Attitude estimation method based on sparse Gaussian process with noise input and attitude estimation system thereof

Abstract

The invention provides an attitude estimation method based on the sparse Gaussian process with noise input and an attitude estimation system thereof. The attitude estimation method comprises the stepsof image feature extraction: the human body attitude contour of an image is extracted through the suppressed background model and the adaptive threshold and the HOG features of the image are extracted; and prediction estimation: the induction input and output set of a sparse algorithm is selected, the Gaussian noise is added to the input set, and the hyper-parameters of the Gaussian process withthe noise input are determined through the maximum likelihood estimation method and the gradient descent method so that the joint probability distribution function of the Gaussian process can be determined, the selected induction input and output set is used as the training set to perform sparse training of the sparse algorithm and the prediction estimation model can be obtained. Great invariancecan be maintained for the geometrical and optical deformation, and the tiny motion of the human body can be omitted without influencing the detection effect so that the method has better fitting degree and robustness, low complexity and great operation efficiency, prediction accuracy and anti-noise performance.

Description

technical field [0001] The present invention relates to the technical field of attitude estimation, in particular to an attitude estimation method and system based on a sparse Gaussian process with noise input. Background technique [0002] For the problem of human pose estimation, the usual solution process is: image acquisition, data preprocessing, feature extraction, machine learning algorithm prediction and estimation. [0003] Through image acquisition technology, an entire video sequence is converted into a single frame image that is easy to identify. Since there is no usable optical flow information in a single frame image, there is no usable background knowledge. However, the background in real scenes is usually relatively complex, so effectively extracting human target features is an essential step for pose estimation. In terms of feature extraction techniques, both 2D pose estimation and 3D pose estimation can be divided into model-based methods and model-free me...

AI Technical Summary

Problems solved by technology

However, human pose estimation usually requires huge data sets and high input and output dimensions. Therefore, using Gaussian processes directly for prediction will result in long running time and a large amount of computing space.

Images