An aperiodic feature representation for gait recognition in cross-view scenarios for unconstrained biometrics

The state-of-the-art gait recognition algorithms require a gait cycle estimation before the feature extraction and are classified as periodic algorithms. Their effectiveness substantially decreases due to errors in detecting gait cycles, which are likely to occur in data acquired in non-controlled conditions. Hence, the main contributions of this paper are: (1) propose an aperiodic gait recognition strategy, where features are extracted without the concept of gait cycle, in case of multi-view scenario; (2) propose the fusion of the different feature subspaces of aperiodic feature representations at score level in cross-view scenarios. The experiments were performed with widely known CASIA Gait database B, which enabled us to draw the following major conclusions, (1) for multi-view scenarios, features extracted from gait sequences of varying length have as much discriminating power as traditional periodic features; (2) for cross-view scenarios, we observed an average improvement of 22 % over the error rates of state-of-the-art algorithms, due to the proposed fusion scheme.