[Paper Review] Visual SLAM algorithms: a survey from 2010 to 2016

논문 : Visual SLAM algorithms: a survey from 2010 to 2016

URL : ipsjcva.springeropen.com/articles/10.1186/s41074-017-0027-2

Visual SLAM algorithms: a survey from 2010 to 2016

저자 : Takafumi Taketomi, Hideaki Uchiyama & Sei Ikeda

Pulish : IPSJ Transactions on Computer Vision and Applications

 

[Introduction]

• SLAM : Simultaneous Localization and Mapping
  • obtaining the 3D structure of an unknown environment
  • obtaining sensor motion in the environment
• 처음에는 여러가지 센서를 합쳐서(fusion) 사용하고자 함 → 너무 복잡하고 어렵다.
• 주된 sensor들을 선정하고 그 센서의 데이터를 기반으로 SLAM을 진행
  • Lidar -Based SLAM : Lidar 기반
  • Visual -Based SLAM : 상업용 카메라 기반
• Lidar - SLAM은 Lidar Sensor에 굉장히 의존적, Lidar 센서의 값이 너무나 비쌈
• 따라서 visual information 만으로 SLAM을 시도해보자 → visual slam (vslam)

[Elements of vSLAM]

• Basic
  • Initialization
    • define a certain coordinate system → global coordinate (initial map)
  • Tracking
    • 2D-3D correspondences between map and image(input)
  • Mapping
    • continuously estimate camera poses (PnP problem,Perspective-n-Point)
    • assume that we know intrinsic camera parameters.
• Additional Modules
  • Relocalization
    • when the tracking is failed(becuase of fast camera motion or some disturbances)
  • Global map optimazation
    • map redefined by considering the consistency
    • pose-graph optimazation
      • optimizing camera poses
    • Bundle adjustment(BA)
      • minimize the reprojection error of the map
  • Loop closing
    • acquire the reference information
    • searched by matching a current image with previously acquired images

[Related Technologies]

• Visual odometry
  • estimate the sequential changes of sensor positions
  • vSLAM = VO + global map optimizations
• Structure from motion
  • estimate camera motion and 3D structure of the environment
  • real time SfM = SLAM

[Feature based VS Direct]

[Feature-based]

 

[Direct]

 

[Open Problem]

• pure rotation
• Map initialization
• Estimating intrinsic camera prarmeters
• Rolling shutter distortion
• Scale ambiguity

[Appendix]

• BA(bundle adjustment)
• EKF & KF
• pose graph optimization
• multi -base line stereo
• photometric consistency