Driver Gaze Zone Estimation Using Deep Neural Network

This chapter delves into the critical role of driver gaze estimation in enhancing road safety, particularly in the context of increasing road accidents in India. The study employs deep learning models, specifically EfficientNet variants, to classify driver gaze into five key zones: forward, rearview mirror, right-wing mirror, left-wing mirror, and center stack. The dataset used, DGW (Driver Gaze in Wild), is one of the largest collected under Indian conditions, featuring variability in lighting and diverse participant demographics. The research compares the performance of EfficientNet-B2 and EfficientNet-B7 models, with the latter achieving the highest accuracy of 76%. The confusion matrix reveals that the right-wing mirror and center stack zones are most accurately classified, while the rearview mirror and forward zones are frequently confused with neighboring zones. The study concludes by discussing the potential applications of driver gaze estimation in understanding driver behavior and improving road infrastructure and safety systems. Future research directions include exploring advanced deep learning architectures and sophisticated CNNs for better feature extraction and gaze zone classification.