Implementasi YOLOv4-tiny dalam Sistem Deteksi Bola Batas dan Halangan Lintasan pada Kapal Tanpa Awak

DIVA DIANSARI H, Ika Candradewi, S.Si, M.Cs.; Muhammad Auzan, S.Si., M.Cs.

2022 | Skripsi | S1 ELEKTRONIKA DAN INSTRUMENTASI

Teknologi Unmanned Surface Vehicle (USV) membutuhkan sebuah acuan seperti sensor kamera untuk penentuan arah geraknya. Teknologi pengolahan citra digital klasikal rentan terhadap perubahan intensitas cahaya dan pengaruh lingkungan perairan sehingga akurasi deteksi objek berkurang. Sistem deteksi yang robust perlu dibangun agar dapat meningkatkan akurasi sistem navigasi kapal. Pada penelitian ini digunakan metode YOLOv4-tiny yang dibandingkan dengan YOLOv3-tiny untuk mendeteksi bola pembatas dan halangan lintasan. Dilakukan analisis hyperparameter subdivision dan learning rate untuk mendapatkan performa model terbaik dalam pelatihan. Model hasil pelatihan digunakan dalam pengujian video pada berbagai waktu pengujian untuk mengetahui ketahanan sistem. Pengujian dilakukan pada perangkat embedded yang banyak digunakan dalam perangkat otonom. Sistem deteksi dengan YOLOv4-tiny robust untuk berbagai waktu pengujian yaitu pagi, siang dan sore dari segi keberagaman intensitas cahaya dan refleksi objek serta kondisi saat ada dan tidaknya gelombang. Sistem mampu bekerja optimal pada siang hari dengan perolehan rata-rata performa F1-score 98,13%, mAP 99,18%, presisi 98,83%, recall 97,45% dan avg IoU 88,75%, sedangkan saat pagi dan sore hari performa sistem tidak menunjukkan penurunan yang signifikan. Dalam kondisi terdapat gelombang sistem menunjukkan performa tidak jauh berbeda saat tidak ada gelombang, yaitu F1-score 94,62%, mAP 91,93%, presisi 96,93%, recall 92,41% dan avg IoU 87,33%. Selain itu, rata-rata kecepatan pemrosesan deteksi video uji pada Jetson Nano 4GB sebesar 14,2 FPS.

Unmanned Surface Vehicle (USV) technology requires a reference such as a camera sensor to determine the direction of its motion. The classical digital image processing technology is susceptible to changes in light intensity and the influence of the aquatic environment so that object detection accuracy is reduced. A robust detection system needs to be built to improve the accuracy of the ship's navigation system. In this study, the YOLOv4-tiny method was used which was compared with YOLOv3-tiny to detect boundary balls and trajectory obstacles. Subdivision and learning rate hyperparameter analysis was performed to get the best model performance in training. The training model is used in video testing at various test times to determine the robustness of the system. The tests were carried out on embedded devices which are widely used in autonomous devices. A detection system with YOLOv4-tiny is robust for various testing times, namely morning, afternoon, and evening in terms of the diversity of light intensity and object reflections as well as conditions when there are waves or not. The system can work optimally during the day with an average performance of 98.13% F1-score, 99.18% mAP, 98.83% precision, 97.45% recall, and 88.75% avg IoU, while in the morning and evening the system performance does not show a significant decrease. In the condition where there is a wave, the system shows that the performance is not much different when there is no wave, namely F1-score 94.62%, mAP 91.93%, precision 96.93%, recall 92.41%, and avg IoU 87.33%. In addition, the average test video detection processing speed on the Jetson Nano 4GB is 14.2 FPS

Kata Kunci : Kapal tanpa awak, YOLOv4-tiny, SBC