Abstract

The advantage of gait over other biometrics such as face or fingerprint is that it can operate from a distance and without sub ject cooperation. However, this also makes gait sub ject to changes in various covariate conditions including carrying, clothing, surface and view angle. Existing approaches attempt to address these condition changes by fea- ture selection, feature transformation or discriminant subspace learning. However, they suffer from lack of training samples from each sub ject, can only cope with changes in a subset of conditions with limited success, and are based on the invalid assumption that the covariate conditions are known a priori. They are thus unable to perform gait recognition under a genuine uncooperative setting. We propose a novel approach which casts gait recognition as a bipartite ranking problem and lever- ages training samples from different classes/people and even from differ- ent datasets. This makes our approach suitable for recognition under a genuine uncooperative setting and robust against any covariate types, as demonstrated by our extensive experiments.