From Accuracy to Trust: Deployment-verified Adaptive Confidence-aware Fusion for Occlusion-resilient Driver Monitoring on CPU-bound Automotive Edge Hardware

Read full text Summarise with VIRAT
Authors: Aruna S, K Sailaja Kumar, Sanathkumar S, Madugundu Chiranjeevi, Muralidhar Billa
Cited by: 0
References: 16

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Fatigue as well as distraction continue being the primary reasons for road traffic accidents, which in turn necessitate real-time driver state inference that is strong enough to work on automotive edge hardware.The existing vision-based Driver Monitoring Systems have often faced difficulties in case of occlusion, eyewear artifacts, temporal instability, and high false-positive rates when moved to CPU-constrained embedded environments.The work presented here is the first to bring forward a deployment-validated Adaptive Confidence-Aware Multimodal Fusion (ACAMF) strategy for detecting driver fatigue and distraction, that encompasses dynamic occlusion confidence weighting, temporal behavioural reasoning, and CPU-optimized inference for edge automotive hardware, all combined.The newly developed Advanced Driver Monitoring System (ADMS) framework is integrating a custom-trained Convolutional Neural Network (CNN) backbone (Yawn-Eye Dataset), fine-tuned InceptionV3 (MRL-Eye Dataset), and classical Machine Learning (ML) classifiers (Support Vector Machine (SVM), Random Forest) working in a unified confidence-weighted ensemble.Temporal dependency modelling is being done across blink rate, head nodding, and yawn duration which makes use of a lightweight Bidirectional Long Short-Term Memory (BiLSTM) decision head.The late-stage fatigue scoring is augmented by adaptive confidence fusion which is able to suppress single-cue failures in case of partial visibility, extreme head pose, low illumination, and reflective eyewear.Comprehensive assessment with methods such as 5-fold cross-validation, Receiver Operating Characteristic -Area Under Curve (ROC-AUC), and F1 reliability scoring coupled with real-world occlusion stress-testing has shown that ACAMF is superior to CNN-only and standard late-fusion baselines.It also has the advantage of reducing false positives by 34% and still having the throughput of 27-30 frames per second (FPS) on quad-core Advanced RISC Machine (ARM) CPUs with no hardware acceleration.The results are the proof of ACAMF being a scalable, trustworthy, and an edgedeployable solution for the next-generation driver safety systems.The findings also indicate that reliability-centric fusion is a more powerful determinant of real-world Driver monitoring system (DMS) trustworthiness than equal-trust late fusion, thereby, pushing forward the current development of uncertainty-aware driver safety frameworks which are ready for industrial automotive deployment.

Fields of Study & Subjects

Sleep and Work-Related Fatigue Human-Automation Interaction and Safety Autonomous Vehicle Technology and Safety

Publication details

Journal / source
International journal of intelligent engineering and systems
Volume
19
Issue
4
Pages
754–765
Publication date
23 March 2026
Language
EN