Attention-based Gait Recognition System

Human identification based on biometrics is a popular research filed in computer vision.Gait recognition is one of the biometric identification technologies,which exerts walking styles of human-beings to verify identities.Gait data is usually collected by cameras from a distance,and this means that identification process is contactless and can be carried out without the awareness of the subject.Besides,combining both static and dynamic information,walking postures are much more difficult to imitate than other static biometics like face and fingerprint images.Due to these pre-mentioned merits,gait recognition gained popularity in academic community,and has already been used in public security and criminal reconnaissance.In this thesis,we will discuss gait recognition systems based on deep learning methods,and propose two attention-based gait recognition systems to deal with cross-view problem.The angle between the camera's view direction and the walking direction is called view angle of gait data.In some conditions,there may be a difference between the view angle of query gait and those in gait database,which is called the view difference.Cross-view problem is the gait recognition with view difference larger than 0.To settle the cross-view problem,we propose an attention based embedding network for cross-view gait recognition.Difference from pair-wise input networks and siamese networks,our method utilize a single-input network which strives to embeds gait data into low dimensional embedding space,such that the similarities can be directly measured by the distance in the embedding space.The whole network is trained in an end-to-end way with both contrastive loss and triplet.For the backbone of the embedding network,we propose two structures:(1)CNN with GEI input;(2)3D-CNN with sequential silhouettes input.Additionally,an attention block is incorporated in the backbone network to locate the most discrimative part of feature map.Existing gait recognition approaches deal with different parts of feature maps indiscriminately due to the weight-shared strategy of CNN.This may cause a loss of discriminative information since different human parts vary in shape,movement range,individual variation and so on.To address such a problem,we devise an spatial attention block to assign different saliency weights to different regions of feature maps at pixel level.This helps us to focus on the most importance pixels in feature maps.In this thesis,we evaluate the performance of our proposed methods on OU-MVLP and OULP datasets,two large-scale datasets for cross-view gait recognition.The results demonstrate that both our proposed methods prevail over the state-of-the-art works on both OULP and MVLP dataset under cross-view conditions.Notablely,on OULP dataset,the fusion of the two methods achieves 2.1% improvement on average and 6.3% improvement under 30?view difference.