Raydar: A Ray-Tracing-Driven Framework Enabling Pattern-Based Recognition in ISAC Radar
2026 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events (PerCom Workshops), PerRad 2026, pp. 328–333
Abstract
This paper investigates the challenges of long-range ISAC radar sensing under angle-of-arrival (AoA) mismatch and introduces Raydar, a lightweight RT-driven simulation and recognition framework. Raydar reproduces the full ISAC sensing chain through six modular components, enabling realistic waveform and range–Doppler (RD) map generation without base-station hardware. Using RT-generated reflections, we show that vehicle targets exhibit strong geometry-driven RCS variations and that AoA mismatch further suppresses matched-filter (MF) responses, making amplitude-based discrimination unreliable beyond 100 m. Nevertheless, the underlying RD-map spatial patterns remain class-distinct. Exploiting this property, we design a lightweight CNN recognizer operating on CFAR-extracted RD patches. Simulations confirm that pattern-based learning substantially outperforms threshold methods, especially for weak-RCS vehicle targets, underscoring its value for practical ISAC sensing where AoA and RCS distortions are unavoidable.
Integrated Sensing and Communication (ISAC) is a core technology for Beyond 5G/6G, where communication waveforms are reused for sensing. Under realistic long-range conditions, recognition performance is strongly affected by angle-of-arrival (AoA) mismatch and by geometry-dependent radar cross section (RCS) variation. Empirical data collection is heavily constrained by regulation and privacy, and existing simplified analytical models cannot faithfully reproduce these distortions.
We introduce Raydar, a ray-tracing (RT)-driven simulation and recognition framework. Raydar reproduces the full ISAC sensing pipeline through six modular components — environment generation, transmit waveform generation, channel construction, received-signal synthesis, Range–Doppler (RD) map formation, and target recognition — and generates realistic waveforms and RD maps without any base-station hardware. Ray tracing brings geometry- and material-dependent reflections directly into the simulation.
Using Raydar, we show that vehicle targets exhibit strong geometry-driven RCS variation and that AoA mismatch further suppresses matched-filter (MF) responses, making amplitude-based discrimination unreliable beyond 100 m. Yet the spatial patterns in the RD map remain class-distinct, and a lightweight CNN trained on CFAR-extracted RD patches substantially outperforms threshold-based methods — particularly for weak-RCS vehicle targets.