Eric T. Psota

Real-Time Stereo Matching on CUDA Using an Iterative Refinement Method for Adaptive Support-Weight Correspondences

High-quality real-time stereo matching has the potential to enable various computer vision applications including semi-automated robotic surgery, teleimmersion, and 3-D video surveillance. A novel real-time stereo matching method is presented that uses a two-pass approximation of adaptive support-weight aggregation, and a low-complexity iterative disparity refinement technique. Through an evaluation of computationally efficient approaches to adaptive support-weight cost aggregation, it is shown that the two-pass method produces an accurate approximation of the support weights while greatly reducing the complexity of aggregation. The refinement technique, constructed using a probabilistic framework, incorporates an additive term into matching cost minimization and facilitates iterative processing to improve the accuracy of the disparity map. This method has been implemented on massively parallel high-performance graphics hardware using the Compute Unified Device Architecture computing engine. Results show that the proposed method is the most accurate among all of the real-time stereo matching methods listed on the Middlebury stereo benchmark.

Average min-sum decoding of LDPC codes

Simulations have shown that the outputs of min-sum (MS) decoding generally behave in one of two ways: either the output vector eventually stabilizes at a codeword or it eventually cycles through a finite set of vectors that may include both codewords and non-codewords. The latter behavior has significantly contributed to the difficulty in studying the performance of this decoder. To overcome this problem, a new decoder, average min-sum (AMS), is proposed; this decoder outputs the average of the MS output vectors over a finite set of iterations. Simulations comparing MS, AMS, linear programming (LP) decoding, and maximum likelihood (ML) decoding are presented, illustrating the relative performances of each of these decoders. In general, MS and AMS have comparable word error rates; however, in the simulation of a code with large block length, AMS has a significantly lower bit error rate. Finally, AMS pseudocodewords are introduced and their relationship to graph cover and LP pseudocodewords is explored, with particular focus on the AMS pseudocodewords of regular LDPC codes and cycle codes.

Analysis of Connections Between Pseudocodewords

The role of pseudocodewords in causing non-codeword outputs in linear programming decoding, graph cover decoding, and iterative message-passing decoding is investigated. The three main types of pseudocodewords in the literature-linear programming pseudocodewords, graph cover pseudocodewords, and computation tree pseudocodewords-are reviewed and connections between them are explored. Some discrepancies in the literature on minimal and irreducible pseudocodewords are highlighted and clarified, and the minimal degree cover necessary to realize a pseudocodeword is found. Additionally, some conditions for the existence of connected realizations of graph cover pseudocodewords are given. This allows for further analysis of when graph cover pseudocodewords induce computation tree pseudocodewords. Finally, an example is offered that shows that existing theories on the distinction between graph cover pseudocodewords and computation tree pseudocodewords are incomplete.

MAP Disparity Estimation Using Hidden Markov Trees

A new method is introduced for stereo matching that operates on minimum spanning trees (MSTs) generated from the images. Disparity maps are represented as a collection of hidden states on MSTs, and each MST is modeled as a hidden Markov tree. An efficient recursive message-passing scheme designed to operate on hidden Markov trees, known as the upward-downward algorithm, is used to compute the maximum a posteriori (MAP) disparity estimate at each pixel. The messages processed by the upward-downward algorithm involve two types of probabilities: the probability of a pixel having a particular disparity given a set of per-pixel matching costs, and the probability of a disparity transition between a pair of connected pixels given their similarity. The distributions of these probabilities are modeled from a collection of images with ground truth disparities. Performance evaluation using the Middlebury stereo benchmark version 3 demonstrates that the proposed method ranks second and third in terms of overall accuracy when evaluated on the training and test image sets, respectively.

The Generator and Parity-Check Matrices of Turbo Codes

Turbo codes are typically represented as parallel concatenated convolutional codes, but will be treated as serially concatenated codes in this paper. Treating turbo codes as serially concatenated codes makes possible the general description of their generator and parity-check matrices. Given the generator and parity-check matrices, turbo codes may be analyzed from a block and low-density parity-check code point of view.

Real-time temporal stereo matching using iterative adaptive support weights

Stereo matching algorithms are nearly always designed to find matches between a single pair of images. A method is presented that was specifically designed to operate on sequences of images. This method considers the cost of matching image points in both the spatial and temporal domain. To maintain real-time operation, a temporal cost aggregation method is used to evaluate the likelihood of matches that is invariant with respect to the number of prior images being considered. This method has been implemented on massively parallel GPU hardware, and the implementation ranks as one of the fastest and most accurate real-time stereo matching methods as measured by the Middlebury stereo performance benchmark.

LDPC decoding and code design on extrinsic trees

Extrinsic tree decoding of low-density parity-check codes operates on modified, finite computation trees created from the Tanner graph of the code. The goal of the extrinsic tree algorithm is to maintain or improve the performance of existing iterative decoders, while providing a decoding algorithm for which upper bounds can be computed. The extrinsic tree algorithm is examined, along with the design of parity-check matrices on which the extrinsic tree decoder performs well.

Iterative Construction of Regular LDPC Codes from Independent Tree-Based Minimum Distance Bounds

An independent tree-based method for lower bounding the minimum distance of low-density parity-check (LDPC) codes is presented. This lower-bound is then used as the decision criterion during the iterative construction of regular LDPC codes. The new construction algorithm results in LDPC codes with greater girth and improved minimum-distance bounds when compared to regular LDPC codes constructed using the progressive edge-growth (PEG) construction and the approximate cycle extrinsic message degree (ACE)-constrained PEG construction. Simulation results of codes constructed with the new method show improved performance on the additive white Gaussian noise channel at moderate signal-to-noise ratios.

Extrinsic tree decoding

A new decoding method, called extrinsic tree decoding, is presented for decoding low-density parity-check codes on modified finite computation trees. The proposed method maintains similar performance to that of existing iterative decoders, while providing a decoding method for which realistic upper bounds can be computed for practical codes.

Smart watch RSSI localization and refinement for behavioral classification using laser-SLAM for mapping and fingerprinting.

As a first step toward building a smart home behavioral monitoring system capable of classifying a wide variety of human behavior, a wireless sensor network (WSN) system is presented for RSSI localization. The low-cost, non-intrusive system uses a smart watch worn by the user to broadcast data to the WSN, where the strength of the radio signal is evaluated at each WSN node to localize the user. A method is presented that uses simultaneous localization and mapping (SLAM) for system calibration, providing automated fingerprinting associating the radio signal strength patterns to the users location within the living space. To improve the accuracy of localization, a novel refinement technique is introduced that takes into account typical movement patterns of people within their homes. Experimental results demonstrate that the system is capable of providing accurate localization results in a typical living space.