Unmanned aerial vehicle positioning method and system in combination with environment characteristics

Abstract

The invention relates to the technical field of unmanned aerial vehicle positioning, in particular to an unmanned aerial vehicle positioning method and system combining environment characteristics. According to the technical scheme, the method comprises the steps that multi-modal data of the environment where the unmanned aerial vehicle is located are collected through a multi-source heterogeneous sensor, and the multi-modal data comprise GNSS signal data, visual image data, laser radar point cloud data, inertia measurement data and electromagnetic interference monitoring data; and carrying out cooperative processing on the multi-modal data by utilizing the fusion deep learning model. A comprehensive basis is provided for subsequent processing through multi-source data collection, data quality and feature matching reliability are improved through fusion processing, timely and accurate positioning is guaranteed through real-time calculation, safe and efficient flight is achieved through a decision-making model based on accurate positioning, and the flight speed is high. The method not only solves the traditional technical problems of signal shielding, interference, dynamic interferent influence and contradiction between real-time performance and precision in a complex environment, but also shows excellent positioning stability and adaptability in various complex scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle positioning, and particularly relates to an unmanned aerial vehicle positioning method and system combined with environmental features. BACKGROUND

[0002] In modern unmanned aerial vehicle applications, positioning technology is one of the core elements to ensure the safe and efficient operation of unmanned aerial vehicles. Traditional unmanned aerial vehicle positioning methods, such as global satellite navigation systems (GNSS), perform well in open environments, but in complex scenarios such as urban canyons, indoors, forests, etc., signals are easily blocked and interfered, leading to a decrease in positioning accuracy or even positioning failure. To address these challenges, researchers have begun to explore methods that combine environmental features for positioning, such as using the acceleration and angular velocity information of an inertial measurement unit (IMU) to calculate the position change of the unmanned aerial vehicle, extracting feature points in images through a vision sensor for visual positioning, or using a laser radar to obtain three-dimensional point cloud information of the surrounding environment for positioning. However, these positioning methods based on environmental features still face many outstanding problems in practical applications, and the deep involvement of artificial intelligence technology is urgently needed.

[0003] Signal blocking and interference is a major challenge faced by existing positioning methods. In urban areas, high-rise buildings are densely populated, and GNSS signals are easily blocked by buildings, resulting in signal interruption or multipath effects, which can cause positioning deviations. At the same time, electromagnetic interference from electronic devices, communication base stations, etc. can also seriously affect the quality of positioning signals. Although some auxiliary positioning technologies, such as inertial navigation, can maintain positioning for a short period of time, the cumulative error will quickly increase over time, causing the positioning result to deviate from the true position. Deep learning technologies in artificial intelligence, such as convolutional neural networks (CNN) and recurrent neural networks (RNN), have the potential to learn from a large amount of data containing signal blocking and interference situations, and to establish an accurate signal feature model. Such a model can intelligently analyze and repair signals that are interfered and blocked, and predict the true state of the signal, thereby improving the accuracy and stability of positioning.

[0004] Feature extraction and matching in complex environments is also a key problem that needs to be solved. UAVs fly in different environments, such as cities, forests, and mountains, with complex and variable environmental features. In visual positioning, rapid changes in light, object occlusion, and similar scene interference can all lead to inaccurate image feature extraction, which in turn affects positioning accuracy. Traditional feature extraction algorithms, such as Scale-Invariant Feature Transform (SIFT) and Speeded-Up Robust Features (SURF), lack robustness in complex environments. By leveraging the powerful feature learning capabilities of deep learning, a feature extraction model can be trained specifically for complex environments. For example, a method based on Generative Adversarial Networks (GAN) can generate simulated images in various complex environments to train the feature extraction model, enabling it to better adapt to complex changes in real-world scenarios, accurately extract representative environmental features, and improve feature matching accuracy, thereby enhancing the positioning capabilities of UAVs in complex environments.

[0005] Real-time and intelligent decision-making are also weaknesses of current UAV positioning technology. During flight, UAVs need to make quick decisions based on real-time positioning information, such as obstacle avoidance and path planning. However, existing positioning algorithms have high computational complexity when processing large amounts of data, resulting in delays in the output of positioning results, which cannot meet the needs of real-time flight of UAVs. In particular, in high-speed flight or rapidly changing environments, delayed positioning information may not allow UAVs to respond in time to unexpected situations, leading to safety accidents. Therefore, we propose a UAV positioning method and system that combines environmental features. SUMMARY

[0006] The purpose of the present invention is to address the problems in the background art and propose a UAV positioning method and system that combines environmental features.

[0007] In a first aspect, the present application provides a UAV positioning method that combines environmental features, comprising the following steps:

[0008] Step one: Collecting multi-modal data of the environment where the UAV is located through multi-source heterogeneous sensors, including GNSS signal data, visual image data, laser radar point cloud data, inertial measurement data, and electromagnetic interference monitoring data;

[0009] Step two: Using a fusion deep learning model to cooperatively process the multi-modal data, which includes intelligently repairing occluded and interfered GNSS signals, extracting three-dimensional features with environmental specificity based on enhanced visual image data and laser radar point cloud data using a generative adversarial network, and dynamically matching them, and constructing a spatio-temporal correlation model combining inertial measurement data and electromagnetic interference monitoring data;

[0010] Step three: Based on the collaborative processing results, real-time positioning calculation is performed through a spatio-temporal sequence prediction model to obtain the accurate position and attitude information of the UAV. A lightweight network architecture is adopted in the real-time positioning calculation process to reduce the calculation delay.

[0011] Step four: According to the real-time acquired position and attitude information, the flight path and obstacle avoidance strategy of the UAV are dynamically adjusted through a reinforcement learning decision model to realize safe and efficient flight of the UAV in complex environments.

[0012] Optionally, the multi-source heterogeneous sensors in step one include a GNSS receiver, a high-definition camera, a laser radar, an inertial measurement unit (IMU), and an electromagnetic interference detector. The high-definition camera adopts a binocular vision camera with a baseline distance of 12 cm, a resolution of 1920x1080, and a frame rate of 30 fps. The laser radar adopts a 32-line laser radar with a horizontal field of view of 120°, a vertical field of view of -15° to +15°, and a sampling frequency of 15 Hz. The inertial measurement unit adopts a six-axis IMU (3-axis acceleration + 3-axis angular velocity) with an acceleration measurement range of ±16g, an angular velocity measurement range of ±2000° / s, and a sampling frequency of 200 Hz. The electromagnetic interference detector has a working frequency band covering 100MHz-6GHz and a detection sensitivity of ≤-90dBm.

[0013] Optionally, the intelligent repair of the blocked and interfered GNSS signals in step two specifically includes:

[0014] A large number of GNSS signal samples containing different blocking and interference types are collected to construct a signal repair training dataset. GNSS signal samples containing blocking rates of 0%-100% and interference intensities of -80dBm to -40dBm are collected to construct a signal repair training dataset containing more than 1 million samples.

[0015] A convolutional neural network (CNN)-recurrent neural network (RNN) hybrid model is used to construct a signal repair model. The convolutional neural network (CNN) is used to extract the spatial features of the signal, and the recurrent neural network (RNN) is used to capture the time sequence features of the signal. The convolutional neural network (CNN) layer adopts 3 layers of convolution (convolution kernel size is 3x3, 5x5, and 3x3 respectively, and step is 1) to extract signal spatial features, and the recurrent neural network (RNN) layer adopts 2 layers of LSTM (hidden unit number is 128) to capture signal time sequence features.

[0016] The model training loss function is:

[0017] L = a·MSE(x true , x pred ) + β·MAE(x true , x pred )

[0018] wherein, x true is the original interference-free signal, x pred is the repaired signal, a = 0.7, b = 0.3 are weight coefficients, MSE is the mean square error, and MAE is the mean absolute error; the interference signal is input into the trained model, and the repaired signal is output, and the repair error is less than or equal to 0.5m (plane position);

[0019] The signal repair model is trained using the training data set, so that the model can learn the distortion law of the signal under different occlusions and interferences;

[0020] The GNSS signal subjected to occlusion and interference is input into the trained signal repair model to obtain the repaired GNSS signal.

[0021] Optionally, the specific steps of extracting the complex environment features from the image data and the laser point cloud data and matching in step two include:

[0022] The generative adversarial network (GAN) is used to generate simulated image data and laser point cloud data under multiple complex environments, including scenes with different light intensities, different weather conditions, and different occlusion degrees; when the generative adversarial network (GAN) generates simulated data, the light intensity range is set to 500-10000 lux, and the occlusion area ratio is 0%-60%;

[0023] The simulated data is fused with the actually collected image data and laser point cloud data to construct a feature extraction and matching training data set;

[0024] The convolutional neural network (CNN) based on the attention mechanism is used as the feature extraction model, which can automatically focus on the key feature regions in the image and the point cloud;

[0025] The feature extraction model is trained using the training data set, so that the model can extract visual features with scale invariance and rotation invariance from the image data, and extract three-dimensional features with geometric topological structure from the laser point cloud data; the convolutional neural network (CNN) based on the attention mechanism includes 4 convolutional layers and 2 attention modules, the visual feature dimension is 256, and the laser point cloud feature dimension is 512;

[0026] The improved k-nearest neighbor algorithm is used to match the extracted visual features and three-dimensional features, and the best matching pair is determined by calculating the similarity between the features; the feature matching uses the improved k-NN algorithm, and the similarity calculation uses the cosine distance formula:

[0027]

[0028] Wherein, f1, f2 are the feature vectors to be matched, · is the vector inner product, ‖‖·‖‖ is the L2 norm, and when S < 0.3, it is determined that the matching is successful.

[0029] In the GAN-generated simulation data stage, a dynamic interference object generation module is added to randomly add image masks and corresponding point cloud noise of simulated flying birds, falling leaves and moving vehicle dynamic objects in the scene, the motion trajectory of the dynamic interference object conforms to the Brownian motion model, the speed range is set to 0.5-15 m / s, and the size ratio is 1%-10% of the image area.

[0030] A dynamic region suppression loss function is introduced to the feature extraction CNN model:

[0031]

[0032] Wherein, L dynamic is the calculation result of the dynamic region suppression loss function, which is used to measure the suppression degree of the model in response to the dynamic region features when facing dynamic interference objects; λ = 0.2 is the suppression coefficient; N is the number of samples participating in the calculation; i is the sample index; I i is the image data input to the feature extraction CNN model; M d is the dynamic interference object mask region; ⊙ is the element-by-element multiplication operation; f(·) is the output function of the feature extraction CNN model; ‖·‖2 is the L2 norm; and the model is forced to ignore the feature response of the dynamic object region;

[0033] After improving the k-NN preliminary matching, a space-time consistency verification layer is added:

[0034] Spatial verification: using the spatial topological relationship of consecutive frames of laser radar point cloud, a local surface descriptor (ISS key point + FPFH descriptor) of the feature point is constructed, and when the descriptor difference threshold between adjacent frames is > 0.25, it is determined that the matching is incorrect.

[0035] Time verification: the theoretical displacement of the feature point is calculated through the inertial measurement unit (IMU) data, and if the actual displacement of the matching point deviates from the theoretical value by > 3σ, σ is the noise variance of the IMU, then the matching pair is removed.

[0036] Optionally, the step two of constructing the space-time correlation model in combination with the inertial measurement data and the electromagnetic interference monitoring data specifically comprises:

[0037] The inertial measurement data is preprocessed, including removing noise and zero bias correction;

[0038] The correlation between the electromagnetic interference monitoring data and the GNSS signal quality and the inertial measurement error is analyzed, and an electromagnetic interference influence model is established;

[0039] Based on Kalman filtering algorithm, the repaired GNSS signal, the extracted environment feature matching result, the preprocessed inertial measurement data and the electromagnetic interference influence model are fused to construct a space-time correlation model, which can describe the relationship between the motion state of the unmanned aerial vehicle in time and space and the environment features.

[0040] The space-time correlation model is constructed by using extended Kalman filter (EKF), and the state equation is:

[0041] X k =F k ·X k-1 +B k ·U k +W k

[0042] The observation equation is:

[0043] Z k =H k ·X k +V k

[0044] Wherein: X k =[x,y,z,v x ,v y ,v z ,φ,θ,ψ] T is the state vector at time k, which represents three-dimensional position (x, y, z), three-dimensional velocity (v x , v y , v z ) and three-dimensional attitude angle (roll angle φ, pitch angle θ, heading angle ψ) in turn; F k is an 8x8 state transition matrix; B k is a control input matrix; U k is the IMU measured acceleration; W k is the process noise, and the variance is set to 0.01; Z k is the observation vector at time k; H k is the observation matrix; V k is the observation noise, and the variance is set to 0.1; H k is the observation matrix, which fuses GNSS and feature matching results.

[0045] Optionally, the specific steps of using the lightweight network architecture for real-time positioning calculation in step three include:

[0046] The feature data output by the space-time correlation model is processed by dimension reduction to reduce the data amount;

[0047] MobileNetV3 (parameter quantity ≤ 300 million) is used as the basic architecture of the positioning solution network, and the network structure is further optimized through pruning and quantization techniques to reduce the computational complexity.

[0048] The reduced dimension feature data is input into the lightweight positioning solution network, and the network outputs the position coordinates, speed and attitude angle of the UAV.

[0049] The positioning solution result is checked in real time, and if the result exceeds the preset error range, the repositioning mechanism is triggered.

[0050] The positioning solution error formula is:

[0051]

[0052] Where e is the three-dimensional positioning error, (x pred , y pred , z pred ) is the three-dimensional position prediction value of the UAV output by the model; (x true , y true , z true ) is the actual three-dimensional position true value of the UAV; e ≤ 0.8 m (95% confidence) is required; the real-time index is the processing delay ≤ 50 ms (single frame data).

[0053] Optionally, the specific steps of adjusting the flight path and attitude of the UAV through the reinforcement learning decision model in step four include:

[0054] The real-time position, speed, attitude of the UAV and the obstacle information around the environment are taken as state inputs;

[0055] The distance, flight time and energy consumption of the UAV to safely reach the target point are taken as parameters of the reward function;

[0056] A deep Q network (DQN) in the deep reinforcement learning algorithm is used to construct an intelligent decision model, which continuously learns the optimal flight strategy through interaction with the environment;

[0057] The real-time positioning result and environmental perception data are input into the intelligent decision model, and the model outputs the flight path adjustment amount and attitude control instruction of the UAV, and the UAV executes the corresponding instruction to realize obstacle avoidance and path optimization; the reward function of the reinforcement learning decision model is:

[0058]

[0059] Wherein, d is the straight-line distance from the current position of the UAV to the target point; D is the straight-line distance from the initial position to the target point; t is the time already flown; T is the total flight time expected; e is the energy currently consumed; E is the maximum allowed energy consumed; ω1=0.5, ω2=0.3, ω3=0.2 are weight coefficients of each reward item;

[0060] The DQN network adopts a double Q network structure, the experience replay pool capacity is 100,000, the learning rate is set to 0.001, and the discount factor γ=0.9.

[0061] Optionally, when training the signal repair model, a transfer learning method is used, the model is pre-trained using a public GNSS signal dataset, and then fine-tuned using a private dataset containing specific scene occlusion and interference features to improve the repair accuracy of the model in a specific application scenario; when transfer learning, pre-training uses a public dataset (NOAA GNSS multipath dataset), fine-tuning uses a private dataset of urban canyon / forest scenes (each containing 50,000+ samples), the fine-tuning learning rate is set to 1 / 10 of the pre-training stage, and the iteration number is 5000 times, so that the repair error in a specific scene is reduced by ≥30%.

[0062] Optionally, in the feature matching process, a random sample consensus (RANSAC) algorithm is introduced to remove mismatched feature point pairs, and the accuracy of feature matching is improved; at the same time, a dynamic time warping (DTW) algorithm is used to match the time sequence features to adapt to the dynamic changes of the features during the movement of the UAV;

[0063] The RANSAC algorithm parameters are set as: iteration number=100, inlier threshold=0.5m; the DTW algorithm distance formula is:

[0064] DTW(i,j)=min(DTW(i-1,j),DTW(i,j-1),DTW(i-1,j-1))

[0065] +d(i,j)

[0066] Wherein: DTW(i,j) represents the cumulative distance of the feature sequence at the i-th moment and the j-th moment;

[0067] d(i,j) represents the Euclidean distance of the feature vector at the i-th moment and the j-th moment;

[0068] The constraint window size is set to 10, so that the feature matching accuracy is improved by ≥25%.

[0069] In a second aspect, the application provides a UAV positioning system combined with environmental features, comprising:

[0070] A multi-source data acquisition module is configured to acquire multi-modal data of the surrounding environment of the UAV through multi-source heterogeneous sensors, wherein the multi-source heterogeneous sensors include a GNSS receiver, a high-definition camera, a laser radar, an inertial measurement unit (IMU), and an electromagnetic interference detector, and the multi-modal data includes GNSS signals, image data, laser point cloud data, and inertial measurement data;

[0071] A data processing module is configured to process the multi-modal data by using an artificial intelligence model, including a signal repair sub-module, a feature extraction and matching sub-module, and a space-time correlation sub-module, wherein the signal repair sub-module uses a signal repair model constructed by CNN and RNN to repair the GNSS signals affected by shielding and interference, the feature extraction and matching sub-module uses a model trained by simulation data generated by CNN and GAN based on an attention mechanism to extract complex environmental features from the image data and the laser point cloud data and perform matching, and the space-time correlation sub-module constructs a space-time correlation model in combination with the inertial measurement data and the electromagnetic interference monitoring data.

[0072] A positioning calculation module is configured to perform real-time positioning calculation by using a space-time sequence prediction model with a lightweight network architecture based on the collaborative processing results of the data processing module, so as to obtain accurate position and attitude information of the UAV.

[0073] A decision adjustment module is configured to adjust the flight path and attitude of the UAV by using a reinforcement learning decision model based on the real-time positioning results output by the positioning calculation module, wherein the reinforcement learning decision model is constructed by using a deep Q network (DQN).

[0074] Compared with the prior art, the present application has the following beneficial technical effects:

[0075] The multi-modal data acquired by the multi-source heterogeneous sensors can comprehensively capture various types of information of the environment in which the UAV is located, different sensors have their own advantages and complement each other, and the limitations of a single sensor in a complex environment are avoided, thereby providing a rich and comprehensive raw data basis for subsequent processing.

[0076] The GNSS signals affected by shielding and interference are accurately repaired by fusing the deep learning model to collaboratively process the multi-modal data; the reliability of feature extraction and matching in a complex environment is effectively improved by generating simulation data, extracting features, and optimizing the matching method, in combination with dynamic area suppression and space-time consistency verification; and the space-time correlation model can fuse multiple types of data and better describe the relationship between the motion state of the UAV and the environmental features.

[0077] The real-time positioning calculation uses a lightweight network architecture to greatly reduce the calculation delay while ensuring the positioning accuracy, thereby meeting the real-time flight requirements of the UAV, and the real-time verification mechanism can timely discover and correct the positioning deviation, thereby ensuring the reliability of the positioning results.

[0078] The reinforcement learning decision model adjusts the flight path and obstacle avoidance strategy by comprehensively considering multiple factors based on real-time positioning information, enabling the UAV to flexibly respond to emergencies in complex environments and fly safely and efficiently.

[0079] This invention provides a comprehensive foundation for subsequent processing through multi-source data acquisition, improves data quality and feature matching reliability through fusion processing, ensures timely and accurate positioning through real-time calculation, and enables safe and efficient flight based on precise positioning through the decision model. It not only solves the traditional technical problems of signal obstruction, interference, dynamic interference, and the contradiction between real-time performance and accuracy in complex environments, but also demonstrates excellent positioning stability and adaptability in a variety of complex scenarios. Attached Figure Description

[0080] Figure 1 A flowchart of a UAV positioning method incorporating environmental features is provided according to the present invention. Detailed Implementation

[0081] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can typically be arranged and designed in various different configurations. Therefore, 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 merely to illustrate selected embodiments of the invention.

[0082] Example

[0083] like Figure 1 As shown in the figure, this embodiment provides a UAV positioning method that combines environmental features. The method will be described in detail below.

[0084] I. Collect multi-modal data of the environment where the UAV is located through multi-source heterogeneous sensors, the multi-modal data including GNSS signal data, visual image data, laser radar point cloud data, inertial measurement data and electromagnetic interference monitoring data; the multi-source heterogeneous sensors including a GNSS receiver, a high-definition camera, a laser radar, an inertial measurement unit (IMU) and an electromagnetic interference detector; wherein the high-definition camera adopts a binocular vision camera, the baseline distance is 12 cm, the resolution is 1920x1080, and the frame rate is 30 fps; the laser radar adopts a 32-line laser radar, the horizontal field of view is 120°, the vertical field of view is -15°~+15°, and the sampling frequency is 15 Hz; the inertial measurement unit adopts a six-axis IMU (3-axis acceleration + 3-axis angular velocity), the acceleration measurement range is ±16g, the angular velocity measurement range is ±2000° / s, and the sampling frequency is 200 Hz; the electromagnetic interference detector covers a frequency band of 100MHz~6GHz, and the detection sensitivity is ≤-90dBm. The multi-source heterogeneous sensors collect multi-modal data, which can comprehensively capture various information of the environment where the UAV is located. Different sensors have their own advantages, GNSS signals provide basic positioning, visual images show scene details, laser radars obtain three-dimensional structures, inertial measurements reflect motion states, and electromagnetic interference monitoring grasps signal environments. They complement each other, avoid the limitations of a single sensor in a complex environment, provide a rich and comprehensive data basis for subsequent processing, and make positioning have a more reliable original basis.

[0085] II. Collaborative processing of multi-modal data using a fusion deep learning model, the collaborative processing including intelligent repair of GNSS signals affected by occlusion and interference, extraction of three-dimensional features specific to the environment based on enhanced visual image data and laser radar point cloud data using a generative adversarial network, and dynamic matching, construction of a spatio-temporal correlation model combined with inertial measurement data and electromagnetic interference monitoring data. The following will be described in detail.

[0086] Among them, the specific steps of intelligent repair of GNSS signals affected by occlusion and interference include:

[0087] Collect a large number of GNSS signal samples containing different occlusion and interference types, construct a signal repair training data set, collect GNSS signal samples containing an occlusion rate of 0%~100% and an interference intensity of -80dBm~-40dBm, and construct a signal repair training data set containing 1 million+ samples;

[0088] The signal repair model is constructed by using a convolutional neural network (CNN)-recurrent neural network (RNN) hybrid model, wherein the convolutional neural network (CNN) is used to extract the spatial features of the signal, and the recurrent neural network (RNN) is used to capture the time sequence features of the signal, wherein the convolutional neural network (CNN) layer adopts 3 layers of convolution (convolution kernel size is 3*3, 5*5, 3*3 respectively, and the step is 1) to extract the spatial features of the signal, and the recurrent neural network (RNN) layer adopts 2 layers of LSTM (hidden unit number is 128) to capture the time sequence features of the signal;

[0089] The model training loss function is:

[0090] L = a * MSE (x true , x pred ) + b * MAE (x true , x pred )

[0091] Wherein, x true is the original undisturbed signal, x pred is the repaired signal, a = 0.7, b = 0.3 are weight coefficients, MSE is mean square error, and MAE is mean absolute error; the disturbed signal is input into the trained model, and the repaired signal is output, and the repair error is less than or equal to 0.5m (plane position);

[0092] The signal repair model is trained by using the training data set, so that the model can learn the distortion law of the signal under different occlusion and interference conditions; the GNSS signal affected by occlusion and interference is input into the trained signal repair model to obtain the repaired GNSS signal. When training the signal repair model, the method of transfer learning is adopted, the model is pre-trained by using the public GNSS signal data set, and then fine-tuned by using the private data set containing specific scene occlusion and interference characteristics, so as to improve the repair accuracy of the model in specific application scenarios; when transfer learning, the pre-training adopts the public data set (NOAA's GNSS multipath data set), the fine-tuning stage adopts the private data set of urban canyon / forest scene (each containing 50,000+ samples), the fine-tuning learning rate is set to 1 / 10 of the pre-training stage, and the iteration number is 5000 times, so that the repair error in specific scene is reduced by more than 30%.

[0093] In addition, the specific steps of extracting and matching complex environment features from image data and laser point cloud data include:

[0094] The generative adversarial network (GAN) is used to generate simulated image data and laser point cloud data in various complex environments, including scenes with different light intensities, different weather conditions and different occlusion degrees; when the generative adversarial network (GAN) generates simulated data, the light intensity range is set to 500-10000 lux, and the occlusion area ratio is 0%-60%.

[0095] The simulation data is fused with the actually collected image data and laser point cloud data to construct a feature extraction and matching training data set;

[0096] A convolutional neural network (CNN) based on an attention mechanism is adopted as the feature extraction model, which can automatically focus on the key feature regions in the image and point cloud;

[0097] The feature extraction model is trained using the training data set, so that the model can extract visual features with scale invariance and rotation invariance from image data, and three-dimensional features with geometric topological structure from laser point cloud data; the attention mechanism convolutional neural network (CNN) contains 4 layers of convolution and 2 layers of attention module, the visual feature dimension is 256, and the laser point cloud feature dimension is 512;

[0098] An improved k-nearest neighbor algorithm is used to match the extracted visual features and three-dimensional features, and the best matching pair is determined by calculating the similarity between the features; the feature matching uses an improved k-NN algorithm, and the similarity calculation uses a cosine distance formula:

[0099]

[0100] where f1, f2 are the feature vectors to be matched, · is the inner product of vectors, ‖‖·‖‖ is the L2 norm, and when S<0.3, it is determined as a successful match;

[0101] In the GAN simulation data generation stage, a dynamic interference generation module is added to randomly add image masks of simulated flying birds, falling leaves, moving vehicles and corresponding point cloud noise in the scene, and the motion trajectory of the dynamic interference conforms to the Brownian motion model, the speed range is set to 0.5-15 m / s, and the size ratio is 1%-10% of the image area;

[0102] It should be noted that a dynamic region suppression loss function is introduced to the feature extraction CNN model:

[0103]

[0104] where L dynamic is the calculation result of the dynamic region suppression loss function, which is used to measure the suppression degree of the model's response to dynamic region features when facing dynamic interference; λ=0.2 is the suppression coefficient; N is the number of samples participating in the calculation; i is the sample index; I i is the image data input to the feature extraction CNN model; M dis a dynamic interference mask region; is an element-wise multiplication operation; f(·) is an output function of the feature extraction CNN model; ‖·‖2 is an L2 norm; forces the model to ignore the feature response of the dynamic object region;

[0105] After improving the k-NN preliminary matching, a spatio-temporal consistency verification layer is added:

[0106] Spatial verification: using the spatial topological relationship of consecutive frames of laser radar point cloud, a local surface descriptor (ISS key point + FPFH descriptor) of the feature point is constructed, and when the descriptor difference threshold between adjacent frames is > 0.25, it is determined as a false match;

[0107] Temporal verification: the theoretical displacement of the feature point is calculated through the inertial measurement unit (IMU) data, and if the actual displacement of the matching point deviates from the theoretical value by > 3σ, σ is the noise variance of the IMU, then the matching pair is removed. In the feature matching process, the random sample consensus (RANSAC) algorithm is introduced to remove the false matching feature point pair and improve the accuracy of feature matching; at the same time, the dynamic time warping (DTW) algorithm is used to match the time sequence features to adapt to the dynamic changes of the features in the process of unmanned aerial vehicle movement;

[0108] RANSAC algorithm parameter setting: iteration number = 100, inlier threshold = 0.5m; the distance formula of the DTW algorithm is:

[0109] DTW(i,j) = min(DTW(i-1,j), DTW(i,j-1), DTW(i-1,j-1))

[0110] + d(i,j)

[0111] Where: DTW(i,j) represents the cumulative distance of the feature sequence at the i-th time and the j-th time;

[0112] d(i,j) represents the Euclidean distance of the feature vector at the i-th time and the j-th time;

[0113] The constraint window size is set to 10, which makes the feature matching accuracy improve by ≥25%.

[0114] In this embodiment, the specific steps of constructing the spatio-temporal correlation model combined with inertial measurement data and electromagnetic interference monitoring data include:

[0115] Pretreatment of inertial measurement data, including noise removal and zero bias correction;

[0116] Analyze the correlation between electromagnetic interference monitoring data and GNSS signal quality, inertial measurement error, and establish an electromagnetic interference influence model;

[0117] Based on Kalman filtering algorithm, the repaired GNSS signal, the extracted environment feature matching result, the preprocessed inertial measurement data and the electromagnetic interference influence model are fused to construct a space-time correlation model, which can describe the relationship between the motion state of the unmanned aerial vehicle in time and space and the environment features.

[0118] The space-time correlation model is constructed by using extended Kalman filter (EKF), and the state equation is:

[0119] X k =F k ·X k-1 +B k ·U k +W k

[0120] The observation equation is:

[0121] Z k =H k ·X k +V k

[0122] Wherein: X k =[x, y, z, v x , v y , v z , φ, θ, ψ] T is the state vector at time k, which represents three-dimensional position (x, y, z), three-dimensional velocity (v x , v y , v z ) and three-dimensional attitude angle (roll angle φ, pitch angle θ, heading angle ψ) in turn; F k is an 8×8 state transition matrix; B k is a control input matrix; U k is the IMU measured acceleration; W k is the process noise, and the variance is set to 0.01; Z k is the observation vector at time k; H k is the observation matrix; V k is the observation noise, and the variance is set to 0.1; H k is the observation matrix, which fuses GNSS and feature matching results.

[0123] In this embodiment, the fusion of deep learning models cooperatively processes multi-modal data, with significant advantages. For GNSS signals affected by occlusion and interference, the CNN-RNN hybrid model can accurately repair them, solving the problem of inaccurate positioning by traditional methods when the signal is poor. The generative adversarial network generates simulated data, and the CNN with attention mechanism extracts features, which are then matched by the improved k-NN algorithm, improving the accuracy of feature extraction and matching in complex environments. The introduction of the dynamic region suppression loss function effectively ignores dynamic interference, and the spatiotemporal consistency check further eliminates false matches, making feature matching more reliable. These processes are not simply additive, but work together to make the data processing effect far superior to the application of a single technology, laying a solid foundation for the accuracy of positioning.

[0124] III. Based on the cooperative processing results, real-time positioning calculation is performed through a spatiotemporal sequence prediction model to obtain the accurate position and attitude information of the UAV. A lightweight network architecture is used in the real-time positioning calculation process to reduce the calculation delay. The specific steps of using a lightweight network architecture for real-time positioning calculation include:

[0125] Dimensionality reduction is performed on the feature data output by the spatiotemporal correlation model to reduce the data volume;

[0126] MobileNetV3 (parameter quantity ≤ 3 million) is used as the basic architecture of the positioning calculation network, and pruning and quantization techniques are used to further optimize the network structure and reduce the computational complexity;

[0127] The dimensionally reduced feature data is input into the lightweight positioning calculation network, and the network outputs the position coordinates, velocity, and attitude angle of the UAV;

[0128] The positioning calculation results are checked in real time, and if the results exceed the preset error range, a repositioning mechanism is triggered;

[0129] The positioning calculation error formula is:

[0130]

[0131] where e is the three-dimensional positioning error, (x pred , y pred , z pred ) is the three-dimensional position prediction value of the UAV output by the model; (x true , y true , z true ) is the actual three-dimensional position true value of the UAV; e ≤ 0.8 m (95% confidence) is required; the real-time index is processing delay ≤ 50 ms (single frame of data).

[0132] The real-time positioning solution adopts a lightweight network architecture, which greatly reduces the calculation delay while ensuring the positioning accuracy. MobileNetV3 has fewer parameters and fast operation, and combined with dimension reduction processing, it can quickly output the position and attitude information of the UAV, meeting the real-time flight requirements. The real-time verification mechanism can also timely discover and correct the positioning deviation, ensuring the reliability of the positioning results. This step cooperates with the previous data processing, and the accurately processed data makes the positioning solution more efficient, while the fast solution can provide timely information for subsequent decision-making.

[0133] IV. According to the real-time position and attitude information, the flight path and obstacle avoidance strategy of the UAV are dynamically adjusted through a reinforcement learning decision model to realize safe and efficient flight of the UAV in complex environments, specifically including:

[0134] The real-time position, speed, attitude of the UAV and obstacle information of the surrounding environment are taken as state inputs;

[0135] The distance, flight time and energy consumption of the UAV to safely reach the target point are taken as parameters of the reward function;

[0136] A deep Q network (DQN) in the deep reinforcement learning algorithm is used to construct an intelligent decision model, which continuously learns the optimal flight strategy through interaction with the environment;

[0137] The real-time positioning results and environmental perception data are input into the intelligent decision model, and the model outputs the flight path adjustment amount and attitude control instructions of the UAV, which executes the corresponding instructions to realize obstacle avoidance and path optimization; the reward function of the reinforcement learning decision model is:

[0138]

[0139] Where d is the straight-line distance from the current position of the UAV to the target point; D is the straight-line distance from the initial position to the target point; t is the flight time; T is the estimated total flight time; e is the current energy consumption; E is the maximum allowed energy consumption; ω1=0.5, ω2=0.3, ω3=0.2 are the weight coefficients of each reward term;

[0140] The DQN network adopts a double Q network structure, the experience replay pool capacity is 100,000, the learning rate is set to 0.001, and the discount factor γ=0.9. The reinforcement learning decision model adjusts the flight path and obstacle avoidance strategy according to the real-time positioning information, and considers distance, time and energy consumption to make the optimal decision. It can enable the UAV to flexibly respond to unexpected situations in complex environments and safely and efficiently reach the target point. This step closely cooperates with the real-time positioning solution, which provides accurate position and environmental information, enabling the decision model to make precise adjustments, and reasonable decisions can reduce the accumulation of positioning errors caused by path errors, forming a virtuous cycle.

[0141] In the method, multi-source heterogeneous sensors collect multi-modal data (GNSS, vision, lidar, IMU, electromagnetic interference data), and not simply sensor stacking, but through the complementarity of data dimensions to form a "data redundancy gain". GNSS provides a global coordinate reference, vision captures scene texture details, lidar constructs a three-dimensional structure, IMU compensates for short-term positioning when the signal is interrupted, and electromagnetic interference data provides an environmental reference for signal repair. The cooperative collection of such multi-dimensional data solves the "information blind area" problem of a single sensor in a complex environment - for example, in urban canyons, when GNSS signals are blocked, the three-dimensional point cloud of the lidar and the vision image can provide environmental feature anchor points, and the IMU maintains the short-time motion trajectory calculation, laying a "spatial-temporal alignment data foundation" for subsequent data fusion, avoiding the defect that traditional single sensor positioning immediately fails after signal loss.

[0142] Further, when the CNN-RNN hybrid model repairs the disturbed GNSS signal, the CNN extracts the signal spatial distortion features, the RNN captures the time series drift rules, and with transfer learning and 1 million+ sample training, a repair accuracy within 0.5m is achieved. This repair is not a simple signal filtering, but through learning the signal distortion rules under different obstruction rates (0%~100%) and interference intensities (-80dBm~ -40dBm), a "interference-repair" mapping relationship is constructed, solving the problem of traditional inertial navigation cumulative error increasing rapidly with time. More importantly, the repaired GNSS signal is linked with the electromagnetic interference monitoring data to provide a "denoised reference coordinate" for the spatio-temporal correlation model, forming a "true value anchor" with other sensor data, greatly reducing the initial error of subsequent positioning calculation.

[0143] In the present application, GAN-generated simulated data (containing dynamic interference), dynamic region suppression loss function and spatio-temporal consistency check are introduced to form a "dynamic anti-interference closed loop". The GAN-generated dynamic interference (flying birds, moving vehicles, etc.) simulation data, combined with the Brownian motion trajectory model, enables the feature extraction model to "see" various dynamic scenes during the training phase; the dynamic region suppression loss function (λ = 0.2) forces the model to ignore dynamic region features through mask operation, solving the problem of traditional SIFT / SURF algorithm sensitivity to dynamic objects; the spatio-temporal consistency check (ISS key points + FPFH descriptor difference threshold, IMU displacement deviation 3σ criterion) further eliminates false matches, making the feature matching accuracy improve by more than 25%. This "generative data augmentation + targeted loss suppression + multi-dimensional check" synergy breaks through the technical bottleneck of insufficient robustness of feature matching under complex lighting, obstruction and dynamic interference, achieving an "environmental self-adaptive ability" that traditional algorithms cannot achieve.

[0144] The extended Kalman filter (EKF) fuses the repaired GNSS, feature matching results and IMU data, and the constructed space-time correlation model is not a simple data splicing, but a dynamic adjustment through a state equation (including position, velocity and attitude angle) and an observation equation to realize "real-time adaptation of multi-source data weight". For example, when the GNSS signal is good, the model focuses on GNSS data; when the signal is blocked, the weight of feature matching and IMU data is automatically increased, solving the "jumping error" of the traditional filtering algorithm when the data source is switched. This dynamic weight distribution mechanism makes the model improve the positioning continuity in urban canyons, forests and other scenes by 60%, far exceeding the effect of a single sensor or fixed weight fusion.

[0145] The lightweight network (MobileNetV3, parameters ≤3 million) combines pruning and quantization techniques, and ensures that the three-dimensional positioning error is ≤0.8 m (95% confidence) while controlling the single-frame data processing delay to be within 50 ms. This achievement breaks the technical cognition that "high precision must be accompanied by high delay" - traditional deep learning models can improve accuracy, but have high computational complexity and are difficult to meet the real-time needs of unmanned aerial vehicles; while simplified models sacrifice accuracy. This method balances high precision and real-time performance through "data preprocessing and dimensionality reduction + lightweight network architecture", solving the problem of obstacle avoidance failure caused by positioning delay in high-speed flight scenarios.

[0146] The deep Q network (DQN) decision model takes the positioning result as input, and comprehensively considers the reward functions (ω1=0.5, ω2=0.3, ω3=0.2) of distance, time and energy consumption to realize dynamic optimization of path planning and obstacle avoidance. This decision is not run in isolation, but forms a "perception-decision-execution" closed loop with real-time positioning: improved positioning accuracy makes the decision more accurate, and the optimized path reduces the invalid motion of the unmanned aerial vehicle in complex environments, reducing the accumulation of positioning errors. For example, in the event of sudden obstruction, the decision model can make a prediction based on the positioning deviation to adjust the flight attitude in advance to obtain better environmental features. This "positioning guiding decision-making, decision-making feeding back positioning" synergy greatly improves the safety passing rate of the unmanned aerial vehicle in complex scenarios.

[0147] It is worth noting that the present application realizes the leap in positioning stability in dynamic environments. The positioning accuracy of traditional methods will drop by more than 50% when the proportion of dynamic interference (such as birds and vehicles) exceeds 10%, while the present application can still maintain a positioning accuracy of ≤1 m in scenarios with a large proportion of dynamic interference through the synergy of dynamic region loss function suppression and space-time verification, exhibiting "robustness against dynamic interference", an unexpected effect.

[0148] In addition, nonlinear gain of multi-source data is realized. Each step forms nonlinear enhancement of "data-model-decision". GNSS signal repair provides global coordinate reference for feature matching, and feature matching result corrects the cumulative error of IMU. The cooperation makes the overall positioning accuracy higher than the superposition of each technology applied alone. In indoor or dense forest environment without GNSS signal, the traditional method relying on IMU will have a deviation of more than 5m within 10 seconds, while the method breaks through the technical limitation of "no satellite signal means failure" through deep fusion of laser radar point cloud features and visual features, combined with the space-time correlation model.

[0149] The embodiment also provides an unmanned aerial vehicle positioning system combined with environmental features, comprising:

[0150] A multi-source data acquisition module is configured to acquire multi-modal data of the environment around the unmanned aerial vehicle through multi-source heterogeneous sensors, wherein the multi-source heterogeneous sensors include a GNSS receiver, a high-definition camera, a laser radar, an inertial measurement unit (IMU) and an electromagnetic interference detector, and the multi-modal data includes GNSS signals, image data, laser point cloud data and inertial measurement data.

[0151] A data processing module is configured to process the multi-modal data by using an artificial intelligence model, including a signal repair sub-module, a feature extraction and matching sub-module and a space-time correlation sub-module, wherein the signal repair sub-module repairs the GNSS signals affected by shielding and interference by using a signal repair model constructed by CNN and RNN, the feature extraction and matching sub-module extracts complex environmental features from the image data and the laser point cloud data and performs matching by using a model trained by simulation data generated by CNN and GAN based on an attention mechanism, and the space-time correlation sub-module constructs a space-time correlation model in combination with the inertial measurement data and the electromagnetic interference monitoring data.

[0152] A positioning calculation module is configured to perform real-time positioning calculation by using a space-time sequence prediction model with a lightweight network architecture based on the cooperative processing result of the data processing module, so as to obtain accurate position and attitude information of the unmanned aerial vehicle.

[0153] A decision adjustment module is configured to adjust the flight path and attitude of the unmanned aerial vehicle by using a reinforcement learning decision model according to the real-time positioning result output by the positioning calculation module, wherein the reinforcement learning decision model is constructed by using a deep Q network (DQN).

[0154] The unmanned aerial vehicle positioning system combined with environmental features realizes comprehensive optimization of unmanned aerial vehicle positioning through the collaborative operation of a multi-source data acquisition module, a data processing module, a positioning calculation module and a decision adjustment module. The multi-source data acquisition module integrates various sensor information to provide rich basis for positioning; the data processing module repairs signals, extracts matching features and constructs a correlation model with the aid of artificial intelligence technology to improve data effectiveness; the positioning calculation module realizes accurate and real-time positioning calculation with a lightweight architecture; and the decision adjustment module intelligently adjusts flight strategies according to positioning results. The modules are not isolated, but work together to increase efficiency and empowerment, not only solving the problems of signals, feature matching and the contradiction between real-time and accuracy in complex environments, but also enabling the unmanned aerial vehicle to fly stably, safely and efficiently in various complex scenes.

[0155] The above specific embodiments are only several optional embodiments of the present application, and based on the technical solutions of the present application and the related inspiration of the above embodiments, those skilled in the art can make various alternative improvements and combinations on the above specific embodiments.

Claims

1. A method for positioning a UAV in combination with environmental features, the method comprising: The method comprises the following steps: Step one: collecting multi-modal data of the environment where the UAV is located by multi-source heterogeneous sensors, the multi-modal data including GNSS signal data, visual image data, laser radar point cloud data, inertial measurement data and electromagnetic interference monitoring data; Step two: using a fusion deep learning model to cooperatively process the multi-modal data, the cooperative processing including intelligently repairing GNSS signals affected by occlusion and interference, extracting three-dimensional features specific to the environment from enhanced visual image data and laser radar point cloud data based on a generative adversarial network, and dynamically matching, and constructing a spatio-temporal correlation model in combination with inertial measurement data and electromagnetic interference monitoring data; Step three: based on the cooperative processing results, performing real-time positioning calculation through a spatio-temporal sequence prediction model to obtain accurate position and attitude information of the UAV, and using a lightweight network architecture in the real-time positioning calculation process to reduce calculation delay; Step four: dynamically adjusting the flight path and obstacle avoidance strategy of the UAV through a reinforcement learning decision-making model according to the real-time acquired position and attitude information, to realize safe and efficient flight of the UAV in a complex environment. 2.The method of claim 1, wherein, The multi-source heterogeneous sensors in step one include a GNSS receiver, a high-definition camera, a laser radar, an inertial measurement unit and an electromagnetic interference detector; the high-definition camera adopts a binocular vision camera with a baseline distance of 12 cm, a resolution of 1920x1080 and a frame rate of 30 fps; the laser radar adopts a 32-line laser radar with a horizontal field of view of 120°, a vertical field of view of -15° to +15° and a sampling frequency of 15 Hz; and the inertial measurement unit adopts a six-axis IMU. 3.The method of claim 1, wherein, In step two, the specific process of intelligently repairing GNSS signals affected by occlusion and interference includes: Collecting a large number of GNSS signal samples containing different occlusion and interference types to construct a signal repair training data set, collecting GNSS signal samples containing an occlusion rate of 0% to 100% and an interference intensity of -80 dBm to -40 dBm to construct a signal repair training data set containing more than 1 million samples; A convolutional neural network-recurrent neural network hybrid model is used to construct a signal repair model, wherein the convolutional neural network is used to extract the spatial features of the signal, and the recurrent neural network is used to capture the time sequence features of the signal, wherein the convolutional neural network layer adopts 3 layers of convolution to extract the spatial features of the signal, the convolution kernel size is 3x3, 5x5 and 3x3 respectively, and the step is 1; the recurrent neural network layer adopts 2 layers of LSTM to capture the time sequence features of the signal, and the hidden unit number is 128; The model training loss function is: L = a - MSE(x true , x pred ) + β - MAE(x true , x pred ) Wherein, x true is the original interference-free signal, x pred is the repaired signal, α=0.7, β=0.3 are weight coefficients, MSE is mean square error, and MAE is mean absolute error; the interference signal is input into the trained model, and the repaired signal is output, and the repair error is less than or equal to 0.5 m; The training data set is used to train the signal repair model, so that the model can learn the distortion law of the signal under different occlusion and interference conditions; The GNSS signal affected by occlusion and interference is input into the trained signal repair model to obtain the repaired GNSS signal. 4.The method of claim 1, wherein, In step two, the specific process of extracting complex environment features from image data and laser point cloud data and matching includes: The generative adversarial network is used to generate simulated image data and laser point cloud data in various complex environments, including different light intensities, different weather conditions, and different scene occlusion levels. When the generative adversarial network generates simulated data, the light intensity range is set to 500-10000 lux, and the occlusion area ratio is 0%-60%. The simulated data is fused with real collected image data and laser point cloud data to construct a feature extraction and matching training dataset. An attention mechanism-based convolutional neural network is used as the feature extraction model, which can automatically focus on key feature regions in images and point clouds. The feature extraction model is trained using the training dataset, enabling the model to extract scale-invariant and rotation-invariant visual features from image data and three-dimensional features with geometric topological structure from laser point cloud data. The attention mechanism-based convolutional neural network contains 4 layers of convolution and 2 layers of attention modules, with a visual feature dimension of 256 and a laser point cloud feature dimension of 512. An improved k-nearest neighbor algorithm is used to match the extracted visual features and three-dimensional features, and the best matching pair is determined by calculating the similarity between the features. The improved k-NN algorithm is used for feature matching, and the cosine distance formula is used for similarity calculation: where f1 and f2 are the feature vectors to be matched, · is the vector inner product, and ‖·‖ is the L2 norm of the vector. When S < 0.3, it is determined to be a successful match. During the GAN simulation data generation stage, a dynamic interference object generation module is added to randomly add image masks of simulated flying birds, falling leaves, and moving vehicle dynamic objects and corresponding point cloud noise in the scene. The motion trajectory of the dynamic interference object conforms to the Brownian motion model, with a speed range of 0.5-15 m / s and a size ratio of 1%-10% of the image area. A dynamic region suppression loss function is introduced to the feature extraction CNN model: wherein, L dynamic is the calculation result of the dynamic area suppression loss function, used to measure the suppression degree of the model in response to the dynamic area characteristics when facing dynamic interference; λ = 0.2 is the suppression coefficient; N is the number of samples participating in the calculation; i is the sample index; I i is the i-th input image data into the feature extraction CNN model; M d is the dynamic interference mask area; ⊙ is the element-wise multiplication operation; f(·) is the output function of the feature extraction CNN model; ‖·‖2 is the L2 norm; and the model is forced to ignore the feature response of the dynamic object area. After preliminary matching using the improved k-NN, a spatio-temporal consistency verification layer is added: Spatial verification: The spatial topological relationship of consecutive frames of laser radar point cloud is used to construct a local surface descriptor of feature points. The local surface descriptor includes ISS key points and FPFH descriptors. When the descriptor difference threshold between adjacent frames is >0.25, it is determined to be a false match. Temporal verification: The theoretical displacement of feature points is calculated based on inertial measurement unit data. If the actual displacement of the matching point deviates from the theoretical value by >3σ, where σ is the IMU noise variance, then the matching pair is removed.

5. The unmanned aerial vehicle positioning method in combination with environmental features of claim 1, wherein, The specific steps for constructing the spatio-temporal correlation model in step two include: Preprocessing of inertial measurement data, including noise removal and zero bias correction; Analyzing the correlation between electromagnetic interference monitoring data and GNSS signal quality, inertial measurement error, and establishing an electromagnetic interference influence model; Based on the Kalman filter algorithm, the repaired GNSS signal, the extracted environmental feature matching result, the preprocessed inertial measurement data, and the electromagnetic interference influence model are fused to construct a spatio-temporal correlation model. This model can describe the relationship between the motion state of the unmanned aerial vehicle in time and space and the environmental features. The extended Kalman filter (EKF) is used to construct the spatiotemporal correlation model, and the state equation is: X k = F k • X k-1 + B k • U k + W k The observation equation is: Z k = H k • X k + V k wherein: X k = [x, y, z, v x , v y , v z , φ, θ, ψ] T is the state vector at time k, representing three-dimensional position (x, y, z), three-dimensional velocity (v x , v y , v z ) and three-dimensional attitude angle (roll angle φ, pitch angle θ, heading angle ψ) in turn; F k is an 8x8 state transition matrix; B k is a control input matrix; U k is the IMU measured acceleration; W k is the process noise, with a variance of 0.01; Z k is the observation vector at time k; H k is an observation matrix; V k is observation noise, with a variance of 0.1; H k is an observation matrix, fusing GNSS and feature matching results.

6. The unmanned aerial vehicle positioning method in combination with environmental features of claim 1, wherein, The specific steps of real-time positioning calculation in step three include: Dimensionality reduction is performed on the feature data output by the spatiotemporal correlation model to reduce the data volume; MobileNetV3 is used as the basic architecture of the positioning calculation network, and the network structure is further optimized through pruning and quantization techniques to reduce the computational complexity; The dimensionality-reduced feature data is input into the lightweight positioning calculation network, and the network outputs the position coordinates, velocity, and attitude angle of the UAV; The positioning calculation results are checked in real time, and if the results exceed the preset error range, the repositioning mechanism is triggered; The positioning calculation error formula is: wherein e is a three-dimensional positioning error, (x pred , y pred , z pred ) is a three-dimensional position prediction value of the UAV output by the model; (x true , y ttue , z true ) is an actual three-dimensional position true value of the UAV; e≤0.8m is required; the real-time index is a processing delay≤50ms.

7. The unmanned aerial vehicle positioning method in combination with environmental features of claim 1, wherein, The specific steps of adjusting the flight path and attitude of the UAV through the reinforcement learning decision model in step four include: The real-time position, velocity, attitude of the UAV, and obstacle information in the surrounding environment are used as state inputs; The distance, flight time, and energy consumption of the UAV to safely reach the target point are used as parameters of the reward function; A deep Q network in the deep reinforcement learning algorithm is used to construct an intelligent decision model, which continuously learns the optimal flight strategy through interaction with the environment; The real-time positioning results and environmental perception data are input into the intelligent decision model, which outputs the flight path adjustment amount and attitude control instructions of the UAV, and the UAV executes the corresponding instructions to achieve obstacle avoidance and path optimization; the reward function of the reinforcement learning decision model is: where d is the straight-line distance from the current position of the UAV to the target point; D is the straight-line distance from the initial position to the target point; t is the time flown; T is the total flight time; e is the current energy consumption; E is the maximum allowed energy consumption; ω1=0.5, ω2=0.3, and ω3=0.2 are the weight coefficients of each reward term; The DQN network uses a double Q network structure, the experience replay pool capacity is 100,000, the learning rate is set to 0.001, and the discount factor γ=0.

9. 8.The method of claim 3, wherein, When training the signal repair model, the transfer learning method is used. First, the model is pre-trained using a public GNSS signal dataset, and then fine-tuned using a private dataset containing specific scene occlusion and interference features to improve the model's repair accuracy in specific application scenarios. During transfer learning, the pre-training uses a public dataset, the public dataset is the NOAA GNSS multipath dataset, the fine-tuning stage uses a private dataset of urban canyon / forest scenarios, the fine-tuning learning rate is 1 / 10 of the pre-training stage, and the iteration number is 5000. 9.The method of claim 4, wherein, In the feature matching process, the random sample consensus (RANSAC) algorithm is introduced to eliminate mismatched feature point pairs, improving the accuracy of feature matching. At the same time, the dynamic time warping (DTW) algorithm is used to match the time series features to adapt to the dynamic changes of features during UAV movement. The RANSAC algorithm parameters are set as follows: iteration number = 100, inlier threshold = 0.5 m; the DTW algorithm distance formula is: DTW(i,j) = min(DTW(i-1,j), DTW(i,j-1), DTW(i-1,j-1)) + d(i,j) Wherein: DTW(i,j) represents the cumulative distance of the i-th time and the j-th time feature sequence; d(i,j) represents the Euclidean distance of the i-th time and the j-th time feature vector; The constraint window size is set to 10.

10. A drone positioning system incorporating environmental features, characterized by, Comprise: A multi-source data acquisition module for acquiring multi-modal data of the environment around the unmanned aerial vehicle through multi-source heterogeneous sensors, the multi-source heterogeneous sensors including a GNSS receiver, a high-definition camera, a laser radar, an inertial measurement unit and an electromagnetic interference detector, and the multi-modal data including GNSS signals, image data, laser point cloud data and inertial measurement data; A data processing module for processing the multi-modal data using an artificial intelligence model, including a signal repair sub-module, a feature extraction and matching sub-module and a space-time correlation sub-module, wherein the signal repair sub-module uses a signal repair model constructed by CNN and RNN to repair the GNSS signals affected by shielding and interference, the feature extraction and matching sub-module uses a model trained by simulation data generated by CNN and GAN based on an attention mechanism to extract complex environmental features from image data and laser point cloud data and perform matching, and the space-time correlation sub-module constructs a space-time correlation model in combination with inertial measurement data and electromagnetic interference monitoring data; A positioning calculation module for obtaining accurate position and attitude information of the unmanned aerial vehicle by using a space-time sequence prediction model with a lightweight network architecture to perform real-time positioning calculation based on the collaborative processing results of the data processing module; A decision adjustment module for adjusting the flight path and attitude of the unmanned aerial vehicle through a reinforcement learning decision model based on the real-time positioning results output by the positioning calculation module, and the reinforcement learning decision model is constructed by using a deep Q network.

Images