Research On Gait Recognition Method Based On Deep Learning

As an important branch of biometric identification,gait recognition is widely used in criminal investigation,identity verification,public security intelligent monitoring and other fields because of its uniqueness,non-contact,and long-distance characteristics.It analyzes the gait contours presented by pedestrians when walking,extracts reasonable gait features from them,and constructs an effective classification and recognition model,thereby realizing the identification of pedestrians.However,in practical applications,it is difficult for the learning model to extract effective gait features due to factors such as the perspective of the scene,pedestrian backpacks,and clothing.Few training samples,unbalanced samples,and improper selection of deep learning network models can also easily lead to failures.When the fitting phenomenon occurs,the accuracy of gait recognition is often unsatisfactory.The thesis aims to improve the accuracy of gait recognition,focusing on the hyperparameter optimization method of deep learning model and avoiding over-fitting strategies,and develops the gait recognition method based on the residual block-based convolutional neural network and the attention mechanism.Research on the gait recognition method of twin neural network,and develop the gait recognition application system.The main contents include:(1)Research on the gait recognition method of convolutional neural network based on residual block.The effects of different parameter initialization methods and activation functions of the convolutional neural network on the convergence speed and recognition rate of the network model are analyzed.In order to avoid the over-fitting phenomenon caused by the deep network,the residual block is used to "jump-connect" the gait features output by the upper convolutional layer to the subsequent convolutional layer for feature fusion.It is used in the traditional convolutional network.In fusion of the residual block model optimization strategy,a convolutional neural network gait recognition method based on the residual block is proposed.The experimental results show that on the CASIA-B data set,Xavier is selected as the parameter initialization method and Re LU as the activation function to improve the convergence speed of the gait recognition method based on the residual block of the convolutional neural network,and the recognition accuracy is improved to a certain extent(2)Research on gait recognition method of twin neural network based on attention mechanism.Deep learning requires a large amount of sample data in the training process,but the current gait recognition data set has less training data for each category,and the number of samples in different categories is not balanced.To solve the above problems,the gait is realized by introducing a twin neural network model.The enhancement of training samples avoids the problem of small and unbalanced sample data sets.At the same time,in order to make more effective use of identity tags,a learning method of dual-supervised twin neural networks is proposed that constrains similarity measures and tags.The attention mechanism is used to improve the characteristic expression ability of the network and further improve the learning ability of the network.Experiments show that the accuracy of the twin neural network gait recognition method based on the attention mechanism has been significantly improved;(3)The design and implementation of the gait recognition system based on deep learning.Using the object-oriented programming method,the functional modules of the gait recognition system are designed.With Open CV,pycharm and QT as the development environment,the gait recognition system based on the twin neural network of the attention mechanism is used to develop the gait recognition system.The system can automatically collect pedestrian videos,detect gait images and preprocess it,extract gait features and recognize gait,and has a better human-computer interaction interface.