Research On Gait Recognition Based On Three-dimension Convolutional Neural Network

Biometrics is a certification mechanism that automatically identifies\validates a person by the unique physical or behavioral characteristics.Because of the higher accuracy of biometric-based recognition techniques,many biometric features have been used in different fields.Gait recognition,one of the behavioral biometric methods,is used to identify people by the way of their walking.It has the advantages of long distance recognition.Gait recognition has been widely used in practical applications.In vision-based gait recognition,the change of view is one of the most challenging factors for gait recognition.In order to address the cross view issues in gait recognition,this thesis proposes a view-invariant gait recognition method by 3D convolutional neural network.First,3D convolutional neural network(3DCNN)is introduced to learn gait feature,which can capture the spatial and temporal information simultaneously on normalized silhouette sequences.Second,a network training method based on cross-domain transfer learning is proposed to solve the problem of the limited gait training samples.The final 3DCNN model can be obtained by pretraining on Sports-1M and fine-tuning on gait training samples.In the recognition stage,the final model can be used to extract gait features and Euclidean distance is used to measure the similarity of gait sequences.In order to further improve the performance,this thesis proposes a gait recognition method based on local and global feature fusion.First,local and global features are extracted by using 3DCNN model.Second,the global and local features are combined appropriately by serial feature fusion.Then,feature selection is made by principal component analysis.The fusion gait feature has better robustness by combining the holistic of the global features and the detailed variations of the local features.Finally,the gait sequence is identified by support vector machine.Sufficient experiments are carried out on the CASIA-B dataset.The experimental results demonstrate that the method based on 3DCNN improves the performance of the existing gait recognition methods,and the method still maintains good performance when the degree of view changes greatly.The method can effectively address the cross view issues in gait recognition.Based on this method,the global and local features are fused and the accuracy is further improved.