关键词：
植株;Kinect;三维点云;配准;信息融合
摘 要：
针对传统点云信息融合需要限制传感器之间位置以及繁杂标定和Kinect传感器室外工作受光照条件影响会出现目标边缘缺失的问题,提出了基于SICK和Kinect相机组合探测的植株点云超限补偿信息融合方法。首先采用SICK二维激光传感器融合实时行进速度传感器,实现对植株三维点云重构,同时通过Kinect传感器获取植株彩色和深度图像合成彩色点云,然后分别对SICK和Kinect异源点云进行阈值滤波预处理和体素栅格下采样,求取各点法线及快速点特征直方图,利用采样一致性初始配准方法使异源点云之间拥有较好的初始位置关系,再进一步使用ICP算法精确配准,通过近似最近邻搜索和超限补偿的方法完成点云信息融合。在超限补偿方法中,通过对比转换后点云间误差,判断数据有效性,实现对数据的最终融合。试验结果表明,本文方法可以有效、准确地实现不同点云之间的信息融合,并能有效抑制阳光的干扰。
译 名：
Plant Point Cloud Information Fusion Method Based on SICK and Kinect Sensors
作 者：
LIU Hui;PAN Chengkai;SHEN Yue;GAO Bin;School of Electrical and Information Engineering,Jiangsu University;
单 位：
LIU Hui%PAN Chengkai%SHEN Yue%GAO Bin%School of Electrical and Information Engineering,Jiangsu University
关键词：
plant;;Kinect;;3D point cloud;;registration;;information fusion
摘 要：
Aiming at solving the location restrict of sensor and complicated calibration problem of traditional point cloud fusion and the problem of missing edges in Kinect outdoor work,a method of plant point cloud information fusion based on SICK and Kinect was put forward. The 3 D accurate reconstruction of data acquired by 2 D laser sensor,SICK LMS151,needed the cooperation of real speed sensor. Due to the short time of obtaining per frame data of SICK and low speed,X-axis data for each column was set the same and the distance between two columns was calculated according to the real speed and sensor working frequency. Original color point clouds of plant were merged by color images and depth images obtained by Kinect. Firstly,preprocessing was carried out to extract point cloud of plant from original point clouds,in which lots of point clouds of background and noise were involved. In order to minimize the amount of points and keep enough characteristics,voxel grid was executed to down sample the plant point cloud.Secondly,normal calculation was executed on each point of plant point cloud to compute feather information by making use of depth features and peripheral point,fast point feature histograms(FPFH)was performed to enrich the feather information,which contained 33 dimensions element for each point.Thirdly,sample consensus-initial alignment(SAC-IA) algorithm,an initial registration algorithm,was applied to register SICK laser point cloud and Kinect point cloud to provide a better spatial mapping relationship for accurate registration. Fourthly,on the basis of initial registration,the iterative closest point(ICP) algorithm was used to refine the initial transform matrix inferred by initial registration.Finally,the information fusion was adopted by ANN algorithm to find corresponding point in Kinect color point cloud from SICK point cloud. However,Kinect point cloud would lose lots of edge information when working under the sun,resulting in the fusion faults. Overlimit compensation would work,when ANN cannot find the corresponding point or the distance between corresponding point and searching point was beyond threshold,the searching point would be considered as corresponding point and found the color information by corresponding function provided by Kinect SDK. Experiments showed that the fusion method can effectively and accurately realize the information fusion between different point clouds and suppress the interference of the sun.
