Date of Award

2017

Document Type

Open Access Dissertation

Degree Name

Doctor of Philosophy in Computer Science (PhD)

Administrative Home Department

Department of Computer Science

Advisor 1

Saeid Nooshabadi

Committee Member 1

Timothy Havens

Committee Member 2

Scott Kuhl

Committee Member 3

Stefaan De Winter

Abstract

With the advent of new commodity depth sensors, point cloud data processing plays an increasingly important role in object recognition and perception. However, the computational cost of point cloud data processing is extremely high due to the large data size, high dimensionality, and algorithmic complexity. To address the computational challenges of real-time processing, this work investigates the possibilities of using modern heterogeneous computing platforms and its supporting ecosystem such as massively parallel architecture (MPA), computing cluster, compute unified device architecture (CUDA), and multithreaded programming to accelerate the point cloud based object recognition. The aforementioned computing platforms would not yield high performance unless the specific features are properly utilized. Failing that the result actually produces an inferior performance. To achieve the high-speed performance in image descriptor computing, indexing, and matching in point cloud based object recognition, this work explores both coarse and fine grain level parallelism, identifies the acceptable levels of algorithmic approximation, and analyzes various performance impactors. A set of heterogeneous parallel algorithms are designed and implemented in this work. These algorithms include exact and approximate scalable massively parallel image descriptors for descriptor computing, parallel construction of k-dimensional tree (KD-tree) and the forest of KD-trees for descriptor indexing, parallel approximate nearest neighbor search (ANNS) and buffered ANNS (BANNS) on the KD-tree and the forest of KD-trees for descriptor matching. The results show that the proposed massively parallel algorithms on heterogeneous computing platforms can significantly improve the execution time performance of feature computing, indexing, and matching. Meanwhile, this work demonstrates that the heterogeneous computing architectures, with appropriate architecture specific algorithms design and optimization, have the distinct advantages of improving the performance of multimedia applications.

Available for download on Wednesday, November 28, 2018

Share

COinS