Chen 2018 SLAM-based dense surface reconstruction in MIS with AR
Authors Chen et al
Abstract
- Intra-operative dense surface reconstruction framework to provide geometry information from only monocular videos
- The proposed framework works well with rapid camera movements, however is not suitable for large deformations
- Only tweaks ORBSLAM to adjust between point density and computational performance
Contents/Chapters
Problems in medical AR:
- tissue surface illumination
- tissue deformation
- rapid movements of the medical tool e.g. endoscope (s. also kidnapped robot problem for relocalisation, tracking mus therefore be robust)
- field of view often very small
“A typical human uses 14 visual cues to perceive depth, only 3/14 are binocular vision related.” s. also: Monocular depth perception in humans
Traditional feature-tracking for AR in MIS:
- SIFT
- SURF
- Optical flow tracking
- Other approaches for soft tissues
However, these are designed for 2D tracking
--> ORB descriptor
- binary feature descriptor
- faster than SURF and SIFT with better accuracy
- invariant to rotation, illumination and scale
ORB-SLAM uses Bag of Words algo. for relocalisation
Task of the SLAM algo in this framework:
- Recovers camera POSE
- Generates a sparse point cloud, based on which 3D geometric information is generated
Inialisation is problematic for monocular SLAM because the depth isn’t recoverable from a single image --> automatic approach in ORBSLAM:
- for planar scenes: calculate homography
- for non-planar scenes: calculate a fundamental matrix dynamically
Takeaway
Not suitable for large deformations