Smart Parking from Existing Cameras

🚗 Inspiration

What really inspired this project is something simple:
I hate trying to find parking on campus.

There’s nothing more frustrating than circling a lot, thinking it might have space, only to discover it’s completely full. It wastes time, creates traffic, and adds stress to something that should be simple.

So I asked myself:

What can I build to fix this — not just for me, but for everyone?

Instead of adding expensive sensors or new infrastructure, I wondered if we could use what already exists: security cameras.

🏗 How I Built It

The core idea was to turn existing parking lot cameras into a smart availability system using computer vision.

1️⃣ Vehicle Detection

I trained a custom YOLO26 object detection model to recognize vehicles in parking lots. For each frame, the model predicts bounding boxes:

$$ \text{Detection} = (x, y, w, h, c) $$

Where:

$(x, y)$ is the center of the vehicle
$w, h$ are width and height
$c$ is the confidence score

Detections are filtered using a confidence threshold:

$$ c > \tau $$

to reduce false positives.

2️⃣ Zone-Based Space Estimation

Instead of manually labeling every parking spot (which doesn’t scale), I built a zone-based system.

Each parking area:

Has a defined capacity $N$
Automatically infers orientation from the zone geometry
Generates a virtual grid of parking spaces

If $D$ vehicles are detected inside the zone:

$$ \text{Open Spaces} = \max(0, N - D) $$

Vehicles are assigned to the nearest virtual space to estimate which spots are occupied.

This hybrid geometric + AI approach dramatically reduces setup effort while maintaining strong accuracy.

3️⃣ Smart Filtering

Parking isn’t just “full or not.” Some spots are:

Handicapped
Metered
Reserved

Users can filter by type, meaning someone can ignore lots that only have metered spots remaining.

Everything is displayed on a live, map-based dashboard for quick decision-making.

🧠 What I Learned

GPU power doesn’t matter if you're memory-limited.
HPC resource allocation can silently kill training jobs.
Early stopping is better than guessing the perfect epoch count.
Infrastructure reuse is often smarter than adding new hardware.
The hardest problems are often UX, not ML.

I also learned how to:

Train YOLO26 efficiently on an H100 GPU
Tune batch size and DataLoader workers in SLURM
Balance inference speed with detection quality

⚙️ Challenges

🔹 Limited Training Resources

Initial training runs were killed due to memory constraints. Increasing RAM allocation and adjusting worker settings solved the issue.

🔹 Proprietary Video Formats

The surveillance system exported .ave files, which had to be converted before processing.

🔹 Avoiding Manual Space Labeling

Most parking AI projects require labeling each space individually. I instead built a geometric reconstruction system using zone capacity and orientation inference.

🌆 Impact

This project demonstrates that smart cities don’t require new hardware — just smarter use of existing systems.

If cities can use cameras they already own to reduce congestion and improve accessibility visibility, we can make everyday frustrations — like finding parking — significantly better.

And it started with one simple thought: