Unmanned aerial vehicle laser point cloud dust removal method

Abstract

The invention discloses an unmanned aerial vehicle laser point cloud dust removal method, which filters dust by combining flight state information of an unmanned aerial vehicle and adopting a method of matching a current frame point cloud with a local point cloud map constructed by a historical frame, and does not need to perform feature analysis and training on the dust in advance. The method has a good removal effect on discrete or structural dust data; and the current frame point cloud after dust removal is updated to the historical frame point cloud set, so that a local point cloud map used by the subsequent frame point cloud can maintain a relatively clean low-dust state for a long time.

Description

Technical Field

[0001] The invention relates to the field of laser point clouds, and in particular to a method for removing dust from laser point clouds of unmanned aerial vehicles. Background Art

[0002] In recent years, with the improvement of performance such as miniaturization and intelligence, the application field of drones is no longer limited to outdoor scenes, but is further expanding to indoor scenes. In particular, closed spaces such as tunnels, mines, and pipe galleries urgently need to change their safety production management from "manual" to "unmanned" due to their harsh environment, complex structure, and high risk factor. Among them, the most common demand is for drones to carry laser radar scanning equipment and other equipment to perform deformation detection, three-dimensional mapping, and autonomous inspections in dangerous areas. However, the working conditions in these closed spaces are very harsh, with gravel, dust, water seepage, etc., especially after the drone takes off, the large amount of dust raised around it will cause a large number of noise points in the point cloud data scanned by the laser radar, which has a great impact on both SLAM real-time positioning and three-dimensional space reconstruction. Therefore, the original point cloud data usually needs to be dust-free.

[0003] Existing LiDAR point cloud dust removal algorithms all treat dust as discrete noise. Commonly used methods include: 1) filtering methods based on statistical principles, such as proposing discrete points based on the distance distribution between neighboring points or the number of scattered points in the spatial voxels, and various modified bilateral filters; 2) classification methods based on model training, such as clustering point clouds, extracting features such as row sparsity, column density non-uniformity and average density, and then classifying them through pre-trained models. However, these methods are mainly effective in removing outlier noise dust with low spatial distribution density. However, when drones fly in closed spaces, due to the narrow space and close to the ground, they often cause high-density structural noise around the drones. The current methods have limited effect on removing this type of dust. Summary of the invention

[0004] The present invention discloses a method for removing dust from a UAV laser point cloud, which can effectively improve the dust interference problem in the laser point cloud collected when the UAV is flying in a closed space. The method comprises the following steps: The drone is equipped with a multi-line laser radar, which periodically obtains the scanned point cloud of the current frame after power-on; Extracting a local point cloud map from the historical frame point cloud set and matching it with the current frame point cloud, calculating the optimal conversion translation T(k) and optimal conversion rotation R(k) of the current frame, where k represents the frame number of the current frame point cloud; The local point cloud map is subjected to concave hull processing, and an effective denoised point cloud is screened from the current frame point cloud in combination with the flight status of the UAV, and the point cloud is added to the historical frame point cloud set.

[0005] Preferably, extracting a local point cloud map from a historical frame point cloud set and matching it with the current frame point cloud comprises the following steps: Part of the historical frame point clouds are selected from the historical frame point cloud set to form the local point cloud map, and the selection condition is: the time interval between the historical frame point cloud and the current frame point cloud or the difference between the optimal conversion translation amount or the optimal conversion rotation amount is less than a set threshold; Filter a set of point pairs to be matched from the current frame point cloud and the local point cloud map, traverse each point in the current frame point cloud, search for the nearest neighbor point in the local point cloud map, and if the Euclidean distance d1 between the point and the nearest neighbor point is less than a set first distance threshold D1, add the point and the nearest neighbor point as a point pair to the set of point pairs to be matched; The set of to-be-matched point pairs is matched, and the matching method may adopt an ICP method, an NDT method, or a method derived therefrom.

[0006] Preferably, performing concave hull processing on the local point cloud map and selecting effective denoised point clouds from the current frame point cloud in combination with the flight status of the drone includes: Calculating the concave hull of the local point cloud map according to the set second distance threshold D2; Convert the current frame point cloud to a relative navigation system according to the optimal conversion translation amount T(k) and the optimal conversion rotation amount R(k), and traverse each point of the current frame point cloud in the relative navigation system; If the point is outside the concave hull of the local point cloud map, the point is added to the effective denoised point cloud; If the point is inside the concave hull of the local point cloud map and the drone is in a ground-locked state, the point is added to the effective denoised point cloud; If the point is inside the concave hull of the local point cloud map and the drone is in a ground unlocked state or a flying state, search for the nearest neighbor point of the point in the local point cloud map. If the Euclidean distance d2 between the point and the nearest neighbor point is less than the set third threshold D3, add the point to the effective denoised point cloud.

[0007] The above technical solution of the present invention has at least the following beneficial effects: Based on the analysis of the causes and characteristics of dust during drone laser point cloud mapping in closed spaces, this method effectively combines the drone flight status information and filters dust by matching the current frame point cloud with the local point cloud map constructed by historical frames. This method does not require prior feature analysis and training of dust, and has a good removal effect on both discrete and structured dust data. Then, by updating the current frame point cloud after dust removal to the historical frame point cloud set, it is ensured that the local point cloud map used by subsequent frame point clouds can maintain a relatively clean and low-dust state for a long time. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 A schematic diagram of a process for removing dust from a drone laser point cloud provided in the first embodiment of the present invention; Figure 2 A schematic diagram of a specific implementation process of step S103 provided in the first embodiment of the present invention; Figure 3 A set of concave hull diagrams of a provided two-dimensional point cloud; Figure 4 A schematic diagram of the corresponding convex hull of a provided two-dimensional point cloud. Implementation

[0009] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0010] Figure 1 The following is a flow chart of a method for removing dust from a drone laser point cloud provided in the first embodiment of the present invention. Figure 1 As shown, the following steps are included: S101: The drone is equipped with a multi-line laser radar, which periodically obtains the scanned point cloud of the current frame after power-on; S102: extracting a local point cloud map from the historical frame point cloud set to match the current frame point cloud, and calculating the optimal conversion translation T(k) and rotation R(k) of the current frame, where k represents the frame number of the current frame point cloud; S103: performing concave hull processing on the local point cloud map, and screening effective denoised point clouds from the current frame point cloud in combination with the flight status of the drone, and adding the result to the historical frame point cloud set.

[0011] In this embodiment, the current frame point cloud is in the laser radar coordinate system, which can be the original point cloud data obtained by the multi-line laser radar scanning, or the feature point cloud data extracted from the original point cloud data. The feature point point cloud data is composed of corner points and plane points, and the curvature value of each point in the original point cloud data is calculated to determine whether it meets the feature point requirements. The calculation formula of the curvature value c(i) of the i-th point of the original point cloud data is:, where n represents the number of neighboring points obtained forward and backward on the same Z-axis plane with the i-th point as the center; represents the ranging value of the i-th point, and L represents the laser radar coordinate system. If the curvature value c(i) is greater than the set first curvature threshold C1, the i-th point is judged to be a corner point; if the curvature value c(i) is less than the set second curvature threshold C2, the i-th point is judged to be a plane point. Here it satisfies: C1>C2.

[0012] In this embodiment, the historical frame point cloud set and the local point cloud map are both in a relative navigation system and correspond to the current frame point cloud: if the current frame point cloud is the original point cloud data, then the points of the historical frame point cloud set and the local point cloud map are all from the original point cloud data; if the current frame point cloud is the feature point cloud data, then the points of the historical frame point cloud set and the local point cloud map are all from the feature point cloud data.

[0013] In this embodiment, the optimal conversion translation T(k) and the optimal rotation R(k) are both a 3*1 vector, and the optimal conversion translation T(k) is composed of the translation components of the X, Y and Z axes of the relative navigation system converted from the laser radar coordinate system where the current frame point cloud is located:; the optimal conversion rotation R(k) is composed of the rotation attitude angles around the X, Y and Z axes of the relative navigation system converted from the laser radar coordinate system where the current frame point cloud is located:.

[0014] Optionally, extracting a local point cloud map from the historical frame point cloud set and matching it with the current frame point cloud in step S102 includes the following steps: The local point cloud map is formed by filtering part of the historical frame point cloud from the historical frame point cloud set, and the filtering condition is: the time interval between the historical frame point cloud and the current frame point cloud or the difference between the optimal conversion translation amount or the optimal conversion rotation amount is less than a set threshold, which is expressed by the formula: or or, where p is the frame number of the historical frame point cloud; Filter the set of point pairs to be matched for the current frame point cloud and the local point cloud map, traverse each point in the current frame point cloud, search for the nearest neighbor point in the local point cloud map, and if the Euclidean distance d1 between the point and the nearest neighbor point is less than a set first distance threshold D1, add the point and the nearest neighbor point as a point pair to the set of point pairs to be matched; The set of to-be-matched point pairs is matched, and the matching method may adopt an ICP (Iterative Closest Point) method, an NDT (Normal Distributions Transform) method, or a method derived therefrom, which is not specifically limited here.

[0015] Furthermore, Figure 2 This is a flowchart of a specific implementation of step S103 provided in the first embodiment of the present invention, see Figure 2 As shown, the concave hull processing of the local point cloud map and the screening of effective denoised point clouds from the current frame point cloud in combination with the flight status of the drone include the following steps: S1031: Calculating the concave hull of the local point cloud map according to the set second distance threshold D2; S1032: converting the current frame point cloud to the relative navigation system according to the optimal conversion translation amount T(k) and the optimal conversion rotation amount R(k), and traversing each point of the current frame point cloud in the relative navigation system; S1033: If the point is outside the concave hull of the local point cloud map, then proceed to S1038, otherwise proceed to S1034; S1034: If the drone is currently in ground lock state, if yes, proceed to S1038, otherwise proceed to S1035; S1035: The drone is in a ground unlocked state or a flying state, searching for the nearest neighbor point of the point in the local point cloud map, and calculating the Euclidean distance d2 between the two; S1036: If d2 is less than or equal to the set third distance threshold D3, if so, proceed to S1038, otherwise proceed to S1037; S1037: Skip this point; S1038: adding the point to the effective denoised point cloud; S1039: After traversing all points of the current frame point cloud, the effective denoised point cloud is added to the historical frame point cloud set.

[0016] In this embodiment, the concave hull refers to a polygon (two-dimensional) or a polyhedron (three-dimensional) created to represent the area occupied by a group of data points. Figure 3 A set of concave hull diagrams of a two-dimensional point cloud is provided. Figure 4 For a schematic diagram of the corresponding convex hull of a two-dimensional point cloud, see Figure 3 and Figure 4 As shown, compared with the convex hull, the concave hull can better describe the specific shape of the point cloud, but it is not unique. For the same input data point, according to the second distance threshold D2, the generated concave hull will have different results, and the larger the second distance threshold D2, the closer the generated concave hull is to the convex hull.

[0017] In this embodiment, the current frame point cloud data will be divided into two parts, inner and outer, according to the concave hull of the local point cloud map: the external point cloud belongs to the newly expanded space, which is far away from the drone and contains less dust in the point cloud data; the internal point cloud includes the space around the drone, which has a high dust content, but the corresponding local point cloud map is extracted from the historical frame point cloud, and the historical frame point clouds are all filtered for dust. Therefore, the internal point cloud is matched with the local point cloud map by the nearest neighbor, and the internal point cloud representing the non-dust valid spatial structure is screened out. The external point cloud and the internal point cloud of the non-dust valid spatial structure are merged to form the effective denoised point cloud, and are updated to the historical frame point cloud set to ensure that the local point cloud map used by subsequent frame point clouds is the same as the internal point cloud map. Figure 1Always maintain a cleaner and low dust state.

[0018] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for removing dust from a drone laser point cloud, comprising the following steps: The drone is equipped with a multi-line laser radar, which periodically obtains the scanned point cloud of the current frame after power-on; Extracting a local point cloud map from the historical frame point cloud set and matching it with the current frame point cloud, calculating the optimal conversion translation T(k) and optimal conversion rotation R(k) of the current frame, where k represents the frame number of the current frame point cloud; The local point cloud map is subjected to concave hull processing, and an effective denoised point cloud is screened from the current frame point cloud in combination with the flight status of the UAV, and the point cloud is added to the historical frame point cloud set.

2. The method according to claim 1, characterized in that The extracting of a local point cloud map from a historical frame point cloud set and matching it with the current frame point cloud comprises the following steps: Part of the historical frame point clouds are selected from the historical frame point cloud set to form the local point cloud map, and the selection condition is: the time interval between the historical frame point cloud and the current frame point cloud or the difference between the optimal conversion translation amount or the optimal conversion rotation amount is less than a set threshold; Filter a set of point pairs to be matched from the current frame point cloud and the local point cloud map, traverse each point in the current frame point cloud, search for the nearest neighbor point in the local point cloud map, and if the Euclidean distance d1 between the point and the nearest neighbor point is less than a set first distance threshold D1, add the point and the nearest neighbor point as a point pair to the set of point pairs to be matched; The set of to-be-matched point pairs is matched, and the matching method may adopt an ICP method, an NDT method, or a method derived therefrom.

3. The method according to claim 1, characterized in that The performing concave hull processing on the local point cloud map and selecting effective denoised point clouds from the current frame point cloud in combination with the flight status of the drone includes: Calculating the concave hull of the local point cloud map according to the set second distance threshold D2; Convert the current frame point cloud to a relative navigation system according to the optimal conversion translation amount T(k) and rotation amount R(k), and traverse each point of the current frame point cloud in the relative navigation system; If the point is outside the concave hull of the local point cloud map, the point is added to the effective denoised point cloud; If the point is inside the concave hull of the local point cloud map and the drone is in a ground-locked state, the point is added to the effective denoised point cloud; If the point is inside the concave hull of the local point cloud map and the drone is in a ground unlocked state or a flying state, search for the nearest neighbor point of the point in the local point cloud map. If the Euclidean distance d2 between the point and the nearest neighbor point is less than a third set threshold D3, add the point to the effective denoised point cloud.

Images