Unmanned aerial vehicle dynamic scene intelligent visual positioning navigation system based on deep learning

Abstract

The invention discloses an unmanned aerial vehicle dynamic scene intelligent visual positioning navigation system based on deep learning, and relates to the technical field of unmanned aerial vehicle intelligent positioning navigation, and the system comprises a multi-source sensor data acquisition module, a carrying binocular camera, a laser radar, an inertial navigation, a millimeter wave radar and a GPS / Beidou module; the data preprocessing module is used for generating a space-time sequence through multi-source fusion; the dynamic feature extraction module is used for constructing a feature map; a deep learning positioning model module; the dynamic path planning module is used for generating a path by combining a layered cost map with an improved A * algorithm; the navigation control module is used for realizing trajectory tracking through PID control and anti-interference compensation; the global map construction module is used for constructing and optimizing a map in an incremental SLAM manner; and the system state monitoring and fault-tolerant module is used for monitoring the state of the sensor and starting a standby strategy. The method improves the dynamic scene positioning precision, strengthens the fault tolerance and long-term operation stability of the sensor, and is suitable for multiple complex scenes.

Description

Technical Field

[0001] This invention relates to the field of intelligent positioning and navigation technology for unmanned aerial vehicles (UAVs), and more particularly to an intelligent visual positioning and navigation system for UAVs in dynamic scenes based on deep learning. Background Technology

[0002] Visual positioning and navigation technology for unmanned aerial vehicles (UAVs) is a core support for achieving autonomous flight and mission execution, and is widely used in scenarios such as logistics delivery, power line inspection, and emergency rescue. In dynamic scenarios, the frequent movement of dynamic targets such as vehicles and pedestrians, as well as complex factors such as changes in lighting and occlusion, place extremely high demands on the real-time performance, robustness, and positioning accuracy of navigation systems. Traditional navigation systems mostly rely on a single visual sensor or GPS, making it difficult to adapt to the complex needs of dynamic environments.

[0003] Existing UAV visual navigation solutions have significant limitations. Some solutions use monocular vision combined with SLAM algorithms to build maps and for localization, but monocular vision lacks scale information, making it prone to feature matching errors in areas with dense dynamic targets, leading to localization drift. Furthermore, it fails to perform feature separation for dynamic targets, resulting in severe interference with static environment feature extraction. Another type of solution introduces multi-source sensor fusion, but sensor data synchronization accuracy is low, and many employ simple weighted fusion strategies, failing to fully explore the spatiotemporal correlations between data. When a sensor fails, localization continuity cannot be guaranteed. Path planning is mostly based on static environment modeling, resulting in insufficient trajectory prediction for dynamic obstacles, delayed obstacle avoidance decisions, and a high risk of collisions.

[0004] In recent years, while the application of deep learning technology has improved feature extraction capabilities, existing solutions still have shortcomings. Dynamic target detection often employs general-purpose target detection algorithms, failing to optimize for small targets and high-speed moving targets from the UAV's perspective, making it difficult to balance detection accuracy and speed. Localization models lack effective suppression of dynamic features, and temporal feature fusion is insufficient, resulting in poor localization stability in fast-moving scenarios. Furthermore, the system lacks a closed-loop optimization mechanism, leading to the accumulation of map building and localization errors with flight distance, and it does not consider the impact of external interference such as wind fields on navigation control, easily resulting in trajectory tracking deviations in practical applications. These problems cause existing systems to fail to meet the requirements of professional applications in terms of localization accuracy, obstacle avoidance timeliness, and system reliability in dynamic scenarios, necessitating the construction of a deep learning navigation system integrating multi-source fusion, dynamic perception, and intelligent decision-making. Summary of the Invention

[0005] The present invention proposes a deep learning-based intelligent visual positioning and navigation system for UAVs in dynamic scenes to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a deep learning-based intelligent visual positioning and navigation system for dynamic scenes of unmanned aerial vehicles, comprising: The multi-source sensor data acquisition module is equipped with a binocular vision camera, LiDAR, MEMS inertial navigation module, millimeter-wave radar and multi-mode GPS / BeiDou module. Each sensor is synchronized through a synchronization trigger board to acquire environmental stereo images, 3D point clouds, motion parameters, obstacle distance and speed and absolute position information. The data preprocessing module includes an image preprocessing unit that performs distortion correction and scale normalization, and uses a blind deconvolution algorithm for blurred images; an inertial navigation preprocessing unit that uses Allan variance analysis to determine the zero bias parameters and Kalman filtering to remove noise; a radar data processing unit that removes ground noise through pass-through filtering and converts radar data into obstacle information; and a multi-source data fusion unit that registers data based on a factor graph model to generate a spatiotemporal synchronization sequence. The dynamic feature extraction module includes a dynamic target detection subunit that uses an improved YOLOv5 algorithm to output target information; a motion parameter calculation subunit that uses the pyramid LK optical flow method to calculate and convert speed; a static environment feature extraction subunit that uses a ResNet18 network to extract semantic features; and a feature fusion subunit that constructs a feature map. The deep learning localization model module uses a Transformer architecture to output the coordinates of matching feature points in the feature matching subnetwork; the pose estimation subnetwork consists of LSTM and fully connected layers, which outputs the UAV's six-DOF pose and combines GPS / BeiDou calibration. The dynamic path planning module uses a hierarchical structure to allocate costs in the cost map construction sub-unit; the path search sub-unit is based on the improved A* algorithm and smooths the generated path using B-spline curves; the dynamic update sub-unit refreshes the path periodically and replans when the trajectory changes exceed a threshold. The navigation control module includes a command conversion subunit that discretizes the planned path into a sequence of waypoints and converts it into motor control commands; a PID control subunit that uses an incremental PID algorithm to adjust the coefficients through a parameter self-tuning algorithm; an anti-interference compensation subunit equipped with a wind speed sensor that outputs compensation commands; and an execution drive subunit that uses a brushless motor controller to convert control commands into speed signals. The global map building module uses an incremental SLAM algorithm to generate a 3D point cloud map in the point cloud fusion sub-unit; the map storage sub-unit uses a dynamic and static separation strategy to store data; the local re-optimization sub-unit uses a sliding window algorithm to optimize pose; and the map update sub-unit fuses a GPS-calibrated map. The system status monitoring and fault tolerance module consists of a sensor monitoring subunit that collects parameters and marks anomalies; a fault tolerance processing subunit that uses a redundancy strategy to locate and correct errors; and an alarm subunit that uses an audible and visual alarm to alert the ground station and transmit fault information.

[0007] Furthermore, it also includes a dynamic obstacle threat assessment unit. This unit first predicts the position sequence of dynamic targets within the next 2-4 seconds using Kalman filtering, and then calculates relative parameters based on the current pose of the UAV. Calculate the threat level, where T is the obstacle threat level; This is the distance weighting coefficient. For speed weighting coefficient, This is the size weighting coefficient, and + + =1; d is the normalized value of the relative distance between the drone and the obstacle; It is the relative velocity; The maximum speed of the drone is 4-6 m / s; s is the ratio of the obstacle size to the drone size; the dynamic path planning module sorts the obstacle avoidance priority according to the T value. When T>0.6, emergency obstacle avoidance is triggered, reducing the drone speed to 1-3 m / s and shifting it laterally by 0.8-1.2 m.

[0008] Furthermore, it also includes a visual positioning accuracy optimization unit, which selects ground landmarks as reference feature points after feature matching, and then... Calculate the positioning error compensation value, where The position error compensation amount is n; n is the number of reference feature points. Let be the confidence level of the i-th feature point; This represents the actual location of the feature point on the global map. The predicted location of the feature point in the current frame; The fully connected layer of the input pose estimation subnetwork performs linear correction on the output six-DOF pose.

[0009] Furthermore, the synchronization trigger board of the multi-source sensor data acquisition module is implemented using an FPGA chip, outputting four synchronization pulse signals; the camera lens is equipped with an electric ND filter, and the ambient light intensity is collected through a light sensor. The lens is equipped with an autofocus module with a focusing distance of 0.3-0.7m to infinity; the GPS / BeiDou module supports multi-satellite positioning, simultaneously receiving GPS L1 / L2 and BeiDou B1 / B2 signals. In obstructed scenarios, inertial navigation-assisted positioning is enabled. When positioning is interrupted, the pose is inferred through inertial navigation. When the interruption is less than 5s, the positioning error is less than 0.3m.

[0010] Furthermore, the blind deconvolution algorithm in the data preprocessing module adopts the Richardson-Lucy iterative method, with the number of iterations set to 10-25. The point spread function is estimated by the gradient of the edge of the blurred region in the image. The direct filtering of the LiDAR point cloud is set to the Z-axis range of 0.05-0.15m to 45-55m to remove ground and high-altitude noise, and then outliers are removed by statistical filtering. The factor map spatiotemporal calibration algorithm uses the data of each sensor as factor nodes and the pose as variable nodes, and optimizes the factor map by the Gauss-Newton method.

[0011] Furthermore, the compression ratio of the SE channel attention module in the dynamic feature extraction module is set to 12-20, and the channel feature weights are obtained through global average pooling, assigning a weight of 1.1-1.6 times to the target feature channel; the pyramid LK optical flow method uses a 4-6 layer Gaussian pyramid, and targets with a motion speed > 4-6 pixels / frame are marked as high-speed dynamic targets, and their trajectories are recorded separately; the ResNet18 network adds a feature visualization interface after the conv4_2 layer, outputting a feature heatmap, and marking regions with feature response values ​​< 0.2 as blurred regions.

[0012] Furthermore, the Transformer architecture of the deep learning localization model module is configured with 6-10 attention heads and uses the ReLU activation function; mutual information entropy is used to calculate SIFT descriptors based on feature points, and similarity is measured by Kullback-Leibler divergence. After matching pairs are filtered, a random sampling consensus algorithm is used to remove false matches; the LSTM gating mechanism outputs 0-1 weights through the sigmoid activation function, and the model is trained using the COCO dataset and a custom dynamic scene dataset.

[0013] Furthermore, the dynamic target prediction in the dynamic path planning module adopts extended Kalman filtering. The state vector includes position (x, y, z) and velocity (vx, vy, vz), and the observation vector is the relative distance and velocity measured by radar. The diagonal elements of the process noise covariance matrix are set to 0.005-0.015, and the diagonal elements of the observation noise covariance matrix are set to 0.04-0.06. B-spline curve smoothing adopts cubic B-splines, and the number of control points is 1.2-1.8 times the number of initial path nodes. Emergency path replanning adopts the fast random tree algorithm with a search radius of 8-12m and a sampling frequency of 8-12Hz.

[0014] Furthermore, the parameter self-tuning algorithm of the navigation control module adopts the relay feedback method. When the attitude error is >4°-6° or the position error is >0.8-1.2m, the dynamic parameters of the system are identified through the relay characteristics, and the PID coefficient is adjusted with an adjustment step size of 0.04-0.06. The wind field interference model is based on the aerodynamic parameters of the UAV and the wind speed and direction to calculate the interference force in the X, Y, and Z axis directions and the interference torque in the roll, pitch, and yaw directions. The compensation command is implemented by superimposing it on the PID output. When the wind speed is >8-12m / s, the wind resistance mode is activated to limit the flight speed to <2-4m / s. The brushless motor controller supports PWM signal input and is equipped with overcurrent and overvoltage protection functions. In case of failure, the power output is automatically cut off.

[0015] Furthermore, the octree compression of the global map construction module divides the 3D space into cubic voxels with side lengths of 0.05-0.15m, storing only voxels containing point clouds, achieving a compression ratio of 8-12:1; the time-series database uses InfluxDB, with a storage frequency of 0.5-1.5Hz, retaining dynamic target trajectory data from the most recent 25-35 days; the sliding window algorithm adaptively adjusts the window size according to the dynamic nature of the scene, setting 12-18 keyframes for static scenes and 4-6 keyframes for dynamic scenes; the bundle adjustment method optimizes variables including the poses of all keyframes and the 3D coordinates of feature points within the window, stopping iteration when the optimization residual is <0.5 pixels; map calibration uses the ICP algorithm to align the SLAM point cloud with GPS markers.

[0016] Compared with existing technologies, the beneficial effects of this invention are: At the multi-source data processing level, the system ensures the spatiotemporal consistency of data from multiple sources, including vision, radar, and inertial navigation, through high-precision hardware synchronization and refined preprocessing. Exposure compensation and deblurring techniques in image preprocessing improve image quality under complex lighting conditions, radar point cloud filtering effectively removes noise interference, and factor graph fusion algorithms fully exploit the value of data correlations, laying a high-quality data foundation for subsequent feature extraction and localization. Synchronization mechanisms and sensor redundancy design solve the problems of data asynchrony and localization interruption caused by single sensor failure in traditional solutions, significantly improving the stability of data input.

[0017] At the dynamic perception and localization level, the improved target detection algorithm, combined with an attention mechanism, enhances the recognition capability of dynamic small targets and high-speed targets. Optical flow method accurately calculates target motion parameters, achieving effective separation of dynamic and static features. The feature matching sub-network of the Transformer architecture and the pose estimation sub-network of the gated LSTM work together to strengthen the correlation and temporal fusion of static environment features, suppressing interference from dynamic targets, and significantly improving localization accuracy and stability. The visual localization accuracy optimization unit further corrects accumulated errors through reference feature point calibration, solving the pain point of easy drift in traditional localization.

[0018] At the path planning and navigation control level, a hierarchical cost map integrates static obstacle information with dynamic target trajectory prediction. An improved A* algorithm combined with B-spline smoothing generates paths that satisfy both obstacle avoidance requirements and kinematic constraints. A dynamic obstacle threat assessment mechanism enables intelligent prioritization of obstacle avoidance, and an emergency replanning algorithm ensures safe flight in unexpected scenarios. PID parameter self-tuning and wind field interference compensation technologies enhance trajectory tracking accuracy and anti-interference capabilities, resolving the problem of traditional control systems being susceptible to external environmental influences.

[0019] In terms of system reliability and practicality, the dynamic and static separation storage of the global map, along with the local re-optimization strategy, balances map accuracy and storage efficiency, supporting error correction during long-term flight. The system status monitoring and fault-tolerant modules enable real-time identification and redundancy switching of sensor faults, while the alarm mechanism ensures timely fault handling. Overall, this system is adaptable to various complex dynamic scenarios, providing core technical support for autonomous flight of UAVs. Attached Figure Description

[0020] Figure 1 This is a schematic block diagram of the intelligent visual positioning and navigation system for UAVs in dynamic scenes based on deep learning proposed in this invention. Figure 2 A comparison chart of the effects of multi-source sensor data fusion; Figure 3 Radar chart comparing dynamic target recognition and obstacle avoidance performance; Figure 4 Performance evolution diagram of deep learning localization model; Figure 5 Comparison chart of robustness for navigation in complex environments. Detailed Implementation

[0021] 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.

[0022] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0023] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0024] Reference Figures 1 to 5 A deep learning-based intelligent visual positioning and navigation system for dynamic scenes of unmanned aerial vehicles (UAVs) includes the following modules: The multi-source sensor data acquisition module is equipped with a binocular vision camera, LiDAR, MEMS inertial navigation module, millimeter-wave radar, and multi-mode GPS / BeiDou module. The binocular vision uses a global shutter CMOS sensor with a sampling frame rate of 25-30fps, a resolution of 1280×720-1920×1080, a baseline distance of 10-15cm, and a lens field of view of 110°-130° horizontally and 60°-80° vertically. The LiDAR is a 16-32 line mechanical radar with a sampling rate of 8-12Hz, a detection range of 0-50m, and an angular resolution of 0.1°-0.3°. ×0.3°-0.5°; MEMS inertial navigation module sampling rate 80-120Hz, accelerometer range ±12-±20g, gyroscope range ±1500°-±2500° / s; millimeter-wave radar detection range 0-20-40m, velocity measurement range -15 to 15m / s; multi-mode GPS / BeiDou module positioning accuracy centimeter level, update rate 8-12Hz, each sensor achieves hardware synchronization through a synchronization trigger board, acquisition time deviation ≤0.5ms, synchronously acquires environmental stereo images, 3D point clouds, motion parameters, obstacle distance and velocity, and absolute position information; The data preprocessing module includes an image preprocessing unit that first performs distortion correction using the camera intrinsic parameter matrix (including focal length, principal point coordinates, and distortion coefficients), then performs exposure compensation based on the Retinex algorithm to correct the brightness of backlight and shadow areas, and uses bilinear interpolation to scale the image to 512×512-640×640 to achieve scale normalization. For motion-blurred images, a blind deconvolution algorithm is used to restore details. The inertial navigation preprocessing unit uses Allan variance analysis to determine the zero-bias parameters, Kalman filtering to remove high-frequency vibration noise, and wavelet transform to decompose the signal and remove low-frequency drift errors. The radar data processing unit converts the lidar point cloud into XYZ coordinate format, removes ground noise through pass-through filtering, and calculates the relative distance, azimuth, and radial velocity of obstacles from the millimeter-wave radar data. The multi-source data fusion unit, based on a factor graph model and using visual data timestamps as a reference, performs temporal interpolation and spatial registration on inertial navigation and radar data to generate a spatiotemporally synchronized observation sequence. The dynamic feature extraction module employs an improved YOLOv5 algorithm in its dynamic target detection subunit. An SE channel attention module is added after the backbone network CSPDarknet53, and a spatial attention mechanism is introduced through the neck FPN+PAN structure. This weighted approach highlights dynamic target features such as vehicles and pedestrians, with a confidence threshold of 0.4-0.6 and a non-maximum suppression IoU threshold of 0.2-0.4. The module outputs the target category, bounding box coordinates, and confidence score. The motion parameter calculation subunit uses the pyramid LK optical flow method to extract 80-200 Shi-Tomasi corner points within the target bounding box. It tracks the displacement of corner points in adjacent frames, filters out outliers with tracking errors greater than 2 pixels, calculates the target pixel motion velocity, and converts it to world coordinate system velocity. The static environment feature extraction subunit uses a ResNet18 network to extract 512-1024-dimensional deep semantic features from the image after removing dynamic targets, including static structural features such as roads, buildings, and trees. The feature fusion subunit associates dynamic target information with static environment features, constructing a feature map containing target ID, motion trajectory, category attributes, and environmental semantic labels. The deep learning localization model module employs a Transformer architecture for its feature matching subnetwork. It takes the feature maps of the current frame and 8-12 historical keyframes as input, calculates intra-frame feature weights through multi-head self-attention, and achieves cross-frame feature association through multi-head cross-attention. It calculates the mutual information entropy between feature points, retains reliable matching pairs with mutual information entropy < 0.3, and outputs the coordinates of the matched feature points. The pose estimation subnetwork consists of an LSTM and fully connected layers. The LSTM takes the temporal sequence of the matched feature point coordinates and inertial navigation motion parameters as input, sets 128-256 hidden units, and suppresses dynamic feature interference through a gating mechanism. It outputs the UAV's six-DOF pose (x, y, z coordinates and roll, pitch, and yaw angles), and combines GPS / BeiDou data to achieve absolute pose calibration. The dynamic path planning module employs a hierarchical structure in its cost map construction subunit. The bottom static cost layer marks obstacle areas based on LiDAR point clouds, with a cost value of 0-255, and a value of 255 for obstacle areas. The middle dynamic cost layer integrates predicted dynamic target trajectories, allocating dynamic costs based on predicted location and timestamps. The cost for collision areas within the next 1 second is set at 180-220, and for 1-3 seconds at 130-170. The top smoothing cost layer allocates costs based on path curvature, increasing the cost by 40-60% for curvature > 1.5 rad / m. The path search subunit uses an improved A* algorithm, incorporating dynamic threat weights through a heuristic function. The search step size is set at 0.4-0.6 m, and after generating the initial path, it is smoothed using B-spline curves, constraining path curvature < 2 rad / m and turning angle < 25°-35° / m to meet UAV kinematic constraints. The dynamic update subunit refreshes the cost map and path every 150-250 ms, triggering emergency path replanning when the dynamic target trajectory change exceeds 0.4-0.6 m. The navigation control module's command conversion subunit discretizes the planned path into a sequence of key waypoints, calculates the desired speed and attitude between adjacent waypoints, and converts them into motor control commands. The PID control subunit employs an incremental PID algorithm with a position loop proportional coefficient of 0.7-0.9, an integral coefficient of 0.1-0.3, and a derivative coefficient of 0.05-0.15, and a speed loop proportional coefficient of 1.0-1.4, an integral coefficient of 0.2-0.4, and a derivative coefficient of 0.1-0.2. The coefficients are dynamically adjusted based on flight errors using a parameter self-tuning algorithm. The anti-interference compensation subunit is equipped with a miniature wind speed sensor to collect wind speed (0-20 m / s) and wind direction (0-360°), establishes a wind field interference model, calculates the interference torque, and outputs compensation commands to correct motor speed and counteract the wind field's influence on attitude. The execution drive subunit uses a brushless motor controller with a response frequency of 40-60 Hz, converting control commands into motor speed signals. The global map construction module employs an incremental SLAM algorithm in its point cloud fusion subunit. Every 15-25 frames of image acquisition triggers a keyframe selection process, where the selection criteria are a pose change >0.2-0.4m or a rotation angle >4°-6°. The keyframe image features are fused with LiDAR point clouds to generate a dense 3D point cloud map. The map storage subunit uses a dynamic and static feature separation strategy. Static structural features are compressed and stored using an octree at a resolution of 0.05-0.15m, while dynamic target trajectories are stored in a time-series database, recording the target ID, timestamp, and location coordinates. The local re-optimization subunit uses a sliding window algorithm, with each window containing 5-15 keyframes. When a new keyframe is added, bundle adjustment is used to optimize the pose and feature point coordinates within the window, correcting accumulated errors. The map update subunit performs coordinate calibration of the global map every 4-6 minutes or when the flight distance exceeds 80-120m, fusing GPS data. The system status monitoring and fault tolerance module includes a sensor monitoring subunit that collects real-time operating parameters of each sensor, including camera frame rate, radar point cloud quantity, inertial navigation data integrity, and GPS signal strength. When these parameters deviate from the normal range (camera frame rate < 20-25 fps, radar point cloud < 800-1200 points / frame, GPS signal strength < 35-45 dB), an anomaly is marked. The fault tolerance processing subunit employs a multi-modal data redundancy strategy. If the binocular camera fails, it enables fusion positioning of the LiDAR point cloud and inertial navigation; if GPS fails, it relies on SLAM for relative positioning. Through multi-source data cross-validation, error correction is triggered when the positioning error is > 0.4-0.6 m. The alarm subunit issues warnings via audible and visual alarms and simultaneously transmits the fault type (sensor failure, positioning drift, etc.) and current pose information to the ground station.

[0025] This invention also includes a dynamic obstacle threat assessment unit. This unit first predicts the position sequence of the dynamic target within the next 2-4 seconds using Kalman filtering, and then calculates relative parameters based on the current pose of the UAV. Calculate the threat level, where T is the obstacle threat level; The distance weighting coefficient is 0.4-0.6. The speed weighting coefficient is 0.2-0.4. The size weighting coefficient is 0.1-0.3, and + + =1; d is the normalized value of the relative distance between the drone and the obstacle (0-1). It is the relative velocity; The maximum speed of the drone is 4-6 m / s; s is the ratio of the obstacle size to the drone size; the dynamic path planning module sorts the obstacle avoidance priority according to the T value. When T>0.6, emergency obstacle avoidance is triggered, reducing the drone speed to 1-3 m / s and shifting it laterally by 0.8-1.2 m.

[0026] This invention also includes a visual positioning accuracy optimization unit, which selects ground landmarks (such as traffic signs and streetlights) as reference feature points after feature matching, and then... Calculate the positioning error compensation value, where is the position error compensation amount; n is the number of reference feature points (≥4-6). The confidence score of the i-th feature point (determined by the matching similarity, 0-1); This represents the actual location of the feature point on the global map. The predicted location of the feature point in the current frame; The fully connected layer of the input pose estimation subnetwork performs linear correction on the output six-degree-of-freedom pose, reducing positioning drift caused by dynamic disturbances.

[0027] In this invention, the synchronization trigger board of the multi-source sensor data acquisition module is implemented using an FPGA chip, outputting four synchronization pulse signals (one each for the camera, LiDAR, inertial navigation, and millimeter-wave radar), with a pulse width of 8-12μs and a period of 30-40ms (matching the camera frame rate). The camera lens is equipped with an electric ND filter, and the ambient light intensity (10-100000 lux) is collected by a light sensor. When the light intensity is <100 lux, the transmittance is adjusted to 70%-90%, and when the light intensity is >10000 lux, it is adjusted to 10%-30%. The lens is equipped with an autofocus module with a focusing distance of 0.3-0.7m to infinity and a response time of <100ms. The GPS / BeiDou module supports multi-satellite positioning and can simultaneously receive GPS L1 / L2 and BeiDou B1 / B2 signals. In obstructed scenarios, inertial navigation is used for assisted positioning. When positioning is interrupted, the pose is inferred through inertial navigation. When the interruption is <5s, the positioning error is <0.3m.

[0028] In this invention, the blind deconvolution algorithm of the data preprocessing module adopts the Richardson-Lucy iterative method, with the number of iterations set to 10-25. The point spread function is estimated by the gradient of the edge of the blurred region in the image. The direct filtering of the LiDAR point cloud is set to the Z-axis range of 0.05-0.15m to 45-55m to remove ground (Z < 0.05-0.15m) and high-altitude noise (Z > 45-55m). Then, outliers are removed by statistical filtering, with the number of neighborhood points set to 10-20 and the standard deviation multiple set to 1.2-1.8. The factor graph spatiotemporal calibration algorithm uses the data of each sensor as factor nodes and the pose as variable nodes. The factor graph is optimized by the Gauss-Newton method to make the time deviation of multi-source data < 0.1ms and the spatial registration error < 0.05m.

[0029] In this invention, the compression ratio of the SE channel attention module in the dynamic feature extraction module is set to 12-20. Channel feature weights are obtained through global average pooling, and the target feature channels are assigned a weight of 1.1-1.6 times. The pyramid LK optical flow method uses 4-6 layers of Gaussian pyramids, with the image size halved for each layer. The feature point tracking window size is 12×12-18×18 pixels. Targets with a motion speed > 4-6 pixels / frame are marked as high-speed dynamic targets, and their trajectories are recorded separately. The ResNet18 network adds a feature visualization interface after the conv4_2 layer to output a feature heatmap for real-time monitoring of the static feature extraction effect. Regions with a feature response value < 0.2 are marked as blurred regions.

[0030] In this invention, the Transformer architecture of the deep learning localization model module has 6-10 attention heads, a hidden layer dimension of 256-768, a feedforward network dimension of 1024-3072, and uses the ReLU activation function. Mutual information entropy is used to calculate SIFT descriptors based on feature points, and similarity is measured by Kullback-Leibler divergence. After matching pairs are filtered, the Random Sampling Consensus Algorithm (RANSAC) is used to remove mismatches, with 80-120 iterations and an inlier threshold of 1-3 pixels. The LSTM gating mechanism outputs 0-1 weights through the sigmoid activation function, assigning low weights of 0.1-0.3 to feature points in dynamic target regions and high weights of 0.7-1.0 to feature points in static regions. The model is trained using the COCO dataset and a custom dynamic scene dataset (containing 80,000-120,000 frames of images and 40,000-60,000 sets of pose data), with a training batch size of 12-20, 40-60 iterations, and a localization error of <0.2m.

[0031] In this invention, the dynamic target prediction of the dynamic path planning module adopts extended Kalman filtering. The state vector includes position (x, y, z) and velocity (vx, vy, vz), and the observation vector is the relative distance and velocity measured by radar. The diagonal elements of the process noise covariance matrix are set to 0.005-0.015, and the diagonal elements of the observation noise covariance matrix are set to 0.04-0.06. The prediction step size is 0.1-0.3s. B-spline curve smoothing adopts cubic B-splines, and the number of control points is 1.2-1.8 times the number of initial path nodes. The deviation between the smoothed path and the initial path is <0.3m. Emergency path replanning adopts the fast random tree algorithm (RRT*), with a search radius of 8-12m and a sampling frequency of 8-12Hz to ensure that an obstacle avoidance path is generated within 80-120ms.

[0032] In this invention, the parameter self-tuning algorithm of the navigation control module adopts the relay feedback method. When the attitude error is >4°-6° or the position error is >0.8-1.2m, the dynamic parameters of the system are identified by the relay characteristics, and the PID coefficient is adjusted with an adjustment step size of 0.04-0.06. The wind field interference model is based on the aerodynamic parameters of the UAV (lift coefficient, drag coefficient) and wind speed and direction to calculate the interference force in the X, Y, and Z axis directions and the interference torque in the roll, pitch, and yaw directions. The compensation command is superimposed on the PID output. When the wind speed is >8-12m / s, the wind resistance mode is activated to limit the flight speed to <2-4m / s. The brushless motor controller supports PWM signal input with a frequency of 40-60Hz and a speed control accuracy of ±5rpm. It is equipped with overcurrent and overvoltage protection functions and automatically cuts off the power output in case of failure.

[0033] In this invention, the octree compression of the global map construction module divides the three-dimensional space into cubic voxels with side lengths of 0.05-0.15m, storing only voxels containing point clouds, achieving a compression ratio of 8-12:1; the time-series database uses InfluxDB, with a storage frequency of 0.5-1.5Hz, retaining dynamic target trajectory data from the most recent 25-35 days, and supporting queries by target ID and time range; the sliding window algorithm adaptively adjusts the window size according to the dynamic level of the scene, setting 12-18 keyframes for static scenes (<2 dynamic targets) and 4-6 keyframes for dynamic scenes (>5 dynamic targets); the bundle adjustment method optimizes variables including the poses of all keyframes and the 3D coordinates of feature points within the window, stopping iteration when the optimization residual is <0.5 pixels; map calibration uses the ICP algorithm to align the SLAM point cloud with GPS markers, with a calibration error of <0.1m.

[0034] The following two examples further illustrate the specific implementation of this system: Example 1: Application of drone inspection in urban blocks This embodiment targets the inspection scenario of urban main roads and commercial areas. In this scenario, dynamic targets are dense (including pedestrians, motor vehicles, and non-motor vehicles), and the lighting changes frequently. It is necessary to focus on ensuring the accuracy of dynamic target recognition and the timeliness of obstacle avoidance. The specific implementation is as follows.

[0035] 1. Detailed Implementation of Technical Solutions Multi-source sensor data acquisition module: Equipped with a binocular vision camera (global shutter CMOS, 28fps, 1920×1080 resolution, baseline distance 12cm, field of view 120°×75°), a 16-line lidar (10Hz sampling rate, 0-50m detection range, 0.2°×0.4° resolution), a MEMS inertial navigation system (100Hz sampling rate, ±16g accelerometer, ±2000° / s gyroscope), a millimeter-wave radar (0-30m detection range, -15 to 15m / s velocity measurement), and multi-mode GPS / BeiDou (10Hz update rate, centimeter-level accuracy). The synchronization trigger board uses an FPGA chip, outputting 4 channels of 8μs pulse signals with a period of 36ms to match the camera frame rate, ensuring an acquisition deviation ≤0.5ms. The camera lens is equipped with a motorized ND filter, with 80% transmittance when illumination is <100lux and 20% when illumination is >10000lux, and an autofocus response time of 90ms.

[0036] Data preprocessing module: Image distortion correction is performed using an intrinsic parameter matrix (focal length 800px, principal point 960×540px), Retinex algorithm is used to compensate for backlight brightness, and bilinear interpolation is used to reduce the image size to 640×640. Motion-blurred images are deblurred using the Richardson-Lucy iterative method (18 iterations). Inertial navigation data is zero-biased using Allan variance analysis, Kalman filtering is used to remove vibration noise, and wavelet transform is used to remove drift. LiDAR point cloud data undergoes a 0.1-50m Z-axis pass-through filter and statistical filtering (15 neighborhood points, 1.5 times the standard deviation) to remove outliers. Multi-source data is fused using factor graphs, and Gauss-Newton optimization is used to ensure a time deviation <0.1ms and a spatial registration error <0.05m.

[0037] Dynamic feature extraction module: Improved YOLOv5's SE channel attention compression rate by 16, assigned a 1.4x weight to the vehicle feature channel, set a confidence threshold of 0.5, and an IoU threshold of 0.3. Pyramid LK optical flow method uses a 5-layer pyramid with a 15×15 tracking window to extract 150 Shi-Tomasi corner points, filtering out error points >2 pixels. ResNet18 extracts 1024-dimensional features, and the conv4_2 layer outputs a heatmap, marking blurred regions with responses <0.2. Feature fusion constructs a map containing target ID, trajectory, category, and road semantics.

[0038] The deep learning localization model module uses a Transformer with 8 attention heads, 512 hidden layers, and a 2048-dimensional feedforward network. It takes the features of the current frame and 10 historical frames as input, calculates the mutual information entropy of the SIFT descriptors, retains matching pairs with a value <0.3, and performs 100 RANSAC iterations to remove incorrect matches. The LSTM has 128 hidden units, assigning a low weight of 0.2 to dynamic region features and a high weight of 0.8 to static regions. The model is trained on a dynamic dataset of 100,000 frames, with a batch size of 16 and 50 iterations.

[0039] Dynamic path planning module: The cost map has a bottom-level obstacle cost of 255, a middle-level cost of 200 for the future 1s region and 150 for the 1-3s region, and a top-level cost increase of 50% for curvature > 1.5 rad / m. Improvements include a search step size of 0.5m, cubic B-spline smoothing (control points 1.5 times the number of nodes), and constraints of curvature < 2 rad / m and turning angle < 30° / m. Dynamic targets are predicted using extended Kalman filtering for a 3s trajectory, with a process noise covariance of 0.01, observation noise of 0.05, and a prediction step size of 0.2s. Threat assessment passed. calculate, =0.5、 =0.3、 =0.2, v max=5m / s, emergency obstacle avoidance when T>0.6 (speed drops to 2m / s, offset 1m). Path refresh every 200ms, trajectory change >0.5m triggers RRT replanning (10m search radius, 10Hz sampling, 100ms path generation).

[0040] Navigation control module: Incremental PID position loop (0.8, 0.2, 0.1), speed loop (1.2, 0.3, 0.15), relay self-tuning (step size 0.05) when attitude error > 5°. Wind speed sensor collects data, builds a model to calculate disturbance torque, and limits speed to 3m / s in wind-resistant mode when > 10m / s. Brushless motor controller with 50Hz response and ±5rpm accuracy.

[0041] Global map construction module: Keyframe selection is triggered at 20 frames (pose change > 0.3m or rotation > 5°), static features are compressed and stored using an octree at 0.1m resolution, and 30 days of dynamic trajectories are stored in a time-series database at 1Hz. A sliding window of 10 keyframes is used, and bundle adjustment optimizes residuals to <0.5 pixels. The map is calibrated using the ICP algorithm at 5 minutes or 100 meters, with an error <0.1m.

[0042] System status monitoring and fault tolerance module: Marks anomalies when camera frame rate <25fps, point cloud <1000 points / frame, and GPS <40dB. Camera failure triggers radar + inertial navigation fusion; GPS failure relies on SLAM; positioning error >0.5m triggers correction, along with audible and visual alarms and transmission of fault information to the ground station. Visual positioning accuracy optimization passed. To achieve this, n=5 traffic sign reference points, To match similarity, the pose drift is corrected.

[0043] 2. Effectiveness verification data

[0044] Table 1 Table 1 shows the data validating the advantages of this system in dynamic urban scenarios. Traditional systems suffer from poor dynamic target recognition and positioning drift, resulting in an accuracy rate of only 78% and a positioning error of 0.8m. This system, with its improved YOLOv5 and attention mechanism, achieves a recognition rate of 95%, while Transformer matching and LSTM positioning reduce the error to 0.15m. Regarding obstacle avoidance, traditional systems with a 300ms response time are prone to collisions; this system's threat assessment formula and RRT* replanning reduce the response time to 90ms. In the event of sensor failure, traditional systems have a fault tolerance rate of 65%, while this system's redundancy fusion strategy achieves 98%. Trajectory tracking deviation has decreased from 0.6m to 0.1m, thanks to PID self-tuning and wind field compensation. The data demonstrates the effectiveness of multi-module collaboration, addressing the pain points of high dynamic interference and difficult positioning in urban scenarios.

[0045] Example 2: Application of Emergency Rescue Scenarios in Factory Areas This embodiment is designed for an emergency rescue scenario in a chemical plant area. This scenario includes static obstacles such as pipelines and storage tanks, dynamic targets such as rescue vehicles and personnel, and is also subject to wind interference and dust obstruction. It is necessary to focus on ensuring anti-interference capability and positioning continuity. The specific implementation is as follows.

[0046] 1. Detailed Implementation of Technical Solutions Multi-source sensor data acquisition module: binocular camera (25fps, 1280×720 resolution, baseline distance 10cm, field of view 110°×65°), 32-line lidar (8Hz sampling rate, 0-45m detection range, 0.1°×0.3° resolution), MEMS inertial navigation (80Hz sampling rate, ±12g accelerometer, ±1500° / s gyroscope), millimeter-wave radar (0-25m detection range, -12 to 12m / s velocity measurement), and GPS / BeiDou (8Hz update rate). Synchronization trigger board pulse 10μs, period 40ms, deviation ≤0.5ms. Camera ND filter has 70% transmittance in dusty environments, autofocus from 0.5m.

[0047] Data preprocessing module: Image intrinsic parameters (focal length 600px, principal point 640×360px) correction, Retinex shadow compensation, and scaling down to 512×512 size. Blurred images are iterated 15 times using the Richardson-Lucy method. Inertial navigation is denoised using Allan variance determination, Kalman filtering, and wavelet transform. LiDAR point cloud Z-axis filtering is performed from 0.08 to 45m, followed by statistical filtering (12 neighborhood points, 1.2 times standard deviation). Factor graph fusion optimizes spatiotemporal bias.

[0048] Dynamic feature extraction module: YOLOv5 with SE compression ratio of 12, assigning 1.3 times weight to the personnel feature channel, confidence level of 0.4, and IoU of 0.25. A 4-layer LK optical flow pyramid with a 12×12 window and 100 corner points is used to filter out error points greater than 2 pixels. ResNet18 extracts 512-dimensional features and marks blurred areas.

[0049] The deep learning localization model module consists of a Transformer with 6 attention heads, 256 hidden layers, and a 1024-dimensional feedforward network. The input consists of the current frame and 8 historical frames, matching pairs with mutual information entropy < 0.3, and RANSAC iterations for 80 times. An LSTM with 128 hidden units is used, with dynamic features assigned a weight of 0.15 and static features assigned a weight of 0.75. The model is trained on an 80,000-frame dataset with a batch size of 12 and 40 iterations.

[0050] Dynamic path planning module: Cost map: bottom layer obstacle 255, middle layer 210 within 1 second, 140 from 1-3 seconds, top layer curvature > 1.5 rad / m, improvement 45%. A step size 0.4m, 3rd order B-spline smoothing, constraint curvature < 2 rad / m, turning angle < 25° / m. Dynamic target extended Kalman filter predicts 2-second trajectory, process noise 0.008, observation noise 0.04, step size 0.15s. Threat assessment ωd=0.6, ωv=0.2, ωs=0.2, v max =4m / s, speed limit 1.5m / s when T>0.6, offset 0.9m. Path refresh every 180ms, RRT triggered when change>0.4m (8m radius, 8Hz sampling, 90ms generation).

[0051] Navigation control module: PID position loop 0.7, 0.15, 0.08; speed loop 1.0, 0.2, 0.1; self-tuning (step size 0.04) when attitude error > 4°. Wind speed > 8m / s, wind-resistant mode speed limit 2.5m / s. Motor controller 45Hz response.

[0052] Global map building module: 18-frame trigger keyframes (pose change > 0.25m or rotation > 4°), octree with 0.08m resolution, and a time-series database storing 28 days of trajectory at 0.8Hz. Eight-frame sliding window with bundle adjustment optimization. 4.5min or 90m ICP calibration.

[0053] System status monitoring and fault tolerance module: Monitors for anomalies when camera frame rate <22fps, point cloud <900 points / frame, and GPS <38dB. Enables redundant fusion for sensor failures and corrects errors >0.4m. Optimizes visual positioning accuracy using n=4 tank marker points, calculating δp to correct pose.

[0054] 2. Effectiveness verification data

[0055] Table 2 Table 2 data highlights the system's advantages in complex factory environments. Traditional systems suffer from dust obstruction, resulting in image quality of only 62% and positioning stability of 68% under wind interference. This system's ND filter and deblurring algorithm improve image quality to 92%, while the wind compensation model enhances stability to 96%. Static obstacle avoidance success rate increases from 80% to 99% due to accurate marking of pipelines and tanks on the layered map. Long-duration flight error decreases from 1.2m to 0.2m thanks to sliding window optimization and map calibration. Emergency replanning success rate reaches 98%, far exceeding the traditional 70%, attributed to the RRT* algorithm and threat assessment mechanism. Data validates the system's anti-interference, stable positioning, and strong obstacle avoidance capabilities, making it suitable for complex factory rescue environments.

[0056] refer to Figure 2Traditional fusion methods suffer from low synchronization accuracy and simple algorithms, resulting in a multi-source full fusion error of 0.32m, which increases dramatically to over 0.8m when a single sensor fails. This invention, through FPGA hardware synchronization and a factor graph fusion algorithm, reduces the full fusion error to 0.09m. Furthermore, in the event of visual or GPS failure, radar-inertial navigation redundancy fusion is enabled, keeping the error below 0.22m. The figure visually demonstrates the advantages of multi-source fusion and fault-tolerant design.

[0057] refer to Figure 3 This diagram comprehensively demonstrates the advantages of the dynamic perception and obstacle avoidance module. Traditional systems, due to unoptimized target detection and a lack of accurate threat assessment, achieve only 72% accuracy in pedestrian recognition, and their obstacle avoidance response time of 300ms is prone to collisions. This invention improves upon YOLOv5 by introducing an attention mechanism, enhancing the extraction of small target features, achieving pedestrian and vehicle recognition rates of over 95%, and using the pyramid LK optical flow method to improve speed measurement accuracy to 94%. Combined with the threat assessment formula, the assessment accuracy reaches 93%, and with RRT* fast replanning, the response time is reduced to 90ms. Data confirms the effectiveness of the dynamic perception chain, solving the problems of delayed obstacle avoidance and inaccurate recognition in dynamic scenes.

[0058] refer to Figure 4 Traditional single-network models suffer from low feature matching accuracy, remaining at 0.45m after 80 rounds due to the lack of separation of dynamic features and insufficient temporal fusion. The two-stage model of this invention achieves accurate cross-frame matching through cross-attention in the Transformer architecture, while gated LSTM suppresses dynamic interference. It converges to 0.2m after 40 rounds, with an error of only 0.1m after 80 rounds. Feature matching accuracy increases to 96% with each training round, demonstrating the effectiveness of the "feature purification + temporal enhancement" model design and solving the core pain points of traditional localization, such as easy drift and low accuracy.

[0059] refer to Figure 5 Traditional systems are affected by strong light and dust, resulting in positioning stability of only 55% and trajectory deviation of 0.8m in mixed environments. This invention improves image quality by using an electric ND filter to adapt to lighting conditions and blind deconvolution to recover blurred images. Combined with a wind field interference compensation model and dynamic feature suppression, the stability in mixed environments reaches 93%. Trajectory tracking deviations are controlled within 0.2m, thanks to PID parameter self-tuning and path smoothing optimization. Data verification validates the effectiveness of the system's end-to-end anti-interference design, solving the problem of poor robustness of traditional navigation systems in complex environments.

[0060] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent visual positioning and navigation system for dynamic scenes of unmanned aerial vehicles (UAVs), characterized in that, include: The multi-source sensor data acquisition module is equipped with a binocular vision camera, LiDAR, MEMS inertial navigation module, millimeter-wave radar and multi-mode GPS / BeiDou module. Each sensor is synchronized through a synchronization trigger board to acquire environmental stereo images, 3D point clouds, motion parameters, obstacle distance and speed and absolute position information. The data preprocessing module includes an image preprocessing unit that performs distortion correction and scale normalization, and enables blind deconvolution algorithm for blurred images; the inertial navigation preprocessing unit uses Allan variance analysis to determine the zero bias parameters and Kalman filtering to remove noise. The radar data processing unit removes ground noise through pass-through filtering and converts radar data into obstacle information; the multi-source data fusion unit registers data based on the factor graph model and generates a spatiotemporal synchronization sequence. The dynamic feature extraction module and the dynamic target detection subunit use an improved YOLOv5 algorithm to output target information; The motion parameter calculation subunit uses the pyramid LK optical flow method to calculate and convert velocity; the static environment feature extraction subunit uses the ResNet18 network to extract semantic features. feature Feature maps are constructed by fusing sub-units; The deep learning localization model module uses a Transformer architecture to output the coordinates of matching feature points in the feature matching subnetwork; the pose estimation subnetwork consists of LSTM and fully connected layers, which outputs the UAV's six-DOF pose and combines GPS / BeiDou calibration. The dynamic path planning module uses a hierarchical structure to allocate costs in its cost map construction sub-unit. The path search sub-unit is based on the improved A* algorithm, and the generated path is smoothed by B-spline curves; the dynamic update sub-unit refreshes the path periodically, and replans when the trajectory change exceeds the threshold. The navigation control module includes a command conversion subunit that discretizes the planned path into a sequence of waypoints and converts it into motor control commands; a PID control subunit that uses an incremental PID algorithm to adjust the coefficients through a parameter self-tuning algorithm; an anti-interference compensation subunit equipped with a wind speed sensor that outputs compensation commands; and an execution drive subunit that uses a brushless motor controller to convert control commands into speed signals. The global map building module uses an incremental SLAM algorithm to generate a 3D point cloud map in the point cloud fusion sub-unit; the map storage sub-unit uses a dynamic and static separation strategy to store data; the local re-optimization sub-unit uses a sliding window algorithm to optimize pose; and the map update sub-unit fuses a GPS-calibrated map. The system status monitoring and fault tolerance module consists of a sensor monitoring subunit that collects parameters and marks anomalies; a fault tolerance processing subunit that uses a redundancy strategy to locate and correct errors; and an alarm subunit that uses an audible and visual alarm to alert the ground station and transmit fault information.

2. The deep learning-based intelligent visual positioning and navigation system for UAVs in dynamic scenes according to claim 1, characterized in that, It also includes a dynamic obstacle threat assessment unit, which first predicts the position sequence of dynamic targets within the next 2-4 seconds using Kalman filtering, and then calculates relative parameters based on the current pose of the UAV. Calculate the threat level, where T is the obstacle threat level; This is the distance weighting coefficient. For speed weighting coefficient, This is the size weighting coefficient, and + + =1; d is the normalized value of the relative distance between the drone and the obstacle; It is the relative velocity; The maximum speed of the drone is 4-6 m / s; s is the ratio of the obstacle size to the drone size; The dynamic path planning module sorts obstacle avoidance priorities according to the T value. When T > 0.6, emergency obstacle avoidance is triggered, reducing the drone speed to 1-3 m / s and shifting it laterally by 0.8-1.2 m.

3. The deep learning-based intelligent visual positioning and navigation system for UAVs in dynamic scenes according to claim 1, characterized in that, It also includes a visual positioning accuracy optimization unit, which selects ground landmarks as reference feature points after feature matching, and then... Calculate the positioning error compensation value, where The position error compensation amount is n; n is the number of reference feature points. Let be the confidence level of the i-th feature point; This represents the actual location of the feature point on the global map. The predicted location of the feature point in the current frame; The fully connected layer of the input pose estimation subnetwork performs linear correction on the output six-DOF pose.

4. The intelligent visual positioning and navigation system for UAV dynamic scenes based on deep learning according to claim 1, characterized in that, The synchronization trigger board of the multi-source sensor data acquisition module is implemented using an FPGA chip and outputs four synchronization pulse signals; the camera lens is equipped with a motorized ND filter, collects ambient light intensity through a light sensor, and the lens is equipped with an autofocus module with a focusing distance of 0.3-0.7m to infinity; The GPS / BeiDou module supports multi-satellite positioning, simultaneously receiving GPS L1 / L2 and BeiDou B1 / B2 signals. In obstructed scenarios, it enables inertial navigation-assisted positioning. When positioning is interrupted, the pose is inferred through inertial navigation. If the interruption is less than 5 seconds, the positioning error is less than 0.3m.

5. The intelligent visual positioning and navigation system for UAV dynamic scenes based on deep learning according to claim 1, characterized in that, The blind deconvolution algorithm in the data preprocessing module adopts the Richardson-Lucy iterative method, with the number of iterations set to 10-25. The point spread function is estimated by the gradient of the edge of the blurred region in the image. The direct filtering of the LiDAR point cloud is set to the Z-axis range of 0.05-0.15m to 45-55m to remove ground and high-altitude noise, and then outliers are removed by statistical filtering. The factor map spatiotemporal calibration algorithm uses the data of each sensor as factor nodes and the pose as variable nodes, and optimizes the factor map by the Gauss-Newton method.

6. The deep learning-based intelligent visual positioning and navigation system for UAVs in dynamic scenes according to claim 1, characterized in that, The SE channel attention module of the dynamic feature extraction module has a compression ratio of 12-20. Channel feature weights are obtained through global average pooling, and the target feature channels are assigned a weight of 1.1-1.6 times. The pyramid LK optical flow method uses a 4-6 layer Gaussian pyramid. Targets with a motion speed > 4-6 pixels / frame are marked as high-speed dynamic targets, and their trajectories are recorded separately. A feature visualization interface is added after the conv4_2 layer of the ResNet18 network to output a feature heatmap and feature response values. Regions with a value <0.2 are marked as fuzzy regions.

7. The intelligent visual positioning and navigation system for UAV dynamic scenes based on deep learning according to claim 1, characterized in that, The Transformer architecture of the deep learning localization model module has 6-10 attention heads and uses the ReLU activation function. Mutual information entropy is used to calculate SIFT descriptors based on feature points, and similarity is measured by Kullback-Leibler divergence. After matching pairs are filtered, a random sampling consensus algorithm is used to remove false matches. The LSTM gating mechanism outputs 0-1 weights through the sigmoid activation function. The model is trained using the COCO dataset and a custom dynamic scene dataset.

8. The intelligent visual positioning and navigation system for UAV dynamic scenes based on deep learning according to claim 1, characterized in that, The dynamic target prediction in the dynamic path planning module uses extended Kalman filtering. The state vector includes position (x, y, z) and velocity (vx, vy, vz), and the observation vector is the relative distance and velocity measured by radar. The diagonal elements of the process noise covariance matrix are set to 0.005-0.015, and the diagonal elements of the observation noise covariance matrix are set to 0.04-0.

06. B-spline curve smoothing uses cubic B-splines, and the number of control points is 1.2-1.8 times the number of initial path nodes. Emergency path replanning uses the fast random tree algorithm with a search radius of 8-12m and a sampling frequency of 8-12Hz.

9. The deep learning-based intelligent visual positioning and navigation system for UAVs in dynamic scenes according to claim 1, characterized in that, The parameter self-tuning algorithm of the navigation control module adopts the relay feedback method. When the attitude error is >4°-6° or the position error is >0.8-1.2m, the dynamic parameters of the system are identified by the relay characteristics, and the PID coefficient is adjusted with an adjustment step size of 0.04-0.

06. The wind field interference model is based on the aerodynamic parameters of the UAV and the wind speed and direction to calculate the interference force in the X, Y, and Z axis directions and the interference torque in the roll, pitch, and yaw directions. The compensation command is implemented by superimposing it on the PID output. When the wind speed is >8-12m / s, the wind resistance mode is activated to limit the flight speed to <2-4m / s. The brushless motor controller supports PWM signal input and is equipped with overcurrent and overvoltage protection functions. In case of failure, the power output is automatically cut off.

10. The intelligent visual positioning and navigation system for UAV dynamic scenes based on deep learning according to claim 1, characterized in that, The octree compression of the global map building module divides the 3D space into cubic voxels with side lengths of 0.05-0.15m, storing only voxels containing point clouds, achieving a compression ratio of 8-12:1; the time-series database uses InfluxDB, with a storage frequency of 0.5-1.5Hz, retaining dynamic target trajectory data from the most recent 25-35 days; the sliding window algorithm adaptively adjusts the window size according to the dynamic nature of the scene, setting 12-18 keyframes for static scenes and 4-6 keyframes for dynamic scenes; the bundle adjustment method optimizes variables including the poses of all keyframes and the 3D coordinates of feature points within the window, stopping iteration when the optimization residual is <0.5 pixels; map calibration uses the ICP algorithm to align the SLAM point cloud with GPS markers.

Images