Cross-view gait identification based on convolution neural network with joint loss function

Abstract

Most current gait recognition approaches based on convolution neural networks (CNNs) do not learn the discriminative features of separable inter-class differences resulting from cross-view data. To improve this discriminative ability, this paper proposes a network that reduces intra-class variation using a center loss function for view-invariant gait recognition. The proposed method achieved 92% accuracy using OU-MVLP, the largest existing gait recognition dataset. Furthermore, a network trained using the OU-MVLP achieved 95% accuracy with the OU-LP. These results demonstrate that the proposed method offers a good generalization performance. ？ ICROS 2019.