Gait Optical Flow Image Decomposition for Human Recognition

As a behavioral biometric, gait recognition has gained an increased interest in recent years because it can operates without subject cooperation and from a distance. This paper presents a novel gait feature extracting approach based on gait optical flow image (GOFI) decomposition. The variation of algebraic sum of all vertical optical flow components is used to detect gait cycles. We calculate sums of horizontal and vertical optical flow components that greater than 0, respectively, for each row and column of GOFI to obtain four feature vectors of the subject. By exploiting principle component analysis (PCA) for the feature vectors to compute a PCA subspace that has the largest variance associated with them, then linear discriminant analysis (LDA) for the subspace to compute a LDA subspace that discriminates among the PCA subspace. Experiments implemented on the CASIA Database B and C demonstrate the approach achieves a 98% recognition rate under normal walking condition, while a promising performance under the influence of other covariate factors.