My topic was fall detection (as in elderly people falling) specifically without using cameras or wearables. The idea was to take the CSI (basically what you see in the image) and just stuff it into some machine learning model to get a prediction as to whether someone fell in a given time frame, so I was trying to classify the signature of the falling "activity". From my literature survey, this has been done successfully with CSI. But as with a lot of research, it typically lacked practicality. Much of my work was implementing the firmware, data recording, processing, and so on. I also had to record a ton of falls (ouch) and label them. I ended up throwing away the CSI approach though, because of the noise reasons I mentioned. That was simply a deal breaker. I went with FMCW radar instead (and it worked pretty good).