Authors: Hitesh Jitendra Jadhav, Santosh Shriram Karvar, Atharv Arun Patil, Gaurav Anil Waje, Gaurav Vijay Barde, Bajirao Subhash Shirole
Abstract: A significant percentage of traffic accidents in the world result from sleepy drivers. Although a number of detection methods have been established, their utility is often problematic. Physiological signals (EEG, ECG) and vision- based behavioral cues (eye closure, yawning) have been studied in the past, and deep learning models such as Convolutional Neural Networks (CNNs) have shown excellent accuracy in controlled settings. Significant gaps still exist, though, especially in the areas of robustness against various lighting conditions and occlusions, validation in on-road scenarios, and non-intrusive, computationally efficient systems appropriate for real-time deployment on mobile platforms. This review highlights the shortcomings of current vision-based approaches while synthesizing and critiquing them. It then suggests a future- focused approach based on a lightweight CNN architecture (like MobileNetV2) optimized for on-device inference with TensorFlow Lite. This work attempts to close the gap between academic research and useful, scalable solutions that can improve road safety by concentrating on a camera-based, non – intrusive system deployable on common Android devices.
DOI: http://doi.org/10.5281/zenodo.20786979
