Gait recognition for human identification using ensemble of LVQ Neural Networks

Usage of gait biometric in individual identification is a rather new and encouraging research area in biometrics. Requiring no cooperation from the observed individual, and functionality from distance, using non-expensive low resolution cameras, are the benefits that have been dragging enormous attention to gait biometric. However, it should be noted that, gait pattern in humans can be greatly affected by changing of clothes, shoes, or even emotional states. This natural variability, which is absent in other biometrics being used for identification, such as fingerprint and iris, decreases the reliability of recognition. In this paper, a mixture of experts, in form of an LVQNN ensemble was employed to improve recognition rate and accuracy. Majority voting fusion method was used to combine the results of LVQNNs. First, local motion silhouette images (LMSIs) were generated from silhouette walking frame sequences. Then using PCA, lower dimensional features were extracted from LMSIs, and were fed to classifiers as inputs. Experiments were carried out using the silhouette dataset A of CASIA gait database, and the effectiveness of the proposed method is demonstrated.