Normalized Facial Features-Based DNN for a Driver's Gaze Zone Classifier Using a Single Camera Robust to Various Highly Challenging Driving Scenarios

Driver inattention is a significant contributor to fatal car crashes, leading to a need for accurate driver gaze zone classification methods. However, these classifications are particularly challenging under unconstrained conditions, such as when a driver's face is partially occluded by masks or scarves, when the environment has significant lighting differences, or when the driver's eyeglasses have reflections. This paper presents a framework that addresses these challenges by combining computer vision techniques and different deep-learning models to robustly recognize a driver's gaze zone under highly unconstrained conditions. The framework uses a Contrast-Limited Adaptive Histogram Equalization (CLAHE) to adjust the color space of the frame, making it easier to recognize features under varying light conditions. It then employs dense landmark detection techniques to achieve robust recognition of the face, eyes, and pupils, including the use of optical flow estimation methods for tracking pupil and eyelid movement. The framework considers two facial poses and trains individual Deep Neural Network (DNN) models for each pose. As facial structure varies among individuals, the feature vector parameters for these DNN models are based on different relations between pupil and eye landmarks proportional to the driver's face. The method has demonstrated its outstanding performance under a dataset involving highly unconstrained driving conditions.