Incremental modeling and its application for driver fatigue estimation

Abstract

This article focuses on an Incremental Learning (IL)-based approach for examining a driver fatigue dataset and compares its performance with classical Machine Learning (ML) and deep learning techniques with respect to generalization capability. Driver fatigue is of significant concern in traditional driving conditions, contributing to many serious and fatal accidents worldwide. While self-driving cars may eventually alleviate this issue, until they are fully developed, “self-driving carsickness” introduces new problems. Various classification methods have been proposed in recent years to distinguish between drowsy and alert states, including both binary and multi-class classifications. However, the generalization capabilities of these methods are underexplored. IL, often seen as a data-hungry approach, optimizes both model parameters and structure. The primary goal is early recognition of fatigue, enabling timely intervention, and ensuring explainability for safety-critical systems. To address these issues, we propose a novel regression approach. We propose a deterministic regression model, guided by binary labels and using an Incremental Modeling (IM) framework, to address the challenges in fatigue recognition. Our model demonstrates superior performance on a challenging dataset, with improvements in accuracy ($+2\dots +7\text{\%}$), parameter count ($-99\dots -100\text{\%}$), speed ($+0.5\dots -0.1$ s), and latency ($-70\mu s$). It also offers flexibility with optional personalization, illustrating the strength and adaptability of IM for fatigue detection in various conditions.