Loading...

Auto-selection of space-time Interest Points for Action Recognition
Application in Fisherposes Method

Ghojogh, Benyamin | 2017

952 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 49714 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Mohammadzadeh, Narges Hoda
  7. Abstract:
  8. In this project, a novel action recognition method, named Fisherposes, is proposed, which is improved by several space-time (spatio-temporal) methods afterwards. The proposed method utilizes skeleton data obtained from Kinect sensor. First, pre-processing is performed in which the scales of bodies are canceled and the skeletons become aligned in order to make the method robust to location, orientation, and scale of people. In Fisherposes method, every action is defined as a sequence of body poses. Using the training samples for the poses, a Fisher subspace is created which we name it Fisherposes. Moreover, a novel distance measuring function, named regularized Mahalanobis distance, is proposed in order to have a behavior between Euclidean and Mahalanobis distances and use their both benefits. Thereafter, a windowing is performed onto the frames to omit the unqualified frames. In addition, using key-frames is proposed as a temporal-points method for sampling frames because of two goals, i.e., extracting key-poses and extracting input frames for hidden Markov models. Afterwards, using the key-poses is proposed which automates the extraction of body poses and their sampling for creating Fisherposes subspace. Furthermore, using key-joints is introduced as a spatial-points method to select the more informative joints for use. Selecting key-joints are performed and analyzed both manually and automatically. Using key-joints, however, introduces the need to several Fisherposes subspaces which is handled by proposing the novel Fisher forest method. In this project, hidden Markov model is used for modeling an action as a sequence of body poses. In addition, we use histogram of trajectories in order not to miss the motion information of body. The experimental results on three datasets, TST, UTKinect, and UCFKinect, determine the power and effectiveness of the proposed method
  9. Keywords:
  10. Motion Detection ; Keyframes Selection ; Fisher Subspace ; Body Poses ; Space-Time Points ; Key Poses ; Key Joints

 Digital Object List

 Bookmark

...see more