Sparse Gaussian Process With Input Noise For Human Pose Analysis

Human pose analysis is an important direction in image processing and pattern recognition,and has wide applications and broad prospects in many fields.Among various algorithms for human pose estimation,Gaussian process regression(GPR)is a common method to solve this problem because of its superiority in dealing with nonlinear,high-dimensional and complex problems.Computational complexity is a significant consideration of Gaussian process regression and it can be reduced by sparse Gaussian algorithm.The Fully Independent Training Conditional(FITC)algorithm is a good method for sparse Gaussian process,and it can be applied in fully-independent input problems.Input noise is another significant consideration of GP regression.Moment matching can be used to solve trial input noise while training input noise can be modeled as output noise to achieve higher accuracy.Probabilistic modelling with neural network architectures constitute a well-studied area of machine learning.Deep models seem to have structural advantages that can improve the quality of learning in complicated data sets associated with abstract information,but corresponding,a large amount of data is also indispensable.Deep Gaussian process(DGP)can solve this dilemma.The single layer structure of DGP comes from the Gaussian process with latent variable model,and the structure can be undertaken in smaller data sets.In this paper,these two algorithms are used to solve two kinds of human pose datasets respectively.One derives from the HumanEva-I dataset,in which inputs are feature vectors extracted from images in video sequences,and outputs are 3D human pose locations.The other dataset derives from space station cabin videos,in which inputs are still image feature vectors,and outputs are human joint angles and comfort information.Compared with others,algorithms in this paper achieve satisfactory results in accuracy,runtime and stability,and it is also feasible in practical engineering applications.