Gait Recognition Based On High-Order Tensor Subspace Learning

Tensor is a multilinear function which can be used to represent a linear relationship between vectors,scalars,and other tensors.The matrix is a representation of the tensor in a specific set of base vectors,and the image and video sequences can be expressed in the form of tensor.Algorithm DTSA performs good in dealing with two-dimensional tensor data,but can not handle high-order tensor data,which is its biggest drawback.Based on this point,this thesis carried out research works.Considering the advantages of multi-linear discriminant analysis in avoiding dimension disaster and solving small sample problems,combined with graph embedding and manifold methods can preserve the local geometric spatial structure of data and so on,this thesis designs and completes the new algorithm and its related experiments and system implementation.The main works are as follows:Firstly,Research on graph embedding method based on high-order tensor.For the DTSA algorithm can only deal with the problem of two-dimensional local reservation projection,this thesis studies the principle and idea of graph embedding and manifold methods,and designs two kinds of weight graphs embedded in graphs based on high-order tensor: intrinsic graph and local penalty graph,and the similarity matrices in the corresponding tensor subspace are obtained,which can retain the local geometric spatial structure,and retain data space information while dimension reducing.Secondly,Research on improved multi-linear discriminant analysis.For the DTSA algorithm,we can only deal with the problem of second order discriminant tensor subspace analysis,this thesis designs and implements a multi-linear discriminant analysis method that can deal with high-order tensor data,this not only can be coded with multi-linear subspace learning with discriminant information,but also has better performance in avoiding dimension disaster and small sample problem.Thirdly,Design and implementation of new algorithm.For the DTSA algorithm can only deal with low-dimensional and can not handle the problem of high-order tensor data reduction and classification.The new algorithm firstly reduces the dimension of the high-order tensor data,and then extracts the features of tensor,after that using the KNN for classfication,and finally achieve the purpose that the gait recognition is performed by processing the high-order tensor data.What's more,the performance of the new algorithmis evaluated and the compared gait recognition experiments are carried out,and the design and implementation of the gait recognition prototype system based on the new algorithm and contrast algorithm is completed.