Cross-View Gait Identification with Embedded Learning

To improve the accuracy of cross-view gait identification, a novel approach is proposed in this paper. We design a triplet-based convolution neural network with embedded learning. Meanwhile, to prevent over-fitting, we adopt a shallow layer convolution neural network. This network model is trained through the triplet loss function, which makes inter-class variations larger than intra-class variations. The network takes triplet samples as input which are consisted of three parts, query sample, positive sample and negative sample. All triplet samples are selected from two kinds of the challenging gait dataset, such as the CASIA-B dataset and OU-ISIR dataset. Extensive evaluations are carried out based on the new network. Experimental results show that the triplet-based convolution neural network performs better than the traditional ones for the gait identification under cross-view angle. Especially for the large scale dataset, this method performs best. Various kinds of evaluations show its great potential for practical applications.