Enhancing In-Cabin Monitoring Performance Using Unity Eyes Generated Data

Research in autonomous driving has been gaining increasingly more attention, since the introduction of electric vehicles. Autonomous vehicles are required to conform to the levels set by the Society of Automotive Engineers (SAE), which made driver monitoring a legal requirement. In the currently allowed level, the car must be in the park gear to access its infotainment system, but to advance to the next level of autonomy, where the gear state is not monitored, the driver’s state must be monitored. Facial scanning is the first step in determining the driver's current state, especially for drowsiness. However, it is difficult to acquire the data required as using such data would be an invasion of privacy. This paper aims to overcome the challenge of acquiring training data by generating the data with Unity Eyes, which would enable enhanced performance. A model with a ResNet50 backbone achieved 66.0% accuracy when trained with a limited real dataset, whereas the same model trained with generated Unity Eyes data achieved 85.3% accuracy. Our ablation study showed that the use of Unity Eyes data is more effective than known pre-trained models as well. This study demonstrates the effectiveness of generated data in situations where large-scale data is impossible and suggests potential future applications in a variety of studies.