Ranganathan A.,Honda Research Institute United States Inc. |
Dellaert F.,Georgia Institute of Technology
International Journal of Robotics Research | Year: 2011
We present a novel algorithm for topological mapping, which is the problem of finding the graph structure of an environment from a sequence of measurements. Our algorithm, called Online Probabilistic Topological Mapping (OPTM), systematically addresses the problem by constructing the posterior on the space of all possible topologies given measurements. With each successive measurement, the posterior is updated incrementally using a Rao-Blackwellized particle filter. We present efficient sampling mechanisms using data-driven proposals and prior distributions on topologies that further enable OPTM's operation in an online manner. OPTM can incorporate various sensors seamlessly, as is demonstrated by our use of appearance, laser, and odometry measurements. OPTM is the first topological mapping algorithm that is theoretically accurate, systematic, sensor independent, and online, and thus advances the state of the art significantly. We evaluate the algorithm on a robot in diverse environments. © 2011 The Author(s).
Pradeep V.,Microsoft |
Lim J.,Honda Research Institute United States Inc.
International Journal of Computer Vision | Year: 2012
We propose a novel minimal solver for recovering camera motion across two views of a calibrated stereo rig. The algorithm can handle any assorted combination of point and line features across the four images and facilitates a visual odometry pipeline that is enhanced by welllocalized and reliably-tracked line features while retaining the well-known advantages of point features. The mathematical framework of our method is based on trifocal tensor geometry and a quaternion representation of rotation matrices. A simple polynomial system is developed from which camera motion parameters may be extracted more robustly in the presence of severe noise, as compared to the conventionally employed direct linear/subspace solutions. This is demonstrated with extensive experiments and comparisons against the 3-point and line-sfm algorithms. © Springer Science+Business Media, LLC 2011.
Ranganathan A.,Honda Research Institute United States Inc. |
Yang M.-H.,University of California at Merced |
Ho J.,University of Florida
IEEE Transactions on Image Processing | Year: 2011
We present a new Gaussian process (GP) inference algorithm, called online sparse matrix Gaussian processes (OSMGP), and demonstrate its merits by applying it to the problems of head pose estimation and visual tracking. The OSMGP is based upon the observation that for kernels with local support, the Gram matrix is typically sparse. Maintaining and updating the sparse Cholesky factor of the Gram matrix can be done efficiently using Givens rotations. This leads to an exact, online algorithm whose update time scales linearly with the size of the Gram matrix. Further, we provide a method for constant time operation of the OSMGP using matrix downdates. The downdates maintain the Cholesky factor at a constant size by removing certain rows and columns corresponding to discarded training examples. We demonstrate that, using these matrix downdates, online hyperparameter estimation can be included at cost linear in the number of total training examples. We describe a robust appearance-based head pose estimation system based upon the OSMGP. Numerous experiments and comparisons with existing methods using a large dataset system demonstrate the efficiency and accuracy of our system. Further, to showcase the applicability of OSMGP to a wide variety of problems, we also describe a regression-based visual tracking method. Experiments show that our OSMGP algorithm generalizes well using online learning. © 2011 IEEE.
Liu Y.,Rice University |
Artyukhov V.I.,Rice University |
Liu M.,Rice University |
Harutyunyan A.R.,Honda Research Institute United States Inc. |
Yakobson B.I.,Rice University
Journal of Physical Chemistry Letters | Year: 2013
Nanomaterials are anticipated to be promising storage media, owing to their high surface-to-mass ratio. The high hydrogen capacity achieved by using graphene has reinforced this opinion and motivated investigations of the possibility to use it to store another important energy carrier - lithium (Li). While the first-principles computations show that the Li capacity of pristine graphene, limited by Li clustering and phase separation, is lower than that offered by Li intercalation in graphite, we explore the feasibility of modifying graphene for better Li storage. It is found that certain structural defects in graphene can bind Li stably, yet a more efficacious approach is through substitution doping with boron (B). In particular, the layered C3B compound stands out as a promising Li storage medium. The monolayer C 3B has a capacity of 714 mAh/g (as Li1.25C3B), and the capacity of stacked C3B is 857 mAh/g (as Li 1.5C3B), which is about twice as large as graphite's 372 mAh/g (as LiC6). Our results help clarify the mechanism of Li storage in low-dimensional materials, and shed light on the rational design of nanoarchitectures for energy storage. © 2013 American Chemical Society.
Ranganathan A.,Honda Research Institute United States Inc.
Autonomous Robots | Year: 2012
A shared vocabulary between humans and robots for describing spatial concepts is essential for effective human robot interaction. Towards this goal, we present a novel technique for place categorization from visual cues called PLISS (Place Labeling through Image Sequence Segmentation). PLISS is different from existing place categorization systems in two major ways-it inherently works on video and image streams rather than single images, and it can detect "unknown" place labels, i.e. place categories that it does not know about. PLISS uses changepoint detection to temporally segment image sequences which are subsequently labeled. Changepoint detection and labeling are performed inside a systematic probabilistic framework. Unknown place labels are detected by using a probabilistic classifier and keeping track of its label uncertainty. We present experiments and comparisons on the large and extensive VPC dataset. We also demonstrate results using models learned from images downloaded from Google's image search. © 2011 Springer-Verlag.