Research Of Gait Recognition Method Based On Deep Learning

Gait recognition is an emerging biometric technology that identifies people based on their style of pedestrian movement.It has the characteristics of long-distance,non-contact,difficult to hide and not easy to disguise.These characteristics make it an important application in the field of security monitoring and information intelligence,and also make it an irreplaceable identity authentication technology.The paper uses the deep learning method to study gait recognition,The main work is as follows.(1)For the noise and background in the gait video which interfere with the training results of the neural network,we extract the gait Gaussian image of the motion process as the training set.Firstly,the mixed Gaussian model is used to model the background of human motion,and then the moving target is extracted by background subtraction.Secondly,the target image is morphologically processed according to the problem of noise and disconnection.Finally,the gait cycle is detected according to the variation of the height and width ratio of the gait contour during the movement,the gait contour image is normalized,and the normalized image is weighted averaged in one cycle to get the gait Gaussian image.(2)In the gait recognition based on convolution neural network,the training of the model requiires a large number of training samples and time,and a gait recognition method based on the amplified sample information is proposed.The method adds Gaussian noise and salt-pepper noise to the original gait Gaussian image to rich sample sets,and then trains the neural network model by means of transfer learning.The experimental results show that the proposed sample set amplification method effectively improves the recognition rate of the model based on the expansion of the sample set,and the training method of migration learning shortens the convergence time of the model.(3)Aiming at the large difference between samples of the same category and the small difference between different types of samples,an improved neural network model of Siamese structure is proposed.A full-scale training set construction method is designed to overcome the problems of error and imbalance in manual labeling.Firstly,the loss function is optimized by combining the cross entropy loss and the contrastive loss,the cross entropy loss is used to constrain the semantic distance between different classes,and the contrastive loss function is used to constrain the mutual relationship between samples,and increase the distance between classes while reducing the distance within the class.Secondly,any two samples in the batch are used as sample pairs to construct the training set,making full use of the similarity information between all samples in the batch.The experimental results show that the method can reduce the intra-class variance while expanding the variance between classes,and it still has better recognition performance when pedestrians wear coats or backpacks.