Bidirectional recognition between motion context on long-term observation and human motion

Abstract

This paper describes a motion recognition method to reduce recognition error, which has two-layered structure; motion recognition is affected by context estimation in the first layer, and context estimation is affected by motion recognition in the second layer. We introduce an algorithm to integrate the motion recognition by conventional HMM and motion label production by the topic model in the first layer. We also introduce particle filter to estimate and update the context based on the result of motion recognition in the second layer. A set of particles present a probabilistic distribution of motion topics, and motion recognition and particle update procedures are performed on each particle. In an evaluation experiment, we used a sequential motion which is a sequential connection of 33 motion primitives as a long-term observation target. The results showed that the proposed method reduced recognition errors and tracked motion context by topic probability compared with conventional methods.