Gesture Spotting in Continuous Whole Body Action Sequences Using Discrete Hidden Markov Models

Gestures are expressive and meaningful body motions used in daily life as a means of communication so many researchers have aimed to provide natural ways for human-computer interaction through automatic gesture recognition. However, most of researches on recognition of actions focused mainly on sign gesture. It is difficult to directly extend to recognize whole body gesture. Moreover, previous approaches used manually segmented image sequences. This paper focuses on recognition and segmentation of whole body gestures, such as walking, running, and sitting. We introduce the gesture spotting algorithm that calculates the likelihood threshold of an input pattern and provides a confirmation mechanism for the provisionally matched gesture pattern. In the proposed gesture spotting algorithm, the likelihood of non-gesture Hidden Markov Models(HMM) can be used as an adaptive threshold for selecting proper gestures. The proposed method has been tested with a 3D motion capture data, which are generated with gesture eigen vector and Gaussian random variables for adequate variation. It achieves an average recognition rate of 98.3% with six consecutive gestures which contains non-gestures.