Optimizing Computational Load and Energy Efficiency in UAV-Based Port Surveillance System

This paper presents a UAV-based surveillance system that enhances maritime safety by detecting unauthorized vessel entries in restricted port areas. The system employs a combination of sensor fusion, image stitching, and deep learning-based object detection (YOLO with MobileNet 0.75) on edge devices like the Sipeed Maixduino K210.

A Pub-Sub communication model using Apache Kafka enables real-time image sharing and vessel coordinate transfer between drones and a centralized Maritime Information System (MIS). The architecture is optimized using a DroneBalance algorithm for efficient energy use and task distribution among drones.

The system achieves a validation accuracy of 78.5% after 90 epochs of training on a custom dataset, demonstrating the potential for real-time, onboard maritime surveillance at scale.

Authors: Hariharan Sureshkumar*, Shardul Gharat, Dhumravarna Ambre, Lavanya Shetty, Aniruddha Kadam, Danish Ansari, Gajanan Birajdar
Download Paper