This was done for all action sequences and for di

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: probability is the highest. The subjects were asked to perform the actions at a comfortable pace. This was done for all action sequences and for di erent subjects. Human motion is periodic; hence, we can divide the entire sequence into a number of cycles of the activity. After observing the various sequences and di erent subjects executing these actions, it was found that on average ten frames were required to completely describe an action. We found this to hold true for all twenty actions that were modeled and tested. Hence, we designed our system to consider only the rst ten frames of each sequence, ignoring the rest. The rate of capture of the images was 2 frames second. Thus we assumed that each action was performed in roughly ve seconds. We also tested our model on action sequences done at a faster rate, for instance, actions that required only ve, six or seven frames. Hence, for an input sequence that has only ve frames, we select only four of the 9 elements of the X and Y feature vectors obtained from the training samples and use them to compute a 4 by 4 covariance matrix. The model was able to recognize the action correctly in most cases. However, for actions that required fewer frames than this, the model was not that successful. For the threshold and the weighting factor we used = 5:2 and Wi = 2:15. COMPUTING A EXTRACTING FEATURE VECTORS STATISTICAL MODEL BASED ON FEATURE VECTORS DETECTING AND SEGMENTING THE HEAD MATCHING AGAINST ACQUISITION CLASSIFYING TYPE OF ACTION MODULE STORED MODELS Figure 2: System Overview. 5 Results This section describes the results obtained from experiments performed on a database of 77 action sequences. Of these, 38 were used for training and 39 were used for testing. Of the 39 test sequences, the system was able to correctly recognize 31, giving a success rate of 79.74. In 6 of the 8 test sequences that were incorrectly classi ed, the system classi ed the action correctly, but as belonging to the wrong eld of view. The system was able to recognize actions for people of varying physi- cal appearances, from tall to short and from slender to fat. Figures 3-6 show sequences of a subject executing di erent actions. Owing to a paucity of space, only ve key frames in each sequence have been shown. In Figure 3a, a person is standing up in the lateral view. In the segmented sequence of the standing up action, Figure 3b, we can see the distinct movement of the head as it moves forward initially, then slightly downward and progressively upward and backward. In Figure 4a, the subject is seen executing a bending over action in the front view. Figure 4b reveals the characteristic downward motion of the head in the front plane. Figure 5a shows the subject executing a sideways bending action in the front view. In Figure 5b, the segmented version of the same, we see the head of the body trace the arc of a circle that has a radius equal to the length of the upper body torso. The center of this arc lies roughly at the center of the body. Finally, in Figure 6a we see the subject hugging another person. Notice, in Figure 6...
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