強化深度學習應用於溫室無人載具導航及物件影像辨識

Abstract

使用無人載具於溫室中以自動化技術來觀察果物資訊，室內導航及物件辨識為兩項關鍵技術。載具導航中，即便許多導航演算法能順利計算出全局路徑規劃，卻因缺乏即時性局部行為規劃的特性，當障礙物分佈出現變化時便失去導航能力。果物辨識中，當果物被些微遮蔽或過於靠近時，辨識之結果往往不如預期。本研究藉由一台履帶載具車結合光達感測資訊實現深度確定性策略梯度 (Deep Deterministic Policy Gradient, DDPG) 強化學習，使其處理載具過於靠近障礙物的情境。在A*全局路徑規劃下，DDPG 能協助動態視窗法(Dynamic Window Approach, DWA) 局部行為規劃演算法的表現，進而使無人載具能適應於動態環境。一旦載具抵達目標座標，再以YOLOv4 深度學習模型使用載具上相機獲取之即時影像進行果物辨識，並且以分群演算法及座標轉換之技術於即時3D地圖中呈現果物資訊。實驗結果得知載具導航在介入距離0.35至0.36 m的DDPG協助之下，能即時調整行為決策，並避免 DWA 演算法於極端情境下導航發生中斷或載具原地打轉之現象，有效降低行走時間約20%。溫室資訊地圖中，YOLOv4與分群演算法處理後，能夠呈現0.2 m誤差的獨立果物座標與85%平均準確度的成熟狀態。此研究為智慧農業的應用上帶來更有效率的管理以及更穩定的系統。

Indoor navigation and object recognition are two key technologies for an unmanned vehicle observing fruit information in the greenhouse. With navigation, even though many navigation algorithms are able to make the global path planning successfully, they usually fail when the distribution of obstacles changes. With object recognition, the recognition results are often not as expected when fruits are slightly obscured or too close to each other. This research dedicated to the implementation of DDPG (Deep Deterministic Policy Gradient), a reinforcement learning method, for the crawler unmanned vehicle with lidar sensor in order to handle the situation where vehicle is too close to obstacles. With global path planning A* algorithm, DDPG can assist the DWA (Dynamic Window Approach) local behavior planning algorithm, so that the unmanned vehicle can adapt to the dynamic environment. Once the vehicle reaches the target coordinate, the deep learning model YOLOv4 identify fruits using real-time image obtained by the webcam. After that, fruits identification results will be clustered and their coordinate will be transformed and finally presented in the real-time 3D information map. The experimental results show that the vehicle can adjust the behavioral policy in real time with 0.35 to 0.36 m intervention distance DDPG involved, this improvement saved 20% of time, and avoided the navigation being interrupted or the vehicle spinning in place only handled by DWA algorithm. Besides, the individual fruits can be clearly displayed on the greenhouse information map with 0.2 m coordinate error and 85% average accuracy maturity status by YOLOv4 and clustering algorithm. This work enhanced the smart agriculture application with more efficient management and more robust system.