An incremental approach towards automatic model acquisition for human gesture recognition

Abstract

The recognition of natural gestures typically involves: the collection of training examples; the generation of models; and the determination of a model that is most likely to have generated an observation sequence. The first step however, the collection of training examples, typically involves manual segmentation and hand labelling of image sequences. This is a time consuming and labour intensive process and is only feasible for a limited set of gestures. To overcome this problem we suggest that gestures can be viewed as a repetitive sequence of atomic movements, similar to phonemes in speech. We present an approach: to automatically segment an arbitrary observation sequence of a natural gesture, using only contextual information derived from the observation sequence itself; and to incrementally extract a set of atomic movements for the automatic model acquisition of natural gestures. Atomic components are modelled as semi-continuous hidden Markov models and the search for repetitive sequences is done using a discrete version of CONDENSATION that is no longer based on factored sampling.