Sensor Fusion

Sensor fusion benefits

  • One type of sensor cannot do the job alone

    • For a significant subset of applications, one positioning sensor is insufficient to meet requirements.

  • More is better or cheaper

    • When one sensor can meet requirements, a fusion of simpler sensors still might cost less

  • Robustness = availability

    • Sensor fusion is particularly attractive when requirements emphasize availability: multiple sensors have different failure modes and properly combined measurements will perform a ‘soft handover’ between sensors

Sensor fusion filter types

Each treats the probability distribution of possible solutions with a different level of compression:

  • MAP filter

    • Estimates full state probability distribution

    • Maximum aposteriori probability reported

  • Particle filter (PF)

    • Approximate probability distribution with discrete points

    • Hybrid PFs now being invented (ex: Rao-Blackwellized PF)

  • Unscented Kalman filter (UKF)

    • 3 points per eigenvector, good to 3rd moment of distribution. Susceptible to local minimum failure. 

    • >>Numerical jacobian (NJ) simplification: eigenvectors replaced with state axes

  • Gaussian-sum Iterated Extended KF (IEKF)

    • Only 2nd moment stats, but deals with multimodal distributions

  • IEKF

    • iterated to linearize better locally, but still only 2nd moment. Susceptible to local minimum failure