GhostVision

Inspiration

It all started when I came across this crazy project where Wi-Fi routers were being used as makeshift cameras. That got me thinking-what if I could take that concept further using radar instead? No visuals, just pure presence detection, but cooler and more practical. As I explored possible use cases, I realized how valuable this could be in places where cameras and wearables fall short. It reminded me of real situations, like when my grandparents had fallen and no one knew until much later. That’s exactly the kind of problem GhostVision could help solve-providing safety without sacrificing privacy. And so, GhostVision was born.

What it does

GhostVision is a radar-based human sensing system that can detect presence and falls without using any traditional cameras or wearables. It turns raw radar signals into heatmaps and uses AI to track people’s movements in a space. If someone falls, the system detects it instantly and can send an alert. Think of it as a camera that respects your privacy-a ghost camera.

How I built it

I used an ESP32-S3 microcontroller paired with a KLC1a radar sensor to gather real-time data. I synced that with images for training, generating heatmaps using OpenCV and Python. The core detection model uses a U-Net architecture to isolate blobs representing human presence, while a 1D CNN model is trained to detect rapid movement drops as falls. All of this is processed and visualized on a local Flask app running on a Raspberry Pi-no cloud, no Wi-Fi dependency.

Challenges I ran into

• Syncing the radar frames and camera frames during data collection was tricky.
• Handling noise and spike artifacts in FFT outputs from the radar.
• Building an efficient fall detection model that works in near real-time.
• Getting consistent performance with low RAM and CPU usage on the Pi.

Accomplishments that I’m proud of

• Created a fully offline fall detection system with no cameras or wearables.
• Achieved blob detection with minimal false positives using radar-only data.
• Designed an intuitive heatmap interface that visualizes presence at ~15 FPS.
• Ran all AI models locally without relying on the cloud.

What I learned

• Radar data is messy but powerful-once cleaned, it can replace cameras in many use cases.
• Real-time edge AI is hard but doable with the right optimizations.
• Privacy-first technology can be both functional and futuristic.

What’s next for GhostVision

• Integrate fall alert SMS or push notifications.
• Expand training datasets with more movement types. Make it more accurate to body shapes.
• Deploy in assisted living centers for testing.
• Experiment with multiple radar units for multi-person tracking.

GhostVision isn’t just tech-it’s peace of mind, reimagined. Welcome to the future where privacy meets protection effortlessly.