Examining the Effect of Noise on Biosignal Estimates Extracted through Spatio-Temporal Video Processing

Spatio-temporal video processing has been used to extract subject vital signals from optical video and, more recently, thermal video. Thermal video, in conjunction with spatio-temporal video processing can extract biosignals optical video cannot; namely temperature data, but also biosignals in poor light conditions. Video processing involves many system parameters that can result in false biosignal reporting. This paper aimed to robustly test spatio-temporal processing algorithms and determine patterns with respect to increasing noise levels. Over 500 simulated thermal videos were generated at 29 different signal frequencies representing heart rates. These videos were contaminated with 18 different levels of Gaussian noise and were used as inputs to the algorithmic system. The algorithmic system processed each video at 6 different filter widths. The results were examined individually and as a collective. Individual results were as expected; the processing resulted in an accurate heart rate estimate if the original signal was inside the filter passband. If the signal was outside of the filter passband, the processing simply amplified noise. These same patterns were observed in the cumulative results, in addition to overarching patterns with respect to noise. Two main patterns were observed; a failure threshold was determined and quantified and a pattern of error behavior beyond this threshold was quantified. The failure threshold occurred at a noise variance of approximately 500, and around this parameter value, all detected signal frequencies were approaching 1.5 Hz (90bpm). This study was able to characterize patterns of failure, which helps to prevent future false reporting.