Research On Key Techniques Of Gait Recognition Based On Deep Learning

Biometrics based automatic human identification is one of the most fundamental and frontier research topic in computer vision.Among the massive biometric authentication traits,human gait gives more remarkable characteristics such as remote accessed,robust and security.Therefore,gait recognition,which essentially aims to discriminate individuals by the way they walk,has gained much attention in the human-centric modern intelligent video surveillance system.However,this problem remains many difficulties,such as the inconspicuous inter-class differences from different people,and the large intra-class variations from the same people(e.g.,different viewpoints,clothing,belongings and walking speed),and still have large space to be improved.To solve these challenges,most existing methods mainly employ the handcrafted gait features.However,those features can extremely hard to break through feature extraction and representation bottleneck when facing with the requirement of fine-grained gait recognition task.In this thesis,we propose a series of novel models and methods for deep learning based gait recognition problem.First,we design a conventional convolutional neural network(CNN)based gait recognition framework.The proposed framework exploits Gait Energy Image(GEI)as the input of the CNN to fine-tune the pre-trained model,which is a good solution for the data limitation problem and to speed up the convergence of new model.Next,we exploit the Siamese neural network with a distance metric learning architecture to learn sufficient feature representations to tackle gait recognition for human identification.Finally,we propose a 3-Dimensional Siamese neural network to further improve the performance of the gait recognition.In this way,spatiotemporal deep features are extracted directly from sequential gait images via 3-Dimensional convolutional neural network without any preprocessing,which is able to take great advantages of periodical dynamic and motion pattern of human gait.In the experiments,the evaluations on the world's largest and most comprehensive gait benchmark dataset demonstrate that our proposed methods can impressively beyond state-of-the-arts in gait-based gender classification problem and with nearly 5%improvement in human identification task.