Particle filters

Parent: Filter localisation methods

Source: Wikipedia Lokalisierung Particle filter / Monte Carlo localisation / sequential Monte Carlo methods

• allow solution of all three localisation problems
• POSE represented by a particle cloud
• Each particle : possible POSE
• The filter checks the plausibility of each particle
  • Increases and decreases the probabilities of each particle accordingly
  • When a lower probability threshold is exceeded, the particle is not considered any longer

Source: Scaradozzi 2018 SLAM application in surgery Particle filters (sequential Monte Carlo)

• Posterior representation: particles (set of random state samples)

• Suitable for implementation with any probabilistic robot model with Markov chain formulation

• Easy to implement

  • no need to linearise
  • closed-form solutions of the conditional probability are irrelevant (as in KF)

• Poor performance in higher dimensional spaces

• Rao-Blackwellized particle filters

  • more efficient solutions, however, prone to significant estimation inconsistencies
  • therefore, as estimation consistency is quite important, led to the adoption of different sampling strategies

• FastSLAM

  • integrates particle filters and EKF
    • particles filters for estimating robot path
    • for each particle, EKF for estimating feature locations
  • mapping problem split
    • one mapping problem for each feature in the map
  • particle approximation doesn’t converge uniformly in time