Abstract
The article describes a reference and training set free incrementally trained deep learning algorithm for camera-based respiration monitoring systems. The algorithm uses a model based discriminator to find salient areas having respiration like periodic motion. It stores the first principle component of the found waveforms into two slowly growing set along with negative, uncorrelated motion patterns. Using these samples, it trains a deep neural network classifier incrementally to recognize respiration from sudden and motion intensive situations. The classifier had no forgetting mechanism and it is able to adapt quickly the changing respiration patterns and conditions. The algorithm has been validated in a total of 24 hours diverse recording captured in the neonatal intensive care unit (NICU) of the I <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">st</sup> Dept. of Pediatrics and, II. Dept. of Obstetrics and Gynecology, Semmelweis University, Budapest, Hungary and in the COHFACE publicly available dataset of adult subjects. The clinical data set evaluation resulted in mean absolute error (MAE) 6.9 and root mean squared error (RMSE) of 9.8 breaths per minute, respectively, the MAE was below 5 breaths per minute for over 50% of the time. The algorithm was assessed in the COHFACE dataset of adult subjects as well with respiration estimation MAE and RMSE values of 0.95 and 1.7 breaths per minute.
Highlights
I N CAMERA based remote vital sign and respiration monitoring algorithms several solutions have been published in the recent decade [1], in the special field of newborn infant monitoring
Our algorithm is a combination of optical flow based source separation and a low complexity convolutional neural network classifier
OPERATION PRINCIPLES The architecture is a combination of a dense rate calculation source separation model and a neural network classifier
Summary
I N CAMERA based remote vital sign and respiration monitoring algorithms several solutions have been published in the recent decade [1], in the special field of newborn infant monitoring. Recent studies relies on either mixed [7], [8] or neural network solutions [9]–[11] for these tasks to increase motion robustness and overall performance. Another interesting approach is to exploit local motion magnification algorithm [12], which in principle relies on regular, periodic. The presented results show low error rate (3.5-4.5 BPM MAE over selected periods) the data preparation, annotation, augmentation efforts is significant and specific to the observed environment, and the employed neural network requires heavy GPU acceleration. The neural network classifier is trained incrementally with low iteration count using the stored positive and negative samples. The classifier performance becomes better to find weak signal sources and due to the stored sets, long term memorization is preserved and no overfitting occurs due to the increasing and versatile training set
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call