| Abstrakt: |
Indonesia has 127 active volcanoes, which have to be monitored continuously in normal, eruption, or after-eruption conditions, to minimize the effects of disaster. Therefore, we have developed a four-wheeled mobile robot for both exploration and monitoring of volcanoes. To finish its mission on uneven terrain full of obstacles, the robot should be able to detect and avoid these obstacles. Therefore, real-time object detection was designed using the YOLOv5s deep learning algorithm, which was implemented on a Raspberry Pi3 for the front camera of the robot. Before it was tested on a real volcano, the model of the algorithm was trained to be able to detect obstacles. The dataset was trained with three variations of epochs (100, 300, and 500) in sixteen batches of YOLOv5s. The last variant yielded the best results, at 63.4% mAP_0.5 and 40.4% mAP_0.5:0.95, with almost zero loss. This model was then implemented on a Raspberry Pi3 to detect trees and rocks captured by camera on Tangkuban Parahu Volcano. Most of the trees and rocks were successfully detected, with 90.9% recall, 79.9% precision, and 91.5% accuracy. Furthermore, the detection error was low, as indicated by low FP and FN numbers. [ABSTRACT FROM AUTHOR] |
