摘要
为了使基于视觉信息的农业机器人在葡萄种植园实现准确导航检测行间道路,将采集的视频分割成图片做成图像集,训练图像集对FCN、SegNet和U-NET三种卷积神经网络进行迁移学习,得到三种分割网络模型,分别用这些模型对测试集中不同环境下葡萄行间路径图像进行分割试验,以人工分割为基准对三种网络的分割效果进行评价,最后对分割前景提取导航基准线。试验表明,三种分割网络测试集分割精度(MCC)分别达到了89.96%、82.42%和75.78%,三种网络测试集上阴天图像的平均MCC分别比晴天高3.78%、0.45%和9.67%。三种网络中,FCN的总体分割效果最优,测试集上的平均分割精度(MCC)分别比SegNet和U-NET高5.87%和17.12%。FCN网络分割精度高,分割边缘清晰,提取的导航线精准,为农业机器人自主行走提供了一种可靠的导航方法。
In order to achieve accurate navigation of agricultural robots based on visual information between rows of grape plantations,it is necessary to detect inter-row paths.The captured video was divided into pictures to form an image set,and the training image set was used to migrate the three convolutional neural networks including FCN,SegNet and U-NET for segmentation.After training,three kinds of segmentation network models were obtained and these models were employed for segmentation testing on the inter-row path images taken in different environments in the test set,with the segmentation effects of the three networks evaluated based on artificial segmentation datum,followed by extraction of the navigation baseline for the segmentation foreground.The experimental results showed that the segmentation accuracy(average MCC)of three networks reached 89.96%,82.42%,and 75.78% respectively.The average MCC of the cloudy images in the test set was 3.78%,0.45%,and 9.67% higher than that on sunny days respectively.Among the three networks,the overall segmentation effect of FCN was best,for which the average MCC of the FCN was 5.87% and 17.12% higher than SegNet and U-NET,respectively.The path segmented by FCN has high precision,clear edges and accurate navigation lines,which can provide a reliable navigation method for agricultural robots to walk autonomously.
作者
宋广虎
冯全
海洋
王书志
SONG Guang-hu;FENG Quan;HAI Yang;WANG Shu-zhi(School of Mechanical and Electrical Engineering,Gansu Agricultural University,Lanzhou Gansu 730070,China;College of Electrical Engineering,Northwest University for Nationalities,Lanzhou Gansu 730070,China)
出处
《林业机械与木工设备》
2019年第7期23-27,共5页
Forestry Machinery & Woodworking Equipment
基金
国家自然科学基金项目(61461005)
甘肃农业大学SRTP项目(20180601、201906008)
关键词
葡萄园
行间路径
图像分割
深度学习
导航基准线
vineyard
inter-row path
image segmentation
deep learning
navigation baseline
