基于多特征融合的树干快速分割算法  

Fast Segmentation Algorithm of Tree Trunks Based on Multi-feature Fusion

作　　者：刘慧[1] 朱晟辉 沈跃[1] 汤金华 LIU Hui;ZHU Shenghui;SHEN Yue;TANG Jinhua(School of Electrical and Information Engineering,Jiangsu University,Zhenjiang 212013,China)

机构地区：[1]江苏大学电气信息工程学院

出　　处：《农业机械学报》2020年第1期221-229,共9页Transactions of the Chinese Society for Agricultural Machinery

基　　金：江苏省重点研发计划项目(BE2018372);江苏省自然科学基金项目(BK20181443);江苏省国际科技合作项目(BZ2017067);江苏高校“青蓝工程”项目;镇江市重点研发计划项目(NY2018001);江苏省三新工程项目(NJ2018-12)

摘　　要：针对传统的树干分割算法存在分割精度低、实时性差的问题,提出了一种融合深度特征和纹理特征的树干快速分割算法。首先,通过Realsense深度摄像头采集树干彩色图像和深度图像;随后,采用超像素算法对彩色图像进行分割,并融合深度和纹理相近的相邻超像素块,最后对深度图像进行宽度检测,并对宽度在阈值范围内的物体所属的超像素块进行色调匹配,区分树干与非树干。在室内和室外植株实验中分别运用本文算法、GrabCut算法与K-均值算法进行树干分割,本文算法的平均召回率和平均准确率分别为87. 6%和95. 0%,GrabCut算法分别为78. 0%和92. 8%,K-均值算法分别为80. 2%和89. 1%;本文算法平均耗时为0. 20 s,GrabCut算法为0. 66 s,K-均值算法为4. 42 s。实验结果表明,本文算法的快速分割效果较好,在保证分割精度的同时,简化了识别过程,加快了分割速度,能够应用于室内和室外树干的分割。Accurate identification of orchard trunks can provide effective information for orchard robot localisation and navigation. The traditional tree trunk segmentation algorithm has low segmentation accuracy and poor real time performance. To solve this problem,a fast segmentation of tree trunks based on depth and texture features was proposed to improve segmentation accuracy and real time performance.Firstly,a Realsense depth camera was used to capture color and depth images of tree trunks. Then,a superpixel segmentation algorithm was proposed to segment color images,and fuse adjacent superpixel blocks with similar depth and texture values. Finally,plant trunks were distinguished from no-trunk targets in candidate superpixel blocks based on trunk width threshold setting in depth images and hue value matching in color images. Both indoor and outdoor experiments were conducted to compare the proposed tree trunk segmentation algorithm with traditional GrabCut algorithm and K-means algorithm.The average recall rate and average accuracy of the new algorithm were 87. 6% and 95. 0%,respectively,while that of the GrabCut algorithm was only 78. 0% and 92. 8%,respectively,and the K-means algorithm was 80. 2% and 89. 1%,respectively. Meanwhile,the average time of the proposed algorithm was 0. 20 s,while the GrabCut algorithm was 0. 66 s,and the K-means algorithm was 4. 42 s.The experimental results showed that the proposed algorithm was effective in fast segmentation,and can be applied to tree trunk segmentation.

关 键 词：树干识别 图像分割 深度特征 纹理特征 简单线性迭代聚类算法 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] S24[自动化与计算机技术—计算机科学与技术]