To prevent accidents associated with a driver’s sleepiness, it is imperative to estimate the driver’s arousal level and manage the arousal level appropriately. Recently, methods to maintain arousal levels have been diversified; particularly, a method through voice interaction with an artificial intelligence agent has been proposed. In such a method, it is thought that it is effective to set the frequency and content of interactions according to arousal level, but the multistage arousal estimation using face images, which are relatively easy to measure, remains challenging. In this study, aiming at application to real-time feedback, using face images, we consider a multistage arousal level estimation that is independent of the individual. Through experiments, we obtained data on the subjective assessment of sleepiness and face images and constructed models via deep learning. From the model evaluation, learning considering timeseries characteristics is effective in multistage arousal level estimation. With 67.1% and 68.6% data, respectively, for long short-term memory and gated recurrent unit methods, the deviation between the actual and estimated levels can be suppressed to ≤1, suggesting that a rough multistage estimation model can be realized with these methods.

View full abstract

To measure the types of error proneness simply and quickly, we investigate the relationship between the line drawing task, which is the task of drawing a straight line for a given point cloud shown on a tablet device, and the error proneness. A measure for a drawn straight line is proposed, and the relationship between the measure and the scores indicating the degree of error proneness is examined. We also clarify what kind of point clouds are meaningful. For 10 kinds of point clouds, 34 participants performed the line drawing task for three times with one day or more opened. For two kinds of point clouds, we found that there is a relationship between the distance (𝑑b) to the barycenter of the point cloud from the drawn line and the action slip (AS, error proneness as a forgetfulness or absent-minded) score (the correlation coefficient: 0.57, 0.39, p<0.05). By performing a multiple regression analysis with the objective variable 𝑑b (the coefficient of determination: 0.25, p<0.01), we created a perceptual and cognitive model of the line drawing task. This study indicates that the line drawing task is related to AS scores, and may also related to cognitive narrowing (the tendency to narrow the range of information that one can perceive and process under a high workload).

View full abstract