlusion scenes Improved to 0.6-0.8. Industrial scenario tests have verified that this range can effectively balance the contribution of multimodal features). Ultrasonic and LiDAR features are then integrated into fusion features through channel splicing: ,in, For the Ultrasound characteristics of the stage; For the LiDAR features of the stage, Represents the splicing operation on the channel dimension; S33. Construct a top-down and bottom-up feature pyramid, specifically including: Top-down fusion: high-level semantic features are upsampled and then concatenated with shallow features: ; Bottom-up enhancement: shallow positioning features are downsampled and then concatenated with high-level features: ; in, In the top-down fusion process, Feature maps of layers; is the convolution operation; is the upsampling operation; For the The feature map of the layer is usually a high-level semantic feature map; for Convolution operation; In the bottom-up enhancement process, Feature maps of layers; is the downsampling operation; For the The feature map of the layer is usually a shallow positioning feature map; deepen the multi-scale feature fusion to output the feature map Corresponding to the characteristic diagrams of small, medium and large scale industrial equipment respectively; S34, in Set up the detection head on the feature map to predict the device category and location: ; in, For the detection head The output result; is the device bounding box; is the confidence level; It is the category probability; construct device semantic constraints based on the detection results; is the feature map input to the detection head, and ; The final output contains a fusion feature vector of geometric features and semantic features, providing dual constraints for positioning: ; in, is the spatial geometric features obtained from the sensor; is the semantic feature vector; is the constraint vector.

[0022] S4. Fusion of multi-source data by constructing a factor graph containing sensor error terms and semantic constraint factors; specifically, including: S41. Define the system state vector ; in, is the three-dimensional position; is the quaternion attitude; Three-dimensional velocity; IMU gyro / accelerometer bias; set the initial state based on the LiDAR SLAM initial frame pose and IMU bias calibration results and the covariance matrix ; S42. Add factor types and build error models and their corresponding information matrices. Factor types include LiDARSLAM factors. INS factor , UWB factor and semantic constraint factors ; Specifically include: LiDAR SLAM Factor The error model and information matrix are: ; ; in, For the point Fitted spatial straight line model; For the point Fitted spatial plane model; It is the sum of the errors of all line features and plane features in the same frame point cloud; is the noise standard deviation; is the identity matrix; INS factor The error model and information matrix of the factor are respectively: ; ; wherein, is the IMU raw measurement value; is the IMU pre-integration calculation, is the state at the last time point, is the current control input; is the noise standard deviation of the IMU angular velocity measurement; is the noise standard deviation of the IMU acceleration measurement; UWB factor The error model and information matrix of the factor are respectively: ; ; wherein, is the UWB measured distance; is the theoretically predicted distance of the state; is the position vector of the reference point; is the non-line-of-sight error compensation term; is the environment distinguishing parameter, if it is a non-line-of-sight environment, then =1, otherwise 0; is the noise standard deviation of the UWB ranging; semantic constraint factor The error model and information matrix of the factor are respectively: ; ; wherein, is the predicted size of the device; is the prior standard size of the device; is the noise standard deviation of the size constraint; is the noise standard deviation of the spatial relationship constraint; is a diagonal matrix.

[0023] S5, based on the fusion data to calculate the positioning and pose of the unmanned aerial vehicle, based on the error model and information matrix in the multi-source information fusion positioning module, the Levenberg-Marquardt algorithm is used to minimize the error of the factor graph: ; wherein, is the unmanned aerial vehicle pose variable to be optimized; the information matrix defines the weight of the error term in the Mahalanobis distance space, With various factors, the small deviation of error term on the high confidence dimension is amplified, thus dominating the optimization direction; the solution result As the positioning and pose result of the unmanned aerial vehicle after fusing multi-source data.

[0024] Embodiment two

[0025] The unmanned aerial vehicle intelligent precise positioning system based on multi-source information fusion comprises: A multi-sensor data acquisition module is used to acquire original physical signals required for positioning and provide basic data for subsequent processing; specifically comprising: A binocular vision unit acquires stereo images through left and right cameras, and is used for three-dimensional point cloud reconstruction, feature extraction and industrial equipment identification (such as edge contour detection of pipes and valves); An RTK unit receives Beidou / GNSS satellite signals and provides centimeter-level absolute coordinates (such as longitude, latitude and elevation) in outdoor or GNSS signal good areas as a global positioning reference; An ultrasonic sensor array unit is arranged around the unmanned aerial vehicle to provide near-distance obstacle detection and relative distance measurement; An IMU inertial measurement unit measures acceleration and angular velocity of the unmanned aerial vehicle, and is used for attitude solution (such as pitch, roll and yaw angle) and motion prediction (compensation of high-frequency vibration in the flight of the unmanned aerial vehicle) in a short time; A barometer assists in measuring height information and is complementary to the ultrasonic sensor array data to improve the vertical direction positioning accuracy.

[0026] Therefore, the unmanned aerial vehicle intelligent precise positioning system and method based on multi-source information fusion are adopted, multi-source information such as binocular vision, RTK and ultrasonic is fused, the positioning problem in a complex industrial environment is solved, and precise identification and positioning of target equipment are realized.

[0027] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application but not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: the technical solutions of the present application can still be modified or replaced by equivalents, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. An unmanned aerial vehicle intelligent precision positioning system based on multi-source information fusion, characterized in that, Comprise: Multi-sensor data acquisition module for obtaining the original physical signal required for positioning; Data preprocessing and synchronization module, the original physical signal is processed to ensure the consistency of multi-source data; Feature extraction and semantic recognition module, environmental features and industrial equipment semantic information are extracted from the original data to provide geometric and semantic constraints for positioning; Multi-source information fusion positioning module, multi-source data are fused by constructing a factor graph containing sensor error terms and semantic constraint factors; Positioning solution and state estimation module, based on the fusion data to calculate the UAV positioning pose.

2. The unmanned aerial vehicle intelligent precise positioning method based on multi-source information fusion, applied to the unmanned aerial vehicle intelligent precise positioning system based on multi-source information fusion in claim 1, characterized in that, Comprise the following steps: S1, obtaining the original physical signal required for positioning, including visible light image and infrared image data collected by binocular vision unit, Beidou / GNSS satellite signal received by RTK unit, data collected by ultrasonic sensor array unit, three-dimensional point cloud data collected by LiDAR point cloud information acquisition unit, UAV acceleration and angular velocity measured by IMU inertial measurement unit and auxiliary height information collected by barometer; S2, the original physical signal is processed to ensure the consistency of multi-source data; S3, environmental features and industrial equipment semantic information are extracted from the original data to provide geometric and semantic constraints for positioning; S4, multi-source data are fused by constructing a factor graph containing sensor error terms and semantic constraint factors; S5, based on the fusion data to calculate the UAV positioning pose. 3.The multi-source information fusion based intelligent precise positioning method of the UAV according to claim 2, characterized in that, Step S2 specifically includes: Time synchronization and registration, linear interpolation method is used to align the sensor data of different sampling frequencies to a unified time reference, ensuring that each frame of data matches in time; Outlier rejection, by setting threshold to detect and eliminate abnormal data beyond the reasonable range; The determination formula is: ； wherein, is the sensor measurement value; is the sliding window mean; is the standard deviation; Missing value completion, forward difference method is used to complete the short-time missing sensor data; Spatial registration, all sensor data are converted to geographic coordinate system: ； wherein is a translation matrix; is a translation vector; The Kalman filter is used to suppress the IMU noise, smooth the angular velocity and acceleration data, and the IMU state update equation is: ; wherein, is a state transition matrix; is a Kalman gain; is an observation matrix; denotes the optimal estimate of the IMU state at time denotes the optimal estimate of the IMU state at time denotes the optimal estimate of the IMU state at time denotes the optimal estimate of the IMU state at time 4. The unmanned aerial vehicle intelligent precise positioning method based on multi-source information fusion according to claim 2, characterized in that, Step S3 specifically includes: S31, respectively extracting features of the visible light image and the infrared image through a double-branch Swin-Transformer structure, and ultrasonic and LiDAR auxiliary feature extraction, wherein the ultrasonic sensor data is used to generate device close-range features , and the LiDAR auxiliary feature is extracted by calculating the curvature of the LiDAR point cloud data to extract line and surface features : ; wherein, is a function of calculating the curvature of the point cloud data; represents a point in the LiDAR point cloud data; is an eigenvalue of the point cloud neighborhood covariance matrix; S32, visible light features output by each stage of the Swin-Transformer with infrared features , and then the ultrasound and LiDAR features are fused into the fusion features through channel splicing. S33, construct a top-down and bottom-up feature pyramid, deepen multi-scale feature fusion, and output multi-scale device feature map ; S34, in the corresponding feature map detection head, predict the device category and location, construct semantic constraints, and finally output the fusion feature vector containing geometric features, semantic features and constraints, providing double constraints for positioning: ; wherein, is the spatial geometric feature obtained from the sensor; is the semantic feature vector; is the constraint vector.

5. The unmanned aerial vehicle intelligent precision positioning method based on multi-source information fusion according to claim 4, characterized in that, Step S32 realizes feature fusion by constructing a three-stage autonomous learning strategy of feature interaction-attention learning-dynamic weighting, and the specific formula is: ； ； ； ； wherein, is an intermediate feature vector obtained by summing the visible light and infrared features, followed by global average pooling and fully connected layer processing; is a fully connected layer; GAP is global average pooling, is an adaptive weight; is a fully connected layer related to the visible light feature; is a fully connected layer related to the infrared feature; is a feature after fusion of the visible light feature and the infrared feature; concatenating the ultrasound and LiDAR features through a channel into a fused feature: wherein, is the ultrasound feature of the first phase; is the LiDAR feature of the first phase, denotes a concatenation operation over the channel dimension. 6.The multi-source information fusion based intelligent precise positioning method of a UAV according to claim 5, characterized in that, The feature pyramid in step S33 specifically includes: Top-down fusion: high-level semantic features are upsampled and spliced with shallow features: ； Bottom-up enhancement: shallow positioning features are downsampled and spliced with high-level features: ； in, In the top-down fusion process, Feature maps of layers; is the convolution operation; is the upsampling operation; For the The feature map of the layer is usually a high-level semantic feature map; for Convolution operation; In the bottom-up enhancement process, Feature maps of layers; is the downsampling operation; For the The feature map of the layer is usually a shallow positioning feature map.

7. The unmanned aerial vehicle intelligent precision positioning method based on multi-source information fusion according to claim 2, characterized in that, The factor graph model construction in step S4 specifically includes: S41, define system state vector ; wherein, is a three-dimensional position; is a quaternion pose; is a three-dimensional velocity; is an IMU gyroscope / accelerometer bias; based on the initial frame pose of LiDAR SLAM and the calibration result of IMU bias, set the initial state and the covariance matrix ; S42, add factor types and construct error model and its corresponding information matrix, factor types include LiDAR SLAM factor , INS factor , UWB factor and semantic constraint factor .

8. The unmanned aerial vehicle intelligent precision positioning method based on multi-source information fusion according to claim 7, characterized in that, Step S42 specifically includes: LiDAR SLAM factor The error model and information matrix of the LiDAR SLAM factor are respectively: ； ； wherein, is a spatial line model fitted to the points ; is a spatial plane model fitted to the points ; is the sum of errors for all line features, plane features within the same frame of point cloud; is the noise standard deviation; is the identity matrix; INS factor The error model and information matrix of the INS factor are respectively ； ； wherein, is the IMU raw measurement; is the IMU pre-integration calculation, is the previous time state, is the current control input; is the noise standard deviation of the IMU angular velocity measurement; is the noise standard deviation of the IMU acceleration measurement; UWB factor The error model and information matrix of the UWB factor are respectively ； ； wherein, is the UWB measured distance; is the state predicted theoretical distance; is the position vector of the reference point; is the non-line-of-sight error compensation term; is the environment distinguishing parameter, if it is a non-line-of-sight environment, then = 1, otherwise 0; is the noise standard deviation of the UWB ranging. semantic constraint factor The error model and information matrix of the semantic constraint factor are respectively ； ； where, Predicted size for the device; Prior canonical size for the device; Noise standard deviation for size constraint; Noise standard deviation for spatial relationship constraint; Diagonal matrix.

9. The unmanned aerial vehicle intelligent precision positioning method based on multi-source information fusion according to claim 8, characterized in that, Step S5 based on the error model and information matrix in the multi-source information fusion positioning module, Levenberg-Marquardt algorithm is used to minimize the factor graph error: ； wherein, is the pose variable of the UAV to be optimized; information matrix defines the error term the weight in the Mahalanobis distance space, traversing various factors, a slight deviation of the error term in the high confidence dimension will be amplified, thus dominating the optimization direction; the solution result as the positioning and pose result of the UAV after fusing multi-source data.

Images