Recovering Sample Diversity in Rao-Blackwellized Particle Filters for Simultaneous Localization and Mapping

摘要

This thesis considers possible solutions to sample impoverishment, a well-known failure mode of the Rao-Blackwellized particle filter (RBPF) in simultaneous localization and mapping (SLAM) situations that arises when precise feature measurements yield a limited perceptual distribution relative to a motion-based proposal distribution. One set of solutions propagates particles according to a more advanced proposal distribution that includes measurement information. Other methods recover lost sample diversity by resampling particles according to a continuous distribution formed by regularization kernels. Several advanced proposals and kernel shaping regularization methods are considered based on the RBPF and tested in a Monte Carlo simulation involving an agent traveling in an environment and observing uncertain landmarks. RIVIS error of range-bearing feature measurements was reduced to evaluate performance during proposal-perceptual distribution mismatch. A severe loss in accuracy due to sample impoverishment is seen in the standard RBPF at a measurement range RMS error of 0.001 m in a 10 m x 10 m environment. Results reveal a robust and accurate solution to sample impoverishment in an RBPF with an added fixed-variance regularization algorithm. This algorithm produced an average 0.05 m improvement in agent pose CEP over standard FastSLAM 1.0 and a 0.1 m improvement over an RBPF that includes feature observations in formulation of a proposal distribution. This algorithm is then evaluated in an actual SLAM environment with data from a Swiss Ranger LIDAR measurement device and compared alongside an extended Kalman filter (EKF) based SLAM algorithm. Pose error is immediately recovered in cases of a 1.4 m error in initial agent uncertainty using the improved FastSLAM algorithm, and it continues to maintain an average 0.75 m improvement over an EKF in pose CEP through the scenario.