Robust Gaze-Based Intention Prediction for Real-World Scenarios

Existing intention prediction scenarios in daily life primarily focus on 2-D screens, while the process of intention expression in 3-D scenarios remains largely unexplored. We first analyze eye-tracking data from both 2-D and 3-D scenarios to reveal differences in cognitive load. To address the increased error and redundant gaze points in 3-D scenarios, we propose a gaze region model combined with a clustering method based on density and ordering principles, providing a robust representation of visual attention. Additionally, we integrate this visual attention representation with advanced classifiers for intention prediction. The results indicate that when intentions are expressed in a 3-D scenario, subjects’ cognitive load is reduced, facilitating their understanding and expression of intentions, ultimately improving the accuracy of intention prediction. Simultaneously, an evaluation of existing visual attention representation models related to intention prediction is conducted. Our proposed 3-D visual attention model, as part of the intention prediction framework, improves accuracy to 94.50%. To validate the theory and model, we introduce the ADLIP Gaze dataset, which consists of data from 102 individuals. These findings are expected to provide theoretical explanations and methods for intention prediction and efficient human–robot interaction in 3-D scenarios.