A Robust and Efficient Rao-Blackwellized Particle Filter for Nonlinear and Nongaussian SLAM Problem

DoctorIntelligent mobility is a fundamental requirement of autonomous robots. To achieve this, mobile robots should learn the model of environments while estimating their poses. Thus, the simultaneous localization and mapping (SLAM) has been a major topic in the robotics application over the last decades. Most of researches presented so far in the SLAM field use laser range finders or vision sensors. The fundamental reason of this is that performance of the SLAM algorithm depends on the sensor performance in principle. A main goal of this thesis is to develop an efficient and robust SLAM algorithm that can be used in a practical application for personal service robots. A consistent SLAM algorithm is usually computationally heavy. On the other hand, the map learning using the noisy and short-range sensor is more difficult than the high performance sensors and it increases the computational cost in general to maintain the filter consistency with estimation accuracy. So, this thesis proposes algorithms to improve both the computational efficiency and the filter consistency, based on the RBPF framework. In advance, this thesis contributes to solve the SLAM problem that specifically addresses the noisy sonar sensor in the RBPF framework that can handle a multi-hypotheses tracking with non-Gaussianity. We first present the straightforward approach to improve a robustness and efficiency of the RBPF framework. To estimate the uncertainty in both the robot's pose and the map precisely, our approach introduces the scaled unscented transformation technique, which is able to estimate the mean and the covariance to a higher order of accuracy than the linearized techniques. Thus, even if there is a large bearing uncertainty, higher order information about the state distribution can be represented well, and this benefit produces the robustness to the sensor noise. On the other hand, since the filter inconsistency of the RBPF based SLAM algorithm due to a lack of hypotheses is the issue, we propose a novel approach considering a loop closing event in the resampling process of the particle filter after defining the problem thoroughly. We also propose an algorithm for speeding up RBPF that uses an optimal proposal distribution. This algorithm can significantly improve the computational efficiency while maintaining a performance of the standard algorithm due to an assistance of the Gaussian mixture filter in the proposal computation. The optimal proposal of RBPF is still widely spread because of the large sensor noise and intermittent nature of the sonar feature. This increases a required number of particles for learning the consistent map. To solve the problem, we incorporate prior information of structured environments so that the feature is constrained. Finally we present a RBPF-SLAM solution based on the ceiling mosaic approach using a single web-cam. The most attractive factor of this approach is its practicability in typical indoors, because the sensor reading from the ceiling is not influenced by moving objects on the ground. Aforementioned our approaches allow a robot to build a consistent map with estimating its pose accurately, even though a low resolution, short-range noisy sensor is used with less number of particles