Human Pose Estimation And3D Recovery Based On Visual Geometric Features And Machine Learning

Human pose estimation has been an intriguing topic in the field of computer vision and artificial intelligence in recent years, with broad and crucial applications such as video surveillance, interactive animation, motion capture and advanced human computer interaction, etc. However, mankind has proven to be a complex object to estimate the pose, due to the appearance variability among bodies, clothing deformation and articulation, illumination changing and background clutter that are common in the real world. In this paper, we utilize the geometric flow of images from the second generation Bandelet transform and presented new feature descriptors to improve the accuracy of human pose estimation with a machine learning scheme.Research has been done primarily focusing on dimensionality reduction of descriptors, improving feature representativeness and enhancing the robustness of pose estimation. It involved the method of machine learning, regression, holistic features and bag-of-words representations. Specifically, the paper proposes three original pose estimation methods relating to image representation:1) Based on the second generation Bandelet transform, we utilized previous fruits and incorporated maximum statistic features to reduce the error of human pose estimation in images. Here only Bandelet statistical values were extracted and modified as the features of human images, combined with regression methods such as GPR, we learned and predicted3D poses in newly received videos.2) Inspired by our previous work, we proposed a histogram of maximum geometric directions as an image descriptor and estimate3D human pose through machine learning. Region segmentation is carried based upon geometric flow, and Bandelet transform is recorded in directions to form bins of histograms on geometric intensity. We performed normalization on resulted histograms, where experiments show the robustness and selectiveness of our method.3) The third method focused on reducing the time complexity by using EMK kernel and learning a linear regression frame. The previously extracted SURF feature is learned with CKSVD and projected to a low dimension kernel feature space, and is learned to recover3D human pose with linear mapping. Both feature extraction and prediction times are reduced.