Abstract
High accuracy human action recognition is becoming an essential part in developing intelligent systems. Due to the enormous number of motion archives, a structured motion database is developed with directional organization independent of the viewpoint. Nevertheless, it had the drawbacks of recognizing the observed motion which is similar to several motions within the motion database. This is called boundary problem that occurs because of the mis-selection of neighboring motions in the retrieval process. The performance may highly drop when this problem exists. In this paper, we propose a novel resolution approach to this problem to improve the system's performance. We propose two sets of B-Tree database structure: the original B-tree and a left- or right-shifted B-tree. The motions are represented as Motion History Image (MHI) and Exclusive-OR (XOR) representations, and the directional eigenspaces are used as motion extracted feature vector space. The improvement in the recognition rate adopting the proposed approach signifies high accuracy human motion recognition.
