FDBR: Ultra-fast and data-efficient behavior recognition of port vessels using a statistical framework

Abstract

Listen

Translate article

Accurate recognition of maritime vessel behavior is essential for ensuring safety and optimizing port operations. Traditional methods often focus on isolated behaviors, like identifying individual stops or avoiding collisions between encountering vessels, without adopting a systematic approach. Single-vessel behavior recognition typically relies on clustering or deep learning techniques, which face challenges in adapting to diverse scenarios or require extensive data annotation. Moreover, current research on multi-vessel behavior tends to overlook cooperative interactions within port areas, which limits the potential for enhancing port efficiency. To address these challenges, this paper introduces a novel ultra-fast and data-efficient systematic statistical framework for port vessel behavior recognition (FDBR). For single-vessel behavior, we present a vectorized stop identification method with sliding window correction, enabling real-time recognition without the need for data annotation. For multi-vessel behavior, we employ a non-linear logistic regression model that incorporates dynamic vessel features to recognize cooperative behaviors with minimal data annotation. Our method achieves 98.738% accuracy in stop recognition, surpassing most deep learning approaches, and 93.930% accuracy in cooperation recognition. Additionally, recognizing different vessel behaviors at the Port of Los Angeles enables accurate extraction of various functional and high-risk areas, contributing to the optimization of port operations and enhancement of maritime safety.