Human gait recognition from motion capture data in signature poses

Most contribution to the field of structure-based human gait recognition has been done through design of extraordinary gait features. Many research groups that address this topic introduce a unique combination of gait features, select a couple of well-known object classiers, and test some variations of their methods on their custom Kinect databases. For a practical system, it is not necessary to invent an ideal gait feature -- there have been many good geometric features designed -- but to smartly process the data there are at our disposal. This work proposes a gait recognition method without design of novel gait features; instead, we suggest an effective and highly efficient way of processing known types of features. Our method extracts a couple of joint angles from two signature poses within a gait cycle to form a gait pattern descriptor, and classifies the query subject by the baseline 1-NN classier. Not only are these poses distinctive enough, they also rarely accommodate motion irregularities that would result in confusion of identities. We experimentally demonstrate that our gait recognition method outperforms other relevant methods in terms of recognition rate and computational complexity. Evaluations were performed on an experimental database that precisely simulates street-level video surveillance environment.