بازشناسي مكان ربات با الگوريتم CRF- Matching نيمه نظارتي

چكيده انگليسي :

Place Recognition using semi supervised CRF Matching Mohammad ali Aref ma aref@ec iut ac ir Date of Submission 3 12 2012 Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M Sc Language FarsiSupervisors Maziar Palhang palhang@cc iut ac ir Mohammad Ali Montazeri montazeri@cc iut ac irAbstract Finding its way in the environment in which a robot operates is a basic problem that must be solved fortrue autonomy There are two main aspects to this problem known as Simultaneous Localization andMapping SLAM 1 the continuous problem of estimating the location of elements of interest for therobot and 2 the discrete problem of finding correspondences between measurements of the sensor that therobot uses to perceive its environment and the elements already in the map Determining a correspondencebetween the observed data and quantities to be estimated is known as the data association problem It is anessential step for the estimation process and it is one of the most difficult problems in simultaneouslocalization and mapping which has been less of concern in recent years This research proposes newstrategies for data association problem and an accurate method for robot place recognition using semisupervised learning with deficiency and problems of the current state of the art data association methods This problem arises in two situations continuous data association or feature tracking and loop closure orthe place recognition problem Continuous data association is considered as a labeling problem which couldbe solved using probabilistically modeling techniques CRF Matching as a continues data association canbe modeled and solved using Conditional random fields models The disadvantage using this method is inthe fully supervised learning of the model which requires all training data to be labeled in advance This ismain reason why we have a semi supervised learning method for training CRF Matching model Modelparameters in the proposed approach are optimized using particle swarm optimization with regard to set oflabeled and unlabeled data Beside no need for having fully labeled set of data semi supervised learningtakes advantage of much accuracy rather than the supervised methods Second view in data associationproblem is specified as loop closing which is known as the place recognition Detecting when a mobilerobot is in a place already visited is fundamental to the SLAM context to recover from failures and to selectpolicies of exploration in active SLAM Since cameras are easily available and provide rich scene detail place recognition using visual information has been a problem of great interest in robotics for some time Most successful methods consider appearance or geometric information or a combination of both In thesecond part of this research an accurate and efficient method for place recognition would be demonstratedappearance information and extracted geometrical features of the current scene The main problem incurrent place recognition methods is the low rate of recall meaning that only a limited number of alreadyobserved places would be recognized First phase of this technique specify some of the observed places asthe loop closing candidates using bag of visual word algorithm and in the next phase place recognitionverification would be analyzed using semi supervised CRF Matching Number of verified places using theproposed method doubles the same number of any previous work Keywords Data association Place Recognition Bag of visual word Conditional random fields Semi supervised CRF Matching

عارف، محمدعلي

بازشناسي مكان ربات با الگوريتم CRF- Matching نيمه نظارتي

ارتباط دهي داده ها , بسته اي از واژگان , ميدان هاي تصادفي مشروط , يادگيري نيمه نظارتي