An Efficient Model for Driving Focus of Attention Prediction using Deep Learning

In this work, we aim to build a real-time prediction model that can predict the driver's Focus of Attention (FoA) based on the image and motion information (RGB and Optical Flow) of the driving environment. A Y-shape-structured Fully Convolutional Neural Network (Y-FCNN) is proposed to learn and predict the FoA. This network first applies convolution and max-pool layers on RGB and Optical Flow respectively to get the low-level feature maps, and then merges the two encoded low-level feature maps together. After that, the Dilated Convolution, which can get a larger receptive field while still keeping high-resolution information, is applied to make the final prediction. The model is trained and tested using the Dr(eye)ve dataset. Experiment results show that our model can predict the FoA with high accuracy at a speed of about 122 frames per second, which outperforms some previous works for FoA prediction.