Abstract: The design and testing of autonomous vehicles crucially depend on an understanding of human driving characteristics due to the coexistence of automated and human-driven vehicles. Previous studies had relied on personalized background vehicle models, which required either calibrating microscopic traffic model parameters or training supervised models. This paper proposes a method that considers human driving behavior as a collection of some empirical driving patterns. A pattern dictionary generation algorithm is presented that distinguishes real-world driving data into pattern representations in continuous space. Driving patterns extraction based on transformer (DPFormer), a self-supervised deep learning-based method, is proposed to embed high-dimensional driving data into a compact form without manually labeled driving maneuvers. The model is optimized using the pretext task which utilizes attention mechanisms to embed the current driving segment as the posterior embedding of driving context. The embedding representation facilitates the analysis of diverse driving behaviors. Furthermore, a driving characteristics analysis method is proposed based on the Wasserstein distance between driving patterns' transition probability matrices to quantify the differences between different drivers. Finally, the effectiveness of this proposed model is validated using a large amount of real-world driving data collected from 171 human drivers.