DESIGN OF REAL TIME OBJECT DETECTION SYSTEM FOR VOLCANO MONITORING APPLICATION

• Main Author: Tanjung Mustikawati, Sekar
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/55164
• Institution: Institut Teknologi Bandung

Indonesia is one of the countries located at the ring of fire with hundreds of active volcanoes. People especially who live near the volcanoes should be cautious of the unpredictable volcanic eruption. Therefore, the volcano should be monitored to predict the eruption earlier and set the risk zones around. For the latter purposes, it is needed a teleoperated monitoring system to prevent humans from being involved with the measurement process. This observation could be performed by a 4 wheeled mobile robot for both exploration and monitoring a volcano. In this research, it is used a 4 wheeled mobile robot called PRAWIRA (Perangkat Kendaraan Tanpa Awak untuk Daerah Rawan – Unmanned Vehicle for Dangerous Area). The robot should have an ability to avoid the obstacles in front of it in this area. In this research, it has been designed a real time object detection system for volcano monitoring application using deep learning from YOLOv5s (You Only Look Once) model for 4 objects (trees, person, stairs and stones) of 484 images in preliminary testing; and only 2 objects (tree and stone) of 400 images for the next one. These train processes were conducted after pre-train process was conducted with several steps, such as object identification; use downloading performed by Google Chrome Extension and Open Images v6 for preliminary testing and the real dataset from Mt. Tangkuban Perahu; image labelling by labelImg tools; augmentation process consists blur and crop, data training for varies optimum epochs and batches by Jupyter Notebook GPU; and applying the weight of train process into Raspberry Pi 3 operating system of the robot for object detection using a USB camera. In the preliminary testing, the batches and epochs were varied: batch=16 with epochs=100, batch=16 with epochs=500, batch=80 with epochs=100 and batch=80 with epochs=500. The second variation of this testing is the best result where batch = 16 and epochs = 500 resulted in 24.4% of mAP_0.5 and 12.6% of mAP_0.5:0.95. Moreover, the epoch was varied (100, 300 and 500) for batch=16 in the next testing, where the second variation yielded the best result for 59.9% of mAP_0.5 and 37.6% of mAP_0.5:0.95. Furthermore, the last testing was conducted at Tangkuban Parahu, one of an active volcano in West Java, Indonesia. The best result was performed by epochs = 500 and batch = 16 for 63.4% of mAP_0.5 and 40.4% of mAP_0.5:0.95. The model of this result was used for laboratory and field testings of object detection, although some objects was not successfully detected. Hereinafter, Raspberry Pi needed 30 second while the PC needed only 3 second to detect the objects.