- 學位論文
電腦視覺應用於綠牆果實成熟度分級與最佳採收路徑設計
Fruit Ripeness Grading and Harvesting Path Optimization of Green Wall Plant Based on Computer Vision
摘要
近幾年隨著氣候變遷，有關環保永續的議題在世界各國皆被列為重要發展目標，城市綠化在許多都市裡已經有成功的案例。本研究與德國柏林工業大學合作，致力於開發應用於垂直綠牆的採收機器人智慧型系統。本研究分成兩個部分，首先是對綠牆上果實的辨識與分級，要能夠用機器人取代傳統人力，必須使機器人也擁有機器視覺。本研究中使用深度學習訓練物件偵測模型，結合電腦視覺技術將綠牆影像中的果實分成未熟、成熟與過熟三類，成熟果實被視為本研究的採收目標。接下來是機器人採收路徑的設計，每顆果實的採收順序不同會影響採收機器人的總移動距離，不良的機器人動線規劃會導致整體採收作業效率低落。因次本研究中研發了一套基於自組織映射網路的最佳化路徑設計演算法，用來找出一條通過所有成熟果實的最短路徑供採收機器人依循。根據實測，成熟果實辨識模型的平均準確率為93.85%，與研究計畫之前的成果相比提升了近10%；在採收路徑設計部分，使用含有7顆、15顆與30顆採收目標的實例實測，所得到的採收路徑長度皆比研究計畫之前使用貪婪演算法求出的路徑減少了20%以上。本研究設計的最佳採收路徑演算法也能應用於旅行推銷員問題，對TSPLIB、VLSI TSPs與National TSPs中的berlin52、xqf131與qa194實例執行此演算法，所得出的路徑長度結果與現今已知最佳解的差距皆在6%以內。
並列摘要
In recent years, with climate change, the issues related to environmental sustainability have been listed as important development goals in countries around the world. There have been successful cases of urban greening in many cities. In cooperation with the Technical University of Berlin, Germany, this research is dedicated to develop an intelligent system for harvesting robot applied to vertical green walls. This research is divided into two parts, the first is the identification and grading of fruits on the green wall. To be able to replace traditional manpower with robots, robots must have machine vision. In this research, deep learning is used to train the object detection model, which combined with computer vision technology to divided the fruits in the green wall image into three categories: unripe, ripe and overripe. The ripe fruit was considered as the harvesting target for this research. Next is the design of the robot harvesting path. The different harvesting order of each fruit will affect the total moving distance of the harvesting robot, then the terrible traffic flow of robot will lead to low efficiency of the overall harvesting task. Therefore, an optimal path design algorithm based on self-organizing map was developed in this research to find the shortest path which go through every ripe fruit for the harvesting robot to follow. According to the actual measurement, the mean average precision of the ripe fruit detection model is 93.85%, which is almost 10% better than the previous research results. Using instances containing 7, 15 and 30 harvesting targets, the length of the harvesting path obtained from the algorithm was more than 20% shorter than the path obtained by the greedy algorithm in previous research. The optimal path design algorithm in this research can also be applied to the traveling salesman problem. Executing this algorithm in the berlin52, xqf131 and qa194 instances in TSPLIB, VLSI TSPs and National TSPs, the resulting path length are within 6% differ from the best-known solution.