Date of Award

5-2013

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Legacy Department

Computer Engineering

Advisor

Birchfield, Stan

Committee Member

Hoover , Adam

Committee Member

Woodard , Damon

Abstract

The goal of this research is to enable a robotic system to manipulate clothing and other highly non-rigid objects using an RGBD sensor. The focus of this thesis is to define and test various algorithms / models that are used to solve parts of the laundry process (i.e. handling, classifying, sorting, unfolding, and folding). First, a system is presented for automatically extracting and classifying items in a pile of laundry. Using only visual sensors, the robot identifies and extracts items
sequentially from the pile. When an item is removed and isolated, a model is captured of the shape and appearance of the object, which is then compared against a dataset of known items. The contributions of this part of the laundry process are a novel method for extracting articles of clothing from a pile of laundry, a novel method of classifying clothing using interactive perception, and a multi-layer approach termed
L-M-H, more specifically L-C-S-H for clothing classification. This thesis describes two different approaches to classify clothing into categories. The first approach relies upon silhouettes, edges, and other low-level image measurements of the articles of clothing. Experiments from the first approach demonstrate the ability of the system to efficiently classify and label into one of six categories (pants, shorts, short-sleeve shirt, long-sleeve shirt, socks, or underwear). These results show that, on average, classification rates using robot interaction are 59% higher than those that do not use interaction. The second approach relies upon color, texture, shape, and edge information from 2D and 3D data within a local and global perspective. The multi-layer approach compartmentalizes the problem into a high (H) layer, multiple mid-level
(characteristics(C), selection masks(S)) layers, and a low (L) layer. This approach produces 'local' solutions to solve the global classification problem. Experiments demonstrate the ability of the system to efficiently classify each article of clothing into one of seven categories (pants, shorts, shirts, socks, dresses, cloths, or jackets). The results presented in this paper show that, on average, the classification rates
improve by +27.47% for three categories, +17.90% for four categories, and +10.35% for seven categories over the baseline system, using support vector machines. Second, an algorithm is presented for automatically unfolding a piece of clothing. A piece of cloth is pulled in different directions at various points of the cloth in order to flatten the cloth. The features of the cloth are extracted and calculated to
determine a valid location and orientation in which to interact with it. The features include the peak region, corner locations, and continuity / discontinuity of the cloth. In this thesis, a two-stage algorithm is presented, introducing a novel solution to the unfolding / flattening problem using interactive perception. Simulations using 3D
simulation software, and experiments with robot hardware demonstrate the ability of the algorithm to flatten pieces of laundry using different starting configurations. These results show that, at most, the algorithm flattens out a piece of cloth from 11.1% to 95.6% of the canonical configuration. Third, an energy minimization algorithm is presented that is designed to estimate the configuration of a deformable object. This approach utilizes an RGBD image to calculate feature correspondence (using SURF features), depth values, and boundary locations. Input from a Kinect sensor is used to segment the deformable surface from the background using an alpha-beta swap algorithm. Using this segmentation, the system creates an initial mesh model without prior information of the surface geometry, and it reinitializes the configuration of the mesh model after a loss of input data. This approach is able to handle in-plane rotation, out-of-plane rotation, and varying changes in translation and scale. Results display the proposed algorithm over a dataset consisting of seven shirts, two pairs of shorts, two posters,
and a pair of pants. The current approach is compared using a simulated shirt model in order to calculate the mean square error of the distance from the vertices on the mesh model to the ground truth, provided by the simulation model.

Share

COinS