Accelerated CGH Generation for AR Head‐Up Display Based on the Novel YOLO‐MiDaS‐DeepCGH Pipeline

ABSTRACT This work presents a prototype augmented reality head‐up display (AR‐HUD) framework that integrates You Only Look Once (YOLO‐based) vehicle detection, monocular depth estimation, and phase‐only hologram generation using the DeepCGH model. Driving‐scenario video is processed frame by frame, where detected vehicle regions are converted into binary target patterns for hologram synthesis, and a monocular depth map is employed to estimate the relative distance of each rectangle, enabling depth‐aware holographic rendering. The predicted depth information is incorporated into the propagation model to assign appropriate reconstruction distances for different targets. The DeepCGH network generates corresponding phase holograms in real time, forming a holographic sequence that mirrors both the motion and depth variation of vehicles in the original video. Optical validation using an LCoS‐based Fourier reconstruction setup confirms temporal consistency and depth correspondence in the reconstructed rectangles. Performance evaluation shows stable per‐frame operation for both YOLO detection and DeepCGH inference, although reconstruction quality degrades in complex or densely populated scenarios. The results demonstrate the feasibility of combining object detection, monocular depth estimation, and computational holography within a unified AR‐HUD pipeline, while highlighting future improvements in latency reduction, depth accuracy, and reconstruction fidelity for practical automotive applications.