摘要
针对复杂城区环境下行道树靶标点云检测难度较大,导致基于激光雷达(LiDAR)的果园对靶施药技术难以推广的问题,提出基于YOLACT的行道树靶标点云分割方法,为行道树对靶施药提供基础数据。首先,应用移动激光扫描(MLS)技术采集街道一侧的三维点云数据;然后,提取深度、回波强度和回波次数信息,建立由二维LiDAR扫描线组成的三通道街道图像;最后,使用图像实例分割算法YOLACT建立行道树靶标分割模型,从街道图像中分割出每一棵行道树靶标。实验采集了一段300 m长街道两侧的点云数据,通过无损图像转换、切片、翻转扩充等处理得到1948张像素720×720的街道点云图像,按照6∶2∶2的比例划分出训练集、验证集和测试集,用来训练和测试行道树靶标分割模型。在386张测试图像上,令检测框与真值框的交并比阈值为0.5~0.9,以0.05为步长增加,得到的平均精确率为0.973,平均召回率为0.985,平均F1分数为0.979,平均每条LiDAR扫描线的处理时间是12.903 ms。实验结果表明,提出的方法能够快速准确分割出行道树靶标,为行道树对靶施药提供实时数据。
Due to the large spacing of street trees, the traditional continuous spraying technologies cause a large amount of liquid to be wasted to the ground or volatilized into air, which seriously pollutes the city environment and affect the residents’ living and working conditions. Targeted spraying technologies using the light detection and ranging(LiDAR) sensors have been successfully applied in orchards to enhance the spraying efficiency and droplet deposition for tree crops. Due to the difficulty of point cloud segmentation of street trees in complex urban environments, it is hard to apply the targeted spraying technologies using the LiDAR sensors for orchards to street tree spraying. To provide basic data for the street tree spraying, this study proposed a point cloud segmentation method for street tree targets based on the YOLACT image instance segmentation algorithm. Firstly, the mobile laser scanning(MLS) technology with a two-dimensional(2 D) push-broom LiDAR was applied to collect the three-dimensional(3 D) point cloud data on one side of the street. Then, the information of depth, echo intensity and echo number were extracted to build a three-channel color image composed of the LiDAR scanning lines. Finally, a segmentation model of street tree target was developed to segment each street tree target from the street image using the YOLACT image instance segmentation algorithm. In the experiments, the point cloud data of a 300-meter-long street were collected, and 1 948 street images with the size of 720×720 were obtained by lossless image conversion, slicing, flip expansion, etc. To train and test the street tree target segmentation model, the images were divided into training set, validation set and test set according to the ratio of 6∶2∶2. The intersection-over-union(IoU) threshold of the prediction box was set and the ground-truth box increased between 0.5 and 0.9 with a step of 0.05, the average precision was 0.973, the average recall was 0.985, the average F1-score was 0.979, and the average processing time for each LiDAR scanning line was 12.903 ms over 386 test images. The experimental results showed that the proposed method can quickly and accurately segment the street tree targets and provide real-time data for street tree targeted spraying.
作者
李秋洁
童岳凯
薛玉玺
徐志强
李相程
刘旭
LI Qiujie;TONG Yuekai;XUE Yuxi;XU Zhiqiang;LI Xiangcheng;LIU Xu(College of Mechanical and Electronic Engineering,Nanjing Forestry University,Nanjing 210037,China)
出处
《林业工程学报》
CSCD
北大核心
2022年第4期144-150,共7页
Journal of Forestry Engineering
基金
国家自然科学基金(31901239)
江苏省基础研究计划(青年基金)项目(BK20170930)。
关键词
对靶施药
行道树
点云分割
实例分割
YOLACT
targeted spraying
street tree
point cloud segmentation
instance segmentation
YOLACT
作者简介
李秋洁,女,副教授,研究方向为林木信息技术。E-mail:liqiujie_1@163.com。
