智能無人機之鯨豚監測技術研究

Abstract

近年政府積極發展離岸風電，風場位址緊鄰中華白海豚棲息地，施工時的打樁噪音可能對仰賴聽覺的海洋哺乳類造成危害。如何監測及掌握其活動習性成為了復育中華白海豚的關鍵之一。 本研究的目標為開發無人機鯨豚監測系統，該系統能夠按照規劃的搜尋區域自主進行鯨豚搜尋任務，結合深度學習的方法即時偵測影像中的目標，並且利用本文提出的鯨豚追蹤演算法進行鯨豚追蹤。 本研究使用臺灣大學鯨豚實驗室提供的2018年中華白海豚資料庫來訓練YOLOv8m模型，並以2022年真實中華白海豚目擊案例做為擬真的實海域實驗，以測試整個鯨豚偵測系統運作，並為未來在實際海域進行實驗準備。 在Gazebo模擬器中，設定虛擬中華白海豚以不同速度、姿態和路徑移動。虛擬無人機於海面上15公尺高，能夠穩定追蹤10節以下移動速度的中華百海豚。並在戶外實驗中測試真實無人機的追蹤性能，無人機能夠穩定追蹤移動速度3.5節以下移動速度的目標。

In recent years, the government has been actively developing offshore wind power, with wind farms located near the habitat of the Chinese white dolphins. The pile driving noise during construction may pose a threat to marine mammals that rely on their sense of hearing. How to monitor and understand their behavioral patterns has become a key aspect of rehabilitating the Chinese white dolphins. The goal of this study is to develop an unmanned aerial vehicle (UAV) dolphin monitoring system. This system is designed to autonomously conduct dolphin search missions within predetermined search areas. It employs deep learning methods to detect targets in real-time from captured images and utilizes the proposed dolphin tracking algorithm to track the dolphins' movements. For this research, the YOLOv8m model was trained using the Chinese white dolphin database provided by the National Taiwan University's Dolphin Lab for the year 2018. Additionally, real Chinese white dolphin sighting cases from 2022 were used for realistic sea-based experiments, testing the functionality of the entire dolphin detection system and preparing for future experiments in actual marine environments. In the Gazebo simulator, virtual Chinese white dolphins were set to move at varying speeds, postures, and paths. The virtual UAV, positioned 15 meters above the sea surface, was capable of effectively tracking Chinese white dolphins moving at speeds below 10 knots. Outdoor experiments were also conducted to test the tracking performance of real UAVs, demonstrating their ability to stably track targets moving at speeds below 3.5 knots.