Low-Latency Deep Learning Based Doppler Velocity Estimation for Integrated Sensing and Communication
In millimeter-wave Integrated sensing and communication (ISAC) systems, a high-resolution Doppler velocity estimation is crucial for localizing multiple tightly coupled mobile users in the environment for subsequent communication. Conventional subspace methods, such as the Estimation of Signal Parameters via Rotational Invariance Technique (ESPRIT), provide high resolution but incur high complexity and latency, limiting their use in real-time hardware. This work proposes a lowcomplexity end-to-end multi-layer perceptron (MLP) for Doppler estimation for IEEE 802.11ad ISAC waveform. The proposed MLP achieves up to 59.8% lower RMSE than classical ESPRIT while using fewer slow-time packets and up to $20.4 \times$ lower latency on the system-on-chip (SoC). A multi-model extension further improves robustness across signal-to-noise ratio (SNR), providing an additional 56% gain at low SNR by adaptively selecting model parameters. These results demonstrate that MLPbased estimation provides a scalable, efficient alternative to classical subspace methods for real-time ISAC systems.