Abstract
Orthogonal Time Frequency Space radar sensing provides a compact delay-Doppler representation for joint range and velocity estimation, but practical indoor measurements are often corrupted by fractional delay-Doppler misalignment, static clutter, and multipath interference. These effects distort the DD response and make classical peak detection unreliable, particularly when target energy leaks across neighboring bins and overlaps with structured interference. This paper formulates DD peak localization as a supervised peak prediction problem and proposes a physics-guided deep learning framework that maps corrupted DD frames to the underlying target peak coordinates. The approach is trained using a ray-tracing-based OTFS simulation pipeline that generates diverse indoor trajectories with realistic multipath and clutter, where fractional interference emerges naturally from continuous motion and finite DD resolution. Experiments demonstrate robust peak localization and strong generalization to unseen motion conditions, and additional validation on real radar measurements confirms the practical effectiveness of the proposed method for interferenceaware OTFS sensing.