Exploration of the Application of Driver's Visual Attention Model in Autonomous Driving System Development

Driver's visual attention modeling has expanded beyond traditional “purely image data-driven” visual tasks, now integrating both endogenous “volition-driven” and exogenous “stimulus-driven” mechanisms. However, the application of the model has primarily focused on inferring potential attention distributions from intermediate visual features through task-based studies, which helps the overall model extract key features more quickly and accurately. In contrast, the application of “pre-trained driver's visual attention models as submodules in autonomous driving system development” remains underexplored. This paper explores the integration of volition-driven driver's visual attention models in autonomous driving system development, examining their impact on saving storage resources, improving computational real-time performance, and enhancing system convergence. The results show that post-application architectures are more effective than pre-application ones, with the optimal attention discretization threshold coefficient ranging from 0.4 to 0.5.