Advanced Ensemble Deep Random Vector Functional Link for Eye-Tracking-based Situation Awareness Recognition

Abstract

Situation awareness (SA) plays a significant role in takeover transitions from autonomous to manual driving. Previous researchers have shown that eye movement signals can be used for SA recognition. Moreover, ensemble deep random vector functional link (edRVFL) has demonstrated its superiority in different applications. Therefore, this work proposes an advanced edRVFL (AedRVFL) to perform eye-tracking (ET)-based SA recognition, including improvements in two aspects. Specifically, pruning in the direct links is implemented to improve the efficiency of linear features. Then, a weighting method based upon the regression errors of samples is developed. The experiment was conducted on a public conditionally automated driving dataset. Results showed that the proposed AedRVFL out-performed the baseline methods, demonstrating the effectiveness of using AedRVFL for ET-based SA recognition. Ablation studies were conducted to validate the improvements in the edRVFL.