A ResNet-Based Classification Method for Ship Viewpoint Estimation

Ship viewpoint estimation enhances perception capabilities, overcoming radar and AIS limitations in motion parameter acquisition. A dataset with diverse ship types and viewpoints was constructed using Unity3D simulation. The proposed method discretizes viewpoint estimation into 360 categories using ResNet, with localization cropping, image pyramids, and a probability-weighted calculation to reduce ambiguity and improve stability. Experiments showed an average error of 1.61° and a 7% accuracy improvement over the baseline, highlighting its practical potential.