Navigating the Waters of Object Detection: Evaluating the Robustness of Real-time Object Detection Models for Autonomous Surface Vehicles

The safe navigation of Autonomous Surface Vehicles (ASVs) critically depends on their ability to detect objects, such as other vessels or obstacles. However, the variable and often harsh environment, characterized by fluctuating weather and image distortions, presents significant challenges for reliable object detection. To ensure a safe system operation, it is imperative to understand the extent of these challenges and identify specific vulnerabilities. In this paper
¹
, we evaluate the corruption robustness of state-of-the-art real-time object detection models for ASVs. We conduct a comprehensive analysis across various model scales, employing three distinct waterborne object detection datasets. By augmenting each test dataset with 15 types of corruption, we investigate model robustness according to two proposed metrics. Our findings reveal that certain corruption types markedly impair object detection performance, which could pose significant safety risks in autonomous shipping. Conversely, some corruption types have minimal effect on performance, regardless of the model or dataset. Furthermore, the results reveal a notable correlation between the scale of object detection model and its robustness, with larger models generally exhibiting higher resilience to corruption.