Coupled Distance Metric Learning Method Research And Its Application In Gait Recognition

With the rapid development of modern society,the traditional identification method can not meet the needs of people because of the poor security.So the biometric identification with universality,stability,and uniqueness is getting more and more attention,and it is widely used in many fields.However,within the biometric identification,the human features are generally obtained under different conditions.At this point,the traditional metric learning is mostly defined in a single set that it is not able to deal with the metric among elements from different sets.Currently,for this problem a better mothod is the coupled distance metric learning.Its goal is to find a coupled distance functions for directly processing the different sets.The core idea of this method is mapping the data from different sets to the same coupled space,and the elements of two sets with correlation are as close as possible after the projection.Then the traditional metric learning is used in the public coupled space.There are many deficiencies in conventional linear coupled distance metric learning that based on similar constraints,locality preserving,or discriminant information.By carrying out related research work,this paper gives a variety of improvement strategies,and it is applied to feature level fusion field.The coupled distance metric learning theory is fully perfected.Then this theory is applied to gait recognition,to overcome the difficulties in the gait recognition.Finally the practical and convenient gait recognition system is created with high stability and high recognition rate.The main contents are summarized as follows:1.Through the analysis of the existing coupled distance metric learning,the separability criterion is introduced to improve the original optimization goal.The separability-criterion coupled distance metric learning is proposed.By using a criterion function,the method makes the square of average distance between the similar samples as small as possible,and the square of the average distance between the unsimilar samples as large as possible.After,the separability of the coupled space is enhanced.The linear coupled distance metric learning has the "curse of dimensionality" issue in the practical application.Three improvement strategies are studied:First,the original high-dimensional data are transformed by PCA,and the coupled distance metric learning is used to the dimensionality reduction characteristic.This method is the coupled distance metric learning based on K-L transform.Second,the original image is divided into sub-blocks,and the coupled distance metric learning is applied to each sub-block.The final characteristics are obtained from all the sub-block.This method is the coupled distance metric learning based on the sub-pattern.Third,two-dimensional image is processed directly.This method is the coupled distance metric learning based on the two-dimensional.2.The linear coupled distance metric learning is difficult to resolve the widespread nonlinear problems.The nonlinear coupled distance metric learning based on the kernel space is studied.First,through the kernel transform the data is mapped to a high-dimensional space,and then the conventional distance metric learning is used in that high-dimensional space.3.When the supervised information is insufficient,coupled distance metric learning will be affected heavily.By using the graph-based semi-supervised learning method,the local neighbors of the sample and the global information of entire sample set are used to extend the supervised information.Based on local and global supervised information,the extended semi-supervised coupled distance metric learning is proposed.And the expansion of this supervised information is applied to the case of sufficient supervised information,making supervised information contains both the sample's class information and the sample itselfinternal information.Based on local and global supervised information,the extended supervised coupled distance metric learning is proposed.4.In the biometric recognition system,using single data to test is of not stable performance and high error rate.The common solution is the data fusion technology.According to the fact that the coupled distance metric learning can directly deal with two sets,a variety of feature level fusion methods based on coupled distance metric learning are proposed.These methods use different strategies to fuse the features of the coupled space.Then more stable features are obtained for classification and recognition.Moreover,a kind method of multi feature level fusion is proposed for the problem of multiple sets.5.Gait recognition has many deficiencies in the practical application.The coupled distance metric learning is used for proposing a constructing method of gait recognition system with high practical value.This method makes the gait image with different conditions unified together by coupled distance metric learning,which would not only improve system stability and recognition performance,but also greatly reduce the system's storage space.