One-Sequence Learning of Human Actions

摘要

In this paper we address the problem of human action recognition from a single training sequence per class using a modified version of the Hidden Markov Model. Inspired by codebook approaches in object and scene categorization, we first construct a codebook of possible discrete observations by applying a clustering algorithm to all samples from all classes. The number of clusters defines the size of the codebook. Given a new observation, we assign to it a probability to belong to every cluster, i.e., to correspond to a discrete value of the codebook. In this sense, we change the 'winner takes all' rule in the discrete-observation HMM for a distributed probability of membership. It implies the modification of the Baum-Welch algorithm for training discrete HMM to be able to deal with fuzzy observations. We compare our approach with other models such as, dynamic time warping (DTW), continuous-observation HMM, Conditional Random Fields (CRF) and Hidden Conditional Random Fields (HCRF) for human action recognition.