番茄採收機械視覺系統之研究

Abstract

目前台灣農業逐漸在轉型，將機器人應用到生物產業上是一主要趨勢。本論文為研究番茄採收機械視覺系統，並搭配宜蘭大學生機系所開發的爪具控制裝置，共同開發適用於設施內果蔬無人化採收之機器人系統。本研究針對溫室內之番茄進行視覺系統的建立，視覺系統包含三部分，影像擷取、影像處理與分析、雙眼立體視覺計算深度資訊。第一部份影像擷取方面，利用影像擷取控制流程獲得良好品質之影像，並且經過白平衡轉換將原有色偏影像做校正。研究結果在不同光環境或光源變化的情況下，皆可獲得曝光適度的影像；另一方面，對數型迴歸白平衡演算法可用於即時應用之色彩校正。經由以上自動曝光與白平衡轉換的結合，系統的影像擷取部分可即時在3~5秒內獲得良好品質之影像。第二部分進行影像處理與分析，程式搭配色調Hue色層將番茄分級，再搜尋各顆番茄位置與外部特徵二維座標，計算出番茄的蒂頭、中心、底部等資訊。另外若番茄有重疊狀況，程式亦可利用影像處理的方式求得各番茄的中心，解決番茄重疊造成的影像處裡問題。第三部分為雙眼立體視覺系統，使用單支攝影機搭配移動平台，進行雙眼視覺的建立，以降低開發成本。而立體視覺獲取番茄各特徵點之實際三維空間座標，可供採摘爪具及行走載具的路徑規劃。結果顯示雙眼視覺在60 cm~ 80 cm的工作範圍下，XY座標最大誤差為7.63 mm，第三維深度距離判斷也有良好的表現，與實際距離誤差小於4.4 mm。進一步將影像整合機器手臂作整合試驗，70 cm的工作距離下，採摘正確率可高達93.3%。因此綜合以上，本研究結果能成功擷取良好影像，並獲得番茄的外部特徵三維座標位置，達成本研究番茄採收機械視覺系統的研究目的。

In recent years, the agriculture of Taiwan was gradually in the process of reforming. The use of robots in agriculture industry has been regarded as an important trend. This study is aimed to develop a machine vision system for robotic fruit harvesting. This study will be integrated with the claw device controller, which has been built by I-Lan University, to develop a fruit harvesting robot in greenhouse. The research aims to construct a machine vision system for tomato picking using a robot in greenhouses. The system is composed of three parts, which are image grabbing, image processing, and dual camera stereo system. In the first part, the system used image grabbing control process to get a good exposure image, and conducted white balance adjustment. The results showed that no matter how the light condition was changed, the developed algorithm was able to obtain a good image. In the other hand, the log-regression-white-balance method could be used in real time situations. After the automatic exposure processing combined with white balance treatment, the image grabbing control process was able to obtain the good quality image in real time within 3~5 seconds. The system performs image processing and analysis in the second part. The program utilized hue index from HIS color system to do tomato grading, then searched the tomato position individually and the external characteristics on two-dimensional coordinates. It also calculated the information of pedicle, center, and bottom on the tomato. In addition, the program could also make use of image processing methods to achieve the center of the tomato if the situation of overlapping tomatoes occurred. The third part is dual camera stereo system. We used a single camera with a mobile platform for the establishment of binocular visionto reduce costs. The stereo vision system obtained actual three-dimensional position of each tomato feature for the claw path planning. Within the work range of 60~80 cm, the results showed the errorwas less than 7.63 mm in X and Y axes. In addition, the depth of the third dimension also had good performance and the error compared to the actual distance was 4.4 mm at most. Furthermore, the integration tests of the machine arm and the image system were done under a distance of 70 cm. The picking accurate rate was as high as 93.3%. In summary, the results of this study could successfully capture good images and calculate tomato’s three-dimensional coordinates of the external characteristics. Therefore, the machine vision system met the purpose of this study toward developing a tomato harvesting robot.