Mobile MIMO Radar Imaging via Synergistic Deep Learning and Wave Physics Modeling

Abstract

This article presents a multiple-input multiple-output (MIMO) radar imaging and reconstruction framework for mobile Internet of Things (IoT) platforms, such as intelligent robots and autonomous vehicles. It addresses the limited adaptability of existing model-based imaging methods in dynamic environments, as well as the substantial computational and training burden of end-to-end deep learning methods, by decomposing the radar imaging pipeline into three stages and synergistically combining data-driven perception with physics-based modeling. In particular, a lightweight range–Doppler segmentation network (RDSegNet) is used to robustly detect valid target regions (VTRs), followed by a modified hypothetical phase compensation (MHPC) algorithm for velocity disambiguation and wavenumber-domain approximation (WNDA) reconstruction for high-fidelity global image generation. Faithfully adhering to the first principles of wave physics, the proposed method jointly achieves high-fidelity environment reconstruction with very low computational complexity and can be deployed across diverse waveforms and MIMO topologies, thereby providing a foundation for real-time sensing and decision-making for mobile IoT platforms that operate under tight constraints on computation, memory, and latency. Experimental results on both simulated and real-world data validate its effectiveness.