Spectral and Spatial Transformer for Multi-Target Estimation in ISAC

Recently, end-to-end ISAC networks have attracted wide attention for their potential in significant performance improvement through joint optimisation. As an important component in ISAC networks, a fast and accurate multi-target detection network is critical. In this paper, we propose a spectral and spatial transformer to exploit the angle information by spectral-wise attention and the range/Doppler information by spatial-wise attention. Regarding multi-target detection as a set prediction problem, we propose a first-match-first-out loss to stabilize the training process. Experiments show our method performs better than two-stage methods and obtain comparable performance with joint estimation methods.