Towards Imbalanced Multiclass Driver Distraction Identification

Abstract

Driver distraction is one of the leading causes of driving-related accidents worldwide. The ability to detect driver distraction preemptively is crucial to reducing the number of such accidents. This paper utilizes a novel multimodal dataset of thermal, visual, near-infrared, and physiological signals recorded from 45 subjects in order to identify distraction. We explore imbalanced distraction identification as a four-class problem across different types of distractions in order to resemble real-life scenarios, where the occurrence of different distractors varies. Our study analyzes the effectiveness of using early fusion across different modalities, a variety of window sizes, and data balancing schemas using synthetic instances. Moreover, we explore the effects of introducing subject-specific knowledge when training such identification models.