| The advances in processing technology and availability of large storage systems have resulted in a broad application of digital video data in the last few years. There has been a growing explosion video data on the Internet, surveillance systems in banks and hospitals. Human-Computer Interfaces and Virtual Reality have attracted the attention of researchers in recent years. Human action recognition as a fundamental aspect of these problems has always remained a highly active research area. Although a large amount of research has been reported on action categorization, complexity of human actions makes the research area challenging. Many works have been done about using static or motion features, which faced some limitations like self-occlusions in feature tracking. While space-time features both contain spatial information about the pose of the human figure as well as the dynamic information, which has been successfully applied to human action recognition. Another problem in this area is that the variety of human action, which makes it impossible to collect all kinds of actions.This thesis applied time series analysis method in human action recognition for the periodicity and self-similarity in human action. Trend, periodicity and self-similarity are extracted and characterize periodic motion. Then these features are used for recognizing human action using KNN with rejection. We set a rejection radius to ignoring points further than the radius, and label them as "unknown". When the amount of "unknown" data become larger then a fixed number, the system automatically classifies them into some smaller clusters which have same actions. Then we can label these clusters by hand and add them into training examples.Experiments show that the features can facilitate the tasks of human action recognition. Confusing the KNN with rejection and cluster method, we built a semi-automated way to expend human action data to the training example. |
PDF Full Text Request