Triplet Loss for Effective Deployment of Deep Learning Based Driver Identification Models.

Recent studies have shown that both the quantity and quality of the driving data are crucial for the success of deep learning based driver identification solutions. This is not only true for the training of deep neural networks, but also for their deployment. In the deployment phase the models are applied to identify new drivers that are different from the drivers in the training phase. New data is required to re-train the models and adjust them to fit the new drivers. While the focus of recent studies has been mostly on the training phase, this study focuses on the deployment phase. Inspired by face and fingerprint recognition solutions, in this paper we propose a solution that does not require model re-training in the deployment phase. Instead, only a small amount of example driving data from each driver is required. More specifically, we propose an approach that utilises the triplet loss function to learn a mapping from vehicular sensor data to an embedding space, where the distance between two embeddings represents the drivers' similarity. The resulting network can be directly used for identifying new drivers without re-training.