Loading...

Multi-Camera Action Recognition with Manifold Learning

Rezaee Taghiabadi, Mohammad Mehdi | 2017

1142 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 49281 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Karbalaee Aghajan, Hamid
  7. Abstract:
  8. Human action recognition is one of the most attended topics in computer vision and robotics.One of the flavors of this problem relates to the situation in which the task of action recognition is carried out by data from several cameras. Different approaches have been proposed for combining information. Various reduction methods have been introduced to decrease the processing load. All of the methods in this particular field of study can be divided into two linear and non-linear methods. In the linear methods, we don’t pay attention to the non-linear structure of the data, and these kind of approaches are not reliable. Furthermore, combining different actions data is done before the dimension reduction.Manifold learning approaches are based on the geodesic distances, by the help of which,retaining the relationship between different action points is much more meaningful and reliable. In this thesis, our aim is to pay attention to the non-linear structures and the different views’ overlap. To reach our goal, first we describe our data with spatio-temporal feature descriptors, and atter that ST-ISOMAP (a non-linear method) helps us visualize different actions in 3D space. In the next steps, using combination of different actions data would rise to action classification. To test our algorithm, we find the appropriate mapping from highdimensional space to the low-dimensional space and then decide whehter the class has been detected correctly. Finally, we evaluate our approach and propose new solutions to improve the method
  9. Keywords:
  10. Human Action Recognition ; Computer Vision ; Nonlinear Method ; High Dimention Data ; Multi Camera ; Manifold-Based Learning

 Digital Object List

 Bookmark

...see more