Uniprojective Features for Gait Recognition

Recent studies have shown that shape cues should dominate gait recognition. This motivates us to perform gait recognition through shape features in 2D human silhouettes. In this paper, we propose six simple projective features to describe human gait and compare eight kinds of projective features to figure out which projective directions are important to walker recognition. First, we normalize each original human silhouette into a square form. Inspired by the pure horizontal and vertical projections used in the frieze gait patterns, we explore the positive and negative diagonal projections with or without normalizing silhouette projections and obtain six new uniprojective features to characterize walking gait. Then this paper applies principal component analysis (PCA) to reduce the dimension of raw gait features. Finally, we recognize unknown gait sequences using the Mahalanobis-distance-based nearest neighbor rule. Experimental results show that the horizontal and diagonal projections have more discriminative clues for the side-view gait recognition and that the projective normalization generally can improve the robustness of projective features against the noise in human silhouettes.