Real-time large object category recognition using robust RGB-D segmentation features

Abstract

This thesis looks at the problem of large object category recognition for use in robotic systems. While many algorithms exist for object recognition, category recognition remains a challenge within robotics, particularly with the robustness and real-time constraints within robotics. Our system addresses category recognition by treating it as a segmentation problem and using the resulting segments to learn and detect large objects based on their 3D characteristics. The first part of this thesis examines how to efficiently do unsupervised segmentation of an RGB-D image in a way that is consistent across wide viewpoint and scale variance, and creating features from the resulting segments. The second part of this thesis explores how to do robust data association to keep temporally consistent segments between frames. Our higher-level module filters and matches relevant segments to a learned database of categories and outputs a pixel-accurate, labeled object mask. Our system has a run time that is nearly linear with the number of RGB-D samples and we evaluate it in a real-time robotic application.