Articulated Object Detection

We develop an object detection algorithm for difficult-to-detect objects with consumerlevel range cameras. The increasing popularity of such devices made the process of obtaining depth data extremely easy. The availability of depth data motivates the need for algorithms that will be able to process it in meaningful ways, so that it then can be used in applications like robot navigation or object retrieval. However the quality of the point clouds obtained using these cameras often leaves much to be desired, which makes the processing stage very challenging. This project aims at developing a method that is able to process the data in a way that will conceptually alleviate the data issues. In particular we consider the problem of detecting articulated objects like laptops in the office-type, indoor environments. This report has two main contributions. The first is a graph-based scene representation that approximates the captured environment using simple, well described geometric primitives (nodes) and describes the geometric relationships between them (edges). A series of experiments to determine set of stable features to detect these relationships has been performed. A graph created in this way serves as a global scene descriptor,and the object detection can be posed as a sub-graph matching problem. Ways of determining node-to-node correspondence between scene and object of interest graph representations were investigated. Simple method to find this node to node correspondence has been developed, exploiting the fact that most of the scenes can be described by small and simple graphs. The problem has been also posed to fit into existing, advanced framework for graph matching. The second contribution is the way of to reason about the missing data. This project focused on the laptop detection, that can suffer from dramatic data loss, due to the fact that laptop screen tend to reflect structured light pattern. The method described is able to cope with such missing data in a robust way that is also able to reason about the data loss due to other issues, i.e. partial occlusion. The system for laptop detection in cluttered office environment has been developed as a proof of concept for the presented approach. It is argued that the graph based approach is a reliable way of performing object detection, that however suffers greatly from the poor data quality, by performing the tests using synthetically generated data as well. The performance of the detection system is presented both for synthetically generated data and the captured real-world data set. The limitations and possibilities for improvements are discussed.