Driver Motion Detection Using Online Sequential Learning

Driver distraction and fatigue are the major factors causing accidents. Constructing head-nodding and shaking models can better detect abnormal driving behavior for early warning in order to enhance road safety. Due to individual difference, traditionally an entirely new model is trained for each individual driver, which requires a large amount of data for each new driver who uses the detection system. In this paper, we employed the online sequential extreme learning machine (OS-ELM), which updates the model parameters for each new driver based on a general model created beforehand, using only small amounts of data from each new driver. Data collected from Google Glass during head-nodding and shaking were used to train drivers' head gesture model, and a small amount of data from a new driver were used to update an individual-specific model. The detection performance model was then tested. The experimental results show that OS-ELM can achieve an average classification accuracy of 92.45%, increased by 4.74% compared to traditional extreme learning machine (ELM). The accuracy of OS-ELM is gradually improving with the increasing number of new data. An online method is proven efficient in dealing with individual differences, and provides useful approaches for real-time learning and prediction.