Towards Imbalanced Multiclass Driver Distraction Identification

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