Skip to the content ForoHackers
Log inJoin

Localization and SLAM: start here

Localization and SLAM
#363 · Ophuscado · · Report

A robot must know where it is. Localization finds its position on a known map; SLAM, simultaneous localization and mapping, builds the map while locating itself in it. Each depends on the other, and both fight drifting sensors, so the heart of the field is probability.

Key concepts

  • Probabilistic estimation: the robot holds a belief over where it might be, not a single point.
  • Filters: the Kalman filter and its variants, and particle filters, update the belief with each motion and measurement.
  • Graph-based SLAM: poses and landmarks as nodes, measurements as constraints, solved by least squares.
  • Data association: deciding whether what you see now is something you saw before.
  • Loop closure: recognising a revisited place, which corrects accumulated drift.

Vocabulary

  • Pose: position and orientation.
  • Landmark: a feature of the world with a known or estimated position.
  • Visual odometry: motion estimated from camera images.
  • Occupancy grid: a map as cells that are free, occupied or unknown.

What to study, in order

  1. Probability, Bayes' rule and the Gaussian distribution.
  2. A one-dimensional Kalman filter on a simulated moving robot.
  3. Odometry and its drift; then a particle filter that localises on a known grid map.
  4. Build a small 2D lidar SLAM, or run an open-source one, on a recorded dataset.
  5. Study pose graphs and loop closure, and measure the error against ground truth.

Related tags

#robotics #computer-vision #autonomous-systems #linear-algebra

Log in or join to reply.