An unmanned aerial vehicle navigation method and system in low-altitude complex environment

Abstract

This invention discloses a navigation method and system for unmanned aerial vehicles (UAVs) in complex low-altitude environments. The method includes: acquiring multi-source sensor data; identifying and eliminating satellite spoofing signals based on signal fingerprint features, and modeling and compensating for distortions in real signals; dynamically adjusting sensor fusion weights based on environmental perception, constructing a factor graph-optimized SLAM model after spatiotemporal alignment, and introducing motion trend and spatial structure consistency constraints in loop closure detection; constructing a topology-metric hybrid map, combining global topology search and local metric optimization for path planning, outputting control commands using an end-to-end navigation network, and achieving autonomous navigation through trajectory prediction verification and correction; parallel processing of SLAM backend optimization and deep learning inference, and verifying the lightweight compressed model through simulation. This invention improves positioning accuracy, robustness, and autonomous decision-making capabilities, and is suitable for complex environments such as urban canyons, forests, and areas with electromagnetic interference.

Description

Technical Field

[0001] In the field of unmanned aerial vehicle (UAV) navigation technology, specifically, it relates to a navigation method and system for UAVs in complex low-altitude environments. Background Technology

[0002] Unmanned aerial vehicles (UAVs) are increasingly used in military reconnaissance, logistics distribution, agricultural plant protection, and environmental monitoring. Their navigation and positioning capabilities in complex low-altitude environments have become a key factor determining mission success. Complex low-altitude environments generally refer to areas below 1000 meters in altitude, characterized by unfavorable conditions such as tall buildings obstructing views, multipath effects, electromagnetic interference, terrain undulations, and varying lighting conditions. Examples include urban canyons, dense forests, mountainous areas, and regions with complex electromagnetic environments. In these environments, satellite signals are easily blocked or subject to multipath distortion; visual sensors have limited sensing range in low light conditions; lidar may fail due to weather or terrain; and a single navigation method is insufficient to meet the accuracy and reliability requirements for positioning and navigation.

[0003] To address these issues, existing technologies typically employ multi-sensor fusion methods, integrating measurement information from various sensors such as Global Navigation Satellite Systems (GNSS), inertial navigation systems, visual sensors, and lidar to improve the robustness and accuracy of positioning. Simultaneously, simultaneous localization and mapping (SLAM) technology is widely used for autonomous localization and environmental modeling of unmanned aerial vehicles (UAVs) in unknown environments, and deep learning technology has been introduced for target recognition, obstacle detection, and path planning to enhance the environmental understanding and decision-making capabilities of UAVs.

[0004] However, existing technologies still have the following shortcomings: First, spoofed satellite navigation signals exist in low-altitude environments, and traditional filtering methods struggle to effectively distinguish between real and spoofed signals, leading to positioning errors. Second, multi-sensor fusion algorithms often employ fixed weights or simple adaptive rules, making it difficult to quickly adjust sensor confidence levels when the environment changes drastically, thus affecting the fusion effect. Third, loop closure detection in synchronous positioning and mapping is susceptible to interference from factors such as changes in illumination and repetitive structures, resulting in a high false detection rate and increased cumulative error. Fourth, path planning relies on detailed metric maps, which incurs significant computational demands in large-scale complex environments, making it difficult to meet the real-time requirements of airborne systems. Fifth, deep learning navigation often employs a sequential "perception-planning-control" architecture, with errors accumulating at each level and lagging in response to dynamic obstacles.

[0005] Therefore, there is an urgent need for a positioning and navigation method that can achieve high precision, high robustness, and real-time autonomous navigation in complex low-altitude environments. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a navigation method and system for unmanned aerial vehicles in complex low-altitude environments.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This application provides a navigation method for unmanned aerial vehicles in complex low-altitude environments, including the following steps: Data is acquired from multiple sources of sensors, including a global navigation satellite system receiver, an inertial measurement unit, a visual sensor, a lidar, and a geomagnetic sensor. Preprocessing of satellite navigation signals, identification and elimination of deceptive signals based on signal features, distortion modeling and error compensation of the retained real signals, and extraction of effective signals; Based on the current environmental perception results, the fusion weights of each sensor are dynamically adjusted. After spatiotemporal alignment of multi-source sensor data, a synchronous localization and mapping model is constructed, and consistency constraints of motion trend and spatial structure are introduced in loop closure detection. Based on the map constructed by the synchronous positioning and mapping model, a hybrid map combining topology and metric is built. Path planning is carried out by combining global topology search and local metric optimization. Autonomous navigation control is achieved by using an end-to-end navigation network combined with a predicted trajectory verification and correction mechanism. The backend optimization of simultaneous localization and mapping (SMR) and deep learning inference are performed in parallel, and the deep learning model is lightweighted and compressed. The algorithm is verified through a simulation platform.

[0008] Optionally, the step of identifying and eliminating spoofing signals based on signal features includes: extracting fingerprint features of satellite signals, matching the extracted fingerprint features with a pre-stored satellite launch source fingerprint database, and eliminating signals that do not match or whose matching degree is lower than a threshold; the fingerprint features include code phase offset features, carrier phase noise features, and signal angle of arrival variation features.

[0009] Optionally, the distortion modeling and error compensation of the retained real signal includes: establishing a signal distortion and distortion model caused by multipath effects and interference factors, analyzing the code ring phase detection error and ranging deviation caused by autocorrelation function deformation, and realizing error compensation based on signal dynamic parameters.

[0010] Optionally, the dynamic adjustment of the fusion weights of each sensor based on the current environment perception results includes: identifying the current environment type based on visual semantic segmentation and lidar point cloud features, retrieving the corresponding initial sensor confidence template according to the environment type, dynamically fine-tuning the weights of each sensor according to real-time signal quality feedback during the fusion process, and using fuzzy logic control to smooth the weight changes.

[0011] Optionally, the inclusion of consistency constraints between motion trends and spatial structure in loop closure detection includes: requiring that the velocity and angular velocity change trends in the time series of candidate loop closure frames are consistent with historical motion trends, and requiring that the spatial topology be consistent with the current location. Figure 1A loop is considered valid only if it simultaneously satisfies both feature similarity and spatiotemporal consistency.

[0012] Optionally, the construction of the hybrid map combining topology and metric includes: constructing a topological map at a global scale, with key regions as nodes and connectivity as edges; constructing a fine-grained metric map at a local scale; the topological nodes are dynamically updated based on environmental semantic information, and the topological nodes and their connectivity are automatically segmented and updated when new obstacles or regional changes are detected.

[0013] Optionally, the autonomous navigation control achieved by combining an end-to-end navigation network with a predicted trajectory verification and correction mechanism includes: constructing an end-to-end navigation network, inputting the fusion features of multimodal perception data, and outputting control commands for the unmanned aerial vehicle; the network simultaneously outputs predicted trajectories for multiple future frames, compares the predicted trajectories with the real-time positioning results of synchronous positioning and mapping, and triggers a correction module to adjust network parameters online or revert to a safe strategy when the deviation exceeds a threshold.

[0014] Optionally, the end-to-end navigation network employs a lightweight deep neural network and incorporates extreme perturbations and dynamic obstacles through adversarial training in a simulation environment to enhance generalization capabilities.

[0015] Optionally, the lightweight compression of the deep learning model includes employing one or more combinations of model pruning, quantization, and knowledge distillation techniques, and deploying the optimized model on an embedded platform to run in parallel with a synchronous localization and mapping backend.

[0016] Secondly, this application provides a navigation system for unmanned aerial vehicles in complex low-altitude environments, including: Multi-source sensor module for acquiring data from global navigation satellite system, inertial measurement unit, visual sensor, lidar and geomagnetic sensor; The signal preprocessing module is used to identify and remove spoofing signals from satellite navigation signals, and to perform distortion modeling and error compensation on real signals. The environmental perception and fusion mapping module is used to dynamically adjust the fusion weights of each sensor based on the environmental perception results, construct a synchronous localization and mapping model after spatiotemporal alignment of multi-source data, and introduce consistency constraints of motion trend and spatial structure in loop closure detection. The navigation planning and control module is used to construct a hybrid map that combines topology and metric. It performs path planning by combining global topology search and local metric optimization, and achieves autonomous navigation control by using an end-to-end navigation network combined with a predicted trajectory verification and correction mechanism. The real-time optimization and verification module is used to perform parallel processing of synchronous localization and mapping backend optimization and deep learning inference, perform lightweight compression of deep learning models, and verify the algorithm through a simulation platform.

[0017] Optionally, the navigation planning and control module includes a prediction-verification-correction unit, which compares the predicted trajectory output by the end-to-end navigation network with the real-time positioning results of synchronous positioning and mapping, and triggers online fine-tuning of network parameters or rollback of safety policies when the deviation exceeds a threshold.

[0018] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the unmanned aerial vehicle positioning and navigation method described above.

[0019] Compared with the prior art, this application has the following beneficial effects: This invention provides a navigation method and system for unmanned aerial vehicles (UAVs) in complex low-altitude environments. It utilizes signal fingerprinting technology to eliminate deceptive signals at the physical layer, significantly improving anti-spoofing capabilities and ensuring positioning security. An environment-adaptive multi-sensor fusion weight adjustment mechanism enables the system to optimize sensor confidence in real-time according to environmental changes, improving fusion accuracy and robustness. A loop closure detection mechanism with spatiotemporal consistency constraints is introduced to reduce false detection rates and effectively suppress accumulated errors in highly dynamic or repetitive scenarios. A topology-metric hybrid map is constructed for path planning, balancing global search efficiency with precise local obstacle avoidance, significantly improving planning speed and meeting airborne real-time requirements. An end-to-end navigation network combined with a prediction-verification-correction closed-loop mechanism achieves rapid response and self-correction capabilities, adapting to highly dynamic environments. Parallel processing and lightweight model optimization optimize system real-time performance, ensuring efficient algorithm operation on embedded platforms and demonstrating engineering practicality. Attached Figure Description

[0020] Figure 1 This schematic diagram illustrates the error in satellite signal propagation in an urban canyon environment in an exemplary embodiment of the present disclosure.

[0021] Figure 2 The schematic diagram illustrates a research approach in an exemplary embodiment of this disclosure.

[0022] Figure 3 This schematic diagram illustrates the overall technical route in an exemplary embodiment of the present disclosure. Detailed Implementation

[0023] 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.

[0024] Furthermore, in this invention, an element referred to as fixed to or disposed on another element may be directly disposed on the other element, or there may be an intermediate element. When an element is considered to be connected to another element, it may be directly connected to the other element, or there may be an intermediate element present simultaneously. The terms vertical, horizontal, left, right, and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0025] Example 1: Navigation method for unmanned aerial vehicles in complex low-altitude environments See Figures 1-3 As shown: This embodiment provides a navigation method for unmanned aerial vehicles in complex low-altitude environments. The specific process is as follows.

[0026] First, multi-source sensor data acquisition is performed. The UAV is equipped with a GNSS receiver, an inertial measurement unit (IMU), visual sensors (monocular / binocular / RGB-D camera), a lidar (LiDAR) system, and a geomagnetic sensor. All sensors acquire data synchronously via hardware, ensuring data timestamp alignment accuracy at the microsecond level. The acquired data includes: GNSS pseudorange and carrier phase observations; IMU three-axis acceleration and three-axis angular velocity; visual image frames; lidar point clouds; and geomagnetic three-axis field strength.

[0027] Next, the satellite navigation signal undergoes preprocessing and spoofing signal identification. The raw signal output from the GNSS receiver first enters the signal preprocessing module. This module extracts the fingerprint features of each satellite signal, including code phase offset features, carrier phase noise features, and signal angle of arrival variation features. The code phase offset feature is extracted by performing correlation operations between the received signal and the local pseudocode to obtain the correlation peak waveform, and extracting the rising slope, falling slope, and peak width parameters of the peak value to construct a feature vector. The carrier phase noise feature extraction method is as follows: Differential processing is performed on the carrier phase observations to obtain the carrier phase noise sequence. Its variance, skewness, kurtosis, and other statistical characteristics are calculated to form a feature vector. The method for extracting signal angle of arrival (AHA) variation features is as follows: The carrier phase change caused by the movement of the receiving antenna array or a single antenna is used to estimate the signal AHA sequence, calculate its rate of change characteristics, and construct a feature vector. The above features are fused into a comprehensive fingerprint feature vector. The current fingerprint features are matched against a pre-established fingerprint database of satellite launch sources. This database is constructed by collecting real satellite signals from different locations and times and extracting features; each satellite has a unique fingerprint template. The matching degree is calculated using the Pearson correlation coefficient. like (For example If the signal value is less than or equal to 0.9, it is considered a deceptive signal and is removed. For genuine signals that match successfully, their distortion and deformation are further analyzed. A signal distortion model caused by multipath effects and interference factors is established, and the deformation of the autocorrelation function (ACF) can be expressed as: in For an ideal autocorrelation function, For multipath reflection coefficient, For multipath delay, For the number of multipaths, This is noise. The code ring phase detection error can be approximated as: in This is the phase detector function. It is based on the signal dynamic parameters (carrier-to-noise ratio). Doppler frequency To compensate for the error and restore the effective ranging information: Based on this, environmental adaptive multi-sensor fusion and SLAM modeling are performed; the preprocessed sensor data enters the environmental perception and fusion mapping module; firstly, the environment classifier identifies the current environment type in real time based on visual semantic segmentation (such as DeepLabV3+) and LiDAR point cloud features (such as PointNet++). ∈ {Urban canyons, forests, electromagnetic interference zones,} Each environment type has a preset initial sensor confidence template. During the fusion process, the weights of each sensor are dynamically fine-tuned based on real-time signal quality feedback. Signal quality metrics include: GNSS carrier-to-noise ratio. Number of visual feature points Laser point cloud matching residual IMU zero bias stability The adaptive weight adjustment employs fuzzy logic control, defining a fuzzy set of input variables (e.g., "high / medium / low carrier-to-noise ratio"), and outputting the weight adjustment amount. Example of a fuzzy rule: If Gao Ze Upright; if low Large negative impact; if the environment type is urban canyon, then increase. and The final fusion weight is And weighted least squares fusion is used: Then, spatiotemporal alignment is performed; hardware timestamps and spatial calibration parameters (camera-IMU extrinsics) are used. LiDAR-IMU external parameters This unifies all data to the same spatiotemporal coordinate system. Time alignment uses interpolation to interpolate observations from each sensor to a common time reference. A SLAM model is constructed based on a factor graph, integrating GNSS, IMU pre-integration, visual features, laser point clouds, and geomagnetic observations. The factor graph consists of variable nodes and factor nodes: variable nodes are state vectors. express Position, velocity, attitude (quaternion), accelerometer bias, gyroscope bias at any given time; factor nodes include IMU pre-integration factors (connecting adjacent state nodes, residual is...) GNSS factor (when the GNSS signal is valid, the residual is...) Visual factors (based on reprojection error) ), LiDAR factor (based on point cloud registration residual) Geomagnetic factor (based on geomagnetic field strength residual) The objective function of factor graph optimization is to maximize the joint probability of all factors, which is equivalent to minimizing the negative log-likelihood. The Gauss-Newton or Levenberg-Marquardt method is used for iterative solution, and the sparse matrix property is utilized to accelerate the calculation.

[0028] Spatiotemporal consistency constraints are introduced into loop closure detection. When a candidate loop closure frame is detected, not only feature similarity (visual bag-of-words model score) is calculated... Laser point cloud NDT registration score Furthermore, it requires consistency in motion trends and spatial topology. Motion trend consistency requires that the velocity and angular velocity trends of the candidate frame and the current frame in the time series be consistent with the historical motion trends. Let the historical motion trend sequence be... The motion trend of candidate frames is ,Require ,in These are historical trend predictions. Spatial topological consistency requires that the relative positions of adjacent landmarks be consistent with the current situation. Figure 1 Let the set of landmarks surrounding the current frame be defined. Landmark set around candidate frames Calculate topological similarity Overall similarity Only when When the consistency of the movement trend is satisfied, it is confirmed as a valid loop and added to the factor graph for optimization.

[0029] Based on this, autonomous navigation and path planning are performed; for the map built based on SLAM, a hybrid topology-metric map is first constructed; at the global scale, the map is divided into several key areas (intersections, near landmarks) as topology nodes. The connectivity between nodes constitutes topological edges. At the local scale, a detailed metric map (occupancy grid map or ESDF map) is maintained around each node. Topology nodes are dynamically updated based on environmental semantic information. When visual SLAM detects new obstacles or changes in the region, it automatically segments and updates the topology nodes and their connectivity relationships, following the following update rules: Path planning is divided into two layers: first, a global topology search is performed, and then... The A* algorithm is used to search for the starting point. To the finish line The node sequence is determined; then local metric optimization is performed between adjacent nodes, using a trajectory planning algorithm that considers optimal energy consumption and obstacle avoidance; and the local path is parameterized based on B-spline curves. ,in Let p be the B-spline basis functions. The control points are defined as follows; the optimization objective is to minimize energy consumption and path length. Constraints include obstacle avoidance constraints ( ), dynamic constraints Endpoint constraints The solution is obtained by using Sequential Quadratic Programming (SQP).

[0030] Autonomous navigation employs an end-to-end navigation network combined with a prediction-verification-correction mechanism. The end-to-end navigation network is a lightweight deep neural network (such as MobileNetV3 with fully connected layers), which takes as input the fused features of multimodal perception data (including images, point clouds, IMU, current pose, etc.) and directly outputs the aircraft's velocity and heading angle commands. The network structure includes: a visual encoder (MobileNetV3 extracts image features). Point cloud encoder (PointNet++ for extracting point cloud features) IMU encoder (LSTM processes IMU sequences to extract motion features) Feature fusion layer (the features are concatenated and passed through a fully connected layer to obtain fused features) Output layer (two fully connected connectors output speed respectively) and heading angle The network simultaneously outputs the predicted trajectory for the next N frames (e.g., N=10). The predicted trajectory is compared with the real-time SLAM localization results to calculate the trajectory deviation. ;like The correction module is triggered to fine-tune the parameters of the last few layers of the network online (based on recent experience replay) or fall back to a safe strategy (such as hovering or returning along the original path); the network is subjected to extreme perturbations and dynamic obstacles through adversarial training in a simulation environment to improve its generalization ability. The objective function of adversarial training is... .

[0031] Finally, real-time system optimization and verification were performed. To meet airborne real-time requirements, SLAM backend optimization and deep learning inference were executed in parallel. The SLAM backend ran on a CPU thread, while deep learning inference ran on a GPU or NPU thread, exchanging data through shared memory. The deep learning model was lightweighted and compressed, including model pruning (removing redundant channels), quantization (quantizing FP32 precision to INT8 precision), and knowledge distillation (using a large model to guide the training of a small model), reducing the model size to less than 1 / 5 of its original size and improving inference speed by more than 3 times. The optimized model was deployed on an embedded platform (such as Jetson OrinNX). The algorithm was verified through digital simulation (MATLAB / Simulink, Gazebo) and hardware-in-the-loop simulation (a hardware-in-the-loop platform equipped with Pixhawk flight controller and Jetson board), evaluating indicators such as positioning accuracy, navigation success rate, and real-time performance.

[0032] Example 2: Navigation System for Unmanned Aerial Vehicles in Complex Low-Altitude Environments This embodiment provides a navigation system for unmanned aerial vehicles (UAVs) in complex low-altitude environments, including a multi-source sensor module, a signal preprocessing module, an environmental perception and fusion mapping module, a navigation planning and control module, and a real-time optimization and verification module. The multi-source sensor module includes a GNSS receiver, IMU, visual camera, LiDAR, and magnetometer. Data acquisition is triggered by a synchronization signal, and the sensors are connected via a hardware synchronization line to ensure data timestamp alignment. The signal preprocessing module identifies and removes spoofing signals, models and compensates for distortion, and outputs clean sensor observations. It includes a fingerprint feature extraction unit, a fingerprint matching unit, and a distortion compensation unit. The environmental perception and fusion mapping module performs environmental classification, adaptive weight fusion, spatiotemporal alignment, SLAM modeling, and loop closure detection. It includes an environmental classifier, a weight adaptive unit, a factor graph optimization engine, and a loop closure detection unit. The navigation planning and control module constructs a hybrid map, plans the path, and performs end-to-end navigation control. It includes a prediction-verification-correction unit, which compares the predicted trajectory output by the end-to-end navigation network with the real-time SLAM positioning results. When the deviation exceeds a threshold, it triggers online fine-tuning of network parameters or a safety strategy rollback. The real-time optimization and verification module is responsible for parallel scheduling, model lightweighting, and simulation verification, and includes a parallel computing scheduler, a model compression unit, and a simulation interface unit. Modules communicate with each other via ROS2 middleware, supporting modular replacement and expansion. The system workflow is as follows: the multi-source sensor module acquires raw data; the signal preprocessing module processes GNSS signals and removes deception interference; the environmental perception and fusion mapping module performs state estimation and environmental mapping; the navigation planning and control module generates paths and control commands; and the real-time optimization and verification module monitors and optimizes system performance.

[0033] Example 3: Computer-readable storage medium This embodiment provides a computer-readable storage medium storing a computer program. When executed by a processor, this program implements the unmanned aerial vehicle (UAV) positioning and navigation method described in Embodiment 1. The computer-readable storage medium includes, but is not limited to, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. When the program is executed by the processor, it performs the following functions: multi-source sensor data acquisition and preprocessing, spoofing signal identification and removal, multi-sensor adaptive fusion and SLAM mapping, topology-metric hybrid map construction and path planning, end-to-end navigation control and prediction-verification-correction closed loop, and system real-time optimization and verification.

[0034] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A navigation method for unmanned aerial vehicles in complex low-altitude environments, characterized in that, Includes the following steps: Data is acquired from multiple sources of sensors, including a global navigation satellite system receiver, an inertial measurement unit, a visual sensor, a lidar, and a geomagnetic sensor. Preprocessing of satellite navigation signals, identification and elimination of deceptive signals based on signal features, distortion modeling and error compensation of the retained real signals, and extraction of effective signals; Based on the current environmental perception results, the fusion weights of each sensor are dynamically adjusted. After spatiotemporal alignment of multi-source sensor data, a synchronous localization and mapping model is constructed, and consistency constraints of motion trend and spatial structure are introduced in loop closure detection. Based on the map constructed by the synchronous positioning and mapping model, a hybrid map combining topology and metric is built. Path planning is carried out by combining global topology search and local metric optimization. Autonomous navigation control is achieved by using an end-to-end navigation network combined with a predicted trajectory verification and correction mechanism. The backend optimization of simultaneous localization and mapping (SMR) and deep learning inference are performed in parallel, and the deep learning model is lightweight and compressed. The algorithm is verified through a simulation platform.

2. The method according to claim 1, characterized in that, The method of identifying and eliminating spoofing signals based on signal features includes: extracting fingerprint features of satellite signals, matching the extracted fingerprint features with a pre-stored satellite launch source fingerprint database, and eliminating signals that do not match or whose matching degree is lower than a threshold; the fingerprint features include code phase offset features, carrier phase noise features, and signal angle of arrival variation features.

3. The method according to claim 1, characterized in that, The distortion modeling and error compensation of the preserved real signal includes: establishing a signal distortion and distortion model caused by multipath effects and interference factors, analyzing the code ring phase detection error and ranging deviation caused by autocorrelation function deformation, and realizing error compensation based on signal dynamic parameters.

4. The method according to claim 1, characterized in that, The dynamic adjustment of the fusion weights of each sensor based on the current environment perception results includes: identifying the current environment type based on visual semantic segmentation and lidar point cloud features; retrieving the corresponding initial sensor confidence template according to the environment type; dynamically fine-tuning the weights of each sensor according to real-time signal quality feedback during the fusion process; and using fuzzy logic control to smooth the weight changes.

5. The method according to claim 1, characterized in that, The inclusion of consistency constraints between motion trends and spatial structure in loop closure detection includes: in loop closure candidate frames, the trends of velocity and angular velocity changes in the time series are required to be consistent with historical motion trends, and the spatial topology is required to be consistent with the current map. Only when both feature similarity and spatiotemporal consistency are satisfied is it confirmed as a valid loop closure.

6. The method according to claim 1, characterized in that, The construction of the hybrid map combining topology and metric includes: constructing a topological map at a global scale, with key regions as nodes and connectivity as edges; constructing a fine-grained metric map at a local scale; the topological nodes are dynamically updated based on environmental semantic information, and the topological nodes and their connectivity are automatically segmented and updated when new obstacles or regional changes are detected.

7. The method according to claim 1, characterized in that, The method of achieving autonomous navigation control by combining an end-to-end navigation network with a predicted trajectory verification and correction mechanism includes: constructing an end-to-end navigation network, inputting the fusion features of multimodal perception data, and outputting control commands for the unmanned aerial vehicle; the network simultaneously outputs predicted trajectories for multiple future frames, compares the predicted trajectories with the real-time positioning results of synchronous positioning and mapping, and triggers a correction module to adjust network parameters online or revert to a safe strategy when the deviation exceeds a threshold.

8. The method according to claim 7, characterized in that, The end-to-end navigation network employs a lightweight deep neural network and incorporates extreme perturbations and dynamic obstacles through adversarial training in a simulation environment to enhance its generalization ability.

9. The method according to claim 1, characterized in that, The lightweight compression of the deep learning model includes one or more combinations of model pruning, quantization, and knowledge distillation techniques, and the optimized model is deployed on an embedded platform to run in parallel with the synchronous localization and mapping backend.

10. A navigation system for unmanned aerial vehicles in complex low-altitude environments, characterized in that, include: Multi-source sensor module for acquiring data from global navigation satellite system, inertial measurement unit, visual sensor, lidar and geomagnetic sensor; The signal preprocessing module is used to identify and remove spoofing signals from satellite navigation signals, and to perform distortion modeling and error compensation on real signals. The environmental perception and fusion mapping module is used to dynamically adjust the fusion weights of each sensor based on the environmental perception results, construct a synchronous localization and mapping model after spatiotemporal alignment of multi-source data, and introduce consistency constraints of motion trend and spatial structure in loop closure detection. The navigation planning and control module is used to construct a hybrid map that combines topology and metric. It performs path planning by combining global topology search and local metric optimization, and achieves autonomous navigation control by using an end-to-end navigation network combined with a predicted trajectory verification and correction mechanism. The real-time optimization and verification module is used to perform parallel processing of synchronous localization and mapping backend optimization and deep learning inference, perform lightweight compression of deep learning models, and verify the algorithm through a simulation platform.

Images