Learning and Recognizing Human Actions Using PCA and 3-D Motion Trajectories

An approach for learning and recognizing human actions in videos is described. The proposed approach learns various actions through principle component analysis (PCA) of motion curves, and recognizes observed actions using features of their motion curves. Trajectories of one or more key points on the human body are created. A curve is fitted to each trajectory to smooth noise and to produce a continuous and smooth motion curve. A motion curve is partitioned into segments, each representing a basic motion. A basic motion is then recognized by projecting its features to an eigenspace created by PCA and following the k-nearest