Noise-Adaptive Soft Sensors in Mobile Robots through Particle Filtering

AbstractThe assumption of constant noise and requirement of prior calibration can limit the effectiveness of sensor fusion in mobile robot applications. This paper uses techniques from particle filtering parameter estimation research to construct a novel noise-adaptive particle filter which can provide on-line estimation of both robot state and noise statistics, in the presence of initially unknown and time-varying noise. The noise-adaptive particle filter augments its state with the noise relevant statistics and then artificially evolves these during on-line estimation to adapt to changing levels of noise. A simulated global localization problem is used to demonstrate the filter effectiveness against a traditional particle filter and results show it has comparable performance while requiring less initial calibration and fewer direct assumptions. Several techniques from parameter estimation literature are then considered, to reduce the filter computational complexity and inherent limitations.