Search

Computer vision has become ubiquitous in today's society, with applications ranging from medical imaging to visual diagnostics to aerial monitoring to self-driving vehicles and many more. Common to many of these applications are visual perception systems which consist of classification, localization, detection, and segmentation components, just to name a few. Recently, the development of deep neural networks (DNN) have led to great advancements in pushing state-of-the-art performance in each of these areas. Unlike traditional computer vision algorithms, DNNs have the ability to generalize features previously hand-crafted by engineers specific to the application; this assumption models the human visual system's ability to generalize its surroundings. Moreover, convolutional neural networks (CNN) have been shown to not only match, but exceed performance of traditional computer vision algorithms as the filters of the network are able to learn "important" features present in the data. In this research we aim to develop numerous applications including visual warehouse diagnostics and shipping yard managements systems, aerial monitoring and tracking from the perspective of the drone, perception system model for an autonomous vehicle, and vehicle re-identification for surveillance and security. The deep learning models developed for each application attempt to match or exceed state-of-the-art performance in both accuracy and inference time; however, this is typically a trade-off when designing a network where one or the other can be maximized. We investigate numerous object-detection architectures including Faster R-CNN, SSD, YOLO, and a few other variations in an attempt to determine the best architecture for each application. We constrain performance metrics to only investigate inference times rather than training times as none of the optimizations performed in this research have any effect on training time. Further, we will also investigate re-identification of vehicles as a separate application add-on to the object-detection pipeline. Re-identification will allow for a more robust representation of the data while leveraging techniques for security and surveillance. We also investigate comparisons between architectures that could possibly lead to the development of new architectures with the ability to not only perform inference relatively quickly (or in close-to real-time), but also match the state-of-the-art in accuracy performance. New architecture development, however, depends on the application and its requirements; some applications need to run on edge-computing (EC) devices, while others have slightly larger inference windows which allow for cloud computing with powerful accelerators.

Within the past decade, there have been substantial leaps in computer architectures to exploit the parallelism that is inherently present in many applications. The scientific community has benefited from the emergence of not only multi-core processors, but also other, less traditional architectures including general purpose graphical processing units (GPGPUs), field programmable gate arrays (FPGAs), and Intel's many integrated cores (MICs) architecture (i.e. Xeon Phi). The popularity of the GPGPU has increased rapidly because of their ability to perform massive amounts of parallel computation quickly and at low cost with an ease of programmability. Also, with the addition of high-level programming interfaces for these devices, technical and non-technical individuals can interface with the device and rapidly obtain improved performance for many algorithms. Many applications can take advantage of the parallelism present in distributed computing and multithreading to achieve higher levels of performance for the computationally intensive parts of the application. The work presented in this thesis implements three applications for use in a performance study of the GPGPU architecture and multi-GPGPU systems. The first application study in this research is a K-Means clustering algorithm that categorizes each data point into the closest cluster. The second algorithm implemented is a spiking neural network algorithm that is used as a computational model for machine learning. The third, and final, study is the longest common subsequences problem, which attempts to enumerate comparisons between sequences (namely, DNA sequences). The results for the aforementioned applications with varying problem sizes and architectural configurations are presented and discussed in this thesis. The K-Means clustering algorithm achieved approximately 97x speedup when utilizing an architecture consisting of 32 CPU/GPGPU pairs. To achieve this substantial speedup, up to 750,000 data points were used with up 30,000 centroids (means). The spiking neural network algorithm resulted in speedups of about 33x for the entire algorithm and 160x for each iteration with a two-level network with 1000 total neurons (800 excitatory and 200 inhibitory neurons). The longest common subsequences problem achieved speedup of greater than 10x with 100 random sequences up to 500 characters in length. The maximum speedup values for each application were achieved by utilizing the GPGPU as well as multi-core devices simultaneously. The computations were scattered over multiple CPU/GPGPU pairs with the computationally intensive pieces of the algorithms offloaded onto the GPGPU device. The research in this thesis illustrates the ability to scale a heterogeneous cluster (i.e. CPUs and GPUs working collaboratively) for large-scale scientific application performance improvements. Each algorithm demonstrates slightly different types of computations and communications, which can be compared to other algorithms to predict how they would perform on an accelerator. The results show that substantial speedups can be achieved for scientific applications when utilizing the GPGPU and multi-core architectures.