A conversive hidden non-Markovian model based structure for discriminating spatio-temporal movement trajectories

Abstract

We present a new modelling approach for spatio-temporal movement trajectories that is based on the stochastic model class called conversive hidden non-Markovian models (CHnMMs). The approach is improving previous work by facilitating the automatic creation of these models from examples. Created models can be utilised for trajectory classification and verification tasks which is explained with a possible procedure. The use of CHnMMs allows for an explicit modelling of temporal dynamics which allows the discrimination of trajectories by shape and execution speed. The presented approach is evaluated with touch gesture recognition experiments and compared to the $1 unistroke recogniser and the dynamic time warping method. The results show better recognition rates for movements that are only discriminable by their temporal behaviour and very good recognition rates, especially regarding the discrimination of similar shaped trajectories that only differ in their temporal dynamics.