A Multistage Adaptive 2D–3D CNN Framework With Modular Sub-Model Fusion For Robust And Real-Time Driver Drowsiness Detection in Iot-Enabled ADAS

Abstract

This study aims to address the persistent limitation of single-stage, monolithic convolutional architectures in driver drowsiness detection, which struggle to balance real-time computational feasibility against temporal sensitivity, and which degrade under variable lighting, occlusion, and eyewear conditions common to real-world driving. The objective is to design and evaluate a multistage adaptive 2D–3D CNN framework that decomposes drowsiness detection into dedicated sub-models for eye closure and occlusion alignment, yawn and mouth activity, head orientation and nodding, and contextual scene classification, before combining their outputs through an adaptive, multimodal fusion strategy optimized for real-time, edge-deployable, IoT-enabled ADAS integration.A study indicates that 2D and 3D CNN pathways serve complementary rather than competing roles, with 2D convolution efficiently capturing static, frame-level cues and 3D convolution capturing dynamic, sequence-level cues such as head nodding and microsleeps. Prior deployment-focused studies further indicate that pruning and quantization substantially reduce inference latency and memory footprint while preserving detection reliability, and that targeted solutions for lighting variability and occlusion improve robustness across diverse driving conditions.These findings support the framework's core premise that drowsiness is a distributed, evolving state better captured through specialized, fused sub-models than through a single generalized classifier, with direct implications for real-time, privacy-conscious, embedded ADAS deployment.