A Driving Position-Sensitive Neural Network for Driver Fixation Prediction

Driver fixation prediction has recently received increasing attention in the field of autonomous and assisted driving, and is becoming an essential research problem in advanced driver assistance systems (ADAS). Since many information sources attract driver's attention in complex traffic scenarios, the model needs to predict multiple salient regions or objects that the driver must pay close attention to. However, there is still a lack of driver fixation prediction models that can precisely predict drivers' multiple fixation regions. In this work, we propose a driving position-sensitive neural network (DPSNN) with coordinate attention module (CAM) for more accurately predicting drivers' multiple fixation areas or objects. In DPSNN, the CAM filters irrelevant information and redundant information from multi-scale feature maps, and generates driving position-sensitive feature maps containing precise positional information. Combining the driving position-sensitive feature maps with the upsampled outputs in expansive path helps the network to identify and locate objects of interest completely and precisely. The experimental results indicate that the proposed DPSNN outperforms the state-of-the-art fixation prediction models and can predict driver's multiple fixation locations more accurately.