Loading...

Human Action Categorization using Spatiotemporal Features

Ghodrati, Amir | 2009

471 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 40201 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Kasaei, Shohre
  7. Abstract:
  8. Recognizing human actions is an important and challenging topic in computer vision, which has important applications such as video surveillance and Indexing. From a computational perspective, actions can be defined as three-dimensional patterns, in space and in time which can be modeled using several representations. Action representations differ in visual information used in spatial dimensions (e.g., shape or appearance) and the representation of dynamics in time. The goal of this thesis is to develop new techniques and improve current results in action categorization. As such, using a general structure, three methods are proposed. In this structure, local spatio-temporal features are extracted from videos to represent actions which are then categorized by a classifier (such as k-nearest neighbor and support vector machine). It was designed by assumption of fixed camera, stationary background and one action per video but it has ability of recognition in different appearance and illumination conditions, variation in view point up to 20 degree and zooming and shrinking of action. In the first method, in order to obtain better discriminating features, three histogram weighting methods are proposed. In the second method, pyramid spatial matching which achieves surprising performance in object recognition is developed and spatio-temporal pyramid matching is proposed for dealing with 3-D actions. In the third method, a novel action representation is proposed to implicitly cooperate geometric structure of motion. These methods improve the performance up to 3% related to current methods which is significant improvement with regard to 90% recognition accuracy of current methods. All of these methods are evaluated on KTH and Weizmann datasets to show the efficiency of the proposed methods.
  9. Keywords:
  10. Term Weighting ; Human Action Categorization ; Local Spatiotemporal Feature ; Words Bag Model ; Spatiotemporal Pyramid Matching

 Digital Object List

 Bookmark

No TOC