Caleb Olson's Baby Sleep Coach 2.0 Is a Private, AI-Powered Monitoring System
By tracking movements and learning over time from a local camera, the setup detects and predicts the best times for your baby's sleep.
The first version
Over half a year ago, maker and YouTuber Caleb Olson created the Baby Sleep Coach system which utilizes a camera pointed at the crib to analyze a baby's facial expressions and pose for alertness. Beyond merely saying whether the baby is awake or not, the Coach also comes with a web application that displays a chart of sleeping times and even forecasts when the next wake-up time will be thanks to a predictive model.
All of this worked great, at least for the first few months, but after starting to move around more frequently and immersing itself in blankets, Olson's baby was too difficult to measure using the pose estimation techniques alone. In response, he Olson wanted to build an additional model that augments the original to detect awake times even if a blank is acting as cover.
Gathering data and training a model
Image classification models require a lot of data to function, and due to it being difficult to obtain in this case, Olson needed a better approach. By using the existing images, he opted to process them using a histogram of oriented gradients algorithm which extracts the general shapes and "motion" of an image. While good in theory, the deployment proved too inaccurate and computationally expensive for a Raspberry Pi 4B, so it was back to the drawing board.
How it works
Rather than collecting all of the data upfront and training once enough images have been gathered, Olson's new solution opts for a continuous system. After drawing a box around the baby portion of a crib, the user is presented with a fresh classification model on the web app. By selecting either "yes" or "no" when asked if the baby is asleep, a new image will be grabbed from the camera's live feed and used to retrain the model, thus increasing its accuracy over time.
Off-the-shelf solutions
The market has several existing baby sleep monitoring solutions to pick from, but they normally range from $250 all the way to $400+ in price while offering less accuracy due to their more general dataset and specific camera placement requirements. And although Olson's approach needs some intervention up-front, the classification model quickly learns and keeps all of the training data on the local network. To see more about how Olson built the Baby Sleep Coach 2.0, you can watch his video here on YouTube or visit the GitHub repository here.
