Unsupervised motion classification by means of efficient feature selection and tracking

We present an efficient technique for human motion recognition; in particular, it is focused on labeling a movement as a walking or running displacement, which are the most frequent type of locomotion. The proposed technique consists of two stages and is based on the study of feature points´ trajectories. The first stage detects peaks and valleys of points´ trajectories, which are used on the second stage to discern whether the movement corresponds to a walking or a running displacement. Prior knowledge of human body kinematics structure together with the corresponding motion model are the basis for the motion recognition. Experimental results with different video sequences are presented.