(Georgia Institute of Technology, 2009-07-06)
Wu, Jianxin
Knowing the semantic category of a robot's current position not only facilitates the robot's navigation, but also greatly improves its ability to serve human needs and to interpret the scene. Visual Place Categorization (VPC) is addressed in this dissertation, which refers to the problem of predicting the semantic category of a place using visual information collected from an autonomous robot platform.
Census Transform (CT) histogram and Histogram Intersection Kernel (HIK) based visual codebooks are proposed to represent an image. CT histogram encodes the stable spatial structure of an image that reflects the functionality of a location. It is suitable for categorizing places and has shown better performance than commonly used descriptors such as SIFT or Gist in the VPC task.
HIK has been shown to work better than the Euclidean distance in classifying histograms. We extend it in an unsupervised manner to generate visual codebooks for the CT histogram descriptor. HIK codebooks help CT histogram to deal with the huge variations in VPC and improve system accuracy. A computational method is also proposed to generate HIK codebooks in an efficient way.
The first significant VPC dataset in home environments is collected and is made publicly available, which is also used to evaluate the VPC system based on the proposed techniques. The VPC system achieves promising results for this challenging problem, especially for important categories such as bedroom, bathroom, and kitchen. The proposed techniques achieved higher accuracies than competing descriptors and visual codebook generation methods.
(Georgia Institute of Technology, 2008-07-01)
Sun, Jie
A fundamental requirement of any autonomous robot system is the ability to predict the affordances of its environment, which define how the robot can interact with various objects. In this dissertation, we demonstrate that the conventional direct perception approach can indeed be applied to the task of training robots to predict affordances, but it does not consider that objects can be grouped into categories such that objects of the same category have similar affordances. Although the connection between object categorization and the ability to make predictions of attributes has been extensively studied in cognitive science research, it has not been systematically applied to robotics in learning to predict a number of affordances from recognizing object categories.
We develop a computational framework of learning and predicting affordances where a robot explicitly learns the categories of objects present in its environment in a partially supervised manner, and then conducts experiments to interact with the objects to both refine its model of categories and the category-affordance relationships. In comparison to the direct perception approach, we demonstrate that categories make the affordance learning problem scalable, in that they make more effective use of scarce training data and support efficient incremental learning of new affordance concepts. Another key aspect of our approach is to leverage the ability of a robot to perform experiments on its environment and thus gather information independent of a human trainer. We develop the theoretical underpinnings of category-based affordance learning and validate our theory on experiments with physically-situated robots. Finally, we refocus the object categorization problem of computer vision back to the theme of autonomous agents interacting with a physical world consisting of categories of objects. This enables us to reinterpret and extend the Gluck-Corter category utility function for the task of learning categorizations for affordance prediction.