Driver Distraction Identification with an Ensemble of Convolutional Neural Networks

Figure 8

System accuracy evaluated at different values of history intervals . At , current camera input image is solely used for distraction classification; no context is used. For , we average predictions of frames to obtain distraction in the current frame. Up to a certain threshold, the more context incorporated in the prediction, the more accurate and robust the system. After that threshold, contextual information in the past is no longer relevant. Best accuracy is achieved at seconds, which is the average of the fitted Gaussian curve.