Visual SLAM Workflow

Mapping

• Usefulness of Map in SLAM
  • Save Map-Save map to place robot on map at next boot
  • Navigation-Control the robot to find and move to the target point
  • Obstacle avoidance- Lack of feature points only, Dense map required
  • 3D restoration-visualization, reconstruction of surrounding environment
  • Interaction-Interaction between people and map (ex. AR, robot command execution)
• Necessity of Dense Map Generation

Dense Map

• Triangulation after moving the camera using a monocular camera(using current and previous frame)
• Calculate the pixel distance using the baseline using a stereo camera
• RGB-D 카메라 사용

Stereo Vision

• Epipolar line상에서 block matching을 통해 corresponding point 결정

Probabilistic Depth Sensor

• Gaussian distribution depth filter
  • When the search range is long, convex functions are generally not obtained.
  • Instead of using a single value to find true depth from multiple peaks, use probability distribution

Probabilistic Depth Sensor - Uncertainty

1. Initialize gaussian distribution to each pixel depth
2. Perform epipolar line and block matching from new frame
3. Calculate depth and uncertainty through triangulation based on geometric relationship
4. Incorporate current observations into previous estimates
5. If it converges, the calculation stops, otherwise it returns to step 2.

Reference

SLAM KR