Active Object Classification from 3D Range Data with Mobile Robots

Abstract

This thesis addresses the problem of how to improve the acquisition of 3D range data with a mobile robot for the task of object classification. Establishing the identities of objects in unknown environments is fundamental for robotic systems and helps enable many abilities such as grasping, manipulation, or semantic mapping. Objects are recognised by data obtained from sensor observations, however, data is highly dependent on viewpoint; the variation in position and orientation of the sensor relative to an object can result in large variation in the perception quality. Additionally, cluttered environments present a further challenge because key data may be missing. These issues are not always solved by traditional passive systems where data are collected from a fixed navigation process then fed into a perception pipeline. This thesis considers an active approach to data collection by deciding where is most appropriate to make observations for the perception task. The core contributions of this thesis are a non-myopic planning strategy to collect data efficiently under resource constraints, and supporting viewpoint prediction and evaluation methods for object classification. Our approach to planning uses Monte Carlo methods coupled with a classifier based on non-parametric Bayesian regression. We present a novel anytime and non-myopic planning algorithm, Monte Carlo active perception, that extends Monte Carlo tree search to partially observable environments and the active perception problem. This is combined with a particle-based estimation process and a learned observation likelihood model that uses Gaussian process regression. To support planning, we present 3D point cloud prediction algorithms and utility functions that measure the quality of viewpoints by their discriminatory ability and effectiveness under occlusion. The utility of viewpoints is quantified by information-theoretic metrics, such as mutual information, and an alternative utility function that exploits learned data is developed for special cases. The algorithms in this thesis are demonstrated in a variety of scenarios. We extensively test our online planning and classification methods in simulation as well as with indoor and outdoor datasets. Furthermore, we perform hardware experiments with different mobile platforms equipped with different types of sensors. Most significantly, our hardware experiments with an outdoor robot are to our knowledge the first demonstrations of online active perception in a real outdoor environment. Active perception has broad significance in many applications. This thesis emphasises the advantages of an active approach to object classification and presents its assimilation with a wide range of robotic systems, sensors, and perception algorithms. By demonstration of performance enhancements and diversity, our hope is that the concept of considering perception and planning in an integrated manner will be of benefit in improving current systems that rely on passive data collection.