Hence we are able to recognize each view by treating

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: of each action are modeled as individual action sequences. Hence, we are able to recognize each view by treating it as a distinct action sequence and without having to incorporate information from the other view. 2.5 Discriminating similar actions For certain actions the head moves in a similar fashion. For instance, when viewed from the front, during squatting, sitting down and bending down, the head moves downward without much sideward deviation. Similarly, during standing up, rising and getting up actions, the head moves upward without much sideward deviation. In order to distinguish these actions from one another, we consider a discriminant number, whose value depends on how low the head goes in the performing of these actions. During bending down, the head goes much lower than in sitting down, and in sitting the head goes lower than in squatting. Let g = maxyinput =maxytraining  12 In general, maxygettingup  maxysitting  maxysquatting  13 where g is the discriminant number obtained as a ratio of the maximum y co-ordinate in the input sequnce to the maximum y co-ordinate in the training sequences, maxygettingup , maxysitting , maxysquatting  are the maximum values of the y co-ordinate of the head in the getting up, sitting and squatting actions in the front view. We compute ggettingup , gsitting , gsquatting , as the discriminant numbers corresponding to the three classes, namely getting up, sitting and squatting in the front views, which are obtained using equation 12. Thus whenever the system nds that the input action is one of the above three, it decides the most likely action by choosing that action which has the maximum discriminant number. A similar process is invoked for the rising from the squatting position, standing and getting up actions. Other actions that are similar with respect to the motion of the head can be distinguished by considering the size of the head in successive frames. Thus, a walking action in the frontal view, which is similiar to the backwards bending action, can be distinguished by making use of the fact that the size of the head increases over successive frames as the subject approaches the camera. = max=min 14 where  is the size of the head in one frame of the action sequence and is the ratio of the maximum and minimum sizes of the head taken over all frames of that action sequence. If , where is a prede ned threshold, then the computed probability for the walking action in the front view is multiplied by a weighting factor Wi . 3 Detection & Segmentation The detection and segmentation of the head is central to the recognition algorithm. We model our system by estimating the centroid of the head in each frame. Many human activity recognition algorithms depend on e cient tracking of a moving body part 5, 9 . Similarly, in our case, the entire recognition algorithm is based on reliably tracking the centroids of the head. At this stage of the project we do the segmentation by hand, isolating the head from the rest of the scene...
View Full Document

This note was uploaded on 04/12/2013 for the course ECON 2781631 taught by Professor Coop during the Spring '13 term at Emmanuel.

Ask a homework question - tutors are online