設施內帶有環境資訊地圖之建立

Abstract

要讓載具在設施栽培時執行高複雜度的移動動作需要使用到室內定位的技術，而即時定位與地圖建構(SLAM)因為其不需要在環境中搭建訊號發射器或軌道，藉由感測器就能將載具定位到特定位置的特性，所以現在於室內定位領域中被廣泛的應用。但傳統SLAM的地圖只限於分辨特定位置是否有存在障礙物，無法提供更多的環境資訊，因此本論文提出藉由深度相機和深度學習模型捕捉環境資訊並映射到地圖的對應位置上的方法，使其可以分辨出作物和其他障礙物差別。實現方法:用SLAM建構出設施地圖，於建構地圖的同時將相機擷取到的影像輸入YOLO V3目標偵測模型中判斷作物相對於載具的分布位置；再經由座標轉換函式將作物分布位置映射到地圖的對應位置上；最後用分群演算法區分同種作物間的差異。實驗結果證實可以用SLAM正確的建構出設施地圖，並將YOLO V3偵測到的作物的真實位置映射到地圖上。於定位誤差的分析中發現相機旋轉角速度會影響誤差，藉由挑選適當的相機旋轉角速度能夠將定位誤差控制在公分等級。此研究建立的環境資訊地圖能夠使載具根據此地圖規劃出高複雜度的移動動作，例如前往指定的作物位置或讓載具在不會撞到作物的前提下盡量靠近作物等，以提高設施作物管理以及摘採、灌溉等工作的效率。

In order to perform high-complexity movements during cultivation in a facility, indoor positioning technology is required. Simultaneous localization and mapping(SLAM) which can locate the vehicle to a specific position by sensor does not need to build a signal transmitter or track in the environment, so it is now widely used in the field of indoor positioning. However, traditional SLAM maps are limited to distinguishing whether there is an obstacle at a specific location, and cannot provide more environmental information. Therefore, this paper proposes a method of capturing environmental information and mapping it to the corresponding position of the map through a depth camera and deep learning model. So that it can distinguish the difference between crops and other obstacles. The implementation method is to use SLAM to build a facility map and at the same time to input the image captured by the camera into the YOLO V3 object detection model. The position of the crop relative to the vehicle can be found, then the position of the crop to the corresponding position on the map through the coordinate transformation function can also be mapped. Finally, the clustering algorithm to distinguish the differences between the same crops is used. The experimental results confirmed that the facility map can be constructed with SLAM correctly and the real position of the crop detected by YOLO V3 is also mapped on the map correctly. In the analysis of the positioning error, it was found that the camera rotation angular velocity will affect the error. By selecting the appropriate camera rotation angular velocity, the positioning error can be controlled at a centimeter level. The map with environmental information created by this paper can enable the vehicle to plan high-complexity movements based on this map, such as going to a designated crop location or allowing the vehicle to be as close to the crop as possible without hitting the crop, so as to improve the efficiency of crop management, picking and irrigation etc. in facilities.