CN-121977575-A - Finite space unmanned aerial vehicle bridge-like inspection path planning method

CN121977575ACN 121977575 ACN121977575 ACN 121977575ACN-121977575-A

Abstract

The invention belongs to the technical field of path planning, and particularly relates to a finite space unmanned aerial vehicle bridge-like inspection path planning method. The method comprises the steps of sensing point cloud information of an unmanned aerial vehicle subscription environment, wherein the point cloud information is derived from a laser radar, a depth camera or fusion point cloud, filtering, performing point cloud noise point self-adaptive multi-level filtering on the point cloud information, performing 9-step self-adaptive noise point filtering, outputting filtered point cloud and free space F, planning an imitation bridge inspection path, generating an initial path by using a narrow channel RRT, constructing a safe flight inspection corridor SFIC, and optimizing in SFIC to generate a central imitation bridge inspection path. The method solves the problems of insufficient environment perception precision, low path planning safety and efficiency, lack of dynamic adaptability in path planning and low path quality existing in the existing limited space inside the bridge during unmanned aerial vehicle inspection.

Inventors

Gong Bingyu
DONG YUNLONG
LEI XIANG
WANG JUNHAO
LI GUANNAN
TAN YUANHUI

Assignees

智洋创新科技股份有限公司

Dates

Publication Date: 20260505
Application Date: 20260310

Claims (6)

1. The method for planning the bridge-like inspection path of the limited space unmanned aerial vehicle is characterized by comprising the following steps of: s1, sensing point cloud information of an unmanned aerial vehicle subscription environment, wherein the point cloud information is derived from a laser radar, a depth camera or fusion point cloud; S2, filtering, namely performing point cloud noise self-adaptive multi-level filtering on the point cloud information, performing 9-step self-adaptive noise filtering, and outputting filtered point cloud And a free space F; s3, planning, namely generating an imitated bridge inspection path, constructing a safe flight inspection corridor SFIC by generating an initial path through using a narrow channel RRT, and optimizing and generating a centered imitated bridge inspection path in SFIC.
2. The method for planning an imitated bridge inspection path of a finite space unmanned aerial vehicle according to claim 1, wherein the step S2 of 9-step adaptive noise filtering specifically comprises: s21, performing voxel gridding modeling, and representing an input original three-dimensional point cloud as: (1) Wherein, the Representing an original three-dimensional point cloud set, Is the first The spatial coordinates of the individual point clouds, The total point number of the point cloud is; space coordinates of each point cloud Mapping to voxel index Voxel resolution is Voxel index: (2) Wherein, the , And Is the minimum coordinate value of the point cloud in each axis direction, A representation down-rounding function for discretizing the continuous coordinates into voxel indices; Controlling the spatial resolution for the voxel side length; For each voxel Maintaining occupancy probabilities And the logarithmic probability thereof: (3) Wherein, the Representing voxels at time t Probability of being occupied by an obstacle, log probability The method is used for stable updating of the numerical value, and underflow caused by probability multiplication is avoided; S22, detecting local geometric consistency, and setting the unmanned plane position of the current frame as Defining a linear distance from a point cloud to a position of an unmanned aerial vehicle In the neighborhood of Defining a consistency score : The neighborhood is expressed as: (4) The consistency score is expressed as: (5) Wherein, the Representing euclidean norms, i.e. point clouds To the unmanned plane position Is a straight line distance of (2); Expressed in terms of Is the center and radius All neighbor point sets within; For measuring point clouds With its adjacent points Depth uniformity of (c); If it is Or (b) Then The determination is made as noise, wherein, For the depth abrupt change threshold value, The minimum neighborhood point threshold number is used for eliminating isolated and mutation points; s23, detecting the continuity of the edge and depth of the field of view, and using a single frame point cloud Centroid of (2) For reference, point cloud Conversion to spherical angle Projected to depth image pixel coordinates Wherein And Respectively azimuth angle and pitch angle, which are used for projecting the three-dimensional point cloud to a two-dimensional depth map; If the depth of the neighborhood pixel jumps Or too close a depth Then The determination is made as field of view edge noise, wherein, Is that The distance value of the location is calculated, Is that The distance value of the adjacent position is calculated, For the depth jump threshold value, The minimum effective depth is used for eliminating the edge distortion of the sensor; S24, carrying out normal vector consistency and support detection, calculating a neighborhood covariance matrix for fitting a local plane, taking a minimum feature vector as a normal vector direction, and defining a plane residual error: (6) wherein, plane residual error Is a pointing cloud The average value of the distances from all points in the neighborhood to the fitting plane is used for reflecting the fitting quality of the local plane, and the larger the value is, the more sharp or outlier points are likely to be; at the same time, counting the number of support points at both sides in the normal direction If (if) Or (b) Then it is determined as noise which, among other things, Is the plane residual threshold value, The number of the support points is the minimum number of the support points, and the support points are used for eliminating unstructured noise points; S25, performing cluster-based small cluster filtering, performing DBSCAN clustering, and if the DBSCAN clustering is the first Cluster size of each cluster Or density of Then the whole cluster is rejected, wherein, The volume of the bounding box is represented, For the minimum cluster size to be the smallest cluster size, Is the minimum density; S26, performing distance filtering of position constraint of the unmanned aerial vehicle according to point cloud Carrying out partition treatment on the linear distance d from the current position of the unmanned aerial vehicle; if the distance is Then the mixture is sent to the steps S22-S25 for fine filtration, if so The structural information is preserved if Then it is discarded directly, wherein, Representing a fine filter distance threshold value, An upper distance limit representing the preservation of the structural information; s27, carrying out multi-frame probability accumulation and aging mechanism, and specifically updating the logarithmic probability: Each single frame observation will be weighted according to the frequency of observation Weight of observation angle And single observation log probability increment And combining voxel visibility markers Calculating the logarithmic probability increment of the current frame through a formula (7) : (7) The delta is then combined with the log-history probability and limited to the clip function of equation (8) In the interval, numerical value overflow is prevented, and stable accumulation of multi-frame probability is realized: (8) Wherein, the Is the time of When a voxel is Log odds of (a); Is the time of When a voxel is Log probability increment of (2); is the minimum and maximum limit of logarithmic probability; is the observed frequency weight; Is the observation angle weight; Is the log probability increment of single observation; If it is Exponential decay is performed when not observed for a long time: (9) Wherein the method comprises the steps of As a forgetting factor, For a time interval, for forgetting obsolete obstructions; S28, performing adaptive threshold judgment and neighborhood consistency smoothing, and firstly calculating entropy Reflecting the uncertainty and entropy of voxel state The calculation formula of (2) is as follows: (10) Wherein, the Is the voxel at time t Entropy of (2); Is the voxel at time t Probability of being occupied by an obstacle; The occupation threshold is improved through the high entropy area, the idle threshold is reduced, and the occupation reference threshold is dynamically adjusted: (11) Wherein, the Is the voxel at time t For determining whether the voxel is an obstacle; Is a reference value of the occupancy threshold; Is the voxel at time t For determining whether the voxel is free space; Is a reference value for the idle threshold; is an adjusting coefficient of the occupied threshold value and controls the lifting amplitude of the occupied threshold value of the high entropy area; is the adjustment coefficient of the idle threshold value, controls the reduction amplitude of the idle threshold value of the high entropy area, ; Neighborhood consistency smoothing is then performed: (12) Wherein, the The smoothing coefficient is used for controlling the weight ratio of the historical value to the neighborhood mean value; Is a voxel For spatial consistency smoothing; to smooth through space consistency Updated attribute values at time t; Is a voxel Is in the 26 neighborhood of (2) The number of voxels contained in the image; Is a voxel Is in the 26 neighborhood of (2) Any one voxel of (a); s29, carrying out sliding window hit statistics and inspection, and counting hit times in a multi-frame window If the ratio is A stable observation is confirmed, wherein, In order to achieve a sliding window size, Is the hit rate threshold.
3. The method for planning an imitated bridge inspection path of a finite space unmanned aerial vehicle according to claim 2, wherein in step S3, the generation of the imitated bridge inspection path specifically comprises: S31, narrow channel RRT sampling strategy: In free space In the method, a RRT algorithm of a limited sampling space and a multi-grid resolution map are adopted, and a discrete path point sequence is output: (13) Wherein, the The start of the sequence is indicated, Indicating a sequence end point; defining a sequence of decreasing safe radius Wherein Satisfies the following conditions , For grid map base resolution, where K represents the total number of layers of the decreasing sequence of safety radii, i.e. the sequence Comprises K elements, K is the index of the sequence, and the value range is For identifying the kth safety radius in a decrementing sequence ; Constrained sampling space : For the starting point And endpoint Direction vector Defining a sampling corridor as: (14) Wherein: representing candidate points in a sampling space, namely the space positions to be sampled by the unmanned aerial vehicle in the path planning process, for judging whether the points fall in a sampling corridor In turn, the sampling range of the RRT algorithm is constrained, L represents the linear distance between the start point s and the end point g, i.e ; The representation being perpendicular to Components of (2); Is a unit vector in the vertical direction; And Corridor width and height, respectively; defining a sampling space For the intersection of the sampling corridor, global map boundary and free space under the current safety radius: (15) Wherein minBound represents the minimum boundary coordinates of the global map on each coordinate axis, and maxBound represents the maximum boundary coordinates of the global map on each coordinate axis; , Representation points At radius of No obstacle exists inside; Initializing radius index The following iterative process is performed: s311 based on the current radius Construction of grid map Resolution is ; S312 in the sampling space The RRT algorithm is executed internally, the objective function is path length minimization: (16) S313, if planning is successful, directly returning to the path; S314 if the programming fails or the start/end point is in Internal unsafe condition If (if) Returning to step S311, otherwise, terminating; S315, optimizing target connection: When the path end point and the target point are distant The method meets the following conditions: (17) Wherein: is a distance threshold; Directly connecting the target points, otherwise, recursively subdividing the path segments and performing radius reduction inspection until the target points are connected to obtain a rough path for generating a subsequent safety corridor; S32, path densification: coarse path densification using cubic spline interpolation, specifically Each section of (a) Generating Uniformly interpolating points so that the distance between the points is smaller than a set target value; s33, constructing a convex polyhedron safety corridor SFIC: safety corridor SFIC, i.e. The generation mode is directly constructed by point cloud: s331, inputting the filtered point cloud and free space , wherein, Representing free space Any spatial point in the (3); the three-dimensional Euclidean space is the space dimension in which the free space is located; representing a filtered set of point clouds, wherein each point Representing effective obstacle points obtained after S2 noise filtering; Representation points And point cloud Some point in (a) Euclidean distance between them; the equivalent radius of the unmanned aerial vehicle is represented by simplifying the three-dimensional shape of the unmanned aerial vehicle into a sphere, and then the radius of the sphere is represented; The safety margin is the safety distance reserved additionally for avoiding the 'edge rubbing' of the unmanned aerial vehicle and the obstacle; S332 slice construction along a reference path In step size Acquiring sampling points At each point Local slice plane of the structure Normal is the tangential direction of the path ; S333, local free point extraction, in local slice plane Thickness of the vicinity Extraction free point: (18) Wherein, the Is to calculate the point p to the plane Is a vertical distance of (2); s334, convex hull expansion, free point At the position of In-plane computation convex hull And is contracted inwards Obtaining a safe area ; S335, generating a three-dimensional convex polyhedron, namely, generating a safety area Stretching along the normal plane of the path Generating a three-dimensional convex polyhedron Ensuring coverage of unmanned aerial vehicle flight envelope; S336, connectivity restoration, if two adjacent convex polyhedrons Then insert the transitional convex polyhedron Wherein Representing a minimum convex set containing the two convex polyhedrons; S337 corridor simplification, merging overlapping degree Outputs the final , wherein, Representing the volume of convexity; s34, global inspection path optimization Vertex matrix for each convex polyhedron Representation of wherein Is the first The number of vertexes of the convex polyhedron and the path point Represented as convex combinations of vertices, i.e. Ensuring that it is located within the polyhedron, wherein: Is a convex combination coefficient vector for representing the path point Is a weighted combination of convex polyhedron vertices; is through vertex matrix Combining coefficient vectors with projections Matrix multiplication of (a) to calculate a path point Coordinates of (c), the operation ensures Is positioned in the ith convex polyhedron; The overall objective function is as follows: (19) Wherein, the A path length term is represented and is used to represent, Representing the term of the centering constraint, Representing a regularization term; The path length term is as follows: (20) Wherein: Representing the first in the path The three-dimensional coordinates of the individual path points, Is small constant, and avoids gradient explosion; The centering constraint is as follows: (21) Wherein, the The weight of the centering constraint term is used for adjusting the duty ratio of the centering constraint in the overall objective function; Is the target safe distance; Is a path point To the ith convex polyhedron Euclidean distance of the boundary; the regularization term is as follows: (22) Wherein: The L2 norm square of the convex combination coefficient vector; Based on the total objective function, the path is optimized and solved by using an L-BFGS algorithm: (23) Wherein, the To optimize the variables, it consists of convex combination coefficient vectors for all path points.
4. The method for planning an imitated bridge inspection path of a finite space unmanned aerial vehicle according to claim 3, wherein in step S315, the recursively subdivided path segment is a path segment with a smaller safety radius by continuously splitting the path segment into subsections, which gradually approaches the target point, and the specific steps are as follows: if the current path end point is The target point is g, the current safety radius is Subdivision threshold Minimum safety radius ; Step a, recursively subdividing path segments Distance from g Then calculate the midpoint Step b is entered; step b, checking the midpoint Whether or not in free space and meeting a safety radius If so, the path segment Split into And Updating This step is repeated until If it does not meet Unsafe, enter step c; Step c, reducing the safety radius check, i.e If the current safety radius is Then the radius is reduced to Returning to step b, re-checking whether the subdivided path segment meets the safety under the new radius, if so If the data is still not satisfied, judging that the local traffic is not available, and returning to failure; The termination condition is when And is also provided with When in direct connection with And g, obtaining a rough path.
5. A method for planning an artificial bridge inspection path of a limited space unmanned aerial vehicle according to claim 3, wherein the safety corridor surrounds a reference path in the filtered free space A series of connected convex polyhedrons is constructed by Multiple convex polyhedrons Construction, i.e. safety corridor The safety corridor fulfils the following four properties: Coverage: i.e. the corridor must completely cover the reference path; Safety: I.e. each polyhedron is entirely located in free space; connectivity: the continuous passable path is ensured; Convexity: the convex set is convenient for track optimization; Wherein: indicating that the reference path instant tdata is always located in free space In which the independent variables T represents time variable, and the value range is For describing the variation of the reference path with time, T represents the reference path I represents the index of the convex polyhedron, and the value range is For identifying an ith convex polyhedron in the security corridor; is the ith convex polyhedron in the safety corridor and is the basic unit for constructing the safety corridor; representing the total number of convex polyhedrons in the security corridor.
6. A method for planning an inspection path of a bridge-like unmanned aerial vehicle in a limited space according to claim 3, wherein the construction of the safety corridor comprises three modes: mode 1, global safety corridor generation during map building: In unmanned aerial vehicle construction operation in the bridge, the concrete flow is as follows: Firstly, the unmanned aerial vehicle performs the point cloud noise filtering in real time in the process of drawing construction, then, based on the obtained real-time pose of the unmanned aerial vehicle, the convex surface SFIC defined in the step S33 is constructed under the current pose, after drawing construction is completed, a complete SFIC model is generated and is stored in a YAML format through an offline storage flow, in the follow-up inspection task, the SFIC model in the YAML format can be directly read, the global inspection path optimization in the step S34 is performed between adjacent waypoints, and the step S31-S33 is not required to be repeatedly performed; Mode 2, global safety corridor generation after drawing: If SFIC is not constructed in the construction process, constructing SFIC through the whole bridge point cloud and all waypoints of the ground station after the construction process, automatically traversing all the waypoints, generating a complete SFIC model through one step of S31-S33, storing the model in a YAML format through an offline storage flow, directly reading the SFIC model in the YAML format like the mode 1 in the follow-up inspection task, and executing global inspection path optimization in the step S34 between adjacent waypoints without repeatedly executing the steps S31-S33; mode 3, real-time local security corridor generation: The method is suitable for real-time inspection planning in the bridge or real-time inspection path planning in non-bridge scenes, and each time a target point is received, the step S31-S34 is carried out.

Description

Finite space unmanned aerial vehicle bridge-like inspection path planning method Technical Field The invention belongs to the technical field of path planning, and particularly relates to a finite space unmanned aerial vehicle bridge-like inspection path planning method. Background When the unmanned aerial vehicle is independently patrolled and examined in the bridge, such as structures of box girder, truss, arch rib and the like, the prior art scheme generally comprises the following core links: Environmental perception, mainly relying on laser radar or depth camera to collect three-dimensional point cloud data to construct a digital model of the environment. Noise filtering, in which a single, static filtering algorithm is commonly used to process the original point cloud, such as statistical outlier rejection (STATISTICAL OUTLIER REMOVAL, SOR) or radius filtering (Radius Outlier Removal, ROR). These methods identify and remove outliers by setting a fixed statistical threshold or neighborhood radius. The main stream algorithm comprises RRT or A algorithm. These algorithms search on a global static map to find a collision-free path from the start point to the end point. Map representation-environmental model is usually simplified into static three-dimensional grid map, obstacle information is built once before planning, and dynamic update mechanism is lacked. The common goal of these schemes is to realize autonomous navigation and obstacle avoidance of unmanned aerial vehicle in complex bridge structure, but its technical means is relatively basic and isolated, failing to form a closed-loop optimization system aiming at narrow, dynamic and high noise environment, the following significant defects exist: The environment perception precision is insufficient, and a single filtering method, such as SOR and ROR, cannot effectively cope with complex noise sources inside a bridge, such as penetration artifacts caused by metal surface reflection, signal scattering caused by dust and water vapor, and distortion of the edge of a sensor field of view. This results in a built environmental model containing a large number of noise points, severely affecting the safety of subsequent obstacle detection and path planning. The path planning safety and efficiency are low, namely, paths generated by the traditional RRT or A algorithm are often clung to the boundary of an obstacle, clear safety margin constraint is lacked, collision is easy to occur under unmanned aerial vehicle dynamic disturbance or modeling error, and the safety is poor. In a narrow channel (the width is often less than 2 meters), the random sampling strategy of the traditional RRT algorithm has extremely low efficiency, a large number of sampling points fall in a infeasible area and easily fall into local optimum, so that the planning time is overlong or the planning failure is caused, and the efficiency is low. The existing scheme is mostly based on static maps, can not update local environment information in real time and conduct rolling re-planning in the moving process of the unmanned aerial vehicle, has poor system robustness, and is difficult to cope with dynamic changes of environments or inaccuracy of initial maps. The quality of the path is low, the generated path generally only meets the obstacle avoidance requirement, and the smoothness and the centering of the path are lacking, namely, the center flight is kept in the channel, so that the high-speed and stable flight of the unmanned aerial vehicle is not facilitated. Disclosure of Invention The invention aims to overcome at least one defect of the prior art, and provides a limited-space unmanned aerial vehicle simulated bridge inspection path planning method, which aims to solve the problems of insufficient environment sensing precision, low path planning safety and efficiency, lack of dynamic adaptability in path planning and low path quality in the existing limited space inside a bridge in the unmanned aerial vehicle inspection process. The detailed technical scheme of the invention is as follows: a method for planning an imitation bridge inspection path of a limited space unmanned aerial vehicle, the method comprising: s1, sensing point cloud information of an unmanned aerial vehicle subscription environment, wherein the point cloud information is derived from a laser radar, a depth camera or fusion point cloud; S2, filtering, namely performing point cloud noise self-adaptive multi-level filtering on the point cloud information, performing 9-step self-adaptive noise filtering, and outputting filtered point cloud And a free space F; s3, planning, namely generating an imitated bridge inspection path, constructing a safe flight inspection corridor SFIC by generating an initial path through using a narrow channel RRT, and optimizing and generating a centered imitated bridge inspection path in SFIC. According to the present invention, the step S2 of 9 adaptive noise filtering specifically includ