CN-121982523-A - Obstacle detection method and system based on laser point cloud identification

CN121982523ACN 121982523 ACN121982523 ACN 121982523ACN-121982523-A

Abstract

The invention discloses an obstacle detection method and system based on laser point cloud identification, wherein the method comprises the steps of acquiring laser point cloud scanned by a laser radar of a target area and multispectral images acquired by a camera; registering laser point clouds with multispectral images to establish a corresponding relation between the laser point clouds and the multispectral images, extracting echo characteristics from the laser point clouds based on a laser radar, continuously analyzing the laser point clouds to obtain candidate vegetation point clouds and candidate building point clouds, extracting geometric characteristics from the candidate vegetation point clouds and the candidate building point clouds, classifying the laser point clouds through the echo characteristics and the geometric characteristics to obtain building point clouds, performing mask analysis on the building point clouds based on the multispectral images to obtain building outlines, and visually judging whether the building is a building based on the building outlines.

Inventors

ZHAO KUNYU
ZHANG YUNPENG
LI BU
HUANG XIANGYI
LIU MENGYI
Wu yaxing
CHEN QIHONG
LIU WEIYU
CHEN SHIYUN
FENG HAO

Assignees

重庆市送变电工程有限公司

Dates

Publication Date: 20260505
Application Date: 20260114

Claims (10)

1. The obstacle detection method based on laser point cloud identification is characterized by comprising the following operation steps: Registering the laser point cloud with the multispectral image to establish a corresponding relation between the laser point cloud and the multispectral image; Extracting echo characteristics from the laser point cloud based on the laser radar, continuously analyzing the laser point cloud to obtain candidate vegetation point cloud and candidate building point cloud, extracting geometric characteristics from the candidate vegetation point cloud and the candidate building point cloud, classifying the laser point cloud according to the echo characteristics and the geometric characteristics to obtain building point cloud, and performing mask analysis on the building point cloud based on the multispectral image to obtain building contours; and visually judging whether the building is based on the outline of the building.
2. The obstacle detection method based on laser point cloud identification according to claim 1, wherein echo characteristics are extracted from the laser point cloud based on the laser radar, and the specific operation steps are as follows: Extracting a record of laser pulses from the laser radar; Recording the echo times of each laser pulse, extracting the echo intensity value of each echo time, and searching the maximum echo intensity value of all echo times in the laser pulse; and calculating the echo elevation difference by utilizing the first echo and the last echo of the laser pulse, and forming echo characteristics by using the maximum echo intensity value, the average value of the echo intensity values of all echo times and the echo elevation difference.
3. The obstacle detection method based on laser point cloud identification according to claim 2, wherein the laser point cloud is subjected to continuity analysis to obtain a candidate vegetation point cloud and a candidate building point cloud, and the specific operation steps are as follows: The method comprises the steps of calculating time difference according to adjacent echo times, eliminating a proportion value of echo intensity between the adjacent echo times which is larger than or equal to a preset adjacent time difference threshold according to the time difference, screening a point cloud minimum Gao Chengyu point cloud maximum elevation of laser point clouds, dividing a plurality of point clouds according to a preset basic layer thickness, judging adjacent distance and point cloud density of the laser point clouds in each point cloud layer, and obtaining candidate building point clouds and candidate vegetation point clouds.
4. The obstacle detection method based on laser point cloud identification according to claim 3, wherein geometrical characteristics are extracted from the candidate vegetation point clouds and candidate building point clouds, the laser point clouds are classified according to the echo characteristics and the geometrical characteristics to obtain building point clouds, mask analysis is performed on the building point clouds based on the multispectral images to obtain building outlines, and the specific operation steps are as follows: connecting each candidate building point cloud with each point cloud in the candidate vegetation point clouds by using a K nearest neighbor algorithm, and connecting the K nearest laser point clouds to form a graph structure; calculating the edge weight of the graph structure according to Euclidean distance between the adjacent laser point clouds and echo characteristics; the total edge weight of K edge weights of each laser point cloud in the graph structure is calculated, and the laser point cloud with the minimum total edge weight and the K nearest laser point clouds are screened; removing the edge with the maximum edge weight in the graph structure to obtain a plurality of point cloud subgraphs; Calculating the density of the point cloud of each point cloud sub-graph, the outline of the sub-graph and the shape of the sub-graph to form geometric features; Constructing a comprehensive feature vector according to the echo features and the geometric features; inputting the comprehensive feature vector into the multi-class classifier to output vegetation point cloud and building point cloud; The method comprises the steps of utilizing a multispectral image to mask vegetation point clouds and building point clouds to obtain vegetation masks and building masks, extracting pixel points corresponding to areas of the vegetation masks and the building masks from the multispectral image to analyze to obtain building areas, screening non-collinear laser point clouds in the building areas to form four corner planes, carrying out coplanarity combination on adjacent four corner planes to obtain combination Ping Miandian Yun Cu, and analyzing the boundaries of the cloud clusters of the combination Ping Miandian to obtain building outlines.
5. The obstacle detection method based on laser point cloud recognition as set forth in claim 4, wherein the multispectral image is used for masking a vegetation point cloud and a building point cloud to obtain a vegetation mask and a building mask, and the specific operation steps are as follows: respectively projecting the vegetation point cloud and the building point cloud onto a two-dimensional plane of the multispectral image to generate a vegetation mask and a building mask; Extracting pixel points of the region corresponding to the vegetation mask from the multispectral image, and extracting reflectivity values of a plurality of wave bands of each pixel point; Calculating the ratio of the reflectivity values of a plurality of wave bands, and calculating a normalized vegetation index according to the ratio of the reflectivity values; extracting pixel points corresponding to the region of the building mask from the multispectral image, extracting reflectance values of a plurality of wave bands of each pixel point, forming a spectrum curve through the reflectance values of the plurality of wave bands, calculating a spectrum derivative on the spectrum curve, presetting a reference spectrum feature library, extracting spectrum features of the region of the building mask, selecting typical spectrum features corresponding to the reference spectrum feature library based on the spectrum features of the region of the building mask, and calculating a spectrum angle; judging whether the spectrum angle is smaller than the similarity threshold value or not; if yes, judging that the spectral characteristics of the areas of the building mask are similar; and screening pixel points in a building mask in a region corresponding to the maximum spectrum angle with similar spectrum characteristics to cluster similar gray values, so as to form a building region.
6. The obstacle detection method based on laser point cloud identification according to claim 5, wherein the building area is screened for non-collinear laser point clouds to form four-corner planes, the adjacent four-corner planes are combined in a coplanar manner to obtain a combination Ping Miandian Yun Cu, and the combination Ping Miandian cloud cluster boundary is analyzed to obtain a building contour, and the method comprises the following specific operation steps: Eliminating discrete points from the laser point cloud in the building area, further screening non-collinear laser point clouds to form four-corner planes, calculating distances from all the laser point clouds to the four-corner planes to obtain internal point clusters, screening seed planes from the four-corner planes by using the internal point clusters, carrying out coplanar analysis on the seed planes and adjacent four-corner planes to obtain a combination Ping Miandian Yun Cu, and carrying out fitting straight lines of main direction analysis on the combination Ping Miandian cloud clusters to form a building contour.
7. The obstacle detection method based on laser point cloud identification according to claim 6, wherein the method is characterized by removing discrete points from the laser point cloud in the building area, further screening non-collinear laser point clouds to form a four-corner plane, calculating the distances from all the laser point clouds to the four-corner plane, and obtaining an inner point cluster, and comprises the following specific operation steps: comparing the building area with a laser point cloud, and eliminating discrete points in the laser point cloud; Randomly screening four non-collinear laser point clouds from the laser point clouds in the building area to form a quadrangle plane; calculating the straight line distance from all laser point clouds in the building area to the four corner planes, presetting an internal distance threshold value, and judging whether the straight line distance is smaller than the internal distance threshold value or not; if yes, judging that the laser point cloud is inside the four-corner plane and is used as an internal point cloud to form an internal point cluster; Repeating the steps to obtain a plurality of inner point clusters.
8. The obstacle detection method based on laser point cloud recognition according to claim 7, wherein the seed planes are screened by using the inner point clusters, and coplanar analysis is performed on the seed planes and the adjacent four-corner planes to obtain a combination Ping Miandian Yun Cu, and the specific operation steps are as follows: screening the maximum number of interior point clouds in the interior point clusters to serve as an optimal point cloud cluster; taking the four corner planes of the optimal point cloud cluster as seed planes, and calculating normal vectors of the seed planes and adjacent four corner planes; calculating a normal vector included angle between the normal vectors of the seed plane and the adjacent four-corner planes; judging whether the normal vector included angle is smaller than an included angle threshold value or not; if so, judging that the directions of the seed plane and the adjacent four-corner planes are consistent, calculating the plane distance between the seed plane and the adjacent four-corner planes, and further calculating the plane average distance; judging whether the plane average threshold is smaller than the plane distance threshold or not, if so, judging that the seed plane and the adjacent four-corner planes are coplanar; And merging the seed plane with the adjacent four corner planes, and repeating the steps to obtain the merged Ping Miandian cloud cluster.
9. The obstacle detection method based on laser point cloud identification as set forth in claim 8, wherein fitting a straight line to a main direction analysis boundary is performed on the merging Ping Miandian cloud clusters to form a building contour, and the specific operation steps are as follows: projecting the merged Ping Miandian cloud clusters to a two-dimensional plane, and converting the merged Ping Miandian cloud clusters into a two-dimensional point set; the method comprises the steps of carrying out convex hull on the two-dimensional point set, presetting a circle with a radius alpha by using an alpha shape algorithm, and sliding the edge of the two-dimensional point set by using the circle with the radius alpha to form a boundary point set; Dividing the boundary point set into a plurality of boundary line segments, and carrying out a fitting straight line of a least square method on each edge line segment; calculating intersection points of adjacent fitting straight lines to obtain initial corner points; Determining the directions of all fitting straight lines, connecting the directions of all fitting straight lines to the main direction and the perpendicular direction of the main direction to form squaring, adjusting the coordinates of initial angular points according to the squaring to enable adjacent fitting straight lines to be perpendicular or parallel, and reconnecting all fitting straight lines to form the building outline.
10. The obstacle detection system based on laser point cloud identification is characterized by comprising an acquisition module, an analysis module, an identification module and a detection module, wherein the acquisition module is used for acquiring the obstacle of the obstacle; The acquisition module is used for acquiring laser point clouds scanned by the laser radar of the target area and multispectral images acquired by the camera; the analysis module is used for extracting echo characteristics from the laser point cloud based on the laser radar, continuously analyzing the laser point cloud to obtain candidate vegetation point cloud and candidate building point cloud, extracting geometric characteristics from the candidate vegetation point cloud and the candidate building point cloud, classifying the laser point cloud according to the echo characteristics and the geometric characteristics to obtain building point cloud, and performing mask analysis on the building point cloud based on the multispectral image to obtain building contours; and the identification module is used for visually judging whether the building is based on the building outline.

Description

Obstacle detection method and system based on laser point cloud identification Technical Field The invention relates to the field of transmission line inspection, in particular to an obstacle detection method and system based on laser point cloud identification. Background Along with the continuous improvement of the operation and maintenance requirements of the power transmission line, the accurate identification and detection of dangerous obstacles such as conventional buildings (or illegal temporary building buildings), tall trees and the like in the line channel become a key task for guaranteeing the safe operation of the power grid. The traditional identification method mainly depends on single data sources such as remote sensing images, laser radar point clouds and the like, but is difficult to realize fine classification in a complex environment, and particularly has obvious limitation when distinguishing objects with similar spatial characteristics (such as building roofs and vegetation), so that the accuracy and timeliness of hidden trouble identification are affected. The laser radar point cloud can provide accurate three-dimensional space information, including geometric features such as height, outline and volume of the ground feature, and is helpful for locating potential tall and large objects. However, when facing objects of some materials or similar structures, such as metal roofs and dense crowns, relying on geometric features alone is prone to misjudgment, affecting effective differentiation of buildings from trees. On the other hand, the multispectral remote sensing image can provide abundant spectral information, including visible light, near infrared and other wave bands, can reflect the spectral reflection characteristics of ground materials, and is beneficial to distinguishing different materials such as vegetation, building roofs and the like. However, the image data itself lacks accurate three-dimensional information, it is difficult to directly acquire the height and spatial morphology of the target object, and particularly in the area where the topography is undulating or the objects overlap, the recognition effect is limited. In order to overcome the problems, in recent years, a technical method for fusing laser radar point cloud and multispectral remote sensing images is gradually applied to obstacle recognition of a transmission line channel. By jointly analyzing the geometric information of the point cloud and the spectral characteristics of the image, information complementation can be realized in two aspects of three-dimensional morphology and material properties, so that the precision of identifying the roof of a building (namely, the outline of the roof is important for determining the safety distance judgment of a power transmission line) and the precision of identifying tall trees are improved. Specifically, the laser point cloud provides the spatial position and the height characteristics of the target object, but the amount of traffic data is large by only adopting point cloud segmentation, and the identification accuracy is not high, and the identification accuracy efficiency is not ideal by singly using multispectral images to help to distinguish categories, such as vegetation (trees) and artificial buildings (roofs) through spectral reflection characteristics. Therefore, how to effectively use two types of data and comprehensively utilize echo intensity, geometric structure and spectral characteristics to realize fine classification becomes a key research direction for identifying dangerous obstacles in the field of operation and maintenance of the power transmission line. Disclosure of Invention The invention aims to provide an obstacle detection method and system based on laser point cloud identification, which solve the technical problems pointed out in the prior art. The invention provides an obstacle detection method based on laser point cloud identification, which comprises the following operation steps: Registering the laser point cloud with the multispectral image to establish a corresponding relation between the laser point cloud and the multispectral image; Extracting echo characteristics from the laser point cloud based on the laser radar, continuously analyzing the laser point cloud to obtain candidate vegetation point cloud and candidate building point cloud, extracting geometric characteristics from the candidate vegetation point cloud and the candidate building point cloud, classifying the laser point cloud according to the echo characteristics and the geometric characteristics to obtain building point cloud, and performing mask analysis on the building point cloud based on the multispectral image to obtain building contours; and visually judging whether the building is based on the outline of the building. Preferably, the echo characteristics of the laser point cloud are extracted based on the laser radar, and the specific operation steps are as foll