Year: 2010
Type: dissertation
Abstract: Intel produced the 8080 processor chip in 1974, which consisted of approximately 4500 transistors.The revolutionary development of silicon-based manufacturing technology led to the possibility to produce integrated circuits (IC) using several billion transistors in a single chip [8].Such a large amount of basic elements opens the way to commercialize chips with thousands of connectedbut independently controlledparallel computation components, called processor cores.From the aspect of computer science it raises a new challenge to formulate the theory of many-core computing: the structure of the communication network to support maximal data transfer rate between computing nodes to keep them busy with new inputs to process and to transfer outputs to their destination.New algorithms are needed to maximally exploit these novel hardware capabilities, and new methods are required to split problems into parts that can be evaluated in parallel.The optimization of data-communication time enforces the precedence of locality and the utilization of cellular structures [9].The obvious efficacy of biological vision systems may motivate engineers to mimic two major characteristics (parallelism and foveal processing) during the design of artificial solutions that are embedded into real world-environments.When many salient regions arise after preprocessing, all of these should be better explored.This is the multi-fovea model.Successful works (for example [10]) highlighted the importance and efficiency of this design concept that called for a unified algorithm description.In my work I primarily focused on data-parallel interpretation and processing of topological problems, in particular image processing tasks, by using cellular processing structures within the Multi-Fovea Architecture and Algorithmic Framework.The research question from the theoretical point of view was to create a unified software model that supports the selection of parallel architecture and also covers hardware specific details for a given device.The particular practical application inspiring the theoretical work was to analyze algorithms for visual navigation systems applicable in mini / micro unmanned air vehicles (UAVs).The elaborated virtual platform consists of separate processor arrays specialized for parallel execution of preprocessing and foveal computation.The proposed heterogeneous structure can fit the special characteristic of various operators processing the highly parallel data input.I gave a design guideline for Multi-Fovea Architecture and presented it by the comparison of 2D registration methods for ground object motion detection from mini unmanned aerial vehicles.After giving an analytic comparison of registration methods, I proposed a novel method exploring the proposed architecture by running a larger percent of the task in parallel and in cellular structures.
Author: Balázs Gergely Soós
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2183146720