This paper presents a safe imitation learning approach for autonomous vehicle\ndriving, with attention on real-life human driving data and experimental\nvalidation. In order to increase occupant's acceptance and gain drivers' trust,\nthe autonomous driving function needs to provide a both safe and comfortable\nbehavior such as risk-free and naturalistic driving. Our goal is to obtain such\nbehavior via imitation learning of a planning policy from human driving data.\nIn particular, we propose to incorporate barrier functions and smooth\nspline-based motion parametrization in the training loss function. The\nadvantage is twofold: improving safety of the learning algorithm, while\nreducing the amount of needed training data. Moreover, the behavior is learned\nfrom highway driving data, which is collected consistently by a human driver\nand then processed towards a specific driving scenario. For development\nvalidation, a digital twin of the real test vehicle, sensors, and traffic\nscenarios are reconstructed toward high-fidelity and physics-based modeling\ntechnologies. These models are imported to simulation tools and co-simulated\nwith the proposed algorithm for validation and further testing. Finally, we\npresent experimental results and analyses, and compare with the conventional\nimitation learning technique (behavioral cloning) to justify the proposed\ndevelopment.\n