An Adam W-Optimized Vision Transformer Framework with Back propagation Training for Driver Drowsiness Detection for Smart Vehicular Safety

DOI

https://www.doi.org/10.59256/indjcst.20260501039

Published In

Volume 5, Issue 1 (2026)

Published Online

January-April 2026

Pages

273-279

Article Type

Original Article

Check for updates

When drivers suffer from fatigue-induced cognitive impairment, the result is a tragic surge in worldwide traffic accidents that claim lives and drain vital economic resources. High-speed travel leaves no room for the slow reaction times that come with exhaustion, yet many operators don't realize they are impaired until it is too late. Because physical symptoms often lag behind actual cognitive decline, reactive safety systems simply aren't enough to prevent crashes. To truly move the needle on road safety, we must find ways to flag the very first physiological and behavioural warning signs. This paper focuses on catching fatigue in its earliest stages, allowing for the kind of proactive intervention that can stop an accident before it ever starts. Current methods, like analysing vehicle data often have reliability problems. Can be affected by environmental noise. Vehicle operators fatigue is a concern and fatigue detection can help prevent accidents. Road safety protocols need to be enhanced to reduce the number of accidents caused by fatigue Conventional detection methods have limitations and new approaches are needed to address this issue. To mitigate these deficiencies, this research proposes a real-time Driver Drowsiness Detection System leveraging optical sensing and deep representational learning architectures. The framework continuously acquires facial telemetry via a video stream, isolating ocular regions through facial landmark regression. An Adaptive Eye Characteristic Ratio (AECR) algorithm is employed to quantify prolonged ocular closure. A cardinal indicator of fatigue. Furthermore, a Vision Transformer (ViT) model analyses global spatial dependencies within the ocular features to categorize the operator's alertness state. Upon detection of somnolence, the system initiates immediate multi-modal alerts. A relational database backend logs temporal fatigue metrics for longitudinal performance analytics. Empirical validation under diverse illumination and pose conditions yielded a classification accuracy of approximately 92% with a false positive rate of 5% and sub-second inference latency. This cost-efficient, non-intrusive solution addresses the limitations of legacy systems and offers scalability for commercial fleets and private transport. Future iterations may integrate yawing analysis and infrared imaging for nocturnal efficacy.