Off-campus UMass Amherst users: To download campus access dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.

Non-UMass Amherst users: Please talk to your librarian about requesting this dissertation through interlibrary loan.

Dissertations that have an embargo placed on them will not be available to anyone until the embargo expires.

Document Type

Open Access Dissertation

Degree Name

Doctor of Philosophy (PhD)

Degree Program

Computer Science

Year Degree Awarded


Month Degree Awarded


First Advisor

Deepak Ganesan

Second Advisor

Benjamin Marlin

Third Advisor

Marco Duarte

Fourth Advisor

Ivan Lee

Subject Categories

Hardware Systems | Other Computer Sciences | Power and Energy


Clinical studies have shown that features of a person's eyes can function as an effective proxy for cognitive state and neurological function. Technological advances in recent decades have allowed us to deepen this understanding and discover that the actions of the eyes are in fact very tightly coupled to the operation of the brain. Researchers have used camera-based eye monitoring technology to exploit this connection and analyze mental state across across many different metrics of interest. These range from simple things like attention and scene processing, to impairments such as a fatigue or substance use, and even significant mental disorders such as Parkinson's, autism, and schizophrenia.

While there is a wealth of knowledge and social benefit to be gained from eye tracking, the field has historically been restricted to laboratory use by crippling technological limitations - most notably, device size and power consumption. These issues primarily stem from the use of high-resolution cameras and heavyweight video-processing algorithms, both of which induce extremely high performance overhead on the eye tracker. To address this problem, we have constructed a lightweight, ultra-low-power eye monitoring device in the form factor of a pair of eyeglasses. The key guiding design principle for its construction was saliency-aware resource minimization. Specifically, our design leverages the fact that close-up images of the eye are characterized by large salient features which provide a high degree of redundant information; we exploit this to heavily subsample the eye image and reduce resource utilization while performing effective eye tracking.

In the first part of this thesis, we present an initial design of a wearable system to enable ubiquitous eye tracking. By exploiting the fact that the eye has several large, visually redundant features such as the iris and pupil, we were able to develop a neural-network-based adaptive-sampling algorithm for predicting the gaze point while sampling a minimal number of pixels from the image. This enabled us to realize a power savings using specialized imaging hardware that would sample only those most salient pixels, which proportionally reduced the power and time cost of reading images for eye tracking. With these optimizations we were able to build a first-of-of its kind wearable eye tracker that consumed 40 mW of power and demonstrated a gaze tracking error of only 3 degrees across multiple subjects. We refer to this device as the iShadow platform.

The second contribution and section of this thesis is a significant improvement upon the original iShadow design for the purpose of improving both power utilization and eye tracking performance. We constructed a new pupil-tracking algorithm based on lightweight computer vision features, which leverages the smoothness of the eye's motion to reduce even further the amount of camera sampling needed. To guard against very infrequent discontinuities resulting from blinks or reflections off the eye, we integrated this model with the previously-used one-shot neural network algorithm. Because the common case (smooth, uninterrupted eye motion) occurs 90% of the time, we were able to realize a dramatic increase in performance due to the efficiency of the smooth tracking algorithm. The new and improved system, labeled CIDER, enabled much more accurate eye tracking - 0.4 degree error - with power consumption as low as 7 mW. This design also enabled a tradeoff between power consumption and eye tracking rate, in which it was also possible to draw higher power of ~30 mW in order to do eye tracking at rates of up to 240 frames per second.

The final contribution of this thesis is a re-designed version of the iShadow glasses hardware that is suitable for ``in-the-wild'' studies on subjects in their daily living environment. A wearable device, especially one that is worn on the head, must be minimally obtrusive in order to be accepted and used in the field by subjects. This design goal conflicts with the ideal placement of cameras that is needed for achieving consistent eye tracking fidelity. We present multiple possible methods we explored for addressing these competing design challenges, and discuss the reasons that many proved infeasible. To conclude, we present a working design solution that appears to optimally trade off user comfort and convenience and against the technical requirements of the system.