面向非结构环境的水果采摘机器人先验视觉识别感知模型与定位控制系统研究

The fruit picking robot have successfully picked fruit, which mainly depends on its vision recognition perception and picking position controllable ability. However, these current researches of the fruit picking robot vision system mainly used the real test of machine vision, and if there are leaves, branches, sky, light, shadow and the fruit object itself to overlap and occlusion in unstructured environments, there will be these issues,which included the low recognition rate, no precise spatial positioning and picking control instability of the fruit object. So, the recognition and picking position control of the fruit object have not reached the goal of the automation and commercialization, which needs to further explore and research. Therefore, on the basis of analyzing and reviewing the recognition perception behaviors and picking position control mechanism of the fruit picking robot vision system, combined with the factors of the fruit object in unstructured environment, taking vision statistical learning as theoretical basis, a fruit object recognition transcendental perception model was constructed by the multivariate statistics of the vision behaviors, and as this model, the recognition, spatial positioning calculation and the picking positioning control of the fruit object were systematically studied by adopting the virtual simulation technology. Expecting to provide some feasible and effective solutions and practices to the existing problems of the fruit picking robot vision system.

水果采摘机器人成功完成采摘，主要取决于其视觉识别感知以及定位采摘控制能力。然而，目前有关水果采摘机器人视觉系统的研究主要是基于机器视觉的实物试验研究，并且在果实对象识别和定位采摘控制方面，如果存在叶子、干枝、天空、光照、阴影和果实对象自身的相互重叠、遮挡等非结构化环境因素下，就会出现果实对象识别率低，果实对象空间定位不精准、采摘控制不平稳等问题，还没有达到自动化、商用化的目标，需要进一步深入探索研究。为此，本课题在分析水果采摘机器人视觉识别感知行为与定位采摘控制机理的基础上，并结合果实对象的非结构环境因素，以视觉统计学习为理论依据，构建一种基于视觉行为多元统计的果实对象识别先验感知模型，并基于此模型，采用虚拟仿真技术对果实对象的图像识别、空间位置定位计算和采摘定位控制等方面进行系统研究。期望，对水果采摘机器人视觉系统现存的问题，提供一些可行的、有效的解决方法和实践借鉴。

本课题围绕着设计一台球形类水果智能采摘机器人物理样机，运用视觉统计学习方面的理论和采用虚拟仿真手段，拟对非结构环境下，水果采摘机器人的先验视觉识别感知模型与采摘定位控制方面进行定性、定量研究。具体研究内容包括：（1）视觉统计学习理论框架问题。（2）果实对象的识别感知模型问题。（3）果实对象采摘点的定位计算问题。（4）果实对象的精准采摘定位控制问题。在水果智能采摘机器人装备领域取得以下具体成果：（1）提出了一种基于改进的YOLOv3模型现场荔枝检测方法。（2）提出了一种基于DeepLabV3+模型的荔枝分支的语义分割方法。（3）提出了一种基于混合蛙跳优化的采摘机器人相机标定方法。（4）发明了一种智能柑橘采摘机及采摘方法。（5）发明了一种基于图像识别的球状水果方位检测及调整的装置 。（6）发明了一种基于图像识别的高效水果采摘装置。并且，集成设计了一台拥有自主知识产权的球形类水果智能采摘机器人物理样机，初步实现果园球形类水果自动化收获。