DEVELOPMENT OF AN INTELLIGENT COMPUTER VISION TO DETECT THE PRESENCE OF VESSELS APPROACHING THE TSUNAMI EARLY WARNING SYSTEM

• Main Author: Wira Yogantara, Wayan
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/84267
• Institution: Institut Teknologi Bandung

The Tsunami Early Warning System (TEWS) works by relying on a buoy on the sea surface, which is connected to a water pressure sensor based tsunami detection system on the seabed via acoustic communication to detect the potential for tsunami waves. However, a significant obstacle to this system is the noise interference caused by ship propellers approaching the Pelampung permukaan, the value of which can exceed the noise value that the Pelampung permukaan's acoustic modem can tolerate. This research seeks a simple sea surface object recognition method that does not require a high computational process and can run on an embedded system with low power supply consumption. Information from this intelligent vision computer will later be used as a warning sign for approaching ships to stay away from the Pelampung permukaan location. It will also provide information to the Tsunami Data Center about the presence of foreign vessels that interfere with the performance of TEWS. A model of the best surface object detection network for Unmanned Surface Vehicles (USV), currently called the Temporal Context Water Segmentation and Refinement Maritime Obsctacle Detection Network (WaSR-T), is analyzed to be used in TEWS as a computer vision system device that can scout approaching ships. Despite USVs' excellent performance in recognizing surface objects for navigation purposes, WaSR-T networks require high computing devices. For TEWS, a lighter network is needed that balances the accuracy of detecting ships approaching TEWS as an integral part of an intelligent computer vision system in the open sea. Based on the analysis, a modifying the WaSR-T network proposed by replacing the most computationally intensive stages in the encoding process with a lighter network called Lightweight WaSR-T. By using a lighter encoder, the computing process can be run on a low memory ressources embedded system or on a single board computer with good performance to carry out the classification process and label each pixel in the recorded image as an approaching ship object, cloud, or sea surface. In this proposed lightweight WaSR-T network, the previous WaSR-T encoder using ResNet-101 is changed to MobilNetV3, and several feature layer maps are reduced as input to the decoder. To train and validate this lightweight WaSR-T network, a data set representing the high seas domain from an open data set was used, and we expanded the dataset with an additional recorded data set from the TEWS installation location in Indonesian waters. Based on quantitative results and computational load evaluation, the sensitivity of TEWS alarm generation for WaSR-T network performance results is 86,24 %, and lightweight WaSR-T is 86.14%. However, the computational load of the light WaSR-T network requires less memory by 4.3 %, and compare to WaSR-T by 13.2%. Therefore, our proposed lightweight WaSR-T network is promising for use as a key part of intelligent computer vision systems in TEWS. Keywords: Tsunami, Buoy, acoustics, USV, Lightweight WaSR-T, computer vision, open ocean, ResNet-101, MobileNetV3. ?