基于多目传感器的山区果园精准施药控制模型及路径规划方法研究

In order to develop the Path of Taihang Mountain and improve the income of the orchardist in mountain area，automation and precision management of orchard planting is a key problem to be solved urgently. In this project, a precise spraying control model and path planning method is presented for the pome and stone fruit orchards in complex environments based on integrated multi-view color/infrared sensor. Firstly, a classification and recognition method for six main crown types is studied by extracting the geometric features of the crown regions. And a multi-class classifier based on DAG-SVM is built for recognizing crown type models. Secondly, for improving spraying precision, the relationship between the control parameters of spraying and the attachment of drug settlement is analyzed, and the optimal spraying control model is established by the Elastic Net algorithm. Finally, for solving the problem that the route of orchard in mountain area is too complex to identify, a route recognition method based on AdaBoost studied by extracting and classifying the distribution features of fruit trees to achieve route correction and navigation function, so that the spraying equipment and the target fruit tree could keep the best spraying distance. The research results of this project are expected to make significant progresses in the precise spraying method and route recognition method of mountainous orchard, and to provide technical supports for developing intelligent, automatic and practical precise spraying equipment.

为了深化拓展“太行山道路”，提高山区果农收入水平，山区果园种植管理的自动化和精准化是亟待解决的关键问题。本课题针对山区复杂环境下的仁果及核果类果园，采用集成多目彩色/红外传感器，开展了基于多目传感器的山区果园精准施药控制模型及路径规划方法研究。首先，深入研究针对六类主要树型的分类识别方法，通过提取树冠区域几何特征，构建基于有向无环图的支撑向量机多分类器识别树冠模型；其次，从提高喷施精准度出发，分析喷施控制参数与药物沉降附着间的关系，结合弹性网回归分析，建立最优喷施控制模型；最后，针对山区果园路径复杂难于识别的问题，通过提取果树分布特征，研究基于AdaBoost的路径识别方法，实现路径偏差矫正和路径导航功能，使喷施设备与靶标果树保持最佳喷施距离。本课题的研究可望在山区果园的精准施药方法和路径识别方法上取得显著的进展，为研制智能化、自动化及实用化的精准施药设备提供坚实的技术支撑。

大幅降低化学农药用量是当前精准农业的主要发展方向。提高果园施药模型的精准性和喷施着药的有效性是发展节能、减排、环保、高效的精准施药技术的关键环节。本项目构建了配准精确的深度-彩色双目摄像机系统；设计了一种可调整喷施距离和喷雾量的施药系统，主要由机器视觉采集及处理模块、喷施控制模块、可控伸缩喷施臂和角度可控喷头等组成；提出了一种改进Mask R-CNN的B-Mask R-CNN树型识别模型，对搜集的复杂背景下矮化密植型、小冠疏层型、自然开心型、自然圆头型以及Y型5种常见修剪树型的果树进行检测识别；提出了基于ROI区域的树冠区域彩色-深度图融合分割方法，获得精确的可行驶区域，并规划出最佳喷施路径。基于实验结果，B-Mask R-CNN果树树冠树型检测模型平均识别精度达到98.7%，与其他检测模型相比树型识别精度更高，对复杂背景下的树型识别具有更好的鲁棒性；通过对比自动规划路径与标注可行驶路径的实验，证明本研究路径规划算法总体准确率达到97.5%。该研究成果可用于山区果园的精准施药方法和路径识别方法的选择和参考，同时也为喷施模式和精准控制的分析及应用提供了技术支撑。