LaserFusion - UW CSE571 Project

Implement KinectFusion SLAM algorithm (KinectFusion ismar2011) for a 2D laser scan dataset.

Team Members

a
b
c

Timeline (dates are proposed deadlines)

  1. Nov 6: Parse the benchmark Intel-01 Oregon dataset (Robotics Data Set Repository)
  2. Nov 13: Set up the main function, implement the measurement() and TSDF updateReconstruction() functions (see KinectFusion paper)
  3. Nov 20: Implement predictSurface() - TSDF raycasting for surface prediction
  4. Nov 27: Implement estimatePose() from robot odometry model
  5. Dec 11: Finalize LaserFusion evaluation and write-up

Progress and Updates

Dec 11, 2016

  • Q.C. Modified isocontour function using marching squares to get the outline of the refined floor map.
    • Floor Map Outline

    Dec 11, 2016

  • Q.C. Added refinement function to refine TSDF by removing holes due to imcomplete fusion. Added smoothness function to remove noises and small isolated regions.
    • Refine TSDF mesh by filling holes Refine TSDF mesh by removing noises

    Dec 9, 2016

  • Q.C. added some changes to fminunc function in order to achieve faster converge speed. Final state at each frame and iterations are recorded.

    • Dec 7, 2016

    Still image of Laser Scan Fusion using 5cm grid discretization

    Dec 2, 2016

  • P.O. Added pose estimation based on ICP with a random error generated in the ground truth at each measurement. The video shows the ICP iterations and map reconstruction for 47 sampled scans. Green represents a positive SDF value and red represents a negative SDF value. The zero crossing is gray. Yellow lines are distances use to compute the cost function for the pose optimization step of ICP. Note that a scan matching failure occurs at k = 43, and subsequently onward.

    • Nov 23, 2016

  • P.O. Updated the TSDF-based reconstruction in updateReconstruction() to incorporate raycasting for a dataset with absolute ground truth at (MIT CSAIL). The video below shows the TSDF update for 100 sampled scans using absolute ground truth robot pose. Green represents a positive SDF value and red represents a negative SDF value. The zero crossing is gray.

    • Nov 22, 2016

  • P.O., Q.C., T.Y. Debugged and finished implementation of TSDF for the run (updateReconstruction)

    • Nov 8, 2016

  • Q.C. Compute TSDF for each given point

    • Nov 7, 2016

  • P.O. Parsed the Intel-01 Laser Scan and pioneer odometry dataset (see video of raw data)
  • P.O. Commited a MATLAB main routine, plotting function and the measurement() function for computing bilaterally-filtered scan data vertices and normals

    • Nov 4, 2016

  • Q.C. Chose Intel-01 dataset from Radish Repo

    • Oct 31, 2016

  • T.Y. Created the gitlab.io page and timeline