自主巡航無人機系統應用於溫室洋香瓜之定位及量測

Abstract

目前全世界存在大量規模龐大的溫室，而溫室的作物生長狀況監測為溫室管理的一大焦點。傳統的人工巡查監測方式耗時且需要投入大量人力資源，使種植者無法快速、即時地了解溫室當前的整體情況。然而，自動化的無人機導航監控系統能夠解決此難題，基於純視覺定位的無人機不需安裝昂貴的傳感器，僅需搭載RGB相機即可執行導航任務。視覺定位無人機自主導航系統成為低成本的溫室自動監測之核心，從技術上改變了智慧農業的樣貌。本研究的無人機自主巡航系統，使用了加入ArUco Marker的ORB-SLAM2，稱為Enhanced ORB-SLAM2。ArUco Marker是一種標誌，具有特定的幾何模式，通常被用作計算機視覺中的參考點，能夠幫助系統進行位置定位和追蹤。ORB-SLAM2則是一種視覺定位與地圖構建演算法，能夠使用相機實現同時定位和地圖構建。實驗驗證了Enhanced ORB-SLAM2在環境存在陰影特徵的定位結果優於原始的ORB-SLAM2。此無人機導航系統可於溫室中自動執行各式飛行任務，且飛行軌跡均方根誤差範圍在30公分以下。此外，地圖校正使用仿設轉換算法，可以使地圖的MapAruco與人工量測的ArUco Marker位置完全貼合。果實偵測使用YOLOv4深度學習模型訓練，果實偵測模型之mAP達到0.96，DeepSORT基於此果實偵測模型運行果實追蹤任務。將DeepSORT的追蹤結果經過三步驟的資料清理後，使假果實實驗中的ID switch數量由平均5.83顆，下降至0顆，達到準確追蹤之目標。溫室果實定位算法基於清理後的果實追蹤結果，並使用三角測量算法計算果實位置，計算求得之果實位置再分別使用地圖校正以及迭代ArUco Marker的校正常數做位置校正，校正後假果實位置的均方根誤差由2.758公尺下降至0.223公尺，此果實追蹤與定位算法也已驗證可應用於真實果實之追蹤與定位。

There are a large number of large-scale greenhouses around the world, and monitoring the growth of crops in these greenhouses is a major focus of greenhouse management. Traditional manual inspection and monitoring methods are time-consuming and require a significant amount of human resources, preventing growers from quickly and instantly understanding the overall situation within the greenhouse. However, automated unmanned aerial vehicle (UAV) navigation and monitoring systems can address this challenge. UAVs equipped with pure visual positioning, without the need for expensive sensors, can perform navigation tasks with an RGB camera. In this study, an autonomous UAV navigation system was developed using Enhanced ORB-SLAM2, which incorporates ArUco Markers. ArUco Markers are specific geometric patterns used as reference points in computer vision, aiding in location estimation and tracking. ORB-SLAM2 is a visual positioning and mapping algorithm that achieves simultaneous localization and mapping using a camera. The experiment demonstrated that Enhanced ORB-SLAM2 outperforms the original ORB-SLAM2 in positioning results when dealing with environments containing shadow features. This UAV navigation system can perform various flight missions within a greenhouse autonomously, with a root mean square error of the flight trajectory within 30 centimeters. Additionally, map calibration using a similarity transformation algorithm ensures the alignment of the MapAruco with manually measured ArUco Marker positions. Fruit detection employs the YOLOv4 deep learning model, achieving a fruit detection model mAP of 0.96. DeepSORT utilizes this fruit detection model for fruit tracking. After three steps of data cleaning on DeepSORT's tracking results, the average ID switch count in false fruit experiments decreased from 5.83 to 0, achieving accurate tracking. The greenhouse fruit localization algorithm utilizes the cleaned fruit tracking results. Using triangulation, fruit positions are calculated and subsequently corrected using map calibration and iterative ArUco Marker calibration. The root mean square error of false fruit position decreased from 2.758 meters to 0.223 meters after correction. This fruit tracking and localization algorithm has also been validated for real fruit tracking and localization applications.