Paper 1: Real-Time Driver’s Focus of Attention Extraction and Prediction using Deep Learning
Abstract: Driving is one of the most common activities in our modern lives. Every day, millions drive to and from their schools or workplaces. Even though this activity seems simple and everyone knows how to drive on roads, it actually requires drivers’ complete attention to keep their eyes on the road and surrounding cars for safe driving. However, most of the research focused on either keeping improving the configurations of active safety systems with high-cost components like Lidar, night vision cameras, and radar sensor array, or finding the optimal way of fusing and interpreting sensor information without consid-ering the impact of drivers’ continuous attention and focus. We notice that effective safety technologies and systems are greatly affected by drivers’ attention and focus. In this paper, we design, implement and evaluate DFaep, a deep learning network for automatically examining, estimating, and predicting driver’s focus of attention in a real-time manner with dual low-cost dash cameras for driver-centric and car-centric views. Based on the raw stream data captured by the dash cameras during driving, we first detect the driver’s face and eye and generate augmented face images to extract facial features and enable real-time head movement tracking. We then parse the driver’s attention behaviors and gaze focus together with the road scene data captured by one front-facing dash camera faced towards the roads. Facial features, augmented face images, and gaze focus data are then inputted to our deep learning network for modeling drivers’ driving and attention behaviors. Experiments are then conducted on the large dataset, DR(eye)VE, and our own dataset under realistic driving conditions. The findings of this study indicated that the distribution of driver’s attention and focus is highly skewed. Results show that DFaep can quickly detect and predict the driver’s attention and focus, and the average accuracy of prediction is 99.38%. This will provide a basis and feasible solution with a computational learnt model for capturing and understanding driver’s attention and focus to help avoid fatal collisions and eliminate the probability of potential unsafe driving behavior in the future.
Keywords: Driving; attention; interesting zones; deep neural network; deep learning; models