a lot has changed
Before you read
If you want the technical details you can skip this part. So it’s been a couple of months since I’ve updated the changes on my thesis. In fact, the semester is over . Anyway, A lot has changed, after the first couple of sessions I had with my thesis advisor helped me better focus my project; since I originally was going to receive 3 different signals from 3 different sources and I also had to process these signals and even maybe add machine learning to predict whether a person needs help or not. It’s a lot to cover in a couple of semesters and by myself. So instead, I decided to just focus on at least one aspect of my whole system. And of course I had to choose the one thing I know almost nothing about and that was computer vision. I settled with Fall detection via computer vision. It also doesn’t help that I have next to no knowledge at all about artificial intelligence, and I mean I barely even touched code or done a simple character recognition algorithm. But I wanted to challenge myself to force myself to learn something new. Anyway, I’ve come a long way since that decision and my advisor has been pretty helpful and has given me guidance on how I should tackle this problem. The first couple of weeks I spent researching the state of the art on fall detection algorithms based on computer vision. Down below is the process.
State of the Art on Fall Detection using Computer Vision
So like I said, I spent the first couple of weeks researching what were the top of the line algorithms or tendencies on fall detection. It was hard, especially considering how little I know about deep learning and similar topics. However, reading these papers helped a lot, down below is a sneak peak of all the papers I presented to my advisor before choosing a way to go. In the end we chose the research done by Berlin, et al. The reason being I used a kind of novel deep learning architecture Siamese Neural Networks. And by the looks on their results it was very promising. If you’d like to check out this google sheet to some research of your own the link is here.
The next step was to propose the model I need to develop. In my case my advisor suggested I made an SNN with an internal architecture based on transformer encoders. I barely knew at the time what that meant or for what purposes. After that, I researched a lot about the origin, an uses of transformers in AI, more importantly in computer vision.
I’d like to make a more dedicated post about Transformers, it’s a very interesting topic and I have a lot of different encoders to test for my thesis. For now I’d like to share my proposed architecture please feel free to let me know what you think. I’m pretty new at this even after 5 months.
Since I’m on a break for a few weeks I’ll try to post more often, I have more blogs about math and science in general if you’d like to see that. On the next posts I’ll be writing about embedded systems and my thesis, I’m trying to prepare for the next semester. Just a sneak peak I’m thinking on taking a course on digital communications, and if anyone could be so kind as to give me some tips I would appreciate it very much :)
