LiDAR target detection and classification for ship situational awareness: A hybrid learning approach

Abstract

In recent years, LiDARs have been used to enhance situational awareness of autonomous vehicles, including in the marine domain, driven by the need for reliable detections in Marine Autonomous Surface Ships and Unmanned Surface Vehicles. Detecting obstacles and targets within point clouds is generally handled by a fully unsupervised learning framework. While effective and simple, this approach cannot classify targets. This paper presents a combined unsupervised/supervised approach for detecting and classifying marine targets and obstacles. The unsupervised detection framework is maintained by incorporating a lightweight supervised module capable of classifying detection outputs without disrupting the workflow. Rather than training on the entire point cloud, the proposed method focuses on selected target features, reducing model size and information exchange. Specifically, a Random Forest Classifier is trained on features extracted from the point-cloud dataset. The acquisition of an ad-hoc training dataset and its statistical analysis are presented to identify key features. The selection, training, and validation processes are outlined. Finally, the supervised model is integrated into a state-of-the-art unsupervised LiDAR detection pipeline and tested in a real scenario. The results demonstrate the hybrid framework’s effectiveness and compliance with real-time constraints.