Ship Length and Velocity Estimation using Range-Doppler Maps and Neural Networks

Accurate estimation of vessel kinematics and lengths is critical for maritime situational awareness, particularly for non-cooperative targets. Traditional radar systems often rely on scan-to-scan tracking filters, which suffer from initialization latency, or ISAR imaging techniques requiring specific movements. This paper proposes a multi-task deep learning framework for estimating a vessel’s velocity and length directly from sequences of range-Doppler (RD) maps. We utilize a convolutional recurrent neural network with an attention mechanism to extract spatio-temporal features from radar returns. If available, the model fuses target aspect angle metadata with these RD signatures to resolve the velocity ambiguity. Trained and validated on a real-world dataset collected from the Oslofjord, the proposed method demonstrates significant performance gains. Our analysis indicates that target angle metadata is primary for accurate kinematic estimation, while temporal integration across RD sequences plays a key role for improved physical dimension estimation.