Studies on interaction detection generally assume that interaction information is the same over time. However, interaction may change, especially when switching time series dynamics is considered. To detect these, this thesis proposes a hierarchical segmentation-based approach to switching interaction detection using Bayesian non-parametrics. First presented is the proposed segmentation method, which combines the segmentation by beta process - autoregressive hidden Markov model (BP-AR-HMM) and the double articulation by nested Pitman-Yor language model (NPYLM). The proposed approach to interaction detection is presented next, where the method of surrogate data is applied to the autoregressive models obtained from the first part. Results of experiments using synthetic toy and real motion datasets indicate that the method segmented the time series in both low and high levels with high accuracy, and inferred interaction information from the time series sequences with good specificity. These results then suggest that switching interaction can be detected using the proposed hierarchical segmentation approach.