Logo Leibniz Universität Hannover
Logo: Institut für Mikroelektronische Systeme
Logo Leibniz Universität Hannover
Logo: Institut für Mikroelektronische Systeme
  • Zielgruppen
  • Suche
 

Evaluation of Approximate and Stochastic Computing Mechanisms for Egomotion Estimation Using SIMD Vector Processors

Student:

Yicheng Yang

Betreuer:

Moritz Weißbrich

Art der Arbeit:

Master-/Diplomarbeit

Abteilung:

Fachgebiet für Architekturen und Systeme

Status:

abgeschlossen

Bild Evaluation of Approximate and Stochastic Computing Mechanisms for Egomotion Estimation Using SIMD Vector Processors

A growing amount of advanced driving assistance systems relies on an accurate three-dimensional perception of the surrounding scene. As a cost-efficient solution, stereo camera systems can be applied to reconstruct the environment from a moving car. By using feature-based egomotion estimation, the relative camera translation and rotation between subsequent frames is calculated only from matching extracted features in moving camera images, so that additional position sensors are not required to obtain a reconstruction of objects and their movement in the environment. Real-time execution of feature extraction algorithms demands for a processor architecture with high processing performance.

At the Institute of Microelectronic Systems, application-specific horizontal and vertical SIMD vector processors are explored. A generic vectorized implementation of the SIFT feature extraction algorithm is available for these architectures. In a prior thesis at the institute (Accuracy Evaluation of Feature-Based Egomotion Estimation Supported by Approximate SIMD Vector Processors), a framework for egomotion estimation has been set up and evaluated using this vectorized SIFT implementation. Approximate Computing paradigms have been applied in addition to algorithmic parallelization potential in order to cope with the performance and energy budget constraints. For this, the multiplication ALUs of the vector processors are modified by a mask-based operand approximation mechanism to trade-off computational accuracy for lower power consumption in error-resilient Computer Vision applications like egomotion estimation.

Based on previous work, Mr. Yang’s task is to refine and optimize the existing architectures and the evaluation framework with more advanced Approximate and Stochastic Computing methods. The mask-based approximation mechanism shall be applied to other arithmetic units in the vector processors, and a fine-grained analysis of the SIFT approximation potential shall be performed for the quality metrics of egomotion estimation. This task includes the full automation of the FPGA-based evaluation framework to explore the Approximate Computing design space by hardware emulation. Additionally, a logical and timing analysis of the mask-based approximation mechanism shall be performed to identify potential for performance gain or energy optimization by Stochastic Computing mechanisms like frequency or voltage overscaling. ASIC synthesis using Synopsys design tools and switching activity simulations at gate-level with an available 45nm semiconductor technology are used to quantify the impact on power consumption of the evaluated mechanisms.


zurück