An abnormal behavior of a moving vehicule or a moving person is characterized by an unusual or not expected trajectory. The definition of exptected trajectories refers to supervised learning where an human operator should define expected behaviors. Conversely, definition of usual trajectories, requires to learn automatically the dynamic of a scene in order to extract its typical trajectories. We propose, in this paper, a method able to identify abnormal behaviors based on a new unsupervised learning algorithm. The original contributions of the paper lies in the following aspects: first, the evaluation of similarities between trajectories is based on string kernels. Such kernels allow us to define a kernel-based clustering algorithm in order to obtain groups of similar trajectories. Finally, identification of abnormal trajectories is performed according to the typical trajectories characterized during the clustering step. The method has been evaluated on a real dataset and comparisons with other state-of-the-arts methods confirm its efficiency.