Date of Award


Document Type

Thesis (Master's)

Department or Program

Computer Science

First Advisor

Nicholas Jacobson

Second Advisor

Andrew Campbell

Third Advisor

Soroush Vosoughi


Heart rate and heart rate variability (HRV) are important metrics in the study of numerous physical and psychiatric conditions. Previously, measurement of heart rate was relegated to clinical settings, and was neither convenient nor captured a patient’s typical resting state. In effect, this made gathering heart rate data costly and introduced noise. The current prevalence of mobile phone technology and Internet access has increased the viability of remote health monitoring, thus presenting an opportunity to substantially improve the speed, convenience, and reliability of heart rate readings. Recent attention has focused on different methods for remote, non-contact heart rate measurement. Of these methods, video presents perhaps the best option for optimizing cost and convenience. This thesis introduces a lightweight architecture for estimating heart rate and HRV using a smartphone camera. The system presented here runs locally on a smartphone, requiring only a phone camera and 15s or more of continuous video of a subject’s face. No Internet connection or networking is necessary. Building the system to run locally in this manner means that this software confers benefits such as greater user privacy, offline availability, reliability, cost effectiveness, and speed. However, it also introduces added constraints on computational complexity. With these tradeoffs in mind, the system presented here is implemented within an Android mobile app. The performance of our approach fell short of that of existing state-of-the-art methods in terms of mean absolute error (MAE) of heart rate estimation, achieving MAE during validation that was over $17x$ greater than other existing approaches. There are a number of factors which may contribute to this performance discrepancy, including limitations in the diversity of the data used with respect to gender, age, skin tone, and heart rate intensity. Further, remote photoplethysmographic (rPPG) signal generated by this architecture contains a large number of noise artifacts which are difficult to consistently remove through signal processing. This noise is the primary reason for the underperformance of this architecture, and could potentially be explained by model and feature engineering decisions which were made to address the risk of overfitting on the limited dataset used in this work.