Motion primitives learning of ship-ship interaction patterns in encounter situations

Abstract

Understanding the ship encounter situations is a tough task, since the uncertain motions of encountering ships and the long-lasting frequent interactions make the encounter processes dynamic and stochastic. If we can decompose these complex encounter processes into motion primitives, which represent the elementary interaction patterns, an encounter can be more easily understood and identified. In this study, we propose a two-stream Long Short-Term Memory-based autoencoder (LSTM-based AE) approach to extract motion primitives from multi-dimensional encounter data without specific rules and prior knowledge. This approach leverages the sequential representation ability of LSTM to learn the temporal dependencies of each encountering ship motion. Then, we developed a fusion gate named dynamic artificial potential field (DAPF) to fuse the outputs of two LSTMs and generated the high-level representation needed to capture the spatiotemporal relationships of the interactions. After that, through a clustering method, the motion primitives were automatically extracted to describe various interaction patterns. The effectiveness of this approach was validated by naturalistic encounter data. The low reconstruction errors of the LSTM-based AE demonstrated that high-level representations captured the relationships of the interactions in both temporal and spatial dimensions. Case studies demonstrated that the utilization of motion primitives provided a semantically interpretable technique to analyze the interaction patterns and encounter process, which is conducive to situation awareness and decision-making for developing intelligent ships.