Deteksi Landing Pad pada Unmanned Aerial Vehicle (UAV) Multicopter Menggunakan YOLO Darknet

Teknologi robot terbang di Indonesia berkembang terbilang sangat pesat, dengan diselenggarakannya Kontes Robot Terbang Indonesa (KRTI) yang diadakan setiap tahunnya. Salah satu kategorinya adalah VTOL (Vertical Take-off and Landing) dengan salah satu misi pendaratan secara otonom yang mengharuskan multicopter mampu mendeteksi lokasi landing dengan tepat. Oleh karena itu, diperlukan pengembangan sistem computer vision untuk mendeteksi objek landing pad dengan keakuratan yang tinggi, sehingga multicopter mampu melakukan misi selanjutnya dengan baik. Penelitian ini menggunakan sebuah UAV Multicopter yang dilengkapi dengan webcam Logitech C920 sebagai platform computer vision. Sistem deteksi yang digunakan adalah YOLO Darknet, sebuah metode dari deep-learning. Jaringan YOLO yang digunakan adalah YOLOv3 Tiny, dan proses pelatihan melibatkan framework Darknet. Pada tahap pelatihan jaringan, diperoleh nilai average loss sebesar 0,0609 dan mean Average Precision (mAP) sebesar 99,4%. Selama pengujian jaringan, sistem berhasil mendeteksi objek dengan label "landing pad" dan menampilkan bounding box serta nilai confidence pada frame gambar yang dihasilkan. Pada pengujian menggunakan validation set didapatkan nilai peforma jaringan dengan rata-rata nilai accuracy sebesar 0.95, precision 0.98, recall 0.96, dan F1-score sebesar 0.96. Untuk pengujian secara real time diperoleh nilai peforma jaringan yaitu nilai accuracy sebesar 0.97, precision 0.97, recall 0.99, F1-score 0.98, dan nilai fps sebesar 20-22. Sedangkan, untuk pengujian secara real time dengan diberikan noise diperoleh nilai peforma nilai accuracy sebesar 0.90, precision 0.91, recall 0.97, F1-score 0.93, dan nilai fps sebesar 19-22.

The flying robot technology in Indonesia has been rapidly developing, as evidenced by the annual Indonesian Flying Robot Contest (KRTI). One of its categories is VTOL (Vertical Take-off and Landing), with one of the autonomous landing missions requiring multicopters to accurately detect the landing location. Therefore, the development of a computer vision system to detect landing pad objects with high accuracy is necessary, so that multicopters can successfully carry out subsequent missions. This research utilizes a UAV Multicopter equipped with a Logitech C920 webcam as a computer vision platform. The detection system used is YOLO Darknet, a method of deep-learning. The YOLO network used is YOLOv3 Tiny, and the training process involves the Darknet framework. During the network training phase, an average loss value of 0.0609 and a mean Average Precision (mAP) of 99.4% were obtained. During network testing, the system successfully detected objects labeled as "landing pad" and displayed bounding boxes as well as confidence values on the resulting image frames. In testing using a validation set, the network performance values were obtained with an average accuracy of 0.95, precision of 0.98, recall of 0.96, and F1-score of 0.96. For real-time testing, the network performance values obtained were an accuracy of 0.97, precision of 0.97, recall of 0.99, F1-score of 0.98, and an fps value of 20-22. Meanwhile, for real-time testing with added noise, the performance values obtained were an accuracy of 0.90, precision of 0.91, recall of 0.97, F1-score of 0.93, and an fps value of 19-22.