Highly automated driving requires precise models of traffic participants. Many state of the art models are currently based on machine learning techniques. Among others, the required amount of labeled data is one major challenge. An autonomous learning process addressing this problem is proposed. The initial models are iteratively refined in three steps: (1) detection and context identification, (2) novelty detection and active learning and (3) online model adaption.