Research On Cross-View Gait Recognition Based On Deep Learning

Identity authentication with gait is an important schema in mid-long distance intelligent visual supervision senario,which has been a hotspot for frontier researches,where cross view relevant issues need further escavation.In recent years,Deep Learning has been going through fast development and yearmed a big number of achievements.It leads gait recognition to a creative path due to its learning capability of low level presentative info and high level sematic info.Based upon Deep Learning techniques,we investigated into cross view issues of gait recognition and yearned following innovative achievements:1.A novel gait recognition algorithm based on encoder-decoder is proposed.The algorithm uses an encoder-decoder as the architecture,and uses a convolutional neural network and a deconvolution neural network to convert gait images of any views into gait images of standard view.The context-aware network is introduced into the encoder-decoder network to learn gait features of different scales.The extracted gait features are sent to the identity classifier for identification of pedestrian identity.The effectiveness of the method was verified on the CASIA-B database.2.A novel gait recognition algorithm based on two-stream generative adversarial nets is proposed.The algorithm is further improved on the encoder-decoder network,it converts the gait image of any view to a standard view.The global stream GAN learns global gait features.The local stream GAN learn local features to grasp the local details of the network.The global-stream generator is an encoder-decoder architecture.Pixel-level constraints are imposed on the network to ensure that the network can recover details at pixel level.Experiments were carried out on the CASIA-B database and the OU-ISIR database,compared with the current advanced methods,and the effectiveness and advancement of the method are proved.