Lightweight Location Estimation of Boats at Sea from Aerial Drone Footage

Abstract

Abstract This thesis presents a lightweight, modular system for real-time location estimation of boats at sea using aerial drone footage on computationally constrained hardware such as a Raspberry Pi 5, specifically designed for autonomous drone landings. The architecture comprises three core microservice modules – object detection, tracking, and localization – communicating via a custom Redis-based utility class. Object detection is performed by a YOLO11n model, optimized through transfer learning and deployed within the NCNN framework for efficient inference. A nonneural network tracking algorithm, incorporating a simplified Kalman filter and the Jonker-Volgenant assignment method, manages object association. Two depth estimation techniques used for localization are implemented and compared: one utilizing the drone’s altitude and another using the known physical width of the boat. The developed system achieves a mean end-to-end latency of approximately 183 milliseconds. Comparative analysis revealed that the width-based localization method offers greater accuracy than the altitude-based approach at distances below 60 meters, even considering a mean 0.88-pixel error in the detected boat width from the object detection module (assuming a 1-meter drone altitude error). Furthermore, a proposed simple compensation technique for perspective distortion caused by droneboat misalignment demonstrated a reduction in mean position error by 73.11%.