3D human motion data are represented as a sequence of poses, where each pose keeps the 3D coordinates of important skeleton joints in a specific time moment. Such spatio-temporal data are often captured without any information about their semantic partitioning. The objective of this thesis is to evaluate different techniques for determining similarity between two unsegmented sequences. In particular, the unsegmented sequences will be transformed into a text-like representation to which well-known text retrieval models can be applied. The techniques will be experimentally evaluated on a real-life dataset from both the effectiveness and efficiency points of view.