DNN-Based Radar Target Detection With OTFS

Orthogonal Time Frequency Space (OTFS) offers significant benefits in high-mobility wireless communication scenarios by modulating information in the delay-Doppler (DD) domain. It also exhibits significant potential in radar sensing due to the inherent connection between OTFS demodulation and radar range-Doppler matrix computation. However, the performance of the classic constant false alarm rate (CFAR) detector is hindered by the masking effect in multi-target detection. Besides, estimating noise distribution and determining suitable detection window sizes greatly impact the performance of the CFAR detector. In this letter, we present a deep neural network (DNN)-based scheme for radar target detection with OTFS. In our proposed method, the deep learning network utilizes the OTFS demodulation signal, processed by two-dimensional (2D) pulse compression and expansion, as the input. The network outputs the proposal regions and the probability of the region containing a target according to the input and prior learning weights. Then a non-maximum suppression (NMS) method is proposed to suppress the false alarm rate. The simulation results illustrate the potential advantages of the proposed scheme, highlighting its superiority in multi-target detection and generalizability at multiple signal-to-noise ratios (SNRs) with OTFS when compared to the classic benchmarks.