Simutaneous Localization Andmapping Method Based On Multi Camera System

Localization inside unknown environment is one of the most important techniques in the research of robotics.Simultaneous Localization and Mapping(SLAM),recognized as the frontier of positioning techniques,has received extensive attentions among scholars and entrepreneurs.in the past 30 years.Earlier SLAM used laser or odometer as input sensors.In recent 10 years,Visual SLAM has become a research hotspot because it is lowcost,easy to mount and can provide plentiful information,and because the boosted computational ability brought from advanced CPU/GPU.Visual SLAM now is widely adopted in field as autonomous investigation,navigations and augment reality.According to the cameras types,visual SLAM is categorized as Monocular,Stereo and RGB-D SLAM.However with the lack of scale in Monocular,the limited field of view and perception distance of Stereo,and the illumination sensitivity of RGB-D,SLAM with such cameras inevitably suffered from the deficiencies in accuracy and stability.In this paper,a visual SLAM method based on multi monocular camera(Multi Camera SLAM System,MCS)is proposed in order to enlarge the FOV of monocular camera and retrieve the metric scale of map.To achieve high accuracy,this algorithm merges the observations from the array of cameras,uses sparse visual features to generate 3D map points,retrieves metric scale in virtue of the extrinsic calibrations,optimizes local map in a graph structure and estimates system status inside the local map.Besides,the relocalizaton and loop closure are also implemented in MCS to make it as a complete SLAM system.The main research in this paper include:1.Proposed a perception model which can be extend from one to many camera/cameras.In MCS,the common view between each camera is not complustory thus the field of view is widened by combing each single camera.2.Revised the existed SLAM system to adapt multi camera model.First,the map scale is retrieved online by extrinsic during the system initialization procedure.Second,the Keyframe based graph optimization pipeline is adjusted to take account of the observation from all cameras.The Multi-Keyframe is proposed.Owing to each camera in MCS is tightly coupled,the reprojection error and Jacobian matrix is redefined to merge all observation.3.Developed a complete multi camera SLAM system(MCS).MCS is divided into front-end odometry and back-end mapping.The subfunctionalities included in odometry are Feature Matching,Mono Initialization,System Initialization,Tracking and Relocalizaton.And those included in mapping are Map Update,Nonlinear Optimization and Loop Closing.The tracking,local mapping and loop closing modules are concurrently performed in three different threads to ensure the accuracy and real-time performance.