A Cognition-Driven Network for Driver Gaze Prediction in Intelligent Vehicles

Driver gaze prediction is a key research area in human-machine co-driving. It enables co-driving systems decide whether to take over control by predicting driver gaze and identifying possible driver distraction. However, most existing methods only focus on bottom-up attention, which solely relies on visual input from the driving scene, neglecting the top-down attention related to the driver cognition of the scene. In this paper, a novel dual-pipeline cognition-driven network named CoGNet is proposed to improve the accuracy of the gaze prediction. Driver scene cognition is derived from a top-down pipeline that takes as input semantic segmentation and internal vehicle information, such as the ego vehicle's speed and angular velocity. A cognitive-visual attention module is then introduced to fuse driver scene cognition with visual features extracted from the bottom-up pipeline. In addition, a comprehensive eye-tracking dataset is constructed with frame-level vehicle motion data such as speed and angular velocity, addressing the absence of such data in existing open-source datasets. Experiments on this dataset show that, CoGNet outperforms state-of-the-art methods across most metrics, resulting in more accurate driver gaze prediction.