Inspiration
PLEASE WATCH VID ON 0.75x speed, we sped it up to meet the time criteria. Either way it sounds funny :)
Current natural disaster drone relief systems are manual, prone to error in high stress conditions, and hard to standardize due to design flaws and provides poor emergency relief response.
What it does
Provides an optimal semi-automatic drone system leveraging live data -> real decisions -> real impact loop. Feeds a drone system responding to a natural disaster.
How we built it
- Drones provide sensor telemetry + footage
- Perceptron Model classifies people AND estimates distance!
- Writes time-series data to TigerData Database
- Renders a 2D map of people in need
- Feeds Multi-variable optimization algorithm (~8x faster median response over brute force on
benchmarks) to decide Next Drone Decision - Drone decision feeds back to #1
Challenges we ran into
the chronological syncing of the cameras it tough, adding confidence intervals of position is difficult, diluting 3d position of cameras to 2d map is tough, optimization alg is advanced and benchmarking on REAL DATA is tough to nail down.
Accomplishments that we're proud of
Shipping an MVP, deriving an optimal alg w latency constraints.
What we learned
It is important to build a product that is not simply an instance of the tool (e.g. our first idea we spent ~3+hrs on was an allergen contamination tracker for restaurants) but rather adds technical complexity. We don't want to fall into the trap of a "GPT-wrapper"-esque idea.
What's next for Drone Perceptron Relief System
Semi-automatic -> fully automatic using Computer Vision similar to Zoox/Waymo cars.
- Infrastructure for scalable system
Log in or sign up for Devpost to join the conversation.