Based on Siamese Network with Self-Attention Model for Gait Recognition

The gait feature is non-invasive remote feature and does not require high-resolution images, which has attracted widespread attention and has great potential as a security recognition. As an emerging biological feature, gait recognition still has many challenges, such as the recognition of complex background, perspective change and occlusion. This paper proposes a neural network based on Siamese structure, which takes the gait energy image (GEI) as the input of the network. At the same time, attention mechanism is added to the network, and self-attention is used to establish the connection between the distant pixels so as to better learn the non-local long-distance information. In the training stage, using Multi-Layer Side-Output (MLSO) as a reference, companion loss is added to the intermediate layer to supervise the middle layer. The experiment is carried out on the CASIA-B database, and it is comprehensively compared with the unimproved network and other methods which work well by setting up different training sets and test sets. Experimental data show that the network proposed in this paper has improved the average recognition rate in gait recognition across perspectives.