Motion retrieval based on Dynamic Bayesian Network and Canonical Time Warping

Abstract

A novel graph-based motion retrieval method is proposed. The method includes the 2 main stages: (1) in stage of learning, firstly, for each of motion in database, using Aligned Cluster Analysis (ACA) to get key frames, extracting body gesture and joint state features as observation signal of graph model, based on graph model theory and statistical learning of key frame, a new Dynamic Bayesian Network (DBN) frame is constructed, which is combination of the Switching Kalman Filtering Model (S-KFM) and the Hidden Markov Model (HMM). The next, a graph-based motion descriptor is built based on DBN inference, and graph-based motion feature database is constructed. (2) In stage of motion retrieval, according to above steps, the graph-based query motion feature can also be obtained, we can recognize category of motion through Canonical Time Warping (CTW) matching results. The experiments results show proposed method is effectiveness.