Please use this identifier to cite or link to this item: https://hdl.handle.net/10356/55241
Title: Collaborative simultaneous localization and mapping for autonomous vehicles
Authors: Diluka Moratuwage
Keywords: DRNTU::Engineering::Electrical and electronic engineering::Control and instrumentation::Robotics
Issue Date: 2013
Source: Diluka Moratuwage. (2013). Collaborative simultaneous localization and mapping for autonomous vehicles. Doctoral thesis, Nanyang Technological University, Singapore.
Abstract: Among today’s autonomous robotic applications, exploration missions in dynamic, unstructured and uncertain environmental conditions are quite common. Autonomous multi-vehicle systems come in handy for such exploration missions involving larger environments due to higher efficiency, reliability and robustness that can be achieved compared to a single autonomous vehicle. In the absence of a-priori feature maps, various simultaneous localization and mapping (SLAM) algorithms are used in order to improve the navigation accuracy of single-vehicle systems as a key prerequisite for exploration and mapping missions. As for multi-vehicle systems, this process is called collaborative multi-vehicle simultaneous localization and mapping (CMSLAM). In multi-vehicle systems, additional scaling problems such as consistent data fusion, bandwidth constraints and computational requirements have to be taken into account in order to be used in practical applications. The most common single-vehicle SLAM (mono-SLAM) formulation is stochastic SLAM, in which the vehicle state and landmark locations are stored in a single vector, which is propagated as a joint posterior via a recursive prediction and update (observation) process, where sensor and vehicle dynamics are modeled in state space form. In this approach, each recursion involves solving certain additional, but crucial sub-problems outside the Bayes recursion, such as clutter filtering, data association and map management. Moreover, these conventional solutions assume several key simplifications which limit their practical applicability. The most notable assumptions are clutter free measurements, ideal landmark detection and data association, and static landmarks. Several CMSLAM algorithms have been developed in the literature by extending such mono-SLAM solutions. Unfortunately almost all those algorithms inherit above mentioned issues from their single-vehicle counterpart. In the recent years an alternative solution based on the finite set statistics (FISST) has been proposed as a radically different alternative solution to the mono-SLAM problem. In this approach, the landmark map and the measurements are represented as random finite sets (RFS), and a joint posterior containing the landmark map and the vehicle state is propagated while catering for landmark detection uncertainty, landmark survival uncertainty, data association, map management and measurement clutter filtering within the SLAM filter framework. In this thesis we propose to extend this framework in order to solve the more general CMSLAM problem. The first contribution of this thesis is two new CMSLAM algorithms, which are developed using the RFS SLAM framework. The proposed formulation is based on the factorization of the full CMSLAM posterior into a product of the vehicle trajectories posterior and the landmark map posterior conditioned on the vehicle trajectories. The vehicle trajectories are estimated using a Rao-Blackwellized particle filter. The landmark map and the measurements are represented as RFSs; their dynamics are appropriately modeled using FISST; and the landmark map posterior conditioned on the vehicle trajectories is estimated using a Gaussian Mixture (GM) implementation of a probability hypothesis density (PHD) filter. The performance of the proposed solutions is evaluated and presented by benchmarking against a FastSLAM based CMSLAM solution using a simulation and a practical experiment. Secondly, a novel algorithm is presented to address the collaborative multi-vehicle SLAM with moving object tracking (CMSLAMMOT) problem by extending the proposed RFS CMSLAM solution. Although numerous SLAM algorithms exist in the robotics literature, all such algorithms assume that the environment is static and the moving objects present in the environment are considered as clutter. As a result, a separate higher level algorithm is used in order to detect and track the moving objects present in the environment for safe navigation. The proposed CMSLAMMOT solution addresses this issue and presents a novel solution which detects and tracks moving objects present in the sensor field of views (FOVs) while simultaneously performing CMSLAM. Performance of the proposed solution is evaluated using a simulation and a practical experiment. As the third contribution, a novel hierarchical collaborative multi-vehicle SLAM algorithm (HCMSLAM) is presented by introducing a RFS mono-SLAM based submapping process in order to reduce the communication bandwidth and heavy computational load at the fusion node. Instead of communicating the control and measurement data from each vehicle at each time step, the resultant sub-map along with the sequence of applied control commands during the sub-mapping process are communicated into a higher level HCMSLAM algorithm in order to jointly propagate the landmark map and the vehicle trajectories. The landmark map and the measurements are modeled as RFSs and the full CMSLAM posterior is evaluated by factorizing into a product of the vehicle trajectories posterior and the landmark map posterior conditioned on the vehicle trajectories. The vehicle trajectories are propagated using a Rao-Blackwellized particle filter and the landmark map conditioned on the vehicle trajectories is propagated using a GM implementation of a PHD filter. Simulation results are used in order to evaluate the performance of the proposed solution.
URI: http://hdl.handle.net/10356/55241
Fulltext Permission: open
Fulltext Availability: With Fulltext
Appears in Collections:EEE Theses

Files in This Item:
File Description SizeFormat 
RFS-CMSLAM.pdf9.92 MBAdobe PDFThumbnail
View/Open

Google ScholarTM

Check

Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.