Application of YOLOv5s Algorithm for Real-Time Object Detection in Mobile Robot for Volcano Monitoring System Maria Evita, Sekar Tanjung Mustikawati, Mitra Djamal
ITB
Abstract
Indonesia, a county with 172 volcanoes and second after Japan for the most eruption event, should monitor and predict the volcano eruption to prevent the effect of this natural disaster. Therefore, we have developed a 4-wheeled mobile robot equipped with monitoring sensors and a Logitech camera for this purpose. The robot should have the ability to detect objects in this extreme environment to avoid collision while moving and monitoring the volcano^s physical parameters. It has been designed a deep machine learning of YOLOv5s algorithm for two objects mostly found at volcano such as trees and stones. After training steps (object identification- dataset downloading (Google Chrome Extension and Open Images v6)- image labelling (LabeImg)- augmentation process (blur and rotation)) had been carried out, the images of the object then be trained in three model variation which resulted in: mAP_0.5 = 51.9%, mAP_0.5:0.95 = 28.6%, 58% of precision and 50% recall with 12 minutes and 33 second of training time for first model (batch=16 and epochs=100)- mAP_0.5 = 59.7%, mAP_0.5:0.95 = 36.6%, 74% of precision and 54% recall with 36 minutes and 4 second of training time for the second model (batch=16 and epochs=300)- mAP_0.5 = 59.9%, mAP_0.5:0.95 = 37.6%, 80% of precision and 55% recall with one hour and 25 second of training time for the last one (batch=16 and epochs=500) as the best model of these variations. Furthermore, these result were displayed for all test images for the best model.
Keywords: mobile robot, object detection, volcano monitoring system, YOLO, deep machine learning
