Abstract
Integrated sensing and communication (ISAC) is a headline feature for the forthcoming IMT-2030 and 6G releases, yet a concrete solution that fits within the established orthogonal frequency division multiplexing (OFDM) family remains open. Specifically, Doppler-induced inter-carrier interference (ICI) destroys sub-carrier orthogonality of OFDM sensing signals, blurring range-velocity maps and severely degrading sensing accuracy. Building on multi-user multi-input-multi-output (MIMO) OFDM systems, this paper proposes Doppler-Correction Filter Network (DCFNet), an AI-native ISAC model that delivers fine range-velocity resolution at minimal complexity without altering the legacy frame structure. A bank of DCFs first shifts dominant ICI energy away from critical Doppler bins; a compact deep learning network then suppresses the ICI. To further enhance the range and velocity resolutions, we propose DCFNet with local refinement (DCFNet-LR), which applies a generalized likelihood ratio test (GLRT) to refine target estimates of DCFNet to sub-cell accuracy. Simulation results show that DCFNet-LR runs $143\times$ faster than maximum likelihood search and achieves significantly superior performance, reducing the range RMSE by up to $2.7 \times 10^{-4}$ times and the velocity RMSE by $6.7 \times 10^{-4}$ times compared to conventional detection methods.