The inter-beat-interval (time period of the cardiac cycle) changes slightly for every heartbeat; this variation is measured as Heart Rate Variability (HRV). HRV is presumed to occur due to interactions between the parasympathetic and sympathetic nervous system. Therefore, it is sometimes used as an indicator of the stress level of an individual. HRV also reveals some clinical information about cardiac health. Currently, HRV is accurately measured using contact devices such as a pulse oximeter. However, recent research in the eld of non-contact imaging Photoplethysmography (IPPG) has made it possible to extract vital sign measurements from the video recording of any exposed skin surface, such as a person's face. Extracting HRV using a camera holds a lot of promise as it opens up the applications of stress, engagement monitoring during interviews and online education. It can also be used as a non-invasive monitoring solution for measuring autonomous nervous system function for diabetic patients who are at high risk for developing diabetic autonomic neuropathy. The current signal processing methods for extracting HRV using peak detection perform well for contact-based systems but have poor performance for the IPPG signals. The main reason for this poor performance is the fact that current methods are sensitive to large noise sources which are often present in IPPG data. Further, current methods are not robust to motion artifacts that are common in IPPG systems. We developed a new algorithm, HRVCam, for robustly extracting HRV even in low SNR such as is common with IPPG recordings. HRVCam combines spatial combination and frequency demodulation to obtain HRV from the instantaneous frequency of the IPPG signal. HRVCam outperforms other current methods of HRV estimation. Ground truth data was obtained from FDA-approved pulse oximeter for validation purposes. HRVCam improves the accuracy by 25% for light skin tones and by 60% for darker skin tones. HRVCam also allows us to measure HRV during high motion scenarios such as talking with an error of 25%. In such scenarios state-of-the-art approaches perform rather poorly.