EYE-On
Real-time person tracking across multiple cameras, even in crowds.
- Role
- Final Year Project
- Client
- FAST NUCES
- Timeline
- Dec 2024 – May 2025
- Type
- Computer Vision · AI · Open source
The problem
Tracking the same person across several camera feeds breaks down in crowded and occluded scenes, where identities get swapped or lost between cameras.
What I built
- 01
A vision pipeline: SAM2, YOLOv8, then ViT re-identification.Built a vision pipeline: SAM2 segmentation, YOLOv8 detection, then Vision Transformer re-identification.
- 02
Matched people across cameras and drew their trajectories.Matched identities across cameras and drew each person's trajectory.
- 03
React frontend where users upload video and pick who to follow.Built a React and Relay/GraphQL frontend with Firebase, where users upload video and pick targets to follow.
- 04
Flask and GraphQL backend running the PyTorch models.Built a Flask and Strawberry GraphQL backend with PyTorch 2.6, Hydra configs and HuggingFace Hub model loading.
Results
- 3
- Vision models in the pipeline
- 6
- Pipeline stages
- 1
- Identity per person across cameras
Outcome
Raw multi-camera footage becomes processed, visual tracking data with consistent identities across cameras.
Next project
Bouncehouse360
A multi-vendor marketplace for party rentals across major US cities.
