画像間差分とCNNを用いた低解像度画像による歩容認証

抄録

Gait-based person recognition has received an increasing amount of attentions for monitoring and surveillance applications. One of issues in gait recognition is that it is difficult to recognize people with high performance, in case that the resolution of captured images is too low. To deal with this problem, this paper presents SFDEINet, which uses Singed Frame Difference Energy Image (SFDEI) as input images. SFDEI has an advantage of explicit representation of motion of walking people, by changing the size of frame difference. To take the size of frame difference into account, SFDEINet is deigned to have convolution layers in parallel, followed by fusing outputs of each layer. We will show the SFDEINet is more suitable for SFDEI than GEINet in experiments.