After having our residences robbed in the past year; and messing around with open pose, we thought an intrusion detection system which recognizes individuals based on limb length instead of facial recognition would be more robust.
What it does
After training on residents and friends of residents, our system can differentiate between intruders and people who belong. If an intruder is detected; the safetrek API is used to send an alert telling the user there is an intruder on their property. This system works independently of unique identifiers like faces, so ski masks would be seen through. We also detect the number of people in a room using openpose, as well as whether doors are open or not using a light sensor and dragon board.
How we built it
Used openpose to determine body proportions; and used this to train NN models to identify unique people. This is all served over a local tunnel, and displayed on a website built using SAP. We used a dragonboard and light sensor to detect if a door is open or closed.
Challenges we ran into
Heavy graphics meant lots of lag; was hard to get FPS up to a decent number. Setting up the dragonboard was extremely time consuming; and took the better part of a teammate's first day.
Accomplishments that we're proud of
Hustled to get aws to expedite our instance creations, and used the extra computing power + newly compressed video to allow for smooth video streaming and processing. We called amazon support over 5 times over the hackathon.
What we learned
Call amazon support if you need more computing power quick.
What's next for OpenSec
Things like gunshot detection, fire detection, fall detection, etc can all be very easily implemented (but not part of the MVP and thus not in this version of the product). All these events could very easily trigger emergency responses via safetrek as well.