近年來隨著無人機的室外自動飛行日趨成熟,也越來越多研究在探討室內飛行的應用,然而在無法獲得GPS信號的室內環境中,如何定位變成最重要的課題。在本研究中,我們訓練Tiny YOLOv3辨識模型當作無人機的大腦,外接紅外線感測模組當作無人機的眼睛來實現避障功能,讓無人機能透過腦與眼的結合得知所處環境位置,並進一步將此技術應用於一個模擬倉儲環境,我們利用此室內自動飛行技術進行貨物盤點,如此一來可有效減少人力資源及降低工作人員進行高處檢查項目時的危險產生。此無人機盤點系統在軟體方面首先利用TF-Keras進行tiny YOLOv3的深度學習訓練客製的模型,並結合無人機的軟體開發套件(SDK)以及QR code掃描技術,在Android Studio平台上運作;在硬體方面,除了UAV上的內建攝像頭之外,我們還添加了五個用於避障的紅外線傳感器和兩個光敏的LED燈,此紅外線傳感器會將無人機與外在障礙物的距離資料透過LoRa及BLE回傳,達成自動避障的功能,而光敏性的LED主要針對低光源檢查項目所開發。無人機則利用Wi-Fi傳遞即時影像到手機中,藉此建立一個即時辨識盤點App。此系統在Android Studio平台上開發,並可於任何Android裝置上應用,故有便攜、易取得、流通性高等特點,藉此降低硬體規格,有利技術推廣,經實驗結果得貨品錯置偵測正確率皆高於86%,平均正確96.67%。
Autonomous flying in outdoor environments have been intensively studied recently. On the other hand, auto- flying in indoor environments, where GPS signals are not available, is more challenging. In this work, we tackle this challenge by redesigning the software and sensing components of off-the-shelf UAV to enable in- door autonomous flying and real-time monitoring. On the software side, we fuse a tiny YOLOv3 module in OpenCV class, a QR code module and an IR sensing module with the original UAV control class in an Android mobile phone for real-time autonomous flying. Combination of computer vision via tiny YOLOv3 and the depth information via IR sensors is developed for UAV self-positioning. On the hardware side, we add five IR sensors for obstacle avoidance and two LEDs for light sensitivity. The communication system includes Wi-Fi for image data and LoRa and BLE for IR sensor data. These enhancements together allow us to demonstrate an indoor application for inventory inspection in a warehouse environment. From the experimental results with 14 cases, the accuracy of misplaced cargoes detection on a shelf can be more than 86% and the average accuracy of misplaced cargoes detection is 96.67%.