Date of Award


Document Type

Campus Access Dissertation

Degree Name

Doctor of Philosophy (PhD)


Computer Science

First Advisor

Bo Sheng

Second Advisor

Honggang Zhang

Third Advisor

Tiago Cogumbreiro


Modern society is witnessing an exponentially increasing number of smart devices in daily life. As the proliferation of mobile devices and edge devices bring convenience to everyone, their close proximity to the user, various embedded sensor readings, and increasingly powerful computability also provide many possible solutions to traditional challenges. In the present technical world, the cloud-based approach dominates the system requiring intensive computation. However, the offloading process introduces non-negligible delay. Also, the data sent to the server may raise people’s worries as privacy becomes a growing concern nowadays.

This dissertation takes the advantage of widespread mobile and edge devices to build an efficient framework to support real-time computation-intensive applications. Specifically, the framework focuses on providing solutions to applications involving two challenging tasks: feature point comparison and machine learning inference. Both of them consume a noteworthy portion of the limited computing resource on current mobile and edge devices, hindering the development of the real-time applications that require performing either of these two jobs frequently.

In the first part of this dissertation, a system to support feature point comparison on mobile and edge devices is presented. The widely deployed feature point comparison algorithms suffer from the view angle change problem. This dissertation devises a novel approach to overcome the view angle challenge with the assistance of the embedded sensors in smartphones. Our solution is practical and efficient without introducing noticeable overhead in our experiment.

In the second part, this dissertation introduces a real-time object detection system running on resource-insufficient devices, e.g., smart camera, Jetson Nano, and Raspberry Pi. Our study shows those edge devices fail to run object detection at an acceptable rate (~30 FPS) independently. Instead of keep running object detection, edge devices can track objects to improve their responsiveness. However, tracking may lead to lower accuracy as time passes by. The time to trigger tracking and detection becomes an important decision to make. Our system introduces a cloud server to assist in making this decision. The experiments validate the great potential in our solution.

The last part focuses on a real-time multi-user mobile AR application that incorporates both machine learning inference and feature point comparison. Our AR application is based on reference object recognition, which requires recognizing the exact object that appeared before. This dissertation designs an innovative approach to achieving this goal with the consideration of resource deficit and temporal requirements. Taking the knowledge that the number of feature points influences the response time significantly and the traditional feature points comparison algorithms fall short of accuracy when viewpoint changes, our solution embraces the mobile onboard sensor readings and machine learning model to decrease the number of feature points and increase the comparison accuracy. The performance strongly supports the superiority of our method over existing solutions.


Free and open access to this Campus Access Dissertation is made available to the UMass Boston community by ScholarWorks at UMass Boston. Those not on campus and those without a UMass Boston campus username and password may gain access to this dissertation through resources like Proquest Dissertations & Theses Global or through Interlibrary Loan. If you have a UMass Boston campus username and password and would like to download this work from off-campus, click on the "Off-Campus UMass Boston Users" link above.