期刊文献+

基于二维激光雷达的自适应阈值聚类分割方法 被引量:12

Adaptive Threshold Clustering Segmentation Method Based on Two-Dimensional Lidar
原文传递
导出
摘要 二维激光雷达广泛应用于室内障碍物检测中,而障碍物的聚类分割是环境感知中的关键技术。环境特征的复杂性和数据密度分布的不均匀性,导致传统聚类方法无法同时对不同距离、不同类型的障碍物实现良好聚类,容易发生漏检和误检。针对室内障碍物的检测需求,分析了激光雷达的数据特点和室内环境的几何特征,提出了一种改进的基于距离和障碍物特征的自适应阈值聚类分割方法,将阈值调整为随目标距离和类内密度变化的自适应参数。在基于激光雷达的智能车感知系统上进行了复杂障碍物的聚类分割实验,结果表明,相比传统方法,本方法可以明显改善不同距离、不同类型障碍物的聚类分割效果,分割准确度可达到92.23%。 Objective With the continuous advancement of science and technology,research into driverless smart cars has increased significantly.To realize environmental perception,driverless smart cars rely on sensors installed in the platform to collect and process surrounding environmental information.Sensors that perceive the external environment primarily include light detection and ranging(lidar),millimeter-wave radar,ultrasound,and cameras.Lidar is widely used in smart car obstacle detection and mapping navigation systems due to its high accuracy,large detection range,anti-interference properties,and the degree of depth information.Currently,two-dimensional lidar is widely used for obstacle detection.Obstacle clustering segmentation is a key technology in environment perception.Due to the complex environment characteristics and uneven distribution of data density,traditional clustering algorithms cannot achieve good clustering of different types of obstacles at different distances simultaneously,and obstacles can be missed and misdetected.Therefore,considering indoor obstacle detection requirements,the characteristics of lidar data and the geometric characteristics of indoor environments are analyzed and an improved adaptive threshold algorithm based on distance and obstacle characteristics is proposed.With the proposed algorithm,the threshold value is adjusted to change with the target distance and intra-class density.This method can effectively improve the correct detection rate of obstacles,and the segmentation accuracy reaches92.23%.Methods The proposed algorithm introduces the cluster density concept of the DBSCAN algorithm and an improved linear threshold algorithm.First,after receiving lidar data,the upper computer preprocesses the data,eliminates unstable points outside the effective range,and converts the remaining points to a rectangular coordinate system.Then,distance and density thresholds are set.The distance threshold changes adaptively with the distance of the received lidar data.The density threshold is the density average of the current class plus a multiple of its density standard deviation,which can vary as the obstacle class inner density changes adaptively.Then,the starting point is classified as category 1.From the second point to the last point,it is determined whether the distance between the current point and the previous point is greater than the current distance threshold.If the distance is less than the threshold,it is classified into the current category.Otherwise,a new class is created,and it is classified into the new class.It is also determined whether the last point and the starting point are less than the threshold.If less,the last category will be classified into the first category,otherwise it will not be included.Results and Discussions The indoor test results are shown in Fig.6.As shown in Fig.6(b),it is difficult for the linear threshold algorithm to consider the sparse scanning points of distant obstacles and obstacles closely distributed at a short distance.It can be seen from Fig.6(c)that the improved DBSCAN algorithm will group different obstacles that are close together into one category.In Fig.6(d),in the vicinity of the same coordinates,the proposed adaptive threshold algorithm based on distance and obstacle characteristics can distinguish smaller object distances at the same distance and can successfully segment objects with different distance data densities.The outdoor test results are shown in Fig.8.Under complex outdoor conditions,the proposed,DBSCAN,and linear threshold algorithms successfully segmented and detected various obstacles at close distances.However,the linear threshold method incorrectly classified walls as noise,and the improved DBSCAN algorithm mistakenly detected walls as two separate classes.For metal aluminum plates,the linear threshold algorithm and the improved DBSCAN algorithm mistakenly detected several types of obstacles.The proposed adaptive threshold algorithm successfully segmented and detected all types of obstacles that are close to each other and did not cause over-segmentation of walls and aluminum plates at a distance.In addition,the positive detection rate is significantly improved.Conclusions The experimental results showed that the improved clustering segmentation algorithm had the following advantages.First,it improved the positive detection rate of obstacle clustering and segmentation.Compared with the linear threshold algorithm,it had stronger environmental adaptability and could adapt to longer environmental distances and smaller object intervals.Second,compared with the global search of the DBSCAN algorithm,run time is reduced and the efficiency of obstacle detection and segmentation is improved.
作者 王祝 王智 张旭 崔粲 王健 Wang Zhu;Wang Zhi;Zhang Xu;Cui Can;Wang Jian(School of Science,Beijing Jiaotong University,Beijing 100044,China;Key Laboratory of Luminescence and Optical Information Technology of Ministry of Education,Beijing Jiaotong University,Beijing 100044,China)
出处 《中国激光》 EI CAS CSCD 北大核心 2021年第16期176-183,共8页 Chinese Journal of Lasers
基金 国家自然科学基金(61571035,61775012) 集成光电子学国家重点联合实验室开放课题(IOSKL2018KF22) 北京市自然科学基金-交控科技轨道交通联合基金(L201021)。
关键词 遥感 激光雷达 障碍物检测 聚类分割 基于密度的含噪声应用空间聚类 线性阈值法 remote sensing lidar obstacle detection clustering segmentation density-based spatial clustering of applications with noise linear threshold method
  • 相关文献

参考文献6

二级参考文献54

  • 1严新平,张晖,吴超仲,毛喆,雷虎.道路交通驾驶行为研究进展及其展望[J].交通信息与安全,2013,31(1):45-51. 被引量:41
  • 2李荣冰,刘建业,孙永荣.MEMS-IMU构型设计及惯性器件安装误差标定方法[J].中国惯性技术学报,2007,15(5):526-529. 被引量:27
  • 3冯少荣,肖文俊.DBSCAN聚类算法的研究与改进[J].中国矿业大学学报,2008,37(1):105-111. 被引量:89
  • 4WANG Z,ZHANG L Q, FANG T,et al.. A muhiscale and hierarchical feature extraction method for terrestrial laser scanning point cloud Classification [ J ]. IEEE Transactions on Geoscience and Remote Sensing, 2015,53 ( 5 ) : 2409-2425.
  • 5WANG H Y, WANG C, LUO H ,et al.. 3-D point cloud object detection based on supervoxel neighborhood with hough forest framework[ J ]. IEEE J. Selected Topics in Applied Earth Observations and Remote Sensing,2015,8 (4) : 1570-1581.
  • 6MINK P, LEE S J, LEE K H. Multi-scale tensor voting for feature extraction from unstructured point clouds [ J ]. Graphical Models ,2012,74 (4) : 197-208.
  • 7WANG L C, HUNG Y C. Hole filling of triangular mesh segments using systematic grey prediction [ J ]. Computer-Aided Design,2012,44(12) : 1182-1189.
  • 8WANG X C,LIU X P, LU L F, et al.. Automatic hole-filling of CAD models with feature-preserving[ J ]. Computers & Graphics ,2012,36 ( 2 ) : 101-110.
  • 9FLOYD R W. Algorithm 97 : shortest path [ J ]. Communications of the Association for Computing Machinery, 1962,5 ( 6 ) : 345.
  • 10ZHAO W, GAO S M, LIN H W. A robust hole-filling algorithm for triangular mesh [ J ]. The Visual Computer,2007,23 (12) :987-997.

共引文献132

同被引文献116

引证文献12

二级引证文献19

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部