Research Of Multi-view Gait Recognition

Biometric technology refers to the technology which identifying individuals by distinguishing the differences of physiological or behavioral characteristics between humans.In recent years,the computing capability of computer is increasing rapidly.The amount of information is growing explosively.The new generation of artificial intelligence technology that based on deep neural network is widely used in the field of biometric recognition.Gait recognition is a new type of biometric technology.Different from fingerprint recognition,face recognition,iris recognition or other recognition technology which needs high quality input data.The quality of input data for gait recognition can be very low.Gait recognition does not need special equipment,recognition mission can be done with ordinary surveillance cameras.Gait recognition has great advantages in security,investigation fields because it can collect data covertly at a long distance without the cooperation of the subjects.At present,surveillance cameras that increasing rapidly provide large number of training data for gait recognition,also provide hardware for gait recognition.This paper proposes a method for multi-view gait recognition,which uses deep neural networks to extract identity information from gait sequence.This method is proposed to enhance the ability of the network to extract identity information from gait sequence,and reduce the impact of occlusion and view-change,The research contents of this paper are introduced as follow:(1)Dynamic gait image is proposed to reduce the impact of occlusion on feature extraction.Dynamic gait image is a method to process the gait silhouette in the preprocessing stage.In dynamic gait image,gait silhouette is divided into dynamic parts and static parts.Dynamic gait image is beneficial to the network to extract dynamic information which is less affected by occlusion.(2)A Bi-Route gait recognition network is proposed to reduce the influence of occlusion by increasing the proportion of dynamic features and diluting the proportion of static features.The network takes the dynamic gait map as the input,uses 2D-Convolution neural network to extract the global level gait features and frame level gait features in the sequence,and uses 3DConvolution neural network to extract the dynamic information from the frame level gait feature map.(3)A gait recognition software is designed.This software improved the process of gait recognition.It first judges the shooting view of the subject,and then matches the identity tag with the most similar feature from the corresponding database.Using this method can reduce the search scope of the database and improve the matching accuracy and matching speed.In this paper,CASIA-B data set is used to train and test the network.The effectiveness of network is verified under the identical-view gait recognition and cross-view recognition.This method is compared with standard-of-the-art methods,the results indicate that the method proposed in this paper performs better in gait feature extraction.