Real-time Fatigue Driving Detection Algorithm Based On The Fusion Of Eye Feature And Head Pose Estimation

In recent years, with the rapid development of vehicle transportation, the corresponding driving safety problem increasingly attracts widespread attention. The annual number of deaths due to traffic accidents worldwide accounts for a large proportion, and this case deteriorates year by year. According to a survey of the National Highway Traffic Safety Administration (NHTSA), a large part of traffic accidents was caused by drivers’fatigue driving. Therefore, it’s necessary to develop a real-time and reliable fatigue driving detection system, which is of great significance for drivers and the whole society.In this paper, we carry on a deep research on fatigue driving and propose a fatigue driving detection system based on the state-of-art achievements. This system is a real-time, non-contact, and computer vision based system, and is consisted of two different kinds of subsystems, which are fatigue driving detection subsystem based on eye feature detection and fatigue driving detection subsystem based on head pose estimation respectively.The subsystem based on eye feature detection is suitable for good external environment which means the probability of catching the eye feature exactly. Firstly, face detection is carried out with the cascaded classifier based on Haar-like features and AdaBoost algorithm and the classifier’s parameters are optimized for this certain application. Secondly, the iris is located with the statistical proportional model whose parameter is adjustable and the corresponding image processing operations. Finally, the area proportion that the iris takes in its minimum circumscribed circle is used as the indicator for judging the state of the eye and the percent of eye closure (PERCLOS) is used as the criteria to determine whether the driver is tired.The subsystem based on head pose estimation is mainly designed for the situation in where there exists a lot of disturbances and it is difficult to identify the facial features. Frontal face detection and profile face detection (if necessary) are carried out with the cascaded classifier based on Haar-like features and AdaBoost algorithm, and the classifier’s parameters are optimized for this certain application. After that, the second order histograms of oriented gradients (HOG) are calculated from the scaled detection results. At last, head pose is estimated with random regression forests and the drivers’ fatigue state can be judged with the difference between the estimation result and the referred head pose.The experimental results in the situated driving environment show that the proposed fatigue driving detection system works out to be good and can monitor the driver’s eye state and head pose in real-time.Meanwhile, for efficiency problems found in the system design process, the fast algorithm of 2-D convolvers based on field-programmable gate arrays (FPGA) is also discussed in this paper. We propose the improved recurrently decomposable (RD) 2-D convolution architecture based on full buffering (FB), single-window partial buffering (SWPB), and multi-window partial buffering (MWPB) schemes. FPGA implementation results show that the RD based structures can greatly save the hardware resources consumed by convolution calculation module, and improve the whole convolver’s area efficiency. In addition, this paper also proposes a new area efficiency metric to guide the selection of convolution architectures for a certain application scenario.