Small Object Detection in Drone Captured Images for Maritime Surveillance

• Alexandre R. de Mello Certi Foundation
• Murilo M. Ozol Certi Foundation
• Bruno Mingoti Certi Foundation
• Leonardo Colussi Certi Foundation
• Kaique V. C. Souza Certi Foundation
• João P. R. Storer Certi Foundation
• Carlos Duek Iana Tecnologia
• Filipe A. N. Verri ITA

Resumo

Maritime surveillance requires timely and accurate information processing to support critical decision-making, particularly in search and rescue (SAR) operations where lives depend on swift action. This paper addresses challenges in automated maritime object detection by introducing both a novel dataset and a model-agnostic framework. We present a novel dataset, Aussys, collected using unmanned aerial vehicles (UAVs) along the Brazilian coastline, specifically designed for maritime object detection in SAR contexts. Concurrently, we introduce a model-agnostic framework for post-flight analysis on small object detection in drone-captured maritime images, leveraging image slicing and advanced inference techniques. We evaluate the performance of three state-of-the-art object detection frameworks, YOLOv5, YOLOv8, and Detectronv2, on this new dataset and assess how their capabilities are enhanced through integration with Slicing Aided Hyper Inference (SAHI) methodology. Our dataset captures the challenging real-world conditions of maritime environments, including variable object scales, poor target-water contrast, and disruptive factors such as glare and wave patterns. Results demonstrate that while all models provide acceptable baseline performance, SAHI implementation significantly improves detection accuracy for small and distant objects critical targets in maritime SAR scenarios. This research advances Geospatial Intelligence capabilities by providing both specialized training resources and comparative analysis of enhanced detection frameworks, with direct applications to SAR operations where improved detection rates could substantially increase survival outcomes.

Palavras-chave: YOLO, Training, Graphics, Accuracy, Surveillance, Pipelines, Boats, Information processing, Object recognition, Hyperspectral imaging