Date of Award

Spring 4-2017

Access Type

Thesis - Open Access

Degree Name

Master of Science in Electrical & Computer Engineering

Department

Electrical, Computer, Software, and Systems Engineering

Committee Chair

Brian Butka

First Committee Member

Eric Coyle

Second Committee Member

Jianhua Liu

Abstract

Object detection through computer vision has traditionally been difficult to reliably implement due to various lighting conditions caused by weather and time of day. Any changes in conditions can be detrimental to the detector’s ability to accurately identify objects. A modern approach implements deep learning techniques to classify and train a neural network. While highly effective, this approach can be cumbersome and computationally intensive. This project will investigate the feasibility of using deep learning to detect, classify, and track objects in near real-time while being processed on a mobile platform. I will investigate the feasibility of these processes on a small embedded system, such as the NVIDIA Jetson TX1. I will investigate several promising algorithms such as Faster R-CNN, TensorBox, DetectNet, and YOLO. This research is beneficial because it will transition deep learning techniques developed primarily for research in a lab environment to a real-world situation in which high accuracy and fast processing are vital. The work solved through this research will greatly benefit platforms that require object detection capabilities, but do not have the space, budget, or power capabilities for large GPUs or GPU clusters.

Download

Share

COinS