การวิเคราะห์ท่าทางของมนุษย์โดยใช้ข้อมูลสีและความลึกจากหลายมุมมอง

Abstract

This research applied image processing and computer vision techniques to contribute three essential perspectives; modeling profile-based human action recognition, people tracking and re-identification, interesting event detection. The interesting event detection consists of falling, hand-waving for asking help, and jumping. This approach is based on overlapped area of interest using multi-view RGB-D. All functions are purposed for health care promotions and surveillance system applications. The action recognition consists of high level and feature level fusion. In high level, the results of action in single- view system are fused for making decision using empirical analysis to weight the most realizable result to be a result of multi-view system. The maximum improvement for some action is up to 97.70% and overall result increases to 16.66% (percentage point) when compared with single-view action recognition. Another in feature level fusion, we proposed a method to build Layer Feature Model that allows to fuse features of depth from multi- view. The experimental results of fusion model are 86.40% in NW-UCLA dataset, 93.00% in i3DPost dataset, and 99.31% in PSU dataset. In addition, we introduce people tracking and people re-identification by using analysis of position and color descriptor. The position and color descriptor are clearly attributes for both tracking in a single-view and matching those views. Moreover, the color descriptor is also used for supporting cursory people re-identification. The precisions of people re-identification are 92.87% in single person entering and 85.50% when 2-person simultaneously entering. In the interesting event detection, falling detection resulted in the average precision of 90.65% that derived from precision of lying. The average of hand-waving detection precision is 92.96%. The jumping detection has sensitivity rate 92.96% and specificity rate 99.31%.