基於PCA點雲分析之水果採摘夾取位置與姿態最佳化

Optimization of Grasping Position and Orientation for Fruit Picking Using PCA Point Cloud Analysis

隨著農業自動化技術迅速進展，如何有效提升水果採摘機器人之夾取成功率與操作穩定性，已成為當前亟待解決的關鍵課題。本文提出一套整合Intel RealSense D405 深度相機、ROS2 框架以及主成分分析（PCA）技術的水果採摘機器人夾取位置與姿態最佳化系統。該系統可透過深度相機即時擷取RGB 與深度影像，並利用RealSense 內建函式庫進行影像分割，有效且準確地擷取出目標水果之點雲資料。接著，透過PCA 技術進行點雲資料的分析，以估測水果相對於相機的三維空間姿態，進一步精準計算出最適夾取位置。實驗以室內金桔盆栽作為對象，進行多次夾取測試。結果顯示，導入PCA 進行姿態估測後的夾取策略，相較於僅使用XYZ 座標的方式，成功率由70% 顯著提升至93.3%，同時在夾取姿態的精準度上亦有明顯改善。本研究成果驗證了所提方法在提升農業採摘機器人穩定性與實用效能方面的可行性與有效性，展現出高度的應用潛力與推廣價值。

Rapid advances in agricultural automation are hampered by the ability of fruit-picking robots to grasp fruits well and consistently. In consideration of this problem, this study proposes a system that optimizes the fruit-picking position and orientation for fruit-picking robots. The system features an Intel RealSense D405 depth camera, runs on the Robot Operating System 2 framework, and is based on principal component analysis (PCA). The depth camera is used to acquire real-time RGB and depth images, and a built-in library in RealSense is used for image segmentation, enabling the target fruit’s point cloud data to be accurately and efficiently extracted. PCA is then used to analyze the spatial distribution of the point cloud to determine the fruit’s three-dimensional pose relative to the camera and the optimal grasping position relative to that pose. The system was evaluated on multiple grasping trials with an indoor potted kumquat plant. The use of PCA in the system led to a success rate of 93.3%, which was higher than the 70% achieved using only x–y–z coordinates, in addition to markedly improved pose accuracy during grasping. The system is thus suited for implementation in fruit-picking robots.