High-Resolution Neural Network for Driver Visual Attention Prediction.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Automotive Safety

Background:

• Safe driving relies heavily on visual information processing by the human visual system.
• Understanding driver visual attention is critical for developing advanced driver-assistance systems (ADAS) and autonomous vehicles (AV).
• Predicting driver attention from images is essential for intelligent vehicle systems to ensure safe navigation.

Purpose of the Study:

• To investigate human driver visual behavior using computer vision techniques.
• To accurately estimate driver attention locations within images.
• To improve the performance of driver attention prediction models for ADAS and AV applications.

Main Methods:

• Employed a deep convolutional neural network (CNN) framework to learn and extract multi-resolution feature representations.
• Fused high-resolution and low-resolution features to enhance prediction accuracy and localization.
• Trained the network using diverse image regions to mitigate center bias prevalent in standard datasets.

Main Results:

• Feature representations at high resolution, when fused with low-resolution features, significantly improve visual attention prediction accuracy.
• The proposed method demonstrates enhanced localization performance for driver attention.
• Training with diverse image regions effectively overcomes the center-bias issue in attention prediction.

Conclusions:

• The developed computer vision framework effectively predicts driver attention locations.
• High-resolution feature integration and diverse training data are key to improving attention prediction accuracy.
• This research contributes to the advancement of intelligent vehicle systems by providing robust driver monitoring capabilities.