Unsupervised Driver Workload Learning through Domain Adaptation from Temporal Signals

Driver workload monitoring is an important component of the intelligent driver assistant systems today. Accurately monitoring a driver's real-time workload is not an easy task. It requires good quality and well-annotated data collected from drivers to train machine learning models. Usually, a model trained for a target driver cannot be applied interchangeably to another target unless data from them have the similar distributions, which is however not commonly seen. This is known as the personal discrepancy problem between individuals. To deal with this problem, one method is to tune an existing model using annotated data collected from the target driver.However, obtaining annotated data from target drivers is a time consuming, labor-costly and sometimes impractical procedure. To cope with this difficulty, we developed an Adversarial Discriminative Neural Network for Multi-Temporal Signals (MTS-ADNN) architecture. With this method, a model can learn transferable features from well-annotated data in source domain and adapts to non- annotated data in target domain, even if data in the two domains have shifted distributions. Different from many existing adversarial learning architectures that aligns only between-domain distributions, the proposed MTS-ADNN can also aligns in-domain classes to ensure in-domain class-conditional distributions are aligned jointly with between-domain distributions. To enhance the performance, we added an entropy-regularizer to target domain sample predictions, and an entropy-aware weight to aggregate the loss of the discriminator. We evaluated the method using a set of workload estimation data collected from real-world diving. We compared its performance with three state-of-the-art unsupervised domain adaptation methods. The results show that the proposed MTS-ADNN outperforms its counterparts.