Baby Image Segmentation

2020 UPDATE: Cradlewise has been featured in TIME magazine’s top 100 inventions of 2020! Congratulations to the excellent engineering and design teams at Cradlewise.

This project was through an internship with a start-up called Cradlewise. Their flagship product is an all inclusive smart cradle. One of the features which I contributed to is an unsafe position detector.

When an infant rolls over on its side, a dangerous position may arise where the baby may accidentally suffocate. The goal was to flag such dangerous positions so that an alarm can be raised.

To classify safe and unsafe positions, I explored image segmentation based approaches using the openCV library in python. For example the watershed algorithm – which assigns the same segment to regions of low gradient.

The challenge in this project was integrating together information available in two forms: A Depth Image (zFrame) and an Intensity Image (iFrame)

Looking for circles in the intensity Frame (iFrame)

Another non segmentation based method was implemented which followed the sequence: Look for circles corresponding to the Head, Torso and Legs –> Check the alignment between the centers of the three circles –> If mis-aligned beyond a certain threshold, declare unsafe.

• A rendered version of a typical depth frame (zFrame)

So in summary I,

• Devised methods to segment body sections in baby images and establish safe and unsafe situations based on the relative positions of the segments associated with the head and torso.
• Utilized three dimensional point-cloud information to improve flagging accuracy.