Bedtime Story: Deep Learning Baby Monitor Keeps an Eye on Your Crib

At six months old, BabbyCam Lead Test Engineer Elise Lui is already hard at work putting deep learning technology through its paces.

It’s technology that her father, NVIDIA engineer and BabbyCam founder Benjamin Lui, has harnessed to give fellow parents peace of mind.

Like most parents, Lui was worried about his new family member’s safety, not least when his daughter would be sleeping by herself in her room.

“I realized that we didn’t have the baby monitor and we needed one, so I started researching,” said Lui. “There are all kinds of gadgets that track the baby’s breathing, the pulse, and then I stumbled upon Sudden Infant Death Syndrome (SIDS) — which was terrifying.”

Bringing up Baby with Deep Learning

Studies have shown that the risk of SIDS increases in infants who sleep on their stomachs, but it’s not always possible for parents to monitor their infant’s sleep. As an engineer in search of a solution, Lui’s instinct was to turn to technology. Using NVIDIA DIGITS, the Caffe deep learning framework and NVIDIA Tesla GPUs, Lui started building BabbyCam, a baby monitor based in deep learning.

Initially, Lui trained the machine to determine if a baby was present in its crib, developing its recognition ability by feeding it thousands of downloaded images. He then trained it to distinguish whether the baby was on its stomach, and from there created the activity labels of “no baby,” “baby asleep,” “baby awake,” “baby crying” and “face covered.”  Parents can enable email and text alerts for any of the latter four.

Out of the box, BabbyCam labels a baby’s activity with a percentage of confidence that it’s correct. If users choose the option to enable deep learning on their device and surveil the user’s baby specifically, the device’s accuracy improves rapidly.

Beyond BabbyCam’s core function, Lui’s favorite feature is its time lapse history, which allows him and his wife to view past images of their daughter sorted by time and date. The time lapse history’s algorithm removes repeated images, providing parents with the highlights of their baby’s activities.

Lui is also working on advanced features made possible through deep learning, including recognition of facial emotions, different types of crying, and spit up.

However, the heart of BabbyCam is keeping babies safe. “I didn’t want to buy one of these expensive baby monitors that would track breathing and heart rate,” he said. “I wanted to catch these problems before they occurred. If the camera is smart enough to detect that your baby’s face is covered, you can catch that before scary stuff starts happening.”

BabbyCam is already popular with his family and friends with babies of their own. One reviewer wrote, “I hate to be so dramatic, but there are realistic situations where BabbyCam can save your baby’s life.”

BabbyCam is available for purchase at www.babbycam.com. It is compatible with iOS, Android, Windows and Mac.

Similar Stories

  • http://github.com/n2itn/rpiviz Zach Estela

    Does the caffe deep learning analysis + live training occur on the raspberry pi3, or is the photo buffer sent to a GPU enabled server?

  • Ben Lui

    The neural network is trained on GPU servers. The network is then deployed on the Raspberry Pi 3 and runs inference locally.

  • Bryan Del Rizzo

    Great story!