Object Localization and Recognition for Mobile Robots with Online Learning based on Mixture of Projected Gaussian
Mauro Antonello · Padua@research (University of Padova) · 2015
One of the primary capabilities required by autonomous robots is recognizing the surrounding environment with high responsiveness, often combined with object recognition and grasping tasks. Moreover robots acting in mutable scenarios are also required to be capable of learning new object models online. Along with peculiar requirements the robotics offers to the object recognition task some unique advantages, such the robot capability to move in the environment. Moreover, usually an autonomous robot can relax the recognition precision obtained at the beginning of its exploration and favour the speed at which this results are obtained. The aim of the work presented in this thesis is to explore a new object recognition method able to exploit this advantages in order to fulfil the features required by autonomous robotics. In order enhance pose estimation the proposed algorithm prioritize the keeping of the geometrical information from the objects shape and texture. Since the object models also need to be as much lightweight as possible this algorithm relies on local 6 DoF features extraction to describe the object appearance without load the final model of unnecessary information. Once the 6 DoF keypoints are obtained, the proposed method makes the use specifically designed probability distribution, namely the the Mixture of Projected Gaussian (MoPG) in order to learn their spatial distribution. A Bag of Words (BoW) technique has been introduced after the feature detection in order make feature descriptors more invariant to small appearance changes, due to light conditions or perspective distortions. The choice of using the MoPG distribution lies in one algebraic property of the Gaussian function, namely its closure over the convolution operator. In this thesis this property is exploited in order to obtain a closed form formula for calculating the cross-correlation of MoPG. The recognition algorithm makes use of the cross-correlation between MoPG in order to both identify and localize objects in the scene. The recognition and localization performances of the proposed technique was validated on two different publicly available datasets, namely the RGB-D Dataset and the BigBIRD Dataset. An analysis of both category and instance recognition results is presented and the emerged advantages or the issues of the proposed technique are discussed. The localization error (2 degrees) and the instance recognition rate (91%) resulted being aligned of the state of art thus justifying a further exploration of the proposed method. The topics presented in this thesis was further explored in some related works. In particular a collaboration with the Intelligent Systems Research Institute (Sungkyunkwan University, Republic of Corea) led an adapted version of the proposed method that has been successfully integrated in an autonomous domestic robot.