Guard was planned to be a program that accepts live video feeds and uses machine learning algorithms to detect drowning individuals.
You can see the code for Guard here.
Guard is able to currently accept videos and use them to identify (with decent accuracy) whether an individual is drowning or not.
Guard was put together at HackPrinceton in Spring 2017. It didn't work so well, though, because I lacked background in machine learning and it was difficult to put together the algorithm given the time constraints of the hackathon. It was also very difficult to find a good set of training data for the program to learn as there aren't many online repositories of drowning individuals.