Abstract
Activity recognition has been applied to many varied applications ranging from surveillance to medical analysis. Interpreting human actions is often a complex problem for computer vision. Actions can be classified through shape, motion or region based algorithms. While all have their distinct advantages, we consider a feature extraction approach using convexity defects. This algorithmic approach offers a unique method for identifying actions by extracting features from hull convexity defects. Specifically, we create a hull around the segmented silhouette of interest in which the regions that exist in the hull are recognized. A feature database is created through a dataset of features for multiple individuals. These feature points are registered between progressive frames and then normalized for analysis. Using Principal Component Analysis (PCA), the feature points are classified to different poses. From there testing and training is performed to observe the classification into major human activities. This approach offers a robust and accurate method to identify actions and is invariant to size and human shape.