N-RTK-based multi-sensor fusion positioning method applied to AMR

Abstract

The invention discloses an N-RTK-based multi-sensor fusion positioning method applied to AMR, and relates to the field of sensor fusion positioning methods. According to the N-RTK-based multi-sensor fusion positioning method applied to the AMR, N-RTK, laser radar, IMU and odometer data are synchronously collected, after strict space-time alignment and preprocessing, a tight coupling fusion positioning frame with a federated Kalman filter as a core is constructed, and the weight of the sensors is dynamically adjusted according to the N-RTK positioning state and environment characteristics, so that the positioning accuracy is improved. According to the method, the limitation problem of a single sensor in a complex scene is effectively solved, the positioning precision of N-RTK can be fully utilized, and complementation and constraint are carried out through high-frequency relative motion observation of laser and IMU, so that continuous, stable and high-precision positioning output can be realized in open, half-sheltered and full-sheltered environments, and the method is suitable for wide, half-sheltered and full-sheltered environments. And the overall positioning error can be controlled within 10cm in a GNSS shielding area.

Description

Technical Field

[0001] This invention relates to the field of sensor fusion localization methods, specifically to an N-RTK-based multi-sensor fusion localization method applied to AMR. Background Technology

[0002] Currently, high-precision positioning solutions for autonomous mobile robots (AMRs) mainly rely on three types of technologies: first, positioning technologies based on Global Navigation Satellite System (GNSS), including point positioning, real-time kinematic (RTK) and its enhanced morphological network RTK (N-RTK); second, simultaneous localization and mapping (SLAM) technology based on LiDAR; and third, loosely coupled or tightly coupled positioning methods based on the fusion of information from inertial measurement units (IMUs), odometry and the aforementioned sensors. Existing technologies typically use GNSS to provide absolute position reference and utilize LiDAR SLAM or inertial navigation systems (INS) to provide relative motion estimation when GNSS signals are poor. The independent pose results calculated by each sensor are weighted or selected through filtering algorithms to achieve positioning continuity in different environments.

[0003] However, the existing technical solutions still have the following problems, specifically: First, in areas where GNSS signals are blocked (such as urban canyons, forest areas, and indoors), the positioning accuracy of conventional GNSS drops sharply or even fails. Solutions relying on laser SLAM face computational efficiency bottlenecks, with point cloud matching taking too long and prone to matching degradation and pose jumps in environments with sparse features or dynamic object interference. Second, most multi-sensor fusion solutions adopt a loosely coupled architecture, where each sensor independently calculates its pose before fusion, failing to effectively model and compensate for the error coupling relationship between sensors. For example, IMU bias drift deteriorates the motion compensation accuracy of laser point clouds, while laser matching errors affect the online estimation of IMU bias, leading to a cyclic amplification of errors within the system. Third, existing fusion strategies are mostly static weight allocations, which are difficult to balance accuracy, robustness, and real-time requirements under complex working conditions. To address the shortcomings of existing technologies, this invention provides an N-RTK-based multi-sensor fusion positioning method applied to AMR to solve the above problems. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an N-RTK-based multi-sensor fusion positioning method for AMR. By simultaneously acquiring data from N-RTK, lidar, IMU, and odometry, and after rigorous spatiotemporal alignment and preprocessing, a tightly coupled fusion positioning framework with a federated Kalman filter at its core is constructed. The sensor weights are dynamically adjusted based on the N-RTK positioning status and environmental characteristics, effectively solving the limitations of single sensors in complex scenarios. This method can fully utilize the positioning accuracy of N-RTK and complement and constrain the high-frequency relative motion observations of lidar and IMU, thereby achieving continuous, stable, and high-precision positioning output in open, partially obscured, and fully obscured environments. The overall positioning error can be controlled within 10 centimeters in GNSS obstruction areas.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a multi-sensor fusion localization method based on N-RTK applied to AMR, the method comprising the following steps:

[0006] Step S1: Simultaneously collect N-RTK positioning data, LiDAR point cloud data, inertial measurement unit data, and odometry data using multiple sensors mounted on the AMR;

[0007] Step S2: Preprocess and synchronize the collected multi-source data to obtain a spatiotemporally aligned sensor observation sequence;

[0008] Step S3: Based on the positioning status and accuracy factor of N-RTK, dynamically adjust the weight allocation strategy of each sensor in the fusion positioning;

[0009] Step S4: Construct and run a federated Kalman filter, using the preprocessed data and dynamic weighting strategy as input, to perform multi-sensor tight coupling fusion and output high-precision position, velocity and attitude information of the AMR in real time.

[0010] Preferably, step S2 includes:

[0011] Step S21: Perform online calibration, sliding window filtering for noise reduction, and gravity compensation on the IMU data to obtain the linear acceleration of the carrier motion;

[0012] Step S22: Perform wheel diameter calibration and slippage error compensation on the odometer data;

[0013] Step S23: Using the PPS pulse signal output by the GNSS receiver as the hardware synchronization reference, trigger the lidar and IMU to synchronize sampling;

[0014] Step S24: Remove outliers and compensate for motion distortion in the lidar point cloud. Motion distortion compensation uses point-by-point linear interpolation of pose transformation obtained by IMU integration between adjacent frames.

[0015] Preferably, in step S3, the dynamic weight allocation strategy adaptively switches according to the N-RTK positioning status and scene features, including:

[0016] When the N-RTK state is a fixed solution and the accuracy factor is higher than the first threshold, the N-RTK data is assigned the dominant fusion weight.

[0017] When the N-RTK state is a floating solution or a single-point solution, it enters the balanced fusion mode and simultaneously increases the fusion weight of the lidar and IMU.

[0018] When the N-RTK state is unsolvable or the signal is lost, it enters the degradation mode and uses the front-end laser odometry and IMU pre-integration results as the main observation sources to perform state recursion.

[0019] Preferably, in step S4, the state vector of the federated Kalman filter includes: the position, velocity, attitude angle of the AMR in three-dimensional space, the zero bias of the IMU gyroscope, the zero bias of the IMU accelerometer, and the odometer calibration factor error.

[0020] Preferably, in the observation update step of the federated Kalman filter, independent observation models and noise covariance matrices are set for different sensors, and the Kalman gain of each observation channel is adjusted in real time according to the dynamic weights determined in step S3.

[0021] Preferably, the observation model of the lidar uses the front-end laser matching pose as the observation input, wherein the initial pose of the laser matching is provided by the N-RTK fixed solution or the filter predicted pose at the current time, so as to accelerate the convergence of the point cloud registration algorithm.

[0022] Preferably, in the balanced fusion mode, the system further fine-tunes the weights based on the feature richness of the laser point cloud: when sparse or degraded environmental geometric features are detected, the weight of laser observation is automatically reduced, and the weights of IMU and odometry are increased accordingly to maintain system robustness.

[0023] Preferably, the method further includes a backend optimization thread that maintains a sliding window factor map containing IMU pre-integration factors, laser odometry factors, and historical N-RTK position factors. When the system detects that the N-RTK signal has recovered or entered an open area, global pose map optimization is initiated to correct the drift error that may accumulate in the federated Kalman filter during signal obstruction.

[0024] Preferably, in the global pose graph optimization, the use of historical N-RTK factors adopts a precision-based weighting strategy, retaining only high-confidence observations with fixed positioning states, and using them to construct strong constraints with the laser sub-map to achieve consistency correction of the global trajectory.

[0025] Preferably, the positioning result output by the method includes a real-time confidence index, which is calculated by combining the posterior covariance matrix of the federated Kalman filter with the observation residuals of each sensor, and is used to provide a reliability reference for the downstream AMR path planning and decision-making module.

[0026] The technical effects and advantages of this invention are as follows:

[0027] 1. This application of an N-RTK-based multi-sensor fusion positioning method on AMR simultaneously acquires data from N-RTK, LiDAR, IMU, and odometry. After rigorous spatiotemporal alignment and preprocessing, a tightly coupled fusion positioning framework centered on a federated Kalman filter is constructed. The sensor weights are dynamically adjusted based on the N-RTK positioning status and environmental characteristics, effectively solving the limitations of single sensors in complex scenarios. This method can fully utilize the positioning accuracy of N-RTK and complement and constrain it through high-frequency relative motion observations of LiDAR and IMU. As a result, continuous, stable, and high-precision positioning output can be achieved in open, partially obscured, and fully obscured environments. The overall positioning error can be controlled within 10 cm in GNSS obstruction areas.

[0028] 2. The proposed N-RTK multi-sensor fusion positioning method for AMR utilizes a dynamic weight allocation strategy. This strategy can adjust the contribution weight of each sensor in state estimation in real time and adaptively based on the N-RTK positioning status and the richness of laser point cloud features. The dynamic weight allocation strategy ensures that absolute position observation dominates when GNSS is of high quality and smoothly transitions to laser and inertial navigation as the main modes when the signal degrades. This avoids the jumping or divergence of positioning results caused by sudden changes in the performance of a single sensor, and significantly improves the adaptability and robustness of the system under sudden environmental changes.

[0029] 3. The N-RTK multi-sensor fusion positioning method applied to AMR introduces a back-end sliding window factor graph optimization thread. This thread can initiate global pose graph optimization when the system detects GNSS signal recovery or entry into an open area. It uses historical high-confidence N-RTK fixed solution observation factors to correct the drift error that may accumulate in the laser odometry during signal obstruction. This not only ensures the consistency of the global trajectory under long-term operation, but also achieves the unity of local high accuracy and global consistency. It overcomes the defect of pure filter methods that cannot close the loop to correct the error accumulation under long-term obstruction. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart of the overall process of the present invention;

[0032] Figure 2 This is a detailed flowchart of the data preprocessing and time synchronization process of the present invention;

[0033] Figure 3 This is the logic diagram for the dynamic weight allocation strategy of this invention;

[0034] Figure 4 This is a flowchart illustrating the workflow of the federated Kalman filter of this invention.

[0035] Figure 5 This is a flowchart of the backend optimization thread of the present invention;

[0036] Figure 6 This is a schematic diagram of the confidence output of the present invention;

[0037] Figure 7 This is a block diagram of the N-RTK multi-sensor fusion positioning logic of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] This embodiment discloses an N-RTK-based multi-sensor fusion localization method applied to AMR, combined with the attached... Figure 1 To be continued Figure 7 As shown, it includes the following steps:

[0040] Step S1: Synchronous data acquisition from multiple sensors:

[0041] The robot receives satellite signals through a dual-frequency or multi-frequency GNSS receiver mounted on the AMR and accesses the N-RTK service to obtain N-RTK positioning results in real time, including position, velocity, attitude angle, positioning status, and accuracy factor information. The accuracy factor information includes PDOP, HDOP, etc., and the positioning status includes Fixed solution, Float solution, Single solution, and None solution.

[0042] Meanwhile, the inertial measurement unit (IMU) on the AMR, such as a six-axis IMU, collects three-axis acceleration and three-axis angular velocity in real time; the relative displacement and velocity information of the AMR is obtained through encoder-type or visual odometry; and environmental point cloud data is collected through a 16-line or 32-line LiDAR at a fixed frequency, such as 10Hz.

[0043] Step S2: Multi-source data preprocessing and time synchronization;

[0044] Online calibration of IMU data is performed, including estimation of bias and scale factor. The bias mean is calculated during the static initialization phase, and high-frequency noise is suppressed during operation using sliding window filtering, such as moving average filtering or low-pass filtering. Gravity compensation is performed by subtracting the component of gravity in the carrier coordinate system from the accelerometer measurements to obtain the linear acceleration of the carrier motion.

[0045] Wheel diameter calibration and system error modeling are performed on the odometer data. For wheeled AMRs, the wheel diameter is calibrated through a closed loop of a known path, and a slippage detection model is established. When slippage occurs, the confidence weight of the odometer data is reduced or outliers are removed.

[0046] Time synchronization employs a combination of hardware and software: the PPS signal output by the GNSS receiver is used as the global time reference, and synchronous sampling of the lidar and IMU is triggered at the rising edge of the PPS signal; for data sources that are not synchronized by hardware, such as odometers, timestamp interpolation is used to align them to the same time reference, ensuring that multi-sensor data achieves millisecond-level synchronization before fusion.

[0047] The laser point cloud is preprocessed, including the removal of outliers using statistical filtering, such as removing points exceeding 3σ based on the distance distribution of neighboring points, and motion distortion compensation: by using the integral of the IMU angular velocity and acceleration between two adjacent laser frames, the pose change of the AMR in that frame is estimated, and motion compensation is performed on each laser point through linear interpolation to eliminate the stretching or distortion of the point cloud caused by the motion of the AMR itself.

[0048] Step S3: Dynamic weight allocation strategy;

[0049] Based on the N-RTK positioning status and accuracy factor, combined with real-time scene features, such as identifying the openness of the environment through laser point cloud geometric features, the weights of each sensor in the fusion process are dynamically adjusted:

[0050] When N-RTK is a fixed solution and the horizontal accuracy factor HDOP ≤ 1.0, it is determined to be an open GNSS area, and N-RTK data is given a fusion weight of ≥ 70%, with laser and IMU / odometer as auxiliary observations;

[0051] When N-RTK is a floating solution or a single-point solution, or when HDOP > 2.0, it is determined to be a semi-obscured area, such as a tree-lined avenue or between buildings, and enters the balanced fusion mode, with the weights of N-RTK, laser and IMU allocated in a 4:3:3 ratio.

[0052] When N-RTK has no solution or the number of consecutive satellite drops exceeds a threshold, such as 5 seconds, it is determined to be a completely obscured area, such as a warehouse or underground passage. It then switches to degradation mode, using the front-end laser odometry and IMU pre-integration as the dominant observation sources, and the N-RTK weight is reduced to 0.

[0053] Step S4: Federated Kalman filter tight-coupled fusion;

[0054] A federated Kalman filter optimization model incorporating N-RTK residuals, laser residuals, and IMU residuals is constructed. The model's functionality primarily consists of two steps: prediction and update. The prediction step mainly involves deriving the next state from the current state using the model. The update step involves updating the model based on the availability of different sensor data. (Further details omitted)

[0055] Step 1: Define the state vector This indicates the system state vector at time k, such as position, velocity, and attitude.

[0056] Specifically, the state vector includes:

[0057] The robot's position in three-dimensional space );

[0058] speed( );

[0059] Attitude angle ( );

[0060] IMU gyroscope zero bias ( );

[0061] IMU accelerometer zero bias ( );

[0062] Odometer scale factor error ( ).

[0063] The state vector is designed as follows:

[0064] [ , .

[0065] Step 2: Define the state equation and obtain the next state based on the current state.

[0066]

[0067] It is a priori estimate of the state at a given time, based on all information prior to that time. That is, the state at time k-1, where f(n) is the nonlinear state transition function. That is, the input vector (such as acceleration, angular velocity, etc.) is controlled at time k.

[0068] The attitude angle, velocity, and position are obtained by integrating the IMU data. Then, the attitude angle, velocity, and position for the next moment are obtained based on the model. However, the data obtained at this time has errors, and the uncertainty increases with time.

[0069] Step 3: Observation update, update based on the predicted value at the previous time and the observed value at the current time.

[0070]

[0071] in It is the deviation of the sensor's real-time observation data. It is the gain that determines whether to trust the predicted value or the observed value more.

[0072] Based on the sensors used, LiDAR and N-RTK data are used to correct the predicted attitude angles, velocities, and positions. N-RTK data can directly provide the current position, attitude, and velocity data. LiDAR data can also be matched with point cloud maps to obtain position and attitude data. However, each type of observation data contains noise, and the confidence level and error vary under different environments. This necessitates the design of the gain K, which determines whether to place more trust in the predicted or observed values. When the N-RTK signal is strong, increasing the gain K utilizes more N-RTK data. When the N-RTK signal is weak, indicating a larger error, decreasing the gain K places more trust in the predicted values. Similarly, when using LiDAR data as observations, increasing the gain when there are many feature points utilizes more LiDAR data, and vice versa. The specific strategy selection is as follows:

[0073] When the N-RTK state is a fixed solution and the accuracy meets the preset threshold, the high-precision position or velocity information calculated by N-RTK is updated as a strong observation.

[0074] When the N-RTK state is a floating solution, a single-point solution, or its accuracy is below a preset threshold but still usable, its position and velocity information are updated as weak observations. Simultaneously, the observation weights of the adaptive laser are determined based on the scene. In this case, the fusion algorithm relies on the prediction model on one hand, and can still utilize N-RTK and laser information for constraints and corrections on the other.

[0075] When the N-RTK state is unsolvable or the signal is completely lost, the state is recursively derived based on prediction, and the lidar data is increased and the gain is increased for strong observation.

[0076] Observation updates are performed channel-wise based on sensor availability:

[0077] N-RTK observation channel: directly provides position and velocity observations, and the noise covariance is dynamically adjusted according to the N-RTK status;

[0078] Laser observation channel: The front-end laser matching pose is used as the observation input. The initial matching value is provided by the current N-RTK fixed solution (if available) or the filter predicted pose, which significantly accelerates the convergence of the Iterative Closest Point (ICP) or Normal Distribution Transform (NDT) matching.

[0079] IMU observation channel: Using IMU pre-integration results as relative motion observations to constrain the continuity of carrier motion.

[0080] The Kalman gain of each channel is adjusted in real time according to the dynamic weight in step S3 to achieve adaptive allocation of sensor trust.

[0081] It outputs real-time, fused, and optimized high-precision AMR position, velocity, and attitude information, along with the confidence level of the positioning results. This provides input for AMR navigation planning, and the fused positioning algorithm has already been applied to inspection and delivery AMRs.

[0082] The state prediction and observation update mechanism of the federated Kalman filter involves two steps: prediction and update. The federated Kalman filter continuously uses new observation data to correct the state estimate, thereby obtaining the optimal state estimate in an uncertain system.

[0083] The prediction step is based on the system model to predict the state and covariance, while the update step uses the observed data to correct these predictions. The Kalman gain determines the weight between the predictions and the observations.

[0084] Step S5: Backend optimization and global consistency maintenance;

[0085] Start a separate backend optimization thread to maintain a sliding window factor graph, with a window length set to 10 to 20 seconds. The factor graph includes:

[0086] IMU pre-integration factor (connecting adjacent keyframes);

[0087] Laser odometry factor (based on laser matching results);

[0088] Historical N-RTK location factors (only high-confidence observations of fixed solutions are retained).

[0089] When the system detects that the N-RTK signal has recovered or entered an open area, it triggers global pose map optimization, which uses historical high-precision N-RTK factors to globally correct the laser sub-map and eliminate drift errors that may accumulate during signal occlusion.

[0090] Step S6: Outputting Location Results and Evaluating Confidence

[0091] The system outputs the fused AMR position, velocity, and attitude information in real time, along with a confidence index C.

[0092]

[0093] in, The posterior covariance matrix is... This represents the observation residual of the i-th sensor, and α and β are normalized weighting coefficients. This confidence level is used in the downstream path planning and decision-making module to trigger deceleration or relocation behavior under high uncertainty.

[0094] Furthermore, the laser matching employs an improved NDT algorithm, which uses the initial pose provided by N-RTK to quickly align the point cloud to the local map before matching, reducing the number of iterations from the conventional 20 to 30 times to 5 to 10 times, and controlling the single-frame matching time to within 30ms.

[0095] Furthermore, the dynamic weighting strategy introduces a laser point cloud feature richness assessment: calculate the number of planar features and edge features of the current frame point cloud. If the number of feature points is lower than a threshold, such as 100, it is determined to be a feature degradation scenario, such as a long corridor or tunnel. The laser observation weight is automatically reduced, and the IMU and odometry weights are increased to over 60%.

[0096] Specifically disclosed, the federated Kalman filter employs a dual-threshold triggering mechanism for N-RTK observations: when N-RTK is a fixed solution and HDOP ≤ 0.8, it is considered a "strong observation" and directly replaces the position and velocity components in the state vector; when HDOP is between 0.8 and 2.0, it is considered a "weak observation" and is only used for error constraints, without directly covering the state estimation.

[0097] Specifically disclosed, the backend optimization thread adopts a keyframe management strategy, inserting new keyframes only when the AMR displacement exceeds 0.5m or the rotation exceeds 10°, to avoid factor graph size expansion. The optimization frequency is set to 1Hz, and it runs in real time on the embedded processor.

[0098] It is particularly important to emphasize that all sensor error parameters (such as IMU zero bias and odometer calibration factor) in the method are estimated and compensated online in real time without prior calibration. The system completes initialization and enters a stable fusion state within the first 1 to 2 minutes of operation.

[0099] It is particularly important to emphasize that the positioning results output by the method can still maintain a positioning error of ≤10cm in a GNSS completely obscured environment, relative to the map, and the system smoothly transitions from obscuration to re-receiving GNSS signals without position jumps, achieving truly seamless positioning.

[0100] Example 1: This example uses a port container handling AGV as an example to illustrate the workflow of the method in alternating open and semi-obstructed environments:

[0101] Data Acquisition: The AGV is equipped with a dual-frequency GNSS receiver, a 32-line lidar, a six-axis IMU, and an encoder odometer, and travels in the port yard.

[0102] Open area positioning: When the AGV travels in the open area of ​​the container yard, N-RTK continuously outputs a fixed solution (HDOP≈0.6). The system assigns 80% weight to N-RTK, and laser and IMU each account for 10%, with a positioning accuracy of 2 to 3 cm.

[0103] Transition to semi-obscured area: When the AGV approaches the container gap or under the crane, the N-RTK state changes to floating solution (HDOP≈2.5), and the system automatically switches to balanced fusion mode (weight allocation 4:3:3). The laser point cloud uses the initial pose for rapid matching to maintain the positioning error ≤8cm.

[0104] Brief full occlusion handling: If the AGV briefly enters the container (GNSS has no solution), the system relies on laser and IMU recursion, while the back-end factor graph records the trajectory of this segment. After it leaves, the accumulated drift is corrected through global optimization.

[0105] Output and Decision: The AGV system outputs the pose with confidence in real time. When the confidence level is lower than the threshold, the AGV control system triggers deceleration or pauses and waits for the positioning to be restored.

[0106] In this scenario, traditional single GNSS solutions can achieve meter-level errors in obstructed areas, while this method achieves centimeter-level positioning throughout the entire process, meeting the high-precision docking requirements of containers.

[0107] Example 2: Warehouse Inspection AMR Dynamic Obstacle Avoidance and Positioning. This example uses an intelligent warehouse inspection robot as an example, combined with the attached... Figure 1 To be continued Figure 7 Detailed workflow description:

[0108] The warehouse environment includes densely shelved areas, glass-partitioned areas, and dynamic pedestrian flow areas, which places high demands on the robustness and real-time performance of the positioning system. The AMR system, equipped with this technology, operates as follows:

[0109] Stable positioning in the shelving area: In the regular shelving aisle, the LiDAR can capture rich planar and corner features. The N-RTK signal is better in the roof skylight area. The system is in a balanced fusion mode. The laser matching adopts the NDT algorithm. The initial pose is provided by the EKF prediction value. The matching time is reduced from the traditional 100ms to less than 30ms, which meets the real-time control requirements of AMR.

[0110] Anti-interference in glass partition area: When passing through the glass curtain wall area, the laser partially penetrates and partially reflects, the point cloud is sparse and contains ghost images. The system detects the point cloud degradation through feature statistics, automatically reduces the laser weight from 0.40 to 0.25, and increases the IMU weight to 0.35. It relies on the IMU to pass through the area with short-term high-precision recursion to avoid positioning jumps.

[0111] Anti-moving targets in dynamic pedestrian flow areas: In areas with frequent pedestrian flow, the laser point cloud contains a large number of dynamic points. The preprocessing module filters out moving target point clouds through multi-frame statistics and motion consistency detection, while retaining static structural features. At the same time, due to frequent switching of N-RTK states caused by signal obstruction, the system quickly switches between equalization mode and degradation mode. The adaptive gain adjustment of the federated EKF maintains trajectory smoothness and eliminates cumulative drift.

[0112] Confidence-guided decision-making: The system outputs positioning confidence in real time. When the confidence drops below 0.5 in areas with complex signals, the navigation module automatically reduces the AMR's speed and triggers a re-initialization process. At cross-regional junctions, the system optimizes and fuses historical high-precision GNSS poses through backend factor graphs to achieve full-coverage positioning. Figure 1 Consistency calibration ensures that the repeatability of the inspection path is better than 5cm.

[0113] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-sensor fusion localization method based on N-RTK applied to AMR, characterized in that, The method includes the following steps: Step S1: Simultaneously collect N-RTK positioning data, LiDAR point cloud data, inertial measurement unit data, and odometry data using multiple sensors mounted on the AMR; Step S2: Preprocess and synchronize the collected multi-source data to obtain a spatiotemporally aligned sensor observation sequence; Step S3: Based on the positioning status and accuracy factor of N-RTK, dynamically adjust the weight allocation strategy of each sensor in the fusion positioning; Step S4: Construct and run a federated Kalman filter, using the preprocessed data and dynamic weighting strategy as input, to perform multi-sensor tight coupling fusion and output high-precision position, velocity and attitude information of the AMR in real time.

2. The N-RTK multi-sensor fusion localization method applied to AMR according to claim 1, characterized in that, Step S2 includes: Step S21: Perform online calibration, sliding window filtering for noise reduction, and gravity compensation on the IMU data to obtain the linear acceleration of the carrier motion; Step S22: Perform wheel diameter calibration and slippage error compensation on the odometer data; Step S23: Using the PPS pulse signal output by the GNSS receiver as the hardware synchronization reference, trigger the lidar and IMU to synchronize sampling; Step S24: Remove outliers and compensate for motion distortion in the lidar point cloud. Motion distortion compensation uses point-by-point linear interpolation of pose transformation obtained by IMU integration between adjacent frames.

3. The N-RTK multi-sensor fusion localization method applied to AMR according to claim 2, characterized in that, In step S3, the dynamic weight allocation strategy adaptively switches according to the N-RTK positioning status and scene features, including: When the N-RTK state is a fixed solution and the accuracy factor is higher than the first threshold, the N-RTK data is assigned the dominant fusion weight. When the N-RTK state is a floating solution or a single-point solution, it enters the balanced fusion mode and simultaneously increases the fusion weight of the lidar and IMU. When the N-RTK state is unsolvable or the signal is lost, it enters the degradation mode and uses the front-end laser odometry and IMU pre-integration results as the main observation sources to perform state recursion.

4. The N-RTK multi-sensor fusion localization method applied to AMR according to claim 1, characterized in that, In step S4, the state vector of the federated Kalman filter includes: the position, velocity, attitude angle of the AMR in three-dimensional space, the zero bias of the IMU gyroscope, the zero bias of the IMU accelerometer, and the odometer calibration factor error.

5. The N-RTK multi-sensor fusion localization method applied to AMR according to claim 4, characterized in that, In the observation update step of the federal Kalman filter, independent observation models and noise covariance matrices are set for different sensors, and the Kalman gain of each observation channel is adjusted in real time according to the dynamic weights determined in step S3.

6. The N-RTK multi-sensor fusion localization method applied to AMR according to claim 5, characterized in that, The observation model of the lidar uses the front-end laser matching pose as the observation input. The initial pose of the laser matching is provided by the N-RTK fixed solution or the filter predicted pose at the current time, so as to accelerate the convergence of the point cloud registration algorithm.

7. The N-RTK multi-sensor fusion localization method applied to AMR according to claim 3, characterized in that, In the balanced fusion mode, the system further fine-tunes the weights based on the feature richness of the laser point cloud: when the environmental geometric features are detected to be sparse or degraded, the weight of laser observation is automatically reduced, and the weights of IMU and odometry are increased accordingly to maintain the robustness of the system.

8. The N-RTK multi-sensor fusion localization method applied to AMR according to claim 1, characterized in that, The method also includes a back-end optimization thread that maintains a sliding window factor graph containing IMU pre-integration factors, laser odometry factors, and historical N-RTK position factors. When the system detects that the N-RTK signal has recovered or entered an open area, global pose graph optimization is initiated to correct the drift error that may accumulate in the federated Kalman filter during signal obstruction.

9. A multi-sensor fusion localization method based on N-RTK applied to AMR according to claim 8, characterized in that, In the global pose graph optimization, the use of historical N-RTK factors adopts a precision-based weighting strategy, retaining only high-confidence observations with fixed positioning states, and using them to construct strong constraints with the laser sub-map to achieve consistency correction of the global trajectory.

10. The N-RTK multi-sensor fusion localization method applied to AMR according to claim 1, characterized in that, The positioning result output by the method includes a real-time confidence index, which is calculated by combining the posterior covariance matrix of the federated Kalman filter with the observation residuals of each sensor.

Images