基于移动激光扫描的行道树靶标实时检测

Real-time detection of street tree targets based on mobile laser scanning

针对行道树靶标实时检测的问题,本研究基于随机森林(random forests, RF)的逐点检测算法,建立一个能够实时且准确检测行道树点云的分类器。本研究所用点云为一段230 m的校园人行道,根据点的三维坐标、回波次数和回波强度信息基于立方体邻域提取宽度、深度、高度、次数、强度、维度和密度特征,然后根据这些特征的贡献度大小保留5个特征。对比了邻域搜索算法、点云特征数量、决策树数量和检测算法,并研究了点云密度对分类器性能的影响。结果表明:采用的立方体邻域特征提取时间比球域缩短了11.85%;在筛选特征过程中检测器性能基本稳定,特征筛选后提取特征时间缩短了65.40%,平均每帧点云的特征提取时间仅为24.72 ms;在考虑误差下降梯度和检测时间的前提下,决策树数量定为13;对比基于Boosting算法的行道树检测算法,本研究方法实时性更强;点云密度不断下降直到原来的1/20过程中,检测器性能保持平稳,表明该方法能够有效识别不同密度的行道树点云。本研究方法能够实现行道树靶标实时检测。

To address problems of real-time detection of street tree targets, this study investigated the random forests(RF) based point-by-point detection algorithm to build a classifier capable of real-time and accurate detection of street tree point clouds using mobile laser scanning(MLS). The point cloud used in this study was obtained by scanning a 230 m section of campus sidewalk with a cart equipped with a light detection and ranging(LiDAR) sensor. The point cloud contained 78 street trees. The point cloud was divided into a training set and a test set based on the road length ratio of 1∶10. The width, depth, height, count, intensity, dimension, and density features were extracted based on the 3 D coordinates, echo count and echo intensity information of the points based on the cubic neighborhood, and then five features were retained according to the magnitude of the contribution of these features. The neighborhood search algorithm, the number of point cloud features, the number of decision trees and the detection algorithm were experimentally compared, and the effect of point cloud density on classifier performance was investigated. The experimental results showed that the cubic neighborhood feature extraction time used in this study was 11.85% faster than the spherical neighborhood. The detector performance was basically stable in the process of screening features, and the feature extraction time was 65.40% faster after the feature screening, and the average feature extraction time for each frame of the point cloud was only 24.72 ms. The number of decision trees in the RF model was set to be 13 under the premise that the error descent gradient was as small as possible and the detection time was as short as possible. Compared to the Boosting algorithm for the street tree detection, the real-time performance of the algorithm in this study was stronger. The detector performance remained stable during the process of sampling the point cloud to reduce the density of the point cloud to 1/20 of the original one, which showed that the method of this study can effectively identify the point clouds of different densities of street trees. The algorithm in this study enabled the real-time detection of street tree targets.