基于改进卷积神经网络算法的路径导航研究

doi:10.13733/j.jcam.issn.20955553.2022.04.021

摘要/Abstract

摘要： 农业机械自主导航技术一直是现代农业发展的关键技术，而已有的机器视觉导航中普遍存在鲁棒性不强、适应性弱等缺点。针对上述问题，提出基于卷积神经网络的田间路径导航算法。根据主流语义分割模型FCNVGG16得到改进分割网络FCNVGG14，用于田间作物行分割任务的预处理，再通过非监督点聚类法进行特征点分类，最后采用改进后的Hough变换（PKPHT）拟合出导航直线。分析结果表明：与U-Net等主流算法相比，基于FCNVGG14网络模型的图像分割算法，IOU（交并比）指标在多通道输入时提升2%，在单通道输入时IOU指标提升4%，取得良好的分割效果。对分割网络FCNVGG14改进相对于传统的图像阈值分割算法，克服作物缺失、田间杂草过多、光照不均等自身缺陷导致的直线拟合时不可避免出现误检测、偏差大等问题。经田间路径导航试验证明，在田间照度符合人眼的情况下，基于FCNVGG16的改进模型路径识别算法检测准确率不低于92%，单帧检测时间在100 ms以内，在作业平台的速度不大于0.5 m/s条件下，最大横向偏差为9.84 cm、平均偏差不超过6.68 cm，说明用于机器视觉导航可行，这为低算力田间视觉导航提供新的方法和思路。

关键词: 视觉导航, 改进语义分割模型, 非监督聚类, 二次阈值分割

Abstract: The autonomous navigation technology of agricultural machinery has always been the key technology in the development of modern agriculture. The existing machine vision navigation had some shortcomings, such as poor robustness and weak adaptability. To solve these problems, a field path navigation algorithm based on the convolutional neural network was proposed. In this study, an improved segmentation network FCNVGG14 was obtained based on the mainstream semantic segmentation model FCNVGG16, which was used for the pretreatment of crop row segmentation tasks in the field. Then, feature points were classified by the unsupervised point clustering method. Finally, navigation lines were fitted by the improved Hough Transform (PKPHT). The analysis results show that, compared with the mainstream algorithms such as U-Net, the IOU index of the image segmentation algorithm based on the FCNVGG14 network model was improved by 2% in the multichannel input and 4% in the singlechannel input, achieving good segmentation effects. Compared with the traditional image threshold segmentation algorithm, it overcame the inevitable problems of false detection and large deviation in the line fitting caused by the defects of crop loss, excessive weeds in the field, and uneven illumination.
The field path navigation experiment proves that the detection accuracy of the improved model path recognition algorithm based on FCNVGG16was no less than 92%, the detection time of single frame was less than 100 ms, and the speed of the operating platform was no more than 0.5 m/s under the condition that the field illumination was consistent with human eyes. The maximum lateral deviation was 9.84 cm and the average deviation was less than 6.68 cm, indicating that it was feasible for machine vision navigation. This provides a new method and idea for field visual navigation with low computational power.

Key words: visual navigation, improved semantic segmentation model, unsupervised clustering, quadratic threshold segmentation

中图分类号:

S—3

黄林林, 李世雄, 谭彧, 王硕. 基于改进卷积神经网络算法的路径导航研究[J]. 中国农机化学报, 2022, 43(4): 146-152.

Huang Linlin, Li Shixiong, Tan Yu, Wang Shuo.. Research on farmland route navigation based on an improved convolutional neural network algorithm[J]. Journal of Chinese Agricultural Mechanization, 2022, 43(4): 146-152.

参考文献

［1］周俊, 姬长英. 基于知识的视觉导航农业机器人行走路径识别［J］. 农业工程学报, 2003, 19（6）: 101-105.

Zhou Jun, Ji Changying. Road recognition for agricultural robot guided by machine vision ［J］. Transactions of the Chinese Society of Agricultural Engineering, 2003, 19（6）: 101-105.

［2］ ID GarcíaSantillán, Montalvo M, Guerrero J M, et al. Automatic detection of curved and straight crop rows from images in maize fields ［J］. Biosystems Engineering, 2017, 156: 61-79.

［3］ ID GarcíaSantillán, Guerrero J M, Montalvo M, et al. Curved and straight crop row detection by accumulation of green pixels from images in maize fields ［J］. Precision Agriculture, 2017.

［4］梁习卉子, 陈兵旗, 姜秋慧, 等. 基于图像处理的玉米收割机导航路线检测方法［J］. 农业工程学报, 2016, 32（22）: 43-49.

Liangxi Huizi, Chen Bingqi, Jiang Qiuhui, et al. Detection method of navigation route of corn harvester based on image processing ［J］. Transactions of the Chinese Society of Agricultural Engineering, 2016, 32（22）: 43-49.

［5］ Sh A, Qz B, Bn C, et al. A guidance directrix approach to visionbased vehicle guidance systems ［J］. Computers and Electronics in Agriculture, 2004, 43（3）: 179-195.

［6］张勤, 黄小刚, 李彬. 基于彩色模型和近邻法聚类的水田秧苗列中心线检测方法［J］. 农业工程学报, 2012, 28（17）: 163-171.

Zhang Qin, Huang Xiaogang, Li Bin. Detection of rice seedlings rows centerlines based on color model and nearest neighbor clustering algorithm ［J］. Transactions of the Chinese Society of Agricultural Engineering, 2012, 28（17）: 163-171.

［7］胡静涛, 高雷, 白晓平, 等. 农业机械自动导航技术研究进展［J］. 农业工程学报, 2015, 31（10）: 1-10.

Hu Jingtao, Gao Lei, Bai Xiaoping, et al. Review of research on automatic guidance of agricultural vehicles ［J］. Transactions of the Chinese Society of Agricultural Engineering, 2015, 31（10）: 1-10.

［8］曾军英, 朱京明, 曾俊博, 等. 用于嵌入式终端的改进ENet街景语义分割［J］. 信号处理, 2021, 37（11）: 2148-2155.

Ceng Junying, Zhu Jingming, Ceng Junbo, et al. Improved semantic segmentation of ENet street view for embedded terminal ［J］. The Signal Processing, 2021, 37（11）: 2148-2155.

［9］陈峰. 基于深度学习的图像语义分割方法及其应用［D］. 南京: 南京邮电大学, 2020.

［10］仇一帆, 柴登峰. 无人工标注数据的Landsat影像云检测深度学习方法［J］. 国土资源遥感, 2021, 33（1）: 102-107.

Chou Yifan, Chai Dengfeng. A deep learning method for Landsat image cloud detection without manually labeled data ［J］. Remote Sensing For Land & Resources, 2021, 33（1）: 102-107.

［11］杜静, 蔡国榕. 多特征融合与残差优化的点云语义分割方法［J］. 中国图象图形学报, 2021, 26（5）: 1105-1116.

Du Jing, Cai Guorong. Point cloud semantic segmentation method based on multifeature fusion and residual optimization ［J］. Journal of Image and Graphics, 2021, 26（5）: 1105-1116.

［12］苟杰松, 蒋怡, 李宗南, 等. 基于Deeplabv3+模型的成都平原水产养殖水体信息提取［J］. 中国农机化学报, 2021, 42（3）: 105-112.

Gou Jiesong, Jiang Yi, Li Zongnan, et al. Aquaculture water body information extraction in the Chengdu plain based on Deeplabv3+ model ［J］. Journal of Chinese Agricultural Mechanization, 2021, 42（3）: 105-112.

［13］沈言善, 王阿川. 基于深度学习的遥感图像地物分割方法［J］. 液晶与显示, 2021, 36（5）: 733-740.

Shen Yanshan, Wang Achuan. Remote sensing image feature segmentation method based on deep learning ［J］. Chinese Journal of Liquid Crystals and Displays, 2021, 36（5）: 733-740.

［14］王雨. 面向自动驾驶场景的高效实时语义分割方法研究［D］. 南京: 南京邮电大学, 2020.

Wang Yu. Research on efficient realtime semantic segmentation methods for autonomous driving scenes ［D］. Nanjing: Nanjing University Of Posts And Telecommunications, 2020.

［15］ Ilyas T , Khan A , Umraiz M , et al. CEDNet: Crops and weeds segmentation for smart farming using a small cascaded encoderdecoder architecture［J］. Electronics, 2020, 9（10）.

［16］ Zhang Xin, Fu Longsheng, Karkee Manoj, et al. Canopy segmentation using ResNet for mechanical harvesting of apples ［J］. IFAC PapersOnLine, 2019, 52（30）.

［17］巩祎明. 农田场景下的图像语义分割方法研究［D］. 咸阳: 西北农林科技大学, 2018.

Gong Yiming. Research on image semantic segmentation method in farmland scene ［D］. Xianyang: Northwest A & F University, 2018.

［18］臧睦君. 图像场景识别的特征研究［D］. 长春: 吉林大学, 2015.

[1]	陈继清;王志奎;强虎;吴家华;赵超阳;谭成志;. 基于机器视觉边缘检测的园林喷药机器人导航线提取[J]. 中国农机化学报, 2021, 42(3): 40-47.
[2]	杨涛;李晓晓;. 机器视觉技术在现代农业生产中的研究进展[J]. 中国农机化学报, 2021, 42(3): 171-181.
[3]	张博立;吴蒙然;温兴;查家翼;李延凯;杨洋;. 基于动态网格和分区域聚类的玉米苗带识别算法研究[J]. 中国农机化学报, 2020, 41(7): 191-196.
[4]	李亦白;陈聪;曹光乔;. 基于计算机视觉的甘蓝生长过程导航线提取算法[J]. 中国农机化学报, 2020, 41(12): 131-138.
[5]	徐雪峰;黄余;. 基于机器视觉的机器人智能采摘实验平台开发[J]. 中国农机化学报, 2019, 40(12): 177-183.