Paper

Street View Cross-Source Point Cloud Matching and Registration

Motivation

Point clouds of reference and target can be generated by some technique (SfM, LiDAR sensor, Kinect range camera…). Cases in which reference and target point clouds generated by different technologies are rarely discussed. However, in real world, the challenges of corss-source point clouds registration are commonly encountered.

Difficulty

  • One popular point cloud registration: ICP (Iterative Closet Point)
    • based on same type of sensor
  • Some challenges for registration of cross-source point cloud
    • different dimension of scale
      • SfM: no real scale information
      • LiDAR: can be calibrated through earth coordinates
    • different density of point cloud
      • density(LiDAR) > density(SfM)
    • different noise distribution
    • point cloud overlapped problem

Main Algorithm

  • Two stages: coarse matching + fine registration

  • coarse matching:
    • apply multi-scale scanning on LiDAR
    • apply ESF-64 on LiDAR and SfM point clouds
    • select k-best similar LiDAR cloud point cloud regions
      in each scale for reference (use cosine similarity)

  • fine registration:
    • alignment for vetical direction
      in a SfM and LiDAR cloud point
      (since objects are vertical in the urban environmnet)
      • reduce the number of initial poses
        to save computing cost and avoid obvious incorrect initial poses,
        thereby improving accuracy
    • apply ICP with multiple initial poses
      (since ICP is easy to stick in local optimum)
    • optimal transformation matrix (RT matrix) is found
    • residual of registration: ICP-score
      • lower ICP-score is the higher ranked candidate
    • adjustment for ICP score
      (since ICP-score favors small scale candicate point clouds)
      • a penalty factor is performed in ajdusted ICP-score
    • sort ajusted ICP-score in ascending order