Conference Paper

Large-scale terrain modeling from multiple sensors with dependent Gaussian processes

Australian Centre for Field Robot., Univ. of Sydney, Sydney, NSW, Australia
DOI: 10.1109/IROS.2010.5650769 Conference: Intelligent Robots and Systems (IROS), 2010 IEEE/RSJ International Conference on
Source: IEEE Xplore


Terrain modeling remains a challenging yet key component for the deployment of ground robots to the field. The difficulty arrives from the variability of terrain shapes, sparseness of the data, and high degree uncertainty often encountered in large, unstructured environments. This paper presents significant advances to data fusion for stochastic processes modeling spatial data, demonstrated in large-scale terrain modeling tasks. We explore dependent Gaussian processes to provide a multi-resolution representation of space and associated uncertainties, while integrating sensors from different modalities. Experiments performed on multiple multi-modal datasets (3D laser scans and GPS) demonstrate the approach for terrains of about 5 km2.

Full-text preview

Available from:
  • Source
    • "Work done in [12], [13], [14], [15] proposes the use of a KD-tree structure for a fast local approximation. For a training set of n observations, the inference process requires calculating and inverting an n × n covariance matrix, which gets computationally expensive as n gets large. "
    [Show abstract] [Hide abstract]
    ABSTRACT: Efficiently monitoring environmental conditions across large indoor spaces (such as warehouses, factories or data centers) is an important problem with many applications. Deployment of a sensor network across the space can provide very precise readings at discrete locations. However, construction of a continuous model from this discrete sensor data is a challenge. The challenge is made harder by economic and logistical constraints that may limit the number of sensor motes in the network. The required model, therefore, must be able to interpolate sparse data and give accurate predictions at unsensed locations, as well as provide some notion of the uncertainty on those predictions. We propose a Gaussian process based model to answer both of these issues. We use Gaussian processes to model temperature and humidity distributions across an indoor space as functions of a 3-dimensional point. We study the model selection process and show that good results can be obtained, even with sparse sensor data. Deployment of a sensor network across an indoor lab provides real-world data that we use to construct an environmental model of the lab space. We seek to refine the model obtained from the initial deployment by using the uncertainty estimates provided by the Gaussian process methodology to modify sensor distribution such that each sensor is most advantageously placed. We explore multiple sensor placement techniques and experimentally validate a near-optimal criterion.
    Full-text · Conference Paper · Sep 2016
  • Source
    • "The GP data fusion problem was cast as a conditional estimation using several Dependent GP's. The formalism could also be used to simultaneously model multiple aspects of the terrain as demonstrated in [15]. The proposed DGP based on the nonstationary (neural-network) kernel performed significantly better than the stationary squared exponential kernel based DGP ([15]) in fusing multiple terrain data sets. "
    [Show abstract] [Hide abstract]
    ABSTRACT: Obtaining a comprehensive model of large and complex terrain typically entails the use of both multiple sensory modalities and multiple data sets. This paper demonstrates the use of dependent Gaussian processes for data fusion in the context of large scale terrain modeling. Specifically, this paper derives and demonstrates the use of a non-stationary kernel (Neural Network) in this context. Experiments performed on multiple large scale (spanning about 5 sq km) 3D terrain data sets obtained from multiple sensory modalities (GPS surveys and laser scans) demonstrate the approach to data fusion and provide a preliminary demonstration of the superior modeling capability of Gaussian processes based on this kernel.
    Preview · Conference Paper · Jun 2011
  • [Show abstract] [Hide abstract]
    ABSTRACT: Real world phenomena often exhibit complex dynamics across space and time. Understanding such dynamics requires deploying sensors that provide relevant information about the observed phenomena. It is imperative to develop an efficient and accurate model, representing the complex spacetime dynamics, that can both inform sensor placement for efficient monitoring as well as help in appropriate knowledge discovery from the collected sensor data. This paper extends a generic non-stationary non-separable space-time Gaussian Process (GP) model, that models complex correlation properties using variable hyper parameters across input space. We provide approaches to effectively observe the hyper parameter space and thus significantly reduce the training cost for the proposed model. We also formalize relationship between degree of non-stationarity of the phenomenon and learning cost of the proposed model. Extensive empirical validation of the proposed model is performed using two real world sensing datasets, exhibiting diverse dynamics across space and time, to demonstrate the real world applicability of the proposed approach.
    No preview · Article · Aug 2012
Show more