Autonomous Surface Vehicle in Search and Rescue Process of Marine Casualty using Computer Vision Based Victims Detection

Abstract

An Autonomous Surface Vehicle (ASV) is proposed to help the search process of maritime accident victims. A catamaran-type hull is implemented on the ASV for its stability. The ASV fix-mounted the electric propulsion system T200 Thruster on the stern of the ASV. Robot Operating System (ROS) is implemented on the ASV main software architecture. The ASV uses sensors such as a Global Positioning System (GPS), compass, Inertial Measurement Unit (IMU), and gyroscope to detect the state of the ASV. The ASV also uses ultrasonic sensors for obstacle avoidance. To interface with the actuators, a microcontroller STM32F4 is used. You Only Look Once (YOLO)v4 as Convolutional Neural Network (CNN) Architecture was used for the victim detection that was running on Nvidia RTX 2060 Mobile. The navigation system of the ASV performs well despite the noise from the sensor. The ASV is also capable of avoiding obstacles when moving at low speed. Dataset annotation was done manually from the images taken in Danau 8 Institut Teknologi Sepuluh Nopember (ITS). YOLOv4 gives an accuracy of 0.840203. Optimizing the YOLOv4 model from the darknet model to TensorRT increases the inference speed from 27 FPS to 85 FPS.