AN EFFICIENT HIERARCHICAL STRUCTURE FOR RECOGNITION OF THE REACHING MOVEMENTS IN 3D SPACE
Abstract
In this research, classification of the 18 reaching movements was done in 3D space using the elbow flexion/extension angle and the upper arm acceleration. For this, a hierarchical structure was proposed where the approximate region of the movements was determined at high level and then, motion types were recognized exactly at low level. To evaluate this hierarchical structure, the flat classifiers were implemented with the same features. The hierarchical classifier improves the formation of the decision boundaries better than the flat classifiers due to its proper structure. It improves the discrimination of the within-group members by applying an algorithm to increase the ratio of the between-class covariance matrix trace to the within-class covariance matrix trace. It also benefits from low dimensional feature space. Recognition accuracy was higher than the flat classifiers. In other evaluation metrics (Recall, Precision and F-score), this structure indicated better performance, especially in those classes with the least recognition percentage.
