Vantage

Inspiration

As computer science and computer engineering students, we wanted to build something that made real use of the hardware in front of us. We got our hands on the NVIDIA Jetson Nano and set a goal: get our hands dirty with edge AI and aim it at a problem where on-device intelligence could have the greatest possible impact.

Disasters like earthquakes and wildfires were the obvious fit. Drones reach the scene before responders can, but the data they gather only matters if it arrives accurate, uncorrupted, and on time, and disasters are exactly when connectivity fails. A drone that depends on the cloud for its intelligence loses its value the moment the signal drops. We wanted a drone that works as the first set of human eyes on the scene, with or without the internet.

What it does

Vantage is an edge-AI layer for drones that turns aerial footage into an actionable search-and-rescue report, even with no connection.

Detects people on-device from an aerial view, running entirely on the Jetson Nano
Computes each person's latitude and longitude from the drone's position and the person's offset from it
Tracks movement over time, so responders can see where groups were and where they went
Switches to edge-device mode automatically when the connection drops, doing locally everything it would have done in the cloud
Syncs to the cloud once back in range, so no data is lost or corrupted
Generates a human-readable report, not just raw footage or a heat map: how many people were found, where (including nearby landmarks), which areas have the highest and lowest density, and how groups moved between observations
Powers a live web app with a density heat map built from the drone's coordinates, so teams know where to go first

Vantage is software, not a drone. It's designed to run on drones agencies already have.

How we built it

Hardware: NVIDIA Jetson Nano mounted on a drone [add drone and camera details]
On-device AI: [detection model] running locally for real-time person detection from aerial frames [add optimizations, e.g. TensorRT, quantization]
Geolocation: converting each detection's position in the frame into latitude/longitude using the drone's GPS, altitude, and camera geometry [adjust to your method]
Offline-first storage: detections and metadata are written to local storage, with a sync queue that uploads when connectivity returns [add DB or file format]
Report generation: [rule-based summarization / LLM / other] turns detections into readable statements about counts, density rankings, landmarks, and movement
Web app: [frontend/backend stack] renders the live heat map and report as data syncs in

Challenges we ran into

Limited compute: fitting real-time detection onto the Jetson Nano's modest hardware [add what you did about it]
Accurate coordinates: translating pixel positions into real-world latitude/longitude, which depends on altitude, camera angle, and GPS precision
Detecting people from above: small, partially occluded targets look very different from typical ground-level training data
Never losing data: handling dropped connections mid-flight and partial uploads, and making sure synced data arrives uncorrupted
Tracking movement over time: deciding whether detections across passes are the same people or new ones
Report quality: turning raw detections into statements responders can trust at a glance

Accomplishments that we're proud of

Running real-time AI entirely on-device, with no cloud dependency
A working offline-to-online pipeline that holds data through signal loss and syncs automatically
Producing a readable report, not just a heat map or raw footage
A live map that makes survivor density clear at a glance
Building a full hardware-to-web system as a team [add anything specific, like a live demo or a measured accuracy number]

What we learned

How to deploy and optimize AI models on edge hardware with tight compute and memory limits
Why offline-first design matters when the environment you're building for is the one where networks fail
How geolocation from aerial imagery works, and how much camera and GPS error matters
That in a crisis, a clear report beats a pile of data
[Add team-specific lessons]

What's next for Vantage

Better accuracy with models trained on aerial and disaster-specific data
Thermal and night imaging to find people in smoke, darkness, or rubble
Multi-drone coordination, so several drones can share and merge their maps
Mesh and relay syncing, so drones can pass data to each other when no internet is available
Integration with responder tools like dispatch systems and existing mapping software
Broader hardware support, so Vantage runs on a range of drones and edge devices