設施內以影像與雷射為基礎之農用載具定位系統

Abstract

臺灣地處亞熱帶，適合農作物生長，但是也容易受到病蟲害、天災等影響，使得露天栽培的農產品減產，生產成本偏高且平均耕地規模小。透過設施農業，我們可以進行生產環境的監控、作業流程自動化以及降低病蟲害以及天災的影響。而農業行為經常需要使用載具來承載相關的儀器、設施抑或是農作物，遵循智慧農業 4.0概念，本論文欲建立一無人農用載具定位系統，藉以輔助路徑規劃以及避障系統。 其中，本論文使用的定位方法綜合了即時定位與地圖建構（SLAM）與視覺定位。在即時定位與地圖建構（SLAM）中，我們使用機器人作業系統（ROS）來建構程式內容，完成了無人載具於溫室內的點對點運輸問題。在視覺定位中，我們提出了兩種不同的監控相機架設模式，第一種為以監測影像平行於地面的方式架設，第二種為透過不平行的方式架設。前者在物件追蹤演算法上較為簡單且穩定，但是監控範圍受限於相機視角；後者在物件追蹤上需要使用深度學習的方法來達成，監控範圍較廣，但是誤差也較大。而論文中也比較了兩種將影像轉換至世界座標的方法，鳥瞰圖轉換法與全連接層轉換法，在三個不同的架設位置中，前者的平均誤差距離為0.10734公尺，後者為0.033525公尺。最後，透過結合即時定位與地圖建構（SLAM）與視覺定位，我們可以得到比原本精度更高的定位資訊，優化比例約3%~5%。

Because Taiwan is located in the subtropical zone, it is suitable for crop growth. However, it is also susceptible to plant disease and natural calamity. Therefore, it reduces the production and then the costs of agricultural products would be risen. Through the establishment of facilities agriculture, we can monitor the production environment, automate the process and reduce the impact of plant disease and natural calamity. Agricultural activities often require the use of vehicles to carry related equipment, facilities or crops. Following the wisdom agriculture 4.0 concept, this paper wants to establish an agricultural vehicle positioning system to assist path planning and obstacle avoidance systems. In this paper, the positioning method combines SLAM and visual orientation. In SLAM, we use the Robot Operating System (ROS) to construct the program and apply it to complete the point-to-point transportation problem in the greenhouse. In visual orientation, we propose two different modes to erect monitoring camera. The first mode is to erect the camera and let the monitoring image parallel to the ground. The second mode is non-parallel. The former is simple and stable on the object tracking, but the monitoring range is limited by the camera angle. The latter needs to use the deep learning method to achieve object tracking, and the monitoring range is wider, but the error is larger than the former. This paper also compares two methods of converting images coordinates to world coordinates, one is bird-eyes view transformation method and the other is fully-connected layer transformation method. Among the three different erection positions, the average error distance of the former is 0.10734 meters, and the latter is 0.033525 meter. Finally, by combining SLAM and visual orientation, we can obtain more accurate positioning information than the original, with an optimization ratio of about 3% to 5%.