A new feature description based on feature relationships for Gait recognition

In order to carry on the gait recognition fast and efficiently, a new representation scheme for feature description is proposed in this paper which utilizes nostationarity in the distribution of feature relationships. Firstly , edge pixels of silhouette are considered as low level features, then relationships among those features are characterized by two attributes, which are label of relative direction of two adjacent edge pixels in 8-Neighborhoods region and distance from edge pixel to shape centroid point. Shape of the joint probability function of the two attributes vividly embodies the inter-feature relations and changes following specifically motion pattern. Secondly, we represent each relational distribution as a point in a space of probability functions(SOPF), which can be setup using the well-defined theory of principal component analysis(PCA). Finally, nearest-neighbor classifier is adopted for classification by normalized Euclidean distance(NED). The proposed approach is tested on NLPR gait database. The experiment result demonstrates that the proposed method has achieved a good recognition performance with relatively lower computational cost.