University of Twente Student Theses

Login

Embedded neural network design on the ZYBO FPGA for vision based object localization

Fatseas, K. (2018) Embedded neural network design on the ZYBO FPGA for vision based object localization.

[img] PDF
17MB
Abstract:During recent years we have been witnessing great advancements in the field of computer vision due to the utilization of machine learning. Those achievements are attributed to the evolution of Convolutional Neural Networks (CNNs) as nowadays they are the basic building block of every state-of-the-art object detector and classifier. There is an abundance of possible applications for CNNs and especially in cyber-physical systems, but the utilization of a CNN comes with a very high computation and memory requirements. Therefore the objective of this work is to study the suitability of machine learning and more specifically of a CNN in a visual servoing application. The application itself consists of two robots where one has the role of prey and the other of the predator. The CNN is defined as a classifier and trained on a generated training dataset in order to avoid manual annotation. Then it is integrated into an object detection pipeline in order to extract the precise location of the prey robot. Furthermore, a pre-trained neural network is utilized so as to improve the overall performance of the object detection pipeline. Finally, an accelerator is developed in order to port the object detector into the ZYBO FPGA board and to meet the real-time constraint. Robust performance from both the object tracking algorithm and the accelerator is reported. Although the generated training dataset is the limiting factor of this work, the methodology overall offers a lot of potential for future improvements.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Embedded Systems MSc (60331)
Link to this item:https://purl.utwente.nl/essays/75891
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page