Research On Key Technologies Of 3D Map Building Based On Graph Optimizition

Recently,major economies of the world have considered the intelligent robot as the direction of significant strategic development of their countries.In the manufacturing field,industrial robots can take over those jobs that are most of labor-intensive,simple repetition and high risk.At the same time,robot technology also gradually steps into various fields including service industry from the filed of industrial manufacturing.Service robots can content to consumers' diverse demands,which involve in sentiment,accompany interaction and life service and so on.However,most of service robots need to equip with the function of autonomous movement.It is the core problem of mobile robots that is how to use the robot itself sensor and environment where they carries out autonomous navigation.In order to realize autonomous navigation,it is important to require the robot to locate its position and routes the route according to the map of the environment in which it is located.This comes down to an important problem of robot field,namely,simultaneous localization and mapping,SLAM.SLAM is considered to be one of the key issues to achieve robot autonomy.At present,although the main solution which can deal with SLAM problem is based on Extended Kalman Filter(EKF)or Particle Filter(PF)algorithm and so on,there are some problems such as linearization and cumulative error,which can not be applied to large scale building.In recent years,SLAM has been in a further study.SLAM method,which is based on the graph optimization theory,can effectively solve the cumulative error when build the graph under a large scale environment.However,there are some problems in terms of existing algorithm,which include a high requirement to environment,poor robustness and inaccuracy.Based on the present situation,this thesis does deep analysis on them including motion estimation,loop detection and graph optimization in SLAM system.What is more,it also designs SLAM system with Kinect as the sensor.The system has made a further improvement in many aspects and enhances the accuracy and robustness of whole algorithm.In the portion of order data correlation,this thesis,taking into account the real surrounding environment,adopts the method of combining the feature matching with ICP algorithm to solving motion estimation.The method can run effectively even when the surrounding environment is too single and the environmental features are insufficient.Besides,this thesis makes a further improvement in view of classical ICP algorithm and advances accuracy and computational efficiency of algorithm.At last,it puts to use sub-optimization by BA(bundle adjustment),which reduces the cumulative error in local scope and obtains more accurate and reliable motion estimation.In the portion of loop detection,this thesis conducts loop detection on account of combining similar appearance with constraints of time and space.It also uses this means which is on the strength of bag of words to determine candidate loop nodes.And then,with regard to candidate loop nodes,there exists perceived ambiguities in few cases.By adding spatial positions and time-interval constraints which mainly contains candidate points that similarity is greater than threshold to increase accuracy of loop detection.Hardly did the bitmap construct when order data correlated and loop detected,the constraints of edge figure are inconsistent,so this thesis refers to correlated nodes which are correspondent error of motion evaluation,establishes information matrix for every edge.In that case,it can obtain more accurate optimization.When we work out the optimization problem of graph,we should take into full account of the features of sparse matrix and take advantage of high-efficient method to solve this non-linear optimization problem.