SViT: A Novel Multimodal Learning Approach for Ship Distance Estimation via Time-Series Data Visualization

초록

Accurately detecting and estimating ship distances is crucial for maritime safety, resource management, and preventing illegal activities. However, traditional Automatic Identification Systems (AIS) are highly vulnerable to signal interference and tampering, making themunreliable. To address this issue, we propose SViT, a novel multimodal learning approach. SViT converts time-series sensor data into2D image representations, allowing Vision Transformer (ViT) models to analyze them effectively. It further enhances computationalefficiency and noise reduction through hierarchical sensor feature selection. To ensure stable ship movement predictions, we introducea novel loss function combining MSE, Smoothness Loss, and Gradient Loss. Experimental results show that SViT outperforms LSTM-basedmodels in training speed and predictive stability. This study presents a robust framework for utilizing complex multisensor data, withpotential applications in security, surveillance, and maintenance.

키워드