Recognizing human actions by their pose

The topic of human action recognition from image sequences gained increasing interest throughout the last years. Interestingly, the majority of approaches are restricted to dynamic motion features and therefore not universally applicable. In this paper, we propose to recognize human actions by evaluating a distribution over a set of predefined static poses which we refer to as pose primitives. We aim at a generally applicable approach that also works in still images, or for images taken from a moving camera. Experimental validation takes varying video sequence lengths into account and emphasizes the possibility for action recognition from single images, which we believe is an often overlooked but nevertheless important aspect of action recognition. The proposed approach uses a set of training video sequences to estimate pose and action class representations. To incorporate the local temporal context of poses, atomic subsequences of poses using ngram expressions are explored. Action classes can be represented by histograms of poses primitive n-grams which allows for action recognition by means of histogram comparison. Although the suggested action recognition method is independent of the underlying low-level representation of poses, representations remain important for targeting practical problems. Thus, to deal with common problems in video based action recognition, e.g. articulated poses and cluttered background, a recently introduced Histogram of Oriented Gradient based descriptor is extended using a non-negative matrix factorization reconstruction.