Human activity recognition in video using two methods for matching shape contexts of silhouettes
Natasha Kholgade, Andreas E. Savakis · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2008
In this paper, activity recognition is performed based on silhouettes of the human figure obtained by background subtraction and characterized by the shape context, a log-polar histogram derived from boundary points. In the first approach each video frame is tagged by the activity corresponding to the closest matches between the query and known shapes. In the second method, the shape context dimensionality is reduced by principal components analysis, and a neural network is used for activity classification of individual frames. The overall decision for an entire video sequence is based on majority vote. Classification of individual frames ranged between 70-90% and overall classification of video sequences was very accurate.