Date of Award

8-2019

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

School of Computing

Committee Member

Amy Apon, Committee Chair

Committee Member

Mashrur Chowdhury

Committee Member

Alexander Herzog

Abstract

This thesis investigates various degrees of freedom and deployment challenges of building an end-to-end intelligent visual inspection system for use in automotive manufacturing. Current methods of fault detection in automotive assembly are highly manual and labor intensive, and thus prone to errors. An automated process can potentially be fast enough to operate within the real-time constraints of the assembly line and can reduce errors. In automotive manufacturing, components of the end-to-end pipeline include capturing a large set of high definition images from a camera setup at the assembly location, transferring and storing the images as needed, executing object detection within a given time frame before the next car arrives in the assembly line, and notifying a human operator when a fault is detected. As inference of object detection models are typically very computing- and memory-intensive, meeting the time, memory and resource constraints requires careful consideration of the choice of object detection model and model parameters, along with adequate hardware and environmental support. Some automotive manufacturing plants lack floor space to set up the entire pipeline on an edge platform. Thus, we have developed a template for Amazon Web Services (AWS) in Python using the BOTO3 libraries that can deploy the entire end-to-end scalable infrastructure in any region in AWS. In this thesis, we design, develop, and experimentally evaluate the performance of system components, including the throughput and latency to upload high definition images to an AWS cloud server, the time required by AWS components in the pipeline, and the tradeoffs of inference time, memory and accuracy for twenty-four popular object detection models on four hardware platforms.

Share

COinS