Human Gait Recognition by Integrating Motion Feature and Shape Feature

Gait is thought to be the most effective feature for human recognition in the distance. For optimal performance, the feature should include as many different types of information as possible, so in this paper, we present an integrated feature, which integrates motion feature and shape feature based on the Bayesian theory. For motion feature, we use shape variation-based frieze pattern (SVB frieze pattern) as the basis, since it can solve the ball or backpack problems very well, then we match the SVB frieze pattern feature by dynamic time warping (DTW). For shape feature, we use gait energy image (GEI) as the basis, since it is less sensitive to the silhouette noise, then we extract further information by histograms of oriented gradients (HOG) and do the dimensionality reduction by coupled subspaces analysis (CSA) and discriminant analysis with tensor representation (DATER). The proposed approach is tested on the CMU MoBo gait database. The result shows that the proposed approach is an efficient way in increasing the accuracy.