A universal HMM-based approach to image sequence classification

Abstract

A universal approach to the classification of video image sequences by hidden Markov models (HMMs) is presented. The extraction of low level features allows the HMM to build an internal image representation using standard training algorithms. As a result, the states of the HMMs contain probability density functions, so called image density functions, which reflect the structure of the underlying images preserving their geometry. The successful application of the approach to both the recognition of dynamic head and hand gestures demonstrates the universal validity and sensitivity of our method. Even sequences containing only small detail changes are reliably recognized.