Machine Learning and Robot Perception PDF
Preview Machine Learning and Robot Perception
BrunoApolloni,AshishGhosh,FerdaAlpaslan,LakhmiC.Jain, SrikantaPatnaik(Eds.) MachineLearningandRobotPerception StudiesinComputational Intelligence,Volume7 Editor-in-chief Prof.JanuszKacprzyk SystemsResearchInstitute PolishAcademyofSciences ul.Newelska6 01-447Warsaw Poland E-mail:[email protected] Furthervolumesofthisseries canbefoundonourhomepage: springeronline.com Vol.1.TetsuyaHoya ArtificialMindSystem–KernelMemory Approach,2005 ISBN3-540-26072-2 Vol.2.SamanK.Halgamuge,LipoWang (Eds.) ComputationalIntelligenceforModelling andPrediction,2005 ISBN3-540-26071-4 Vol.3.Boz˙enaKostek Perception-BasedDataProcessingin Acoustics,2005 ISBN3-540-25729-2 Vol.4.SamanHalgamuge,LipoWang(Eds.) ClassificationandClusteringforKnowledge Discovery,2005 ISBN3-540-26073-0 Vol.5.DaRuan,GuoqingChen,EtienneE. Kerre,GeertWets(Eds.) IntelligentDataMining,2005 ISBN3-540-26256-3 Vol.6.TsauYoungLin,SetsuoOhsuga, Churn-JungLiau,XiaohuaHu,Shusaku Tsumoto(Eds.) FoundationsofDataMiningandKnowledge Discovery,2005 ISBN3-540-26257-1 Vol.7.BrunoApolloni,AshishGhosh,Ferda Alpaslan,LakhmiC.Jain,SrikantaPatnaik (Eds.) MachineLearningandRobotPerception, 2005 ISBN3-540-26549-X Bruno Apolloni Ashish Ghosh Ferda Alpaslan Lakhmi C. Jain Srikanta Patnaik (Eds.) Machine Learning and Robot Perception ABC ProfessorBrunoApolloni ProfessorLakhmiC.Jain DepartmentofInformationScience SchoolofElectrical&InfoEngineering UniversityofMilan UniversityofSouthAustralia ViaComelico39/41 Knowledge-BasedIntelligentEngineering 20135Milan MawsonLakesCampus Italy 5095Adelaide,SA E-mail:[email protected] Australia E-mail:[email protected] ProfessorAshishGhosh ProfessorSrikantaPatnaik MachineIntelligenceUnit DepartmentofInformation IndianStatisticalInstitute andCommunicationTechnology 203BarrackporeTrunkRoad F.M.University Kolkata700108 VyasaVihar India Balasore-756019 E-mail:[email protected] Orissa,India E-mail:[email protected] ProfessorFerdaAlpaslan FacultyofEngineering DepartmentofComputerEngineering MiddleEastTechnicalUniversity-METU 06531Ankara Turkey E-mail:[email protected] LibraryofCongressControlNumber:2005929885 ISSNprintedition:1860-949X ISSNelectronicedition:1860-9503 ISBN-10 3-540-26549-XSpringerBerlinHeidelbergNewYork ISBN-13 978-3-540-26549-8SpringerBerlinHeidelbergNewYork Thisworkissubjecttocopyright.Allrightsarereserved,whetherthewholeorpartofthematerialis concerned,specificallytherightsoftranslation,reprinting,reuseofillustrations,recitation,broadcasting, reproductiononmicrofilmorinanyotherway,andstorageindatabanks.Duplicationofthispublication orpartsthereofispermittedonlyundertheprovisionsoftheGermanCopyrightLawofSeptember9, 1965,initscurrentversion,andpermissionforusemustalwaysbeobtainedfromSpringer.Violationsare liableforprosecutionundertheGermanCopyrightLaw. SpringerisapartofSpringerScience+BusinessMedia springeronline.com (cid:1)c Springer-VerlagBerlinHeidelberg2005 PrintedinTheNetherlands Theuseofgeneraldescriptivenames,registerednames,trademarks,etc.inthispublicationdoesnotimply, evenintheabsenceofaspecificstatement,thatsuchnamesareexemptfromtherelevantprotectivelaws andregulationsandthereforefreeforgeneraluse. Typesetting:bytheauthorsandTechBooksusingaSpringerLATEXmacropackage Printedonacid-freepaper SPIN:11504634 89/TechBooks 543210 Preface This book presents some of the most recent research results in the area of machine learning and robot perception. The book contains eight chapters. The first chapter describes a general-purpose deformable model based object detection system in which evolutionary algorithms are used for both object search and object learning. Although the proposed system can handle 3D objects, some particularizations have been made to reduce computational time for real applications. The system is tested using real indoor and outdoor images. Field experiments have proven the robustness of the system for illumination conditions and perspective deformation of objects. The natural application environments of the system are predicted to be useful for big public and industrial buildings (factories, stores), and outdoor environments with well-defined landmarks such as streets and roads. Fabrication of space-variant sensor and implementation of vision algorithms on space-variant images is a challenging issue as the spatial neighbourhood connectivity is complex. The lack of shape invariance under translation also complicates image understanding. The retino-cortical mapping models as well as the state-of-the-art of the space-variant sensors are reviewed to provide a better understanding of foveated vision systems in Chapter 2. It is argued that almost all the low level vision problems (i.e., shape from shading, optical flow, stereo disparity, corner detection, surface interpolation etc.) in the deterministic framework can be addressed using the techniques discussed in this chapter. The vision system must be able to determine where to point its high-resolution fovea. A proper mechanism is expected to enhance image understanding by strategically directing fovea to points which are most likely to yield important information. In Chapter 3 a discrete wavelet based model identification method has been proposed in order to solve the online learning problem. The vi Preface method minimizes the least square residual parameter estimation in noisy environments. It offers significant advantages over the classical least square estimation methods as it does not need prior statistical knowledge of measurement of noises. This claim is supported by the experimental results on estimating the mass and length of a nonholonomic cart having a wide range of applications in complex and dynamic environments. Chapter 4 proposes a reinforcement learning algorithm which allows a mobile robot to learn simple skills. The neural network architecture works with continuous input and output spaces, has a good resistance to forget previously learned actions and learns quickly. Nodes of the input layer are allocated dynamically. The proposed reinforcement learning algorithm has been tested on an autonomous mobile robot in order to learn simple skills showing good results. Finally the learnt simple skills are combined to successfully perform more complex skills called visual approaching and go to goal avoiding obstacles. In Chapter 5 the authors present a simple but efficient approach to object tracking combining active contour framework and the optical- flow based motion estimation. Both curve evolution and polygon evolution models are utilized to carry out the tracking. No prior shape model assumptions on targets are made. They also did not make any assumption like static camera as is widely employed by other object tracking methods. A motion detection step can also be added to this framework for detecting the presence of multiple moving targets in the scene. Chapter 6 presents the state-of-the-art for constructing geometrically and photometrically correct 3D models of real-world objects using range and intensity images. Various surface properties that cause difficulties in range data acquisition include specular surfaces, highly absorptive surfaces, translucent surfaces and transparent surfaces. A recently developed new range imaging method takes into account of the effects of mutual reflections, thus providing a way to construct accurate 3D models. The demand for constructing 3D models of various objects has been steadily growing and we can naturally predict that it will continue to grow in the future. Preface vii Systems that visually track human motion fall into three basic categories: analysis-synthesis, recursive systems, and statistical methods including particle filtering and Bayesian networks. Each of these methods has its uses. In Chapter 7 the authors describe a computer vision system called DYNA that employs a three- dimensional, physics-based model of the human body and a completely recursive architecture with no bottom-up processes. The system is complex but it illustrates how careful modeling can improve robustness and open the door to very subtle analysis of human motion. Not all interface systems require this level of subtlety, but the key elements of the DYNAarchitecture can be tuned to the application. Every level of processing in the DYNAframework takes advantage of the constraints implied by the embodiment of the observed human. Higher level processes take advantage of these constraints explicitly while lower level processes gain the advantage of the distilled body knowledge in the form of predicted probability densities. Chapter 8 advocates the concept of user modelling which involves dialogue strategies. The proposed method allows dialogue strategies to be determined by maximizing mutual expectations of the pay-off matrix. The authors validated the proposed method using iterative prisoner's dilemma problem that is usually used for modelling social relationships based on reciprocal altruism. Their results suggest that in principle the proposed dialogue strategy should be implemented to achieve maximum mutual expectation and uncertainty reduction regarding pay-offs for others. We are grateful to the authors and the reviewers for their valuable contributions. We appreciate the assistance of Feng-Hsing Wang during the evolution phase of this book. June 2005 Bruno Apolloni Ashish Ghosh Ferda Alpaslan Lakhmi C. Jain Srikanta Patnaik Table of Contents 1 Learning Visual Landmarks for Mobile Robot Topological Navigation 1 Mario Mata, Jose Maria Armingol, and Arturo de la Escalera 2 Foveated Vision Sensor and Image Processing – A Review ............... 57 Mohammed Yeasin andRajeev Sharma 3 On-line Model Learning for Mobile Manipulations ............................ 99 Yu Sun, Ning Xi, and Jindong Tan 4 Continuous Reinforcement Learning Algorithm for Skills Learning in an Autonomous Mobile Robot ............................................................... 137 Mª Jesús López Boada, Ramón Barber, Verónica Egido, and Miguel Ángel Salichs 5 Efficient Incorporation of Optical Flow into Visual Motion Estimation in Tracking ......................................................................................... 167 Gozde Unal, Anthony Yezzi, and Hamid Krim 6 3-D Modeling of Real-World Objects Using Range and Intensity Images .......................................................................... 203 Johnny Park and Guilherme N. DeSouza 7 Perception for Human Motion Understanding ................................... 265 Christopher R. Wren 8 Cognitive User Modeling Computed by a Proposed Dialogue Strategy Based on an Inductive Game Theory ................................................. 325 Hirotaka Asai, Takamasa Koshizen, Masataka Watanabe, Hiroshi Tsujin and Kazuyuki Aihara 1 Learning Visual Landmarks for Mobile Robot Topological Navigation Mario Mata1, Jose Maria Armingol2, Arturo de la Escalera2 1. Computer Architecture and Automation Department, Universidad Europea de Madrid, 28670 Villaviciosa de Odon, Madrid, Spain. [email protected] 2. Systems Engineering and Automation Department. Universidad Carlos III de Madrid, 28911 Leganés, Madrid, Spain. {armingol,escalera}@ing.uc3m.es 1.1 Introduction Relevant progress has been done, within the Robotics field, in mechanical systems, actuators, control and planning. This fact, allows a wide applica- tion of industrial robots, where manipulator arms, Cartesian robots, etc., widely outcomes human capacity. However, the achievement of a robust and reliable autonomous mobile robot, with ability to evolve and accom- plish general tasks in unconstrained environments, is still far from accom- plishment. This is due, mainly, because autonomous mobile robots suffer the limitations of nowadays perception systems. A robot has to perceive its environment in order to interact (move, find and manipulate objects, etc.) with it. Perception allows making an internal representation (model) of the environment, which has to be used for moving, avoiding collision, finding its position and its way to the target, and finding objects to manipulate them. Without a sufficient environment perception, the robot simply can’t make any secure displacement or interaction, even with extremely efficient motion or planning systems. The more unstructured an environment is, the most dependent the robot is on its sensorial system. The success of indus- trial robotics relies on rigidly controlled and planned environments, and a total control over robot’s position in every moment. But as the environ- ment structure degree decreases, robot capacity gets limited. Some kind of model environment has to be used to incorporate percep- tions and taking control decisions. Historically, most mobile robots are based on a geometrical environment representation for navigation tasks. This facilitates path planning and reduces dependency on sensorial system, but forces to continuously monitor robot’s exact position, and needs precise M.Mataetal.:Learning Visual Landmarks for Mobile Robot Topological Navigation,Studies inComputationalIntelligence(SCI)7,1–55(2005) www.springerlink.com (cid:1)c Springer-VerlagBerlinHeidelberg2005
The list of books you might like
