Recognition of Manipulation Sequences by Human Hand Based on Support Vector Machine

This paper describes a method of recognizing a manual task executed by a human hand by using the support vector machine (SVM). We define several task states which are segmented from the continuous motion of human fingers in the context of an object manipulation. Based on margins of SVMs, the method constructs a binary decision tree which most effectively classifies and symbolizes the task state from joint angle trajectories of human fingers as input. The binary decision tree constructed by our method has been evaluated through experiments of recognizing the task states during a valve manipulation.