A tethered unmanned aerial vehicle autonomous obstacle avoidance optimization method based on laser radar

Abstract

The application relates to the technical field of unmanned aerial vehicle autonomous navigation and obstacle avoidance, and discloses a tethered unmanned aerial vehicle autonomous obstacle avoidance optimization method based on a laser radar. The method performs multi-scale voxelization processing on laser radar point clouds, and utilizes an adaptive clustering unit to analyze point cloud division states in continuous multiple periods. The system identifies state transition points of the point cloud data in adjacent periods, constructs a cross-period correlation data chain, generates a stability quantitative coefficient of each clustering state by analyzing the topological characteristics, recalibrates the geometric boundaries of each cluster in the current frame according to the coefficient, thereby realizing reliable determination of obstacles, and planning a flight path according to the determination result. The method improves the accuracy of obstacle identification and the decision robustness in a dynamic and noise-containing environment.

Description

Technical Field

[0001] This invention relates to the field of autonomous navigation and obstacle avoidance technology for unmanned aerial vehicles (UAVs), specifically to an optimized method for autonomous obstacle avoidance of tethered UAVs based on lidar. Background Technology

[0002] In lidar-based autonomous obstacle avoidance systems for unmanned aerial vehicles (UAVs), real-time processing of environmental point cloud data and accurate segmentation and identification of obstacles are crucial. Current mainstream technologies typically rely on real-time analysis of the spatial geometric features of single-frame lidar point clouds. While this method performs adequately in static or simple environments, its processing results exhibit significant limitations in dynamic, complex, or noisy real-world scenarios. Disturbances in point cloud data caused by sensor noise, partial object occlusion, lighting variations, or dynamic target motion can lead to irregular jumps in clustering results between adjacent timeframes. The same physical entity may be segmented into different clusters in different frames or merged with neighboring entities, reducing the reliability of obstacle identification.

[0003] This frame-by-frame independent processing method lacks the ability to assess the continuity and consistency of obstacle presence over time. The system struggles to effectively distinguish between point cloud sets generated by real, stable obstacles and instantaneous, unstable point cloud distributions formed by temporary interference, dynamic target edges, or noise. Therefore, obstacle presence determinations and subsequent trajectory planning decisions based on unstable clustering results lack robustness and reliability, easily leading to accidental or missed obstacle avoidance by the UAV, posing a flight safety hazard. The purpose of this invention is to address the problems of unreliable obstacle recognition and poor anti-interference capabilities in existing technologies due to the lack of temporal stability assessment and dynamic optimization capabilities for point cloud clustering results. Summary of the Invention

[0004] The purpose of this invention is to provide an optimized method for autonomous obstacle avoidance of tethered unmanned aerial vehicles based on lidar, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides an optimization method for autonomous obstacle avoidance of tethered unmanned aerial vehicles (UAVs) based on lidar, the method comprising:

[0006] Receive environmental point cloud data generated by lidar scanning and perform multi-scale voxelization processing to generate multi-level point cloud subsets;

[0007] A multi-level point cloud subset is input into an adaptive clustering unit, which is a point cloud recognition unit with a dynamic hierarchical structure.

[0008] Obtain the point cloud partitioning state of the adaptive clustering unit in multiple consecutive processing cycles;

[0009] Based on the point cloud partitioning status of the current processing cycle and previous processing cycles, identify and label specific point cloud data points with state transitions in multi-level point cloud subsets.

[0010] Based on the specific point cloud data points with state transitions, establish an associated data chain that spans different processing cycles and belongs to different point cloud partition states;

[0011] The topological features of the associated data chain are analyzed, and a stability coefficient for each point cloud partition state is generated accordingly.

[0012] Based on the stability coefficient of each point cloud partition state, the boundaries of each point cloud partition state in the current processing cycle are dynamically recalibrated.

[0013] Based on the point cloud segmentation state after dynamic recalibration, the existence of obstacles is determined;

[0014] Based on the determination of the existence of obstacles, the autonomous flight path planning process is initiated.

[0015] Preferably, the step of inputting a multi-level point cloud subset into an adaptive clustering unit, wherein the adaptive clustering unit is a point cloud recognition unit with a dynamic hierarchical structure, specifically includes:

[0016] Each point cloud data point in the multi-level point cloud subset is used as the initial node;

[0017] Calculate the spatial association strength between each initial node and its neighboring initial nodes, and construct the connecting edges between nodes based on the spatial association strength.

[0018] Modularity optimization is performed on the network composed of all initial nodes and connecting edges. The network subgraph that meets the modularity optimization criterion is defined as the basic identification unit of the adaptive clustering unit.

[0019] Based on the spatial distance and density relationship between basic identification units, multiple basic identification units are merged into high-level clustering units to form the point cloud identification unit with the dynamic hierarchical structure.

[0020] Preferably, obtaining the point cloud partitioning state of the adaptive clustering unit in multiple consecutive processing cycles specifically includes:

[0021] Record the set of member data points of each point cloud recognition unit obtained by dividing the multi-level point cloud subset at the end of each processing cycle of the adaptive clustering unit;

[0022] Cross-cycle identity matching is performed on point cloud recognition units in each processing cycle. For point cloud recognition units whose member data point sets overlap exceeds a preset matching threshold in different processing cycles, the same cross-cycle identity is assigned.

[0023] With a point cloud identification unit that is assigned a cross-period identity as the core, a temporal state sequence containing a set of member data points, spatial centroid, and volume features is established. The temporal state sequence is the point cloud partitioning state.

[0024] Preferably, the identification and labeling of specific point cloud data points with state transitions within a multi-level point cloud subset specifically includes:

[0025] Compare the set of member data points of the point cloud recognition unit with the same cross-cycle identity in the current processing cycle with the set of member data points in its previous processing cycle;

[0026] Identify data points that are newly added to the member data point set of the point cloud recognition unit in the current processing cycle and mark them as positive transition points;

[0027] Data points that disappear from the set of member data points of the point cloud recognition unit in the current processing cycle but have a membership relationship in the historical processing cycles are identified and marked as negative transition points;

[0028] The positive and negative transition points are merged and collectively referred to as the specific point cloud data points with state transitions.

[0029] Preferably, establishing a correlated data chain spanning different processing cycles and belonging to different point cloud partitioning states specifically includes:

[0030] Starting from any specific point cloud data point with state transition, retrieve the point cloud identification units to which the specific point cloud data point has belonged in all historical processing cycles, and obtain the cross-cycle identity identifier of the point cloud identification unit.

[0031] Using cross-period identity identifiers as nodes and the transition time sequence of specific point cloud data points between point cloud recognition units with different cross-period identity identifiers as the connection relationship, a directed temporal path is constructed.

[0032] The directed temporal paths corresponding to all specific point cloud data points are aggregated, and the directed temporal paths that share the same starting point and ending point cross-period identity identifiers are merged to generate a directed weighted network with cross-period identity identifiers as nodes and path weights as the number of transition data points. The directed weighted network is the related data chain.

[0033] Preferably, the step of parsing the topological features of the associated data chain and generating a stability coefficient for each point cloud partition state specifically includes:

[0034] In the aforementioned related data chain, the path weights of all outgoing edges originating from the cross-cycle identity identifier of a certain point cloud identification unit in the current processing cycle are summed to obtain the output strength.

[0035] The input strength is obtained by summing the path weights of all incoming edges that terminate at the cross-cycle identity identifier.

[0036] Calculate the centrality of the cross-period identity across all nodes in the entire associated data chain;

[0037] The reciprocal of the output intensity, the ratio of the input intensity to the total number of member data points of the point cloud recognition unit in the current processing cycle, and the intermediate centrality are weighted and fused together, and the result is used as the stability coefficient of the point cloud partitioning state corresponding to the point cloud recognition unit.

[0038] Preferably, the step of dynamically recalibrating the boundaries of each point cloud partitioning state in the current processing cycle based on the stability coefficient of each point cloud partitioning state specifically includes:

[0039] For each point cloud recognition unit in the current processing cycle, obtain the stability coefficient of its point cloud partitioning state;

[0040] Point cloud recognition units with stability coefficients lower than a preset stability threshold are identified and marked as units to be calibrated.

[0041] For each unit to be calibrated, calculate the stability coefficient of all other point cloud recognition units in its spatial neighborhood, and set the neighborhood unit with the highest stability coefficient as the target neighborhood unit.

[0042] Extract the data points of the members in the cell to be calibrated that are farthest from its centroid, and calculate the centroid distance between these data points and the target neighboring cell;

[0043] If the distance is less than the maximum internal distance of the unit to be calibrated, then these member data points are removed from the member data point set of the unit to be calibrated and temporarily assigned a pending state.

[0044] Preferably, the determination of the existence of the execution obstacle specifically includes:

[0045] After completing the dynamic recalibration, obtain the final set of member data points of all point cloud recognition units in the current processing cycle;

[0046] For each final set of member data points, calculate its volume expansion rate, which is the ratio of the outer envelope volume of the point cloud recognition unit in the current processing cycle to the outer envelope volume in the previous processing cycle.

[0047] Point cloud recognition units whose volume expansion rate exceeds the dynamic expansion threshold are identified as dynamic object units.

[0048] For member data points marked as pending, they are included in the nearest dynamic object unit. If there is no dynamic object unit that meets the distance condition, the member data points marked as pending are treated as isolated noise points and removed.

[0049] Dynamic object units are mapped to obstacles in the environment, and their position and contour information are recorded.

[0050] Preferably, the autonomous flight path planning process initiated based on the obstacle presence determination result specifically includes:

[0051] Obtain the position, contour information, and stability coefficient of the corresponding point cloud recognition unit for all obstacles identified in the current processing cycle;

[0052] Based on the stability coefficient, the potential location distribution area of ​​each obstacle in the next few processing cycles is calculated, and a dynamic threat field of obstacles is generated.

[0053] By combining the real-time location of the tethered UAV, the preset flight mission path, and the current flight environment constraints, multiple collision-free candidate flight paths are generated in the flight airspace.

[0054] The final flight path is selected based on the cumulative risk value of each candidate flight path crossing the dynamic threat field of the obstacle, and the degree of similarity between each candidate flight path and the preset flight mission path.

[0055] Preferably, the step of receiving environmental point cloud data generated by lidar scanning and performing multi-scale voxelization processing to generate multi-level point cloud subsets specifically includes:

[0056] Based on the spatial bounding box of the original point cloud data of the lidar, the original point cloud data is divided into fine-grained voxel sets, medium-grained voxel sets and coarse-grained voxel sets using three different voxel sizes with different side lengths.

[0057] For each voxel set at each granularity level, calculate the average coordinates of all point cloud data points within each voxel, and use the average coordinates as the representative point of the voxel.

[0058] For fine-grained voxel sets, representative points of all voxels are directly used to form a first-level point cloud subset.

[0059] For medium-grained voxel sets, based on the spatial inclusion relationship between fine-grained voxel sets and their spatial inclusion relationships, voxels with insufficient representativeness in medium-grained voxels are split into multiple fine-grained virtual voxels and representative points are generated. These representative points are then merged with the original representative points to form a second-level point cloud subset.

[0060] For coarse-grained voxel sets, the same method as for medium-grained voxel sets is used to process them, generating third-level point cloud subsets.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] By receiving continuous, multi-cycle LiDAR point cloud data and tracking the historical changes in the clustering states of each data point, a correlated data chain spanning time frames is established. Analyzing the topological features of this data chain allows for the calculation of a quantified stability coefficient for each point cloud clustering state identified in the current frame. This coefficient reflects the consistency and persistence of the cluster in its recent history, thus distinguishing between stable, real-world obstacle clusters that remain stable over time and temporary, volatile point cloud clusters caused by noise, dynamic targets, or changes in viewpoint. This process enables online evaluation and measurement of the reliability of obstacle recognition results.

[0063] Based on the calculated stability coefficients of each cluster, the boundary judgment conditions of these clusters are dynamically adjusted and recalibrated. For clusters with high stability coefficients, their boundary judgment conditions can be adaptively relaxed, enabling them to form more complete and coherent geometric contours, characterizing their properties as reliable obstacles. For clusters with low stability coefficients, stricter boundary constraints are implemented to suppress their expansion, or they are broken down into smaller units, thereby filtering out unstable point cloud regions caused by transient disturbances. This mechanism, which uses historical stability feedback to guide the dynamic adjustment of clustering parameters in the current frame, ensures that the boundary of obstacle segmentation is no longer a static product under a fixed geometric threshold, but rather an adaptive result that dynamically evolves with scene credibility. Attached Figure Description

[0064] Figure 1 This is a schematic diagram illustrating the working principle of the tethered UAV autonomous obstacle avoidance optimization method based on lidar described in this invention.

[0065] Figure 2 A flowchart for constructing a dynamic hierarchical point cloud recognition unit;

[0066] Figure 3 A flowchart for obtaining the point cloud partitioning state in multiple consecutive processing cycles;

[0067] Figure 4 A dual-axis composite image of the dynamic calibration effect of the point cloud recognition unit;

[0068] Figure 5 A bar chart comparing two indicators for the hierarchical structure of adaptive clustering units. Detailed Implementation

[0069] 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.

[0070] Please see Figure 1 This invention provides an autonomous obstacle avoidance optimization method for tethered unmanned aerial vehicles (UAVs) based on lidar. The method includes: receiving environmental point cloud data generated by lidar scanning; performing multi-scale voxelization processing to generate multi-level point cloud subsets; inputting the multi-level point cloud subsets into an adaptive clustering unit with a dynamic hierarchical structure; obtaining the point cloud partitioning state of the adaptive clustering unit in multiple consecutive processing cycles; identifying and labeling specific point cloud data points with state transitions in the multi-level point cloud subsets based on the point cloud partitioning states of the current processing cycle and previous processing cycles; establishing a correlation data chain spanning different processing cycles and belonging to different point cloud partitioning states based on these specific point cloud data points; parsing the topological features of the correlation data chain to generate a stability coefficient for each point cloud partitioning state; dynamically recalibrating the boundaries of each point cloud partitioning state in the current processing cycle based on the stability coefficient of each point cloud partitioning state; performing obstacle existence determination based on the dynamically recalibrated point cloud partitioning state; and finally initiating an autonomous flight path planning process based on the obstacle existence determination result.

[0071] In one embodiment of the present invention, see [reference] Figure 2In a specific implementation, the environmental point cloud data generated by lidar scanning undergoes multi-scale voxelization processing to generate multi-level point cloud subsets. The process uses the spatial bounding box of the original point cloud data as a reference, dividing the original point cloud into three voxel sizes with side lengths of 0.1 meters, 0.3 meters, and 0.6 meters, respectively, to generate fine-grained, medium-grained, and coarse-grained voxel sets. For each voxel, the arithmetic mean of the three-dimensional coordinates of all point cloud data points within it is calculated, and this mean is used as the coordinates of the representative point of that voxel. In a specific implementation, for the fine-grained voxel set, the representative points of all voxels are directly collected to form the first-level point cloud subset. In some embodiments, processing of medium-grained voxel sets requires establishing a spatial mapping relationship between them and fine-grained voxel sets. When it is identified that the number of fine-grained voxels contained within a certain medium-grained voxel exceeds a preset threshold, this medium-grained voxel is deemed insufficiently representative and is split into multiple virtual fine-grained voxels with a side length of 0.1 meters. The representative point of each virtual fine-grained voxel is calculated, and these newly generated representative points are merged with the original representative points of the remaining unsplit medium-grained voxels to form a second-level point cloud subset. It can be understood that for coarse-grained voxel sets, the same processing logic as for medium-grained voxel sets is used, employing the fine-grained voxel set as a reference for representativeness evaluation and virtual splitting operations, ultimately generating a third-level point cloud subset.

[0072] In a specific implementation, the generated multi-level point cloud subset is input to an adaptive clustering unit, which is a point cloud recognition unit with a dynamic hierarchical structure. During implementation, each point cloud data point in the multi-level point cloud subset is considered an initial node. In some embodiments, the spatial association strength between each initial node and all other initial nodes in its spatial neighborhood is calculated. The spatial association strength can be calculated based on factors such as the point spacing and the angle between the point normal vectors. The spatial association strength can be quantified using a formula of the following form:

[0073]

[0074] in: Represents the initial node With the initial node Spatial correlation strength between them Represents the initial node With the initial node The Euclidean distance between them It is a proportional parameter used to control the range of influence of distance. Represents the initial node With the initial node The angle between the estimated normal vectors is calculated. After calculation, connection edges between initial nodes are constructed based on the spatial association strength value. Typically, connection edges are only established between node pairs whose spatial association strength exceeds a preset connection threshold. In specific implementation, modularity optimization is performed on the complex network composed of all initial nodes and constructed connection edges. The goal of modularity optimization is to divide the network so that the internal connections of subgraphs are dense while the connections between subgraphs are sparse. Each network subgraph that meets the modularity optimization criteria after optimization is defined as the basic recognition unit of the adaptive clustering unit. Optionally, above the basic recognition units, a merging operation is performed based on the spatial distance and point cloud density relationship between each basic recognition unit. Basic recognition units with spatial distances below the merging threshold and similar density features are aggregated into a higher-level clustering unit, thereby forming a complete dynamic hierarchical structure point cloud recognition unit.

[0075] In one embodiment of the present invention, see [reference] Figure 3 In a specific implementation, the point cloud partitioning state of the adaptive clustering unit in multiple consecutive processing cycles is obtained. This process requires recording all point cloud identification units and their corresponding member data point sets generated by the adaptive clustering unit partitioning the multi-level point cloud subset at the end of each processing cycle. In some embodiments, to construct cross-cycle temporal associations, the system matches the point cloud identification units of the current processing cycle with the point cloud identification units of historical processing cycles to assign cross-cycle identities. During matching, the degree of overlap between the member data point sets of the point cloud identification units in different processing cycles needs to be calculated. The overlap of member data point sets can be quantified using a formula of the following form:

[0076]

[0077] in: Point cloud recognition unit With point cloud recognition unit The overlap of the member data point sets, Point cloud recognition unit The set of member data points, Point cloud recognition unit Member data point set, operators This indicates the number of data points in the set. This represents the intersection of two sets of member data points. This represents the union of two sets of member data points. It can be understood that when the calculated overlap... When the preset matching threshold is exceeded, the system determines the point cloud recognition unit. With point cloud recognition unit Representing the state of the same entity at different times, and serving as a point cloud recognition unit. With point cloud recognition unit Assign the same cross-cycle identity identifier. In specific implementation, the preset matching threshold can be set to 0.6, which means that the member data point sets of two point cloud recognition units are considered to be the same entity only if they have more than 60% similarity.

[0078] In a specific implementation, for each point cloud recognition unit successfully assigned a cross-cycle identity, the system establishes a temporal state sequence centered on it. The temporal state sequence is an ordered data structure that records the state snapshots of the point cloud recognition unit corresponding to the cross-cycle identity in each cycle, according to the processing cycle sequence. Each state snapshot contains at least the following elements: the set of member data points of the point cloud recognition unit at the end of the processing cycle, the spatial centroid coordinates calculated based on the member data point set, and the volumetric feature characterizing the space occupied by the point cloud recognition unit. In a specific implementation, the spatial centroid coordinates are obtained by calculating the average of the three-dimensional coordinates of all data points in the member data point set. The volumetric feature is calculated by constructing the three-dimensional convex hull or axial bounding box of the point cloud recognition unit and calculating its volume. This temporal state sequence constitutes the point cloud partitioning state for tracking and describing the state of a potential environmental entity.

[0079] In a specific implementation, based on the established point cloud partitioning state with cross-period identifiers, specific point cloud data points with state transitions within multi-level point cloud subsets are identified and labeled. During implementation, for each point cloud identification unit with cross-period identifiers, the system compares its member data point set in the current processing cycle with its member data point set in the previous processing cycle. In some embodiments, the system uses set operations to identify data points that exist only in the member data point set of the current processing cycle but not in the member data point set of the previous processing cycle; these data points are marked as positive transition points. It can be understood that positive transition points represent data points that have recently entered the spatial region represented by the point cloud identification unit. In a specific implementation, the system also uses set operations to identify data points that exist only in the member data point set of the previous processing cycle but have disappeared from the member data point set of the current processing cycle, but which belonged to the point cloud identification unit in an earlier historical processing cycle; these data points are marked as negative transition points.

[0080] In one embodiment of the present invention, a relational data chain spanning different processing cycles and belonging to different point cloud partitioning states is established. This process begins with the identification and labeling of specific point cloud data points exhibiting state transitions. Specifically, taking any specific point cloud data point exhibiting a state transition as the starting point for retrieval, the system searches the stored time-series records for point cloud identification units to which this specific point cloud data point has belonged in all historical processing cycles, and obtains the cross-cycle identity identifiers corresponding to these point cloud identification units. In some embodiments, the retrieval scope can cover ten consecutive processing cycles prior to the current processing cycle. For each retrieved point cloud identification unit, its cross-cycle identity identifier is a unique string or number encoding used to uniquely identify a tracked potential environmental entity in the time series. In the specific implementation, the cross-period identity identifiers obtained by retrieval are used as nodes, and the time sequence of a specific point cloud data point transitioning from the point cloud identification unit corresponding to one cross-period identity identifier to the point cloud identification unit corresponding to another cross-period identity identifier is used as the connection relationship. For each transition, a directed edge is established from the source node to the target node, thereby constructing a directed temporal path that reflects the historical affiliation change of a single data point.

[0081] In a specific implementation, the system aggregates all directed temporal paths corresponding to specific point cloud data points with state transitions. It can be understood that different specific point cloud data points may have the same or similar transition paths. The system merges all directed temporal paths that share the same starting point cross-period identity identifier and the same ending point cross-period identity identifier. The fusion method involves combining these paths into a single directed edge, and setting the path weight of this directed edge to the sum of the number of specific point cloud data points corresponding to all merged paths. In practice, after path fusion processing of all specific point cloud data points, a directed weighted network is finally generated, with all relevant cross-period identity identifiers as nodes, the merged directed edges as connections, and the number of transition data points as path weights. This directed weighted network is defined as an associative data chain, which macroscopically depicts the dynamic associations between different point cloud identification units caused by the flow of point cloud members.

[0082] In a specific implementation, the topological features of the correlated data chain are parsed, and a stability coefficient for each point cloud partitioning state is generated accordingly. The topological features of the correlated data chain include the output strength, input strength, and center centrality of each node. Specifically, for a node in the correlated data chain representing a point cloud identification unit in the current processing cycle, its output strength... The calculation method is to use this node Sum the path weights of all outgoing edges from the starting point. Node Input intensity The calculation method is to use this node The path weights of all incoming edges to the destination are summed. In some embodiments, the center centrality of a node is... Measured the nodes The degree to which a node acts as a "bridge" in the entire interconnected data chain reflects the node's role in the overall data flow. The ability to control the flow of information or data, and the calculation of intermediate centrality. The following formula can be used for quantification:

[0083]

[0084] in: Represents a node The centrality of the center, and Indicates that in a relational data chain, excluding nodes Any two distinct nodes other than Indicates from node To the node The total number of shortest paths, Indicates all slave nodes To the node In the shortest path, the nodes are... The number of paths. Normalizing the summation result ensures its value falls between 0 and 1. This can be understood as the output intensity. The reciprocal of the value reflects the degree of "lost" points in the point cloud recognition unit corresponding to the node, and the input intensity. The total number of member data points of the point cloud recognition unit in the current processing cycle The ratio reflects the relative proportion of "new" points.

[0085] See Figure 4 This is a dual-axis composite chart illustrating the dynamic calibration effect of point cloud recognition units. It shows the correlation between the volume change and stability coefficient of point cloud units before and after dynamic recalibration in LiDAR point cloud processing. The "volume after calibration" for all units is smaller than the "volume before calibration," indicating that dynamic recalibration effectively reduces the redundant space of point cloud units and improves the accuracy of unit boundaries. Units 1 and 4 have stability coefficients below the threshold, corresponding to a greater reduction in volume after calibration, reflecting the logic of "prioritizing the calibration of low-stability units." This type of chart is a core analytical tool for optimizing LiDAR point cloud processing, intuitively verifying the effectiveness of dynamic recalibration; it correlates stability coefficients with calibration priority, assisting in optimizing calibration strategies; and it identifies highly redundant / low-stability point cloud units, providing more accurate basic data for subsequent obstacle detection.

[0086] In one embodiment of the present invention, the boundaries of each point cloud partition state in the current processing cycle are dynamically recalibrated based on the stability coefficient of each point cloud partition state. For each point cloud recognition unit in the current processing cycle, the stability coefficient of its point cloud partition state needs to be obtained. The stability coefficient of the point cloud partition state is calculated by the method described in the embodiment. The system presets a stability threshold to distinguish between stable and unstable units, identifies point cloud recognition units whose point cloud partition state stability coefficient is lower than the preset stability threshold, and marks these point cloud recognition units as units to be calibrated. In a specific implementation, the preset stability threshold can be set to 0.5. Referring to Table 1, a simplified example containing state data of five point cloud recognition units is shown.

[0087] Table 1: Point Cloud Recognition Unit Status Table

[0088] Point cloud recognition unit ID Stability coefficient Marking status Target Neighborhood Unit ID Maximum internal distance (meters) Unit_001 0.72 Stablize - 1.2 Unit_002 0.43 Need to be calibrated Unit_001 0.9 Unit_003 0.81 Stablize - 2.1 Unit_004 0.38 Need to be calibrated Unit_003 1.5 Unit_005 0.65 Stablize - 0.8

[0089] In a specific implementation, for each point cloud recognition unit marked as a unit to be calibrated, the system calculates the stability coefficient of all other point cloud recognition units within its spatial neighborhood. The spatial neighborhood can be defined as a spherical region with radius R centered on the centroid of the unit to be calibrated, and the neighborhood unit with the highest stability coefficient is designated as the target neighborhood unit. The system extracts a subset of member data points from the unit to be calibrated that are furthest from its own centroid. In some embodiments, this subset of member data points can be sorted by distance and the top 20% of data points are extracted. The distance between these member data points and the centroid of the target neighborhood unit is calculated. If the calculated distance is less than the maximum internal distance of the unit to be calibrated, these member data points are removed from the set of member data points of the unit to be calibrated, and a temporary "pending state" label is assigned to them. The maximum internal distance refers to the maximum value among all the distances between the member data points of the unit to be calibrated and their centroids.

[0090] In a specific implementation, the determination of the existence of obstacles needs to be performed after dynamic recalibration. The system acquires the final member data point set of all point cloud recognition units in the current processing cycle and calculates the volume expansion rate for each final member data point set. The volume expansion rate is the ratio of the outer envelope volume of the point cloud recognition unit in the current processing cycle to the outer envelope volume of the previous processing cycle. The following formula can be used for quantification:

[0091]

[0092] in: Indicates the volume expansion rate. This represents the outer envelope volume of the point cloud recognition unit in the current processing cycle. This represents the outer envelope volume of the point cloud recognition unit in the previous processing cycle. The outer envelope volume can be obtained by constructing the smallest axial bounding box of the point cloud recognition unit and calculating its volume. In some embodiments, the system presets a dynamic expansion threshold. For point cloud recognition units whose volume expansion rate exceeds the dynamic expansion threshold, they are determined to be dynamic object units. The dynamic expansion threshold can be set to 1.3, indicating that the volume expansion exceeds 30%. It can be understood that for member data points marked as pending state, the system calculates their spatial distance to all those already determined as dynamic object units and includes them in the set of member data points of the nearest dynamic object unit. If the distance between a member data point marked as pending state and all dynamic object units exceeds a preset fusion distance threshold, then the data point is regarded as an isolated noise point generated by sensor noise or instantaneous reflection and is discarded. In a specific implementation, the finally determined dynamic object units are mapped to obstacles in the environment, and the three-dimensional position coordinates and contour information of each obstacle are recorded. The contour information can be described by the outer envelope shape formed by the set of member data points of the dynamic object units.

[0093] In a specific implementation, the autonomous flight path planning process is initiated based on the obstacle existence determination result. The system acquires the position and contour information of all obstacles identified in the current processing cycle, as well as the stability coefficient of the corresponding point cloud recognition unit. In this implementation, the obstacle's position information is given by the spatial centroid coordinates of its corresponding dynamic object unit, the contour information is described by the vertex coordinates of the three-dimensional convex hull or axial bounding box formed by the set of data points of the dynamic object unit members, and the stability coefficient of the point cloud recognition unit is calculated from the previous steps. Based on the stability coefficient of the point cloud recognition unit, the potential location distribution area of ​​each obstacle in the next few processing cycles is calculated. A lower stability coefficient of the point cloud recognition unit indicates that the corresponding obstacle has active state changes and high position uncertainty. A future time window is preset during the calculation, such as the next five processing cycles. For each obstacle, extrapolation is performed along the direction of its current motion velocity vector based on its current position. The stability coefficient of the point cloud recognition unit is used to dynamically adjust the radius of the extrapolation area. The lower the stability coefficient of the point cloud recognition unit, the larger the radius of the extrapolation area is set to cover greater position uncertainty. The superposition of the extrapolation areas of all obstacles in three-dimensional space constitutes the obstacle dynamic threat field. The obstacle dynamic threat field is a probabilistic occupancy grid representation that evolves over time.

[0094] In some embodiments, multiple collision-free candidate flight paths are generated in the airspace by combining the real-time position of the tethered UAV, a preset flight mission path, and current flight environment constraints. The real-time position of the tethered UAV is provided by the fusion of onboard GPS and inertial measurement unit. The preset flight mission path is a series of pre-loaded three-dimensional waypoints. The current flight environment constraints include the length of the tether cable and physical motion constraints, the minimum turning radius of the UAV, the maximum rate of climb, and airspace altitude restrictions. A sampling-based path planning algorithm is used to generate candidate flight paths. In the space starting from the real-time position of the tethered UAV and ending at the next key waypoint of the preset flight mission path, a large number of path nodes are randomly generated and connected to form collision-free paths. Collision-free means that the path does not intersect with the grid cells marked as "occupied" in the obstacle dynamic threat field at the current moment. Several lower-cost paths are selected from all feasible paths as candidate flight paths.

[0095] It is understandable that the final flight path is selected based on the cumulative risk value of each candidate flight path traversing the dynamic threat field of obstacles and the degree of closeness of each candidate flight path to the preset flight mission path. When calculating the cumulative risk value of each candidate flight path, the path is discretized into a series of dense path points. For each path point, its risk probability value in the corresponding spatiotemporal grid within the obstacle dynamic threat field is queried. The risk probability value is positively correlated with the estimated probability that the grid cell is occupied by an obstacle. The cumulative risk value... It is an integral approximation of the risk probability values ​​of all discrete points on the track, used to calculate the cumulative risk value. The following formula can be used for quantification:

[0096]

[0097] in: Indicates candidate flight paths The cumulative risk value, This represents the total number of segments after discretizing the track. Indicates the first Spatial coordinates of each track point This indicates that the drone is estimated to reach the [number]th [location]. The time of each track point Indicates the location and time The risk probability density is obtained from the dynamic threat field of obstacles. Indicates the first The length of the flight path segment. The closeness between the candidate flight path and the preset flight mission path is measured by calculating the average distance between the two paths in three-dimensional space; the smaller the average distance, the closer the closeness. In practice, selecting the final flight path requires constructing a comprehensive evaluation function, which also considers the cumulative risk value. The proximity metric, for example, uses a weighted summation method, assigning a higher weight to the cumulative risk value to prioritize safety, and a lower weight to the proximity metric to ensure mission execution efficiency. The trajectory with the smallest comprehensive evaluation function value is selected from all candidate flight trajectories as the final flight trajectory and output to the flight control system of the tethered UAV for execution.

[0098] See Figure 5 This is a bar chart comparing two indicators of the hierarchical structure of adaptive clustering units, used to display the modularity score and quantity characteristics of units at different levels during the LiDAR point cloud clustering process. From "basic identification unit" to "dynamic hierarchical structure," the number of clustering units decreases (50→20→10), reflecting the hierarchical aggregation logic "from fine-grained to coarse-grained." The modularity score of each level is close to 0, indicating that the internal compactness of the current clustering result is relatively weak, and the spatial association strength calculation rules of the clustering algorithm may need to be optimized. This type of chart is a core tool for evaluating the performance of adaptive clustering units, intuitively displaying the aggregation effect of the hierarchical structure; and providing basic hierarchical information for the subsequent stability analysis of the point cloud partitioning state.

[0099] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0100] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A tethered unmanned aerial vehicle autonomous obstacle avoidance optimization method based on laser radar, characterized in that, Perform the following steps: Receive environmental point cloud data generated by lidar scanning and perform multi-scale voxelization processing to generate multi-level point cloud subsets; A multi-level point cloud subset is input into an adaptive clustering unit, which is a point cloud recognition unit with a dynamic hierarchical structure. Obtain the point cloud partitioning state of the adaptive clustering unit in multiple consecutive processing cycles; Based on the point cloud partitioning status of the current processing cycle and previous processing cycles, identify and label specific point cloud data points with state transitions in multi-level point cloud subsets. Based on the specific point cloud data points with state transitions, establish an associated data chain that spans different processing cycles and belongs to different point cloud partition states; The topological features of the associated data chain are analyzed, and a stability coefficient for each point cloud partition state is generated accordingly. Based on the stability coefficient of each point cloud partition state, the boundaries of each point cloud partition state in the current processing cycle are dynamically recalibrated. Based on the point cloud segmentation state after dynamic recalibration, the existence of obstacles is determined; Based on the determination of the existence of obstacles, the autonomous flight path planning process is initiated; The process of parsing the topological features of the associated data chain and generating a stability coefficient for each point cloud partition state specifically includes: In the aforementioned related data chain, the path weights of all outgoing edges originating from the cross-cycle identity identifier of a certain point cloud identification unit in the current processing cycle are summed to obtain the output strength. The input strength is obtained by summing the path weights of all incoming edges that terminate at the cross-cycle identity identifier. Calculate the centrality of the cross-period identity across all nodes in the entire associated data chain; The reciprocal of the output intensity, the ratio of the input intensity to the total number of member data points of the point cloud recognition unit in the current processing cycle, and the intermediate centrality are weighted and fused together, and the result is used as the stability coefficient of the point cloud partitioning state corresponding to the point cloud recognition unit. The step of dynamically recalibrating the boundaries of each point cloud partitioning state in the current processing cycle based on the stability coefficient of each point cloud partitioning state specifically includes: For each point cloud recognition unit in the current processing cycle, obtain the stability coefficient of its point cloud partitioning state; Point cloud recognition units with stability coefficients lower than a preset stability threshold are identified and marked as units to be calibrated. For each unit to be calibrated, calculate the stability coefficient of all other point cloud recognition units in its spatial neighborhood, and set the neighborhood unit with the highest stability coefficient as the target neighborhood unit. Extract the data points of the members in the cell to be calibrated that are farthest from its centroid, and calculate the centroid distance between these data points and the target neighboring cell; If the distance is less than the maximum internal distance of the unit to be calibrated, then these member data points are removed from the member data point set of the unit to be calibrated and temporarily assigned a pending state.

2. The method for optimizing autonomous obstacle avoidance of tethered unmanned aerial vehicles based on lidar according to claim 1, characterized in that, The step of inputting a multi-level point cloud subset into an adaptive clustering unit, wherein the adaptive clustering unit is a point cloud recognition unit with a dynamic hierarchical structure, specifically includes: Each point cloud data point in the multi-level point cloud subset is used as the initial node; Calculate the spatial association strength between each initial node and its neighboring initial nodes, and construct the connecting edges between nodes based on the spatial association strength. Modularity optimization is performed on the network composed of all initial nodes and connecting edges. The network subgraph that meets the modularity optimization criterion is defined as the basic identification unit of the adaptive clustering unit. Based on the spatial distance and density relationship between basic identification units, multiple basic identification units are merged into high-level clustering units to form the point cloud identification unit with the dynamic hierarchical structure.

3. The method for optimizing autonomous obstacle avoidance of tethered unmanned aerial vehicles based on lidar according to claim 1, characterized in that, The acquisition of the point cloud partitioning state of the adaptive clustering unit in multiple consecutive processing cycles specifically includes: Record the set of member data points of each point cloud recognition unit obtained by dividing the multi-level point cloud subset at the end of each processing cycle of the adaptive clustering unit; Cross-cycle identity matching is performed on point cloud recognition units in each processing cycle. For point cloud recognition units whose member data point sets overlap exceeds a preset matching threshold in different processing cycles, they are assigned the same cross-cycle identity. With a point cloud identification unit that is assigned a cross-period identity as the core, a temporal state sequence containing a set of member data points, spatial centroid, and volume features is established. The temporal state sequence is the point cloud partitioning state.

4. The method for optimizing autonomous obstacle avoidance of tethered unmanned aerial vehicles based on lidar according to claim 3, characterized in that, The identification and labeling of specific point cloud data points with state transitions within a multi-level point cloud subset specifically includes: Compare the set of member data points of the point cloud recognition unit with the same cross-cycle identity in the current processing cycle with the set of member data points in its previous processing cycle; Identify data points that are newly added to the member data point set of the point cloud recognition unit in the current processing cycle and mark them as positive transition points; Data points that disappear from the set of member data points of the point cloud recognition unit in the current processing cycle but have a membership relationship in the historical processing cycles are identified and marked as negative transition points; The positive and negative transition points are merged and collectively referred to as the specific point cloud data points with state transitions.

5. The method for optimizing autonomous obstacle avoidance of a tethered unmanned aerial vehicle based on lidar according to claim 4, characterized in that, The establishment of a correlated data chain spanning different processing cycles and belonging to different point cloud partitioning states specifically includes: Starting from any specific point cloud data point with state transition, retrieve the point cloud identification units to which the specific point cloud data point has belonged in all historical processing cycles, and obtain the cross-cycle identity identifier of the point cloud identification unit. Using cross-period identity identifiers as nodes and the transition time sequence of specific point cloud data points between point cloud recognition units with different cross-period identity identifiers as the connection relationship, a directed temporal path is constructed. The directed temporal paths corresponding to all specific point cloud data points are aggregated, and the directed temporal paths that share the same starting point and ending point cross-period identity identifiers are merged to generate a directed weighted network with cross-period identity identifiers as nodes and path weights as the number of transition data points. The directed weighted network is the related data chain.

6. The method for optimizing autonomous obstacle avoidance of tethered unmanned aerial vehicles based on lidar according to claim 1, characterized in that, The determination of the existence of the execution obstacle specifically includes: After completing the dynamic recalibration, obtain the final set of member data points of all point cloud recognition units in the current processing cycle; For each final set of member data points, calculate its volume expansion rate, which is the ratio of the outer envelope volume of the point cloud recognition unit in the current processing cycle to the outer envelope volume in the previous processing cycle. Point cloud recognition units whose volume expansion rate exceeds the dynamic expansion threshold are identified as dynamic object units. For member data points marked as pending, they are included in the nearest dynamic object unit. If there is no dynamic object unit that meets the distance condition, the member data points marked as pending are treated as isolated noise points and removed. Dynamic object units are mapped to obstacles in the environment, and their position and contour information are recorded.

7. The method for optimizing autonomous obstacle avoidance of tethered unmanned aerial vehicles based on lidar according to claim 1, characterized in that, The autonomous flight path planning process, initiated based on the obstacle presence determination result, specifically includes: Obtain the position, contour information, and stability coefficient of the corresponding point cloud recognition unit for all obstacles identified in the current processing cycle; Based on the stability coefficient, the potential location distribution area of ​​each obstacle in the next few processing cycles is calculated, and a dynamic threat field of obstacles is generated. By combining the real-time location of the tethered UAV, the preset flight mission path, and the current flight environment constraints, multiple collision-free candidate flight paths are generated in the flight airspace. The final flight path is selected based on the cumulative risk value of each candidate flight path crossing the dynamic threat field of the obstacle, and the degree of similarity between each candidate flight path and the preset flight mission path.

8. The method for optimizing autonomous obstacle avoidance of tethered unmanned aerial vehicles based on lidar according to claim 1, characterized in that, The process of receiving environmental point cloud data generated by lidar scanning and performing multi-scale voxelization processing to generate multi-level point cloud subsets specifically includes: Based on the spatial bounding box of the original point cloud data of the lidar, the original point cloud data is divided into fine-grained voxel sets, medium-grained voxel sets and coarse-grained voxel sets using three different voxel sizes with different side lengths. For each voxel set at each granularity level, calculate the average coordinates of all point cloud data points within each voxel, and use the average coordinates as the representative point of the voxel. For fine-grained voxel sets, representative points of all voxels are directly used to form a first-level point cloud subset. For medium-grained voxel sets, based on the spatial inclusion relationship between fine-grained voxel sets and their spatial inclusion relationships, voxels with insufficient representativeness in medium-grained voxels are split into multiple fine-grained virtual voxels and representative points are generated. These representative points are then merged with the original representative points to form a second-level point cloud subset. For coarse-grained voxel sets, the same method as for medium-grained voxel sets is used to process them, generating third-level point cloud subsets.

Images