Clothing-invariant human gait recognition using an adaptive outlier detection method

Abstract

Human gait as a behavioral biometric identifier has received much attention in recent years. But there are some challenges which hinder using this biometric in real applications. One of these challenges is clothing variations which complicates the recognition process. In this paper, we propose an adaptive outlier detection method to remove the effect of clothing on silhouettes. The proposed method detects the most similar parts of probe and each gallery sample independently and uses these parts to obtain a similarity measure. Towards this end, the distances of the probe and a gallery sample are calculated row by row which are then used to obtain an adaptive threshold to determine valid and invalid rows. The average distance per valid rows is then considered as dissimilarity measure of samples. Experimental results on OU-ISIR Gait Database, the Treadmill Dataset B and CASIA Gait Database, Dataset B, show that this method efficiently detects and removes the clothing effect on silhouettes and reaches about 82 and 84% successful recognition respectively.