Vision and Depth Based Computerized Anthropometry and Object Tracking

Abstract

The thesis has two interconnected parts: Computerized Anthropometry and RGBD (RGB plus Depth) object tracking. In the first part of this thesis, we start from the mathematical representation of the human body shape model. It briefly introduces prior arts from the classic human body models to the latest deep neural network based approaches. We describe the performance metrics and popular datasets for evaluating computerized anthropometry estimation algorithms in a unified setting. The first part of this thesis is about describing our contribution over two aspects of human body anthropometry research: 1) a statistical method for estimating anthropometric measurements from scans, and 2) a deep neural network based solution for learning anthropometric measurements from binary silhouettes. We also release two body shape datasets for accommodating data driven learning methods.

In the second part of this thesis, we explore RGBD object tracking. We start from the current state of RGBD tracking compared to RGB tracking and briefly introduce prior arts from engineered features based methods to deep neural network based methods. We present three deep learning based methods that integrate deep depth features into RGBD object tracking. We also release a unified RGBD tracking benchmark for data driven RGBD tracking algorithms. Finally, we explore RGBD tracking with deep depth features and demonstrate that depth cues significantly benefit the target model learning.