A Deterministic Real-Time System for Monocular Depth Estimation on a Low-Cost Edge AI Platform: From Kernel to Inference

The rapid growth of autonomous systems in automotive and industrial applications demands perception pipelines that are both accurate and temporally deterministic. Although deep learning has significantly advanced monocular depth estimation, deploying such models on resource-constrained edge devices while meeting real-time safety requirements remains challenging. Conventional general-purpose operating systems often introduce unpredictable latency, making them unsuitable for safety-critical control loops.This paper presents an end-to-end, low-cost real-time edge AI framework implemented on the Raspberry Pi 5. A custom Linux kernel with the PREEMPT-RT patch is employed to achieve hard real-time scheduling with microsecond-level determinism. On this platform, a lightweight convolutional neural network, ADALITE, is introduced for monocular depth estimation. The model combines an efficient encoder–decoder architecture with a simplified multi-head self-attention mechanism and is optimized through knowledge distillation from a state-of-the-art teacher network. Experimental results demonstrate worst-case scheduling latencies below 200 μs under heavy system load and competitive depth estimation accuracy on the KITTI dataset. The proposed system achieves real-time inference suitable for collision mitigation, demonstrating that reliable and deterministic edge AI is feasible on affordable consumer hardware.