Deep Learning–Aided Frequency-Modulated Continuous-Wave Radar for Around-the-Corner Non-Line-of-Sight Perception at Urban Intersections

Urban intersections contain severe blind zones where buildings and roadside obstacles block line-of-sight sensing, limiting the ability of autonomous vehicles to anticipate hidden hazards. This paper presents an urban-intersection-oriented non-line-of-sight (NLOS) perception framework that exploits specular reflections from building surfaces using 77 GHz frequency-modulated continuous-wave (FMCW) automotive radar. All evaluations are conducted in a MATLAB-based simulation environment that models intersection geometry, building-induced occlusions, and specular reflection-assisted propagation, and generates 77-GHz FMCW radar echoes under controllable interference; real-world validation with measured radar data and richer multipath/material modeling is planned as future work. To improve robustness under noisy intersection interference, we propose a deep-learning-based mitigation module that restores corrupted radar echoes at the chirp level using a compact AlexNet-derived 1D regression backbone, with minimal architectural changes that insert a residual block after conv2 and apply batch normalization to enhance training stability and suppress interference while preserving informative echo characteristics. The restored echoes are then processed by conventional estimation steps to obtain range and azimuth-related angles. Under severe interference (Noise Factor = 3.0), unmitigated measurements exhibit large errors (root-mean-square error (RMSE) = 5.48 m/18.95°/10.77° for range/angle/azimuth deviation). Conventional AlexNet-based mitigation reduces these errors to 0.75 m/0.83°/0.93°, while the proposed improved AlexNet further reduces them to 0.56 m/0.46°/0.73°. The results demonstrate improved signal stability and measurement accuracy, supporting the potential practicality of low-cost NLOS perception in simulation for safety-critical autonomous driving at occluded urban intersections, subject to future real-world validation.