HammerDrive: A Task-Aware Driving Visual Attention Model

We introduce HammerDrive, a novel architecture for task-aware visual attention prediction in driving. The proposed is learnable from data and can reliably infer the current focus of driver real-time, while only requiring limited easy-to-access telemetry vehicle. build on two core concepts: 1) driving be modeled as collection sub-tasks (maneuvers), 2) each sub-task affects way allocates resources, i.e., their eye gaze fixation. HammerDrive comprises networks: hierarchical monitoring network forward-inverse model pairs recognition an ensemble task-dependent convolutional neural modules modeling. assess ability to we collected 20 experienced drivers virtual reality-based simulator experiment. evaluate accuracy our show that it effective light-weight reliable real-time tracking maneuvers with above 90% accuracy. Our results outperforms comparable state-of-the-art deep learning numerous metrics ~13% improvement both Kullback-Leibler divergence similarity, demonstrate task-awareness beneficial prediction.