The Study Of Retrieval And Segmentation Algorithm For Multi-Dimensional Time-series Based On Human Motion Capture Data

As the digitalization technology and database technology advanced recent years, data mining techniques that focus on multi-dimensional time series attracts more and more researchers. However, the complexity of data mining algorithms for multi-dimensional time series dramatically increases because of the dimension curse. Therefore it is impossible to directly apply the classical algorithm to multi-dimensional time series. To partially attack these problems, a similarity retrieval algorithm and statistic learning based segmentation algorithm is proposed and applied to human motion capture database. Besides, dimension reduction, multi-dimensional index structure and learning algorithm are extensively studied.For the purpose of dimension reduction, a short survey of classical algorithm for dimension reduction is presented. Based on the characteristic of human motion, an algorithm for key joints extraction based on entropy is also proposed. The algorithm effectively reduces the retrieval cost without losing precisionFor the purpose of indexing multi-dimensional time series, a precise similarity measurement DTW and an index data structure R-Tree are studied. An extension algorithm is also presented based on Keogh distance to accommodate multi-dimensional situation. In the end a prototype is developed to implement similarity retrieval for human motion capture data clips.For the purpose of segmentation, a statistics machine learning model conditional random field is studied. Based on the human motion analysis, the non-bias characteristic of CRF is made use of to segment motion clip. The segmentation algorithm can be used as preprocess algorithm for similarity retrieval.