基於機器手臂和視覺反饋控制的表型分析系統

Abstract

種植一個具有高經濟價值的農作物往往需要投入大量的人力和物力，農夫需要對作物的生長資訊進行記錄以監控植物生長狀況與調整種植策略，需要對影響植物產量的因素進行干預，例如植物的病蟲害的檢測。儘管農夫可以很好地完成大部分的工作，但人類無法全天候地工作，開發一個機器人輔助植物檢測系統，不僅能夠節省農夫的時間，對相關問題進行標準化處理，還能夠在任何時間對植物進行檢測。 然而，任何高階任務都需要具有魯棒性的基礎信息來完成，例如，植物葉片病蟲害檢測首先需要拍攝到葉片的全貌，即需瞭解葉片在空間的位置，切除受病的葉片和摘取水果也需要事先知道葉柄和果實的位置，檢測植物的高度與節點數量、分割葉片葉柄需要事先對植物點雲進行重建。已知植物點雲將為上述高階任務提供操作對象的基礎。 本文針對植物點雲重建，利用以標記為基礎的空間植物定位方法，提出了一個多階層植物點雲重建策略，包含對植物的快速掃描和主蔓遮擋重建算法。在流程内，開發了包含但不限於植物點雲濾波算法，多點雲曡合算法，主蔓提取算法，遮擋檢測算法和無遮擋視角優化器。由於主蔓連接這葉片和葉柄，所以主蔓的完整性影響著植物點雲分割的質量，故主蔓遮擋重建算法將對被遮擋的主蔓以最優視角進行重建。 實驗結果證明，本文開發的基於點云關係的主蔓提取算法具有較高的魯棒性，所提出的流程能夠完成定位、植物重建、最優視角重新規劃與植物點雲分割。

Planting a fruit with high economic value often requires a large amount of labor and material input. Farmers need to record information about the growth of the crop in order to monitor the growth status of the plants and adjust planting strategies, and need to intervene in factors that affect crop yield, such as the detection of plant diseases and pests. Although farmers can do most of the work well, humans are unable to work around the clock. Developing a robot assistance plant detection system not only saves the farmer's time, standardizes the handling of related issues, but also can inspect the plants at any time. However, any high-level task requires robust basic information to be completed, such as, plant leaf pest detection first requires capturing the overall appearance of the leaf, that is, knowing the position of the leaf in 3D space. Cutting off the infected leaf and harvesting the fruit also requires knowing the position of the petiole and fruit beforehand. Detecting the height and number of nodes of the plant, and segmenting the leaf petiole, require reconstructing the plant point cloud in advance. It is known that the plant point cloud will provide the basis for the above high-level tasks to operate. This paper proposes a multi-stage plant point cloud reconstruction strategy based on a spatial plant localization method based on fiducial marker, including a fast scan and main vine occlusion reconstruction algorithm for plants. In the process, a plant point cloud filtering algorithm, a pointcloud alignment algorithm, a main stem extraction algorithm, , an occlusion detection algorithm and the occluding-free viewpoint optimizer are developed, but are not limited to these. The experimental results show that the main vine extraction algorithm based on point cloud relationships developed in this paper is highly robust, and the proposed process can complete positioning, plant reconstruction, optimal view angle re-planning and plant point cloud segmentation. For the segmented point cloud, the robot arm can plan a trajectory to touch and simulate subsequent tasks.