An effective approach for real-time drowsy driving prediction using quantized fisher-Gabor features and latent-dynamic conditional random fields

Driver drowsiness or fatigue is among the most important factors that cause traffic accidents; therefore, a monitoring system is necessary to detect the state of a driver drowsiness or fatigue. In this paper, an automated vision-based system for real-time prediction of driver drowsiness or fatigue is presented, in which multiple visual ocular features such as eye closure, eyebrow shape, eye blinking, and other perfectly defined geometric facial features are employed as robust cues for driver's drowsiness. In addition, an efficient scheme is applied to extract local Gabor features of driver images based on Fisher's quantum information. A novel Fisher-Gabor descriptor (FGD) is then constructed from the extracted features, which is invariant to scale and rotation and also robust to changes in illumination, noise, and minor changes in viewpoint. The normalized FGDs are ultimately fed to a Latent Dynamic Conditional Random Field (LDCRF) classification model to predict whether the driver is drowsy/fatigued and a warning signal is thus issued (if required). A series of intensive experiments conducted on the benchmark NTHU-DDD video dataset show that the proposed system can predict driver drowsiness or fatigue effectively and efficiently, exceeding several state-of-the art alternatives by achieving a competitive detection accuracy of 97.6%, while still preserving stringent real-time guarantees.