Object recognition and video analysis on infants

Hello!

I have this project in which I need to analyse infants' movements from RGB and/or depth videos (they are taken by the same camera so as soon as I can with one video, I can track the pixels in the other one). The thing is that I can't find any dataset of images of babies that young (around 6 months old) to train a Neural Network or a Tree to recognize the infant*, so I thought of selecting the desired object (the baby, lol) from several frames of the video itself and then using that information to track him or her along the entire video.

I have no idea how to do that, though: I have thought of using Motion Analysis algorithms, or features extractions... but I am quite lost, to be honest.

Also, the child lies on a hospital cradle, which means that it doesn't move from that fixed area (good thing) but also that I can't use background removal techniques to segment the silhouette (I think! but I'm not so sure about that).

*infant silhouette is very different from an adult's, moreover they are lying and most datasets contain images of standing adult humans.

Thank you for any help you may provide!

Answer 1

[Raunak Gupta]

Hi,

As you mentioned the dataset is not available to get started, I would recommend labelling some of the images from data. This will help to train a classification network to correctly detect the baby in the image. However, for motion estimation and object(baby) tracking there are whole set of functions available here Tracking and Motion Estimation. If the frontal face is present in video frame, face detection can also be used as a way to track. Below are some references which can build an understanding in those workflows.

Face Detection and Tracking Using Live Video Acquisition

Face Detection and Tracking Using the KLT Algorithm

[Alessandro Cacciatore]

Thank you very much for you suggestions! I'll let you know as soon as I'll try them.