Handling the computational complexity of optimisation-based SLAM

Parent: SLAM Index Source: Forster 2017 IMU Preintegration

Complexity of nonlinear batch optimisation

• The trajectory and the map, which comprise the states, grow with time
• The larger the SLAM problem, the less feasible it is to perform the optimisation in real-time

Solutions to improve computational efficiency

• Keyframe-based methods: discard frames except for a few selected keyframes
• Run the optimisation parallelly (e.g. tracking and mapping threads)
• Fixed-lag smoothing: Use of a local map of fixed size, with marginalisation of the old states (summarise the old states into a prior term)
  • Filtering is a special case of this: window of size 1, i.e. only the latest sensor state is optimised
  • [-] commits to a linearisation point when marginalising
  • linearisation errors will eventually accumulate, can lead to drift or other inconsistencies
• Incremental smoothing
  • In between filtering and batch optimisation
  • Usage of factor graphs
  • Identifies and updates only a small subset of variables affected by a new measurement (“new constraint?")