An unmanned aerial vehicle autonomous navigation and map updating method in a high dynamic scene

Abstract

This invention relates to the field of UAV navigation technology, specifically to a method for autonomous navigation and map updating of UAVs in highly dynamic scenarios, comprising the following steps: constructing a dynamic target traction graph with moving targets as nodes and trajectory guidance weights as edges, and outputting the target traction graph structure; planning a predicted path through one or more dynamic target clusters based on the trajectory guidance weights, calculating the dynamic stability index of each segment of the predicted path, and generating a task alignment tension distribution map; for segmented paths that simultaneously satisfy low dynamic stability and high tension score, performing high-frequency local map resampling and relocalization, prioritizing the correction of unstable areas with severe target deviation; and retaining the current map state for segmented paths with high dynamic stability or low tension score. This invention significantly improves map updating efficiency and navigation direction consistency on UAV platforms with limited computing resources.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) navigation technology, and in particular to a method for autonomous navigation and map updating of UAVs in highly dynamic scenarios. Background Technology

[0002] As the reliance on autonomous drones for tasks such as smart city development, emergency response, and inspection of complex scenarios increases, the demand for drones to perform navigation and map maintenance tasks in dynamic environments with dense populations, complex traffic, or rapidly changing targets is becoming increasingly urgent. In order to ensure the stability and safety of operations in such highly dynamic scenarios, drone systems must simultaneously possess high real-time navigation capabilities and a map building mechanism with controllable accuracy, and be able to respond quickly, locally, and differently to environmental changes.

[0003] However, most existing navigation and mapping solutions are based on path planning methods using static landmarks or dense grid maps. These methods assume that the environment is static and perform path search after building a fixed map. However, in dynamic scenarios, landmarks are easily obscured or moved, leading to path failure or drift accumulation, which cannot adapt to real-world scenarios where moving targets change frequently.

[0004] Furthermore, existing SLAM or VIO systems typically update the entire map at a uniform frequency, lacking the ability to identify locally highly variable areas. This approach is resource-intensive, involves a lot of redundant processing, and is difficult to adapt to lightweight UAV platforms with limited computing and energy consumption.

[0005] Traditional systems often treat path planning and map maintenance as two independent modules, lacking a feedback loop. The stability or deviation of the current path cannot substantially affect the map update priority, resulting in lagging navigation strategies and a lack of adaptive map evolution. Summary of the Invention

[0006] This invention provides a method for autonomous navigation and map updating of unmanned aerial vehicles (UAVs) in highly dynamic scenarios. This method has dynamic guidance capabilities, measurable path quality, selective and self-optimizing map updates, enabling highly stable autonomous flight in complex dynamic scenarios.

[0007] A method for autonomous navigation and map updating of unmanned aerial vehicles (UAVs) in highly dynamic scenarios includes the following steps: S1. Construct a dynamic target traction graph structure: Collect environmental visual flow and pose data through the UAV's sensing unit, identify moving targets and static landmarks, and construct a dynamic target traction graph with moving targets as nodes and trajectory guidance weights as edges based on the target trajectory direction, frequency of occurrence and task relevance of the moving targets and the UAV navigation task, and output the target traction graph structure. S2. Execute the expected path generation based on the traction map: Spatially map the target traction map structure with the current navigation target area, identify multiple dynamic target clusters with navigation reference value through the path generation engine, plan an expected path that runs through one or more dynamic target clusters according to the trajectory guidance weight, calculate the dynamic stability index of each segment of the expected path based on the spatiotemporal stability of each node constituting the expected path and its trajectory guidance weight, and at the same time, based on the task alignment tension evaluation model, score the tension of the direction consistency between different segments in the path and the final task point, and generate a task alignment tension distribution map. S3. Multi-scale map update mechanism driven by path segment stability and tension index: The expected path is input into the map update module. For the segment path that simultaneously meets the requirements of low dynamic stability and high tension score, high frequency local map resampling and relocation are performed, and unstable areas with serious target deviation are corrected first. For the segment path with high dynamic stability or low tension score, the current map state is retained.

[0008] Optionally, the sensing unit includes a camera and an inertial measurement unit. The camera acquires the environmental visual flow, and the inertial measurement unit acquires the UAV pose data in real time. The environmental visual flow and pose data are spatiotemporally aligned to generate spatiotemporally consistent environmental perception data. The spatiotemporally consistent environmental perception data is input to the target recognition module. The target recognition module distinguishes between dynamic objects and static backgrounds through a semantic segmentation network, and identifies independent moving targets from dynamic objects through optical flow and point cloud clustering techniques, while extracting static landmarks with stable geometric features from the static background.

[0009] Optionally, for each identified moving target, its target trajectory direction and frequency of occurrence within the observation time window are calculated, and its task relevance is calculated in conjunction with the target point location of the UAV navigation task; based on the calculated target trajectory direction, frequency of occurrence and task relevance, a weighted fusion algorithm is used to generate the trajectory guidance weight corresponding to the moving target.

[0010] Optionally, S1 further includes constructing the dynamic target traction graph with each moving target as a node and the trajectory guidance weight as an edge, and outputting a target traction graph structure containing the nodes and their trajectory guidance weights.

[0011] Optionally, the spatial mapping includes mapping the nodes in the target traction map structure to the global coordinate system of the current navigation target area to form a spatialized navigation map; and using a density-based clustering algorithm, identifying multiple dynamic target clusters composed of nodes with track guidance weights greater than a preset clustering threshold in the spatialized navigation map, as dynamic target clusters with navigation reference value.

[0012] Optionally, S2 further includes planning a prospective path that runs through one or more of the dynamic target clusters on the spatialized navigation map, starting from the current position of the UAV and ending at the final task point of the navigation mission; for each segment of the prospective path, extracting the trajectory guidance weights of all nodes constituting the segment path, and calculating the statistical variance of the weights as the dynamic stability index of the segment path.

[0013] Optionally, the task alignment tension assessment model specifically includes, for each segment of the expected path, calculating the cosine of the angle between the average direction vector of the segment and the direction vector from the midpoint of the segment to the final task point, using the reciprocal of the cosine as the tension score of the current segment, and generating a task alignment tension distribution map covering the entire expected path.

[0014] Optionally, in step S3, the dynamic stability index of the expected path and its segmented paths, as well as the task alignment tension distribution map, are input into the map update module. The map update module sets a dynamic stability threshold and a tension score threshold. For segmented paths that simultaneously satisfy the condition that the dynamic stability is lower than its dynamic stability threshold and the tension score is higher than its tension score threshold, they are marked as key update segments. High-frequency local map resampling and feature-matching-based relocalization are performed on the physical space region corresponding to the key update segments to prioritize the correction of unstable areas in the environmental map where the target is severely deviated. For segmented paths that have a dynamic stability higher than their dynamic stability threshold or a tension score lower than their tension score threshold, they are marked as observation segments, and their corresponding current map state is maintained.

[0015] Optionally, S3 further includes using the repositioning accuracy of the key update segments generated during this map update process and the map state maintenance duration of the observation segments as map update results, and feeding them back to the construction process of the target traction map structure in S1 for weight correction.

[0016] Optionally, the weight correction includes dynamically reducing the trajectory guidance weight of the corresponding moving target in the key update segment based on the map update result through a weight correction function, and positively enhancing the trajectory guidance weight of the corresponding moving target in the observation segment that maintains stability, thereby completing the adaptive adjustment of the dynamic target traction map structure in the next perception cycle and starting a new round of navigation cycle.

[0017] The beneficial effects of this invention are: 1. This invention constructs a dynamic target traction graph structure and introduces a weighted fusion mechanism of three elements: target trajectory direction, frequency of occurrence, and task relevance. This gives each moving target node a quantifiable trajectory guidance weight. This mechanism can transform moving structural information in the environment into navigation constraints, enabling UAVs to use dynamic targets as temporary reference anchors for navigation. Compared to traditional path design methods that rely on static landmarks or dense grids, incorporating the dynamic target itself into the navigation decision system allows UAVs to obtain more stable guidance directions and more reliable path trend judgments in highly dynamic scenarios such as densely populated areas and intersecting vehicles.

[0018] 2. This invention proposes a dual-index system for path quality quantification: on the one hand, it evaluates the spatiotemporal consistency of the guiding target within the path through a segmented path dynamic stability index; on the other hand, it measures the consistency between the path segment and the global task target direction through a task alignment tension score. These two metrics together constitute the path segment identification standard, enabling precise identification of which path segments are unstable and exhibit significant task deviations during navigation, and marking them as critical update segments. High-frequency local map resampling and feature-matching-based relocalization are performed in these critical update segments, quickly correcting geometric drift, sparse feature mismatch, and dynamic target interference in the local map. This dual-index driven strategy avoids the high cost of indiscriminate full-map updates in traditional SLAM systems, transforming map maintenance from globally uniform updates to target-oriented, focused updates, thereby significantly improving map update efficiency and navigation direction consistency on UAV platforms with limited computing resources.

[0019] 3. This invention introduces two key feedback variables—repositioning accuracy and map state maintenance duration—during the map update phase, corresponding to the key update and observation segments respectively. A weight correction function dynamically adjusts the node weights in the traction graph. For areas with large repositioning errors, the guidance weight of the corresponding nodes is automatically reduced, weakening or even eliminating unreliable moving targets in the next path generation round. For areas that remain consistently stable, the node weights are gradually increased, giving reliable targets higher priority in path design. This adaptive feedback mechanism enables the system to continuously optimize its guidance graph structure through multiple rounds of perception, navigation, and update cycles, achieving an effect similar to cognitive convergence. Ultimately, the UAV can improve its navigation decision quality round by round in complex dynamic environments, evolving from a reliance on instantaneous perception to a cumulative self-learning navigation structure, significantly improving its long-term robustness, disturbance resistance, and environmental understanding capabilities. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the update method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the expected path generation based on the traction graph in an embodiment of the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0023] like Figures 1-2 As shown, a method for autonomous navigation and map updating of unmanned aerial vehicles (UAVs) in highly dynamic scenarios includes the following steps: S1. Construct a dynamic target traction graph structure: Collect environmental visual flow and pose data through the UAV's sensing unit, identify moving targets and static landmarks, and construct a dynamic target traction graph with moving targets as nodes and trajectory guidance weights as edges based on the target trajectory direction, frequency of occurrence and task relevance of the moving targets and the UAV navigation task, and output the target traction graph structure.

[0024] S11, Multi-source data fusion perception: Environmental visual flow is acquired via a camera, and UAV pose data is obtained in real time via an inertial measurement unit; the environmental visual flow and pose data are spatiotemporally aligned to generate spatiotemporally consistent environmental perception data, specifically: set up: The sequence of images continuously captured by the camera is ; The pose data output by the inertial measurement unit is (Includes attitude quaternions and position vectors); The spatiotemporally aligned environment perception data is then represented as: ;in, For spatiotemporally consistent environmental perception data This represents a spatiotemporal synchronization function used for linear interpolation or extrinsic parameter calibration alignment of visual frame timestamps and pose timestamps.

[0025] S12, Moving Target and Static Landmark Recognition: S11 yielded spatiotemporally consistent environmental perception data, formed by fusing camera image sequences (visual flow) and IMU pose data (UAV position and orientation) through time alignment, coordinate transformation, and other processing. The goal here is to identify two types of targets: Moving targets: Moving objects, such as vehicles, people, and animals, will affect the drone's path selection.

[0026] Static landmarks: Geometrically stable stationary objects, such as corners, lampposts, and tree trunks, can serve as anchor points and positioning references in the map. The process involves using a neural network model to perform semantic segmentation on each frame of the image, essentially labeling each pixel in the image as a pedestrian, road, building, or tree. This step divides the entire image into dynamic and static regions. Some objects may be classified as dynamic, but they are not actually moving in that particular frame. Conversely, sometimes the background may show false motion due to drone movement. To avoid misjudgments, optical flow is used to estimate the direction and speed of pixel movement in two consecutive frames. Only those regions that also appear to be moving visually are further identified as moving target candidates. The dynamic regions (pixels) in the image are then reconstructed into 3D point clouds based on camera parameters and pose, truly restoring the spatial structure from the 2D image. Clustering is then used in these 3D point clouds to identify groups of points that are clustered together and moving simultaneously as independent moving targets. Each clustered group of points is a candidate graph node, which will be used to construct a dynamic traction map later. In the semantic segmentation above, you have obtained static regions. Extract geometrically stable feature points from these regions, such as corners, obvious edges, and planar intersections. These are the most suitable landmarks for localization reference. Save them for subsequent map building and UAV autolocalization. Ultimately, you will obtain two types of recognition results: A set of independent moving targets (to be used later for composition and path guidance); A stable set of static landmarks (used for subsequent map building and relocation).

[0027] S12 specifically includes: inputting the spatiotemporally consistent environmental perception data into the target recognition module. This target recognition module distinguishes dynamic objects from static backgrounds through a semantic segmentation network, and identifies independent moving targets from dynamic objects through optical flow and point cloud clustering techniques, while extracting static landmarks with stable geometric features from the static background. Specifically, it includes: S121, Semantic Segmentation: ;in For pixel-level category masks (dynamic / static classes), For semantic segmentation tasks, it employs a mature deep convolutional neural network structure to classify pixels in the input image into two main categories: dynamic objects (pedestrians, vehicles, animals, etc.) and static background (such as ground, buildings, trees, etc.). Its encoder extracts the semantic features of the image, while the decoder restores the spatial resolution layer by layer and outputs a segmentation mask map of the same size as the original image, providing a category basis for subsequent separation of dynamic / static objects.

[0028] S122, Dynamic Region Optical Flow Estimation: ;in, Optical flow field, used to identify pixel motion. This represents the time difference between two adjacent frames. It is used to detect pixel movement trajectories. By analyzing the changes in pixel brightness over time in adjacent frames, it estimates the speed and direction of movement of objects in the image, and helps to determine whether the target is a dynamic object. It uses the classic dense optical flow method to estimate the motion vector of each pixel in consecutive frames, and combines semantic masking to extract significant optical flow response regions only in regions that are judged to be dynamic, thereby identifying target individuals with independent movement behavior.

[0029] S123, Dynamic Target Extraction: Perform point cloud clustering on dynamic regions: ;in, For the first Each independent movement goal This indicates the number of moving targets identified. This represents a 3D projection restoration function that combines camera intrinsic and extrinsic parameters. Based on camera intrinsic parameters (focal length, principal point, distortion coefficients) and extrinsic parameters (pose matrix), it is an image pixel back-projection function that projects pixels in the image, especially those in segmented dynamic regions, onto the world coordinate system to generate dense or sparse 3D point cloud data. Semantic segmentation and optical flow detection results are then synchronously mapped to space along with pose information to support subsequent point cloud clustering and target trajectory direction calculation. This represents the dynamic region selected by the semantic segmentation mask. The clustering function is called Cluster. Cluster is a target segmentation algorithm that clusters based on the spatial features of three-dimensional point clouds. It restores the dynamic region of a two-dimensional image into a three-dimensional point cloud by using the camera's intrinsic and extrinsic parameters, and uses DBSCAN's density clustering to spatially divide the continuously moving point cloud segments, thereby distinguishing multiple independent moving targets into multiple node candidates. These targets will serve as graph nodes for constructing the traction graph structure.

[0030] S124, Static Landmark Extraction: ;in, This represents the static region selected by the semantic segmentation mask. Indicates the first A static landmark, For the number of landmarks, This represents a static landmark extraction operator based on geometric stability, which includes extracting stable landmark information from a static background area to construct an environmental geometric structure reference, extracting dense descriptors based on features such as SIFT and ORB, filtering and matching a set of key points with high stability and small pose changes, and using them as static anchor points or landmarks input to the map module for localization and map updates.

[0031] S13, Node Attribute Calculation and Graph Structure Construction: The main goal is to turn each identified moving target into an intelligent navigation node on the graph and assign it a guiding capability weight with navigation value. Finally, a dynamic target traction graph structure is constructed. In the previous stage S12, several independent moving targets were identified, such as pedestrians walking, vehicles driving, or other moving objects. Each moving target has a continuous spatial position record, and we also know where the drone is at this moment and where the target point is. The navigation value of each moving target needs to be described: whether its direction of movement is consistent with the mission objective, how frequently it appears, and how close it is to the target. This involves scoring each target (calculating the trajectory guidance weight) and constructing a graph structure. This step includes extracting the overall direction of movement from the position sequence of the moving target in the most recent frames, like drawing a vector to indicate where it is moving. This directional information is crucial: if a target is moving in the direction of the mission objective, it may be worth following. A time window is set, such as the past 5 seconds. How many times has the target been identified during this period? The more frequently it is identified, the more stable its path and the better its trackability. This helps to determine whether the target is reliable. For example, a briefly appearing object may be a false detection and cannot be used as a navigation reference. The distance from the target's current position to the mission point is calculated and decayed using an exponential function. The closer the distance, the higher the relevance. This design helps the system to more favor reference targets that are closer to the mission point. Using a weighted fusion formula, the three indicators—consistency with the mission direction, frequency of occurrence, and mission relevance—are proportionally combined into a single value called the trajectory guidance weight. This weight measures the guidance value of the target node during navigation. The target is then transformed into a graph node, forming a navigation traction graph. Each moving target is treated as a node in the graph, carrying a trajectory guidance weight. These nodes are then connected based on spatiotemporal proximity, forming a graph structure. This final graph is the dynamic target traction graph. Through S13, traditional landmarks are transformed into moving targets, considering not only geometric paths but also movement trends and mission intent, laying the foundation for guidance nodes and weights in path generation. Specifically: for each identified moving target The system calculates the target trajectory direction and frequency of occurrence within the observation time window, and, combined with the target point location of the UAV navigation task, calculates its mission relevance. Based on these three attributes, a weighted fusion algorithm is used to generate trajectory guidance weights, and a dynamic target traction map structure is constructed. Specifically, this includes... S131, Trajectory Direction Calculation: Assume target in the past The spatial location of the frame is The predicted trajectory direction is: ; Indicate the target At any moment The three-dimensional position, Indicate target The unit vector of the trajectory direction of , This indicates the length of the historical frame window used for trajectory direction calculation.

[0032] S132, frequency calculation: ;in, For the goal The number of times it appears within the time window. The length of the time window; S133, Task Relevance Calculation: Let the final navigation target point of the UAV be... The current position of the moving target is The task relevance is then: ;in, This is the task relevance attenuation coefficient; S134, Track Guidance Weight Calculation ;in, For the goal of sports The weight of the trajectory guidance, This represents the unit direction vector from the current position of the drone to the mission point. These are the weighted fusion coefficients (corresponding to trajectory direction, frequency of occurrence, and task relevance, respectively). This indicates the current location of the drone.

[0033] S135, Traction Graph Structure Construction: Define a set of nodes with each moving target as a node: ; Construct a dynamic target traction graph using trajectory guidance weights as edges: ;in, This represents the potential spatiotemporal adjacency relationships between nodes. This represents the trajectory guidance weight corresponding to each node, and the final output is the target traction graph structure. .

[0034] S2. Execute the expected path generation based on the traction map: Spatially map the target traction map structure with the current navigation target area, identify multiple dynamic target clusters with navigation reference value through the path generation engine, plan an expected path that runs through one or more dynamic target clusters according to the track guidance weight, calculate the dynamic stability index of each segment of the expected path based on the spatiotemporal stability of each node constituting the expected path and its track guidance weight, and at the same time, based on the task alignment tension evaluation model, score the tension of the directional consistency between different segments in the path and the final task point, and generate a task alignment tension distribution map.

[0035] S21, Graph Structure Spatial Mapping and Cluster Recognition: The goal is to select a group of truly valuable targets for navigation guidance from all moving targets identified by the UAV, and cluster them into dynamic target clusters to serve subsequent path planning. Specifically, this includes each target node... spatial location and its trajectory guidance weight Combined, they form a new structure. That is, the target traction diagram structure all nodes Mapping this onto the global coordinate system of the current navigation target area yields a spatialized navigation map, represented as follows: ;in, Indicate target Spatial position vector, This corresponds to the trajectory guidance weight.

[0036] A density-based clustering algorithm is performed on the spatialized navigation map, based on node weights. Several dynamic target clusters were identified under the premise of: This formula represents: in the node set of the target traction diagram In the process, select all trajectory guidance weights that satisfy their nodes. Greater than the set threshold subsets Each such subset is a dynamic target cluster, and all such clusters form a set. In other words, it defines how to select a group of targets with high navigation value from a graph structure, provided that all targets (nodes) in the group have a weight higher than [a certain value]. They can then be grouped into a cluster. Multiple such clusters constitute This is an important reference for path planning. The clustering threshold for trajectory guidance weights, with a value ranging from [0.6, 0.8]. Indicates the first A dynamic target cluster.

[0037] The idea behind density clustering algorithms is that if a group of data points is spatially dense enough, they are grouped into one class (a cluster), while sparse and isolated points are considered noise or outliers. The execution process of the density clustering algorithm in this invention is as follows: 1. Input data: All data that meet the trajectory guidance weights Position coordinates of the moving target node ; 2. Density Definition: A radius is defined with each target as the center. The number of neighbors within this radius is used as a density indicator.

[0038] 3. Core point identification: If a target is located around... A point is considered a core point if it has more than MinPts neighbors within its radius.

[0039] 4. Cluster expansion: Starting from each core point, recursively connect all densely packed points in its neighborhood to expand into a complete cluster.

[0040] 5. Clustered Output: All densely connected target points form a dynamic target cluster. Points that cannot be grouped into any cluster are considered outliers and do not participate in subsequent path planning.

[0041] S22, Expected Path Planning and Stability Assessment: This part assesses the dynamic stability of each segment of the expected path, that is, to what extent the path is composed of targets with stable guidance capabilities. A stable path means less interference and higher predictability in navigation, which is beneficial for the UAV's autonomous flight in dynamic environments. The previously constructed spatial navigation map, built using a graph search algorithm, has planned an expected path starting from the current position, passing through one or more dynamic target clusters, and finally reaching the mission target point. This path can be represented as: starting from the UAV's current position, passing through a series of intermediate nodes (these nodes are previously identified moving targets), and finally reaching the mission target point. This step further analyzes this expected path, specifically dividing the entire path into several segments and then assessing the degree of fluctuation in guidance capability for each segment. This guidance capability is the trajectory guidance weight of each path node. If the guidance capabilities of all nodes on a path are relatively similar, it indicates that the path is stable and reliable. Conversely, if some nodes have strong guidance capabilities while others have weak ones, the path may have breakpoints or interference changes.

[0042] S22 specifically includes: on the spatialized navigation map, using the current position of the UAV... Starting point, navigate to the target point Using a graph search algorithm as the endpoint, plan a desired path that traverses multiple dynamic target clusters. , is represented as: ; the entire expected path It is divided into several consecutive small segments, each segment consisting of consecutive nodes, and each segment can contain 3 to 5 nodes. Each segment is called a... That is, the first Segment path, let the first segment be... The set of nodes of the segment path is Extract the guiding weights of all nodes along this path. Then, the statistical variance of these weights is calculated to see how large the fluctuations between these weights are, which serves as a dynamic stability index. Therefore, the dynamic stability index of this segmented path is: ;in, Indicates the first The dynamic stability of the path segment. Represents a node The weight of the trajectory guidance, This represents the average weight of all nodes in this segment. Indicates the first Number of nodes in the segment path. If A small value indicates that the target's navigational value along this path is stable and suitable for flight; if... The large value indicates significant changes in guidance capabilities along this path, potentially posing risks.

[0043] S23, Task Alignment Tension Analysis: Establish a task alignment tension evaluation model for each segment of the expected path. Calculate separately: S231. The average direction vector of this segmented path: This formula represents the segmented path. The average direction of the path segment is obtained by averaging the unit direction vectors of the adjacent nodes. Its function is to extract the overall travel trend or heading direction of the path segment.

[0044] S232. The direction vector from the midpoint of this segment to the final task point: This formula represents the unit direction vector formed from the midpoint of the current segment path to the final target point of the navigation task, describing the direction in which the task target is located.

[0045] S233. Tension score (reciprocal of the cosine of the direction angle): This formula evaluates the consistency between the current path segment and the mission objective direction by calculating the cosine of the angle between the path direction vector and the mission direction vector, and taking its reciprocal as the tension score.

[0046] If the directions are consistent (small angle), the cosine value is close to 1, and the score Tm is close to 1, indicating low tension and good alignment; If the direction deviates (the included angle is large), the cosine value will be close to 0, and the score Tm will increase, indicating that the tension is large and the deviation is serious.

[0047] in, Indicates the first Segmented path number The position coordinate vector of each node Indicates the first Segmented Path The midpoint position, This indicates the task alignment tension score for that path segment. This represents the angle between two direction vectors. This represents the dot product of vectors.

[0048] Finally, all segmented tension scores are summarized. Generate a task alignment tension distribution map for the entire path. ,in This represents the number of path segments.

[0049] S23 measures the deviation between the UAV's current planned path and the mission objective direction. It analyzes which part of the path is closer to the target direction and which part has deviated. This deviation is represented by tension. In summary, S23 consists of three parts: 1. First, calculate the average direction of this path itself; 2. Next, calculate the direction from the midpoint of this path to the task objective; 3. Finally, calculate the angle between the two; 4. A tension score is calculated for each segment of the final output path. Summarizing all scores yields a task alignment tension distribution map for the entire path, which can be seen on the map: Which segments are in the correct orientation (low tension); Which segments deviate significantly (high tension)?

[0050] This provides an important reference for subsequent map updates: if a path has high tension, it means that it has deviated significantly from the mission direction. Even if there seems to be a target on the path, it is not worth taking. In the map update mechanism, these areas with high tension are prioritized for updates to optimize navigation paths. This also helps drones maintain a "global sense of direction" in complex and dynamic environments.

[0051] S3. Multi-scale map update mechanism driven by path segment stability and tension index: The expected path is input to the map update module. For segment paths that simultaneously meet the requirements of low dynamic stability and high tension score, high-frequency local map resampling and relocation are performed, prioritizing the correction of unstable areas with severe target deviation. For segment paths with high dynamic stability or low tension score, the current map state is retained, and the map update result is fed back to S1 for use in correcting the trajectory guidance weight of the dynamic target in the next perception cycle, thereby starting a new round of navigation cycle.

[0052] S31, Multi-scale Updated Decision-Making and Execution: Expected Path Set of segmented dynamic stability indices Tension distribution map aligned with task Input map update module, set stability threshold With tension scoring threshold , ,like The node guidance capabilities are very consistent and usually do not require updates. If the weight fluctuates significantly, it indicates that the navigation reference target of this path segment is unstable. The intermediate value can be selected as the update trigger point to balance sensitivity and stability. When the path segment deviates from the target direction by 45°, The tension score is approximately Over 60° The score is then 2; set As a typical judgment line, it indicates that the angle between the forward direction and the target direction of the path segment exceeds 60°, and correction is required. Path segments are defined that satisfy the following conditions. For critical update sections: ; the physical space regions corresponding to these key update segments Mark the area as requiring priority updates and perform the following actions: S311. High-frequency local map resampling: Re-acquiring visual stream and IMU data at high frequency for regional... Reconstructing the map using a raster or sparse feature points within the map. This includes performing the following operations: The camera image frame rate is increased from the original 10-15 frames / second to 30-60 frames / second, while the attitude data of the inertial measurement unit (IMU) is collected simultaneously to ensure that the spatial attitude matches the visual information.

[0053] b. Continuous mapping within a time window: Set a short time window, and accumulate continuous image frames and IMU data within the window to form a local visual stream sequence.

[0054] c. Map raster reconstruction: Input the local visual sequence into a sparse SLAM or voxel mapping module, and... The existing sparse landmarks or dense grids in the area will be updated or replaced.

[0055] d. Extract sparse feature points from the re-acquired image and combine them with the relative pose calculated by the IMU to perform position correction and credibility assessment on the features in the original map, so as to eliminate the accumulation of errors caused by occlusion, mismatch, etc.

[0056] S312. Feature-based relocalization: Using image keypoint descriptors to compare with historical maps, local feature alignment is performed to update the geometric and semantic information of the region in the current map. This includes performing the following: a. Feature point detection and description: The ORB algorithm is used to extract the key point locations and their descriptors in the current image frame to obtain a sparse feature representation of the local image.

[0057] b. Historical map matching: Extract the set of historical features of the corresponding area from historical map data, and perform feature matching through descriptor distance (Euclidean distance).

[0058] c performs robust filtering on the matching results, eliminates erroneous matches, and estimates the local rigid transformation matrix (affine or homography matrix). Based on the matching relationship and the estimated spatial transformation, the coordinates of feature points in the map of the region are adjusted to correct the map geometric drift caused by dynamic targets or environmental disturbances.

[0059] S313. For other path segments that do not meet the above joint conditions. ,Right now: Mark this as an observation segment and maintain its original map state without updating. High dynamic stability indicates that the navigation value of targets along this path is consistent and highly referential, meaning these targets exist stably in space and time, appear frequently, and have consistent directions. Even if the tension score is average, these targets can still be used as temporary relay navigation points, and the map state does not need to be updated frequently, otherwise it may cause waste of system resources and unnecessary map jitter. Low tension score indicates that even if the stability is average, this path is still moving towards the mission objective. This path may not be the optimal path, but its direction is correct and it is an acceptable navigation scheme. Maintaining the map state can reduce the frequency of navigation switching and avoid frequent relocation due to slight deviations.

[0060] The above map update strategy is designed based on two core uncertainties faced by UAV navigation in dynamic environments: first, whether there are areas in the path where local guidance information fluctuates greatly and is unstable; and second, whether these areas deviate significantly from the final mission direction.

[0061] Map updates are only performed when both low dynamic stability and high tension score are met, for the following reasons: 1. Low dynamic stability indicates that the guidance capability of target nodes on the path segment fluctuates greatly. The track guidance weight of each node represents its reliability for navigation, whether the direction is consistent, whether it appears frequently, and whether it is close to the target. If the weights of each node on the path segment are very different, that is, the variance is high, it means that this path is pieced together by unstable targets, with low navigation reference value and a high probability of misleading or interfering. However, low stability alone cannot mean that the segment is necessarily harmful to the mission.

[0062] 2. A high tension score indicates that the path deviates from the mission direction. The tension score is calculated by the angle between the path direction and the mission direction. The higher the score, the more the path deviates from the target. Even if a path is stable (all nodes are consistent), it is inappropriate if it moves in the wrong direction. However, a high tension score alone cannot indicate that there are correctable details within the path. Sometimes it is just that the target density is insufficient.

[0063] Only when both problems occur simultaneously (instability within the path + severe deviation from the direction) is it worthwhile to perform map resampling and relocation at a high cost.

[0064] Otherwise, it might update areas that are not actually important or abandon a stable path too early.

[0065] High-frequency local map resampling combined with feature relocalization is suitable for the above-mentioned areas: these areas exhibit drift and deviation in navigation maps, often caused by temporary occlusion, disappearance of moving targets, or sudden target interference. High-frequency map resampling can re-collect a large amount of perception data in a short time to correct the accumulation of errors. Feature relocalization matches known static landmarks in historical maps, which helps to correct the current position and map structure. The combination of these two is an effective means of handling short-term uncertain areas, without requiring global updates, only reconstructing local reliable structures.

[0066] S32, Navigation Feedback and Weight Correction: Record each path segment in the map update process based on its update behavior and extract the following two core feedback metrics: 1. Repositioning accuracy (For critical update segments): This indicates the spatial error between reconstructed feature points and historical feature points in the local map. It shows how well the map update for this segment is performed. If the relocation error is large, it indicates a serious problem with the original target guidance for this segment, and it should not be relied upon heavily in subsequent updates.

[0067] 2. Duration of map status maintenance (For the observation segment): Indicates the duration for which this segment of the map does not need to be updated during consecutive navigation cycles. This indicates that the area has not needed updating for several consecutive cycles, suggesting that this path is highly reliable and stable.

[0068] These two metrics are fed back as map update results to graph structure construction (S13) to guide the weights of the corresponding target node tracks. Make corrections: 1. For target nodes in critical update segments: map resampling and relocation are performed, and the guiding weights of nodes on that path are "decayed" based on their positioning accuracy error, as shown below: If the error is large, it means that the navigation reference capability of the route segment is poor, so multiply it by a small factor to reduce it significantly; if the error is small, it means that although it has been updated, the map is accurate and the changes are not significant, so the weight is only slightly reduced. This can reduce the dependence on these unreliable targets in the next round of map construction and gradually eliminate the interference of disturbance points on the route. 2. Observe the target node in the segment. If it has not been updated for this path, but has not been updated for a long time (i.e., it is continuously stable), its guidance weight can be increased, and it should be given priority in relying on these stable navigation reference points, as shown below: The longer the time, the more reliable the map segment; gradually increasing the guiding weight of these nodes can increase their probability of being selected in the next path generation. in, It is a correction coefficient used to decay the guiding weight of the updated regional nodes, with a value range of [0.3, 0.7]. If you are very sensitive to relocation error, you can choose a larger value to quickly remove inferior nodes. If you can tolerate a certain error (such as walking through a crowd), you can choose a slightly smaller value. You can adjust the parameter by evaluating the correlation between error and path offset in simulation or real-world scenarios, or by using an online learning mechanism to dynamically adjust it according to the continuous error trend. This is a correction coefficient used to enhance the guiding weight of nodes in stable regions, with a value range of [0.01, 0.05]. This value should be much smaller than... To ensure that penalties for instability take precedence over incentives for stability, each enhancement should be gradual to avoid rapidly solidifying guidance weights and affecting flexibility; the duration of map state maintenance can be adjusted. Limit the observation to a fixed window, such as 10 perception cycles, and then adjust the corresponding weight accumulation effect. This ensures that the weight increase does not exceed the set upper limit (1.2 times the original value).

[0069] After the above modifications are completed, a new set of track guidance weights is formed. In the next perception cycle, the target traction map structure will be reconstructed. This enables adaptive navigation closed loop.

[0070] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0071] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for autonomous navigation and map updating of unmanned aerial vehicles (UAVs) in highly dynamic scenarios, characterized in that, Includes the following steps: S1. Collect environmental visual flow and pose data through the UAV's sensing unit, identify moving targets and static landmarks, and construct a dynamic target traction graph with moving targets as nodes and trajectory guidance weights as edges based on the target trajectory direction, frequency of occurrence and task relevance of the moving targets and the UAV navigation task, and output the target traction graph structure. S2. Spatial mapping of the target traction map structure with the current navigation target area, identification of multiple dynamic target clusters with navigation reference value through the path generation engine, planning an expected path through one or more dynamic target clusters based on the trajectory guidance weight, and calculating the dynamic stability index of each segment of the expected path based on the spatiotemporal stability of each node constituting the expected path and its trajectory guidance weight. At the same time, based on the task alignment tension evaluation model, tension scoring is performed on the directional consistency between different segments in the path and the final task point, and a task alignment tension distribution map is generated. S3. Input the expected path into the map update module. For segmented paths that simultaneously meet the requirements of low dynamic stability and high tension score, perform high-frequency local map resampling and relocation, and prioritize the correction of unstable areas with serious target deviation. For segmented paths with high dynamic stability or low tension score, retain the current map state.

2. The method for autonomous navigation and map updating of unmanned aerial vehicles in high dynamic scenarios according to claim 1, characterized in that, The sensing unit includes a camera and an inertial measurement unit. The camera acquires the environmental visual flow, and the inertial measurement unit acquires the UAV pose data in real time. The environmental visual flow and pose data are spatiotemporally aligned to generate spatiotemporally consistent environmental perception data. The spatiotemporally consistent environmental perception data is input to the target recognition module. The target recognition module distinguishes between dynamic objects and static backgrounds through a semantic segmentation network, and identifies independent moving targets from dynamic objects through optical flow and point cloud clustering techniques, while extracting static landmarks with stable geometric features from the static background.

3. The method for autonomous navigation and map updating of unmanned aerial vehicles in high dynamic scenarios according to claim 2, characterized in that, For each identified moving target, calculate its trajectory direction, frequency of occurrence within the observation time window, and, in conjunction with the target point location of the UAV navigation task, calculate its task relevance. Based on the calculated target trajectory direction, frequency of occurrence, and task relevance, a weighted fusion algorithm is used to generate the trajectory guidance weights corresponding to the moving target.

4. The method for autonomous navigation and map updating of unmanned aerial vehicles in high dynamic scenarios according to claim 1, characterized in that, S1 further includes constructing the dynamic target traction graph with each moving target as a node and the trajectory guidance weight as an edge, and outputting a target traction graph structure containing the nodes and their trajectory guidance weights.

5. The method for autonomous navigation and map updating of unmanned aerial vehicles in high dynamic scenarios according to claim 1, characterized in that, The spatial mapping includes mapping the nodes in the target traction map structure to the global coordinate system of the current navigation target area to form a spatialized navigation map; and using a density-based clustering algorithm, identifying multiple dynamic target clusters composed of nodes with track guidance weights greater than a preset clustering threshold in the spatialized navigation map, which are then used as dynamic target clusters with navigation reference value.

6. The method for autonomous navigation and map updating of unmanned aerial vehicles in high dynamic scenarios according to claim 5, characterized in that, The S2 further includes planning a prospective path that runs through one or more of the dynamic target clusters on the spatialized navigation map, starting from the current position of the UAV and ending at the final task point of the navigation mission; for each segment of the prospective path, extracting the trajectory guidance weights of all nodes constituting the segment path, and calculating the statistical variance of the weights as the dynamic stability index of the segment path.

7. The method for autonomous navigation and map updating of unmanned aerial vehicles in high dynamic scenarios according to claim 1, characterized in that, The task alignment tension assessment model specifically includes, for each segment of the expected path, calculating the cosine of the angle between the average direction vector of the segment and the direction vector from the midpoint of the segment to the final task point, using the reciprocal of the cosine as the tension score of the current segment, and generating a task alignment tension distribution map covering the entire expected path.

8. The method for autonomous navigation and map updating of unmanned aerial vehicles in high dynamic scenarios according to claim 1, characterized in that, In step S3, the dynamic stability index of the expected path and its segmented paths, as well as the task alignment tension distribution map, are input into the map update module. The map update module sets a dynamic stability threshold and a tension score threshold. For segmented paths that simultaneously satisfy the condition that the dynamic stability is lower than its dynamic stability threshold and the tension score is higher than its tension score threshold, they are marked as key update segments. High-frequency local map resampling and feature-matching-based relocalization are performed on the physical space region corresponding to the key update segments to prioritize the correction of unstable areas in the environmental map where the target is severely deviated. For segmented paths that have a dynamic stability higher than their dynamic stability threshold or a tension score lower than their tension score threshold, they are marked as observation segments, and their corresponding current map state is maintained.

9. A method for autonomous navigation and map updating of unmanned aerial vehicles in high dynamic scenarios according to claim 1, characterized in that, S3 also includes the process of feeding back the repositioning accuracy of the key update segments generated during this map update process and the map state maintenance duration of the observation segments as map update results to the construction process of the target traction map structure in S1 for weight correction.

10. A method for autonomous navigation and map updating of unmanned aerial vehicles in a high-dynamic scenario according to claim 9, characterized in that, The weight correction includes dynamically reducing the trajectory guidance weight of the corresponding moving target in the key update segment based on the map update result through a weight correction function, and positively enhancing the trajectory guidance weight of the corresponding moving target in the observation segment that maintains stability, thereby completing the adaptive adjustment of the dynamic target traction map structure in the next perception cycle and starting a new round of navigation cycle.

Images