We present in this paper an augmented reality guidance system for liver thermal ablation in interventional radiology. To show the relevance of our methodology, the system is incrementally evaluated on an abdominal phantom and then on patients in the operating room. The system registers in a common coordinate system a preoperative image of the patient and the position of the needle that the practitioner manipulates. The breathing motion uncertainty is taken into account with a respiratory gating technique: the preoperative image and the guidance step are synchronized on expiratory phases. In order to fulfil the real-time constraints, we have developed and validated algorithms that automatically process and extract feature points. Since the guidance interface is also a major component of the system effectiveness, we validate the overall targeting accuracy on an abdominal phantom. This experiment showed that a practitioner can reach a predefined target with an accuracy of 2 mm with an insertion time below one minute. Finally, we propose a passive evaluation protocol of the overall system in the operating room during five interventions on patients. These experiments show that the system can provide a guidance information during expiratory phases with an error below 5 mm.