frontend_view
test_data
test_data
training_graphs

🚀 HACKSMU 2024: Accessible Parking Detection

Inspiration

We were inspired by ParkHub’s prompt to create a system that helps detect accessible vehicles in parking areas using computer vision. We wanted to make parking more inclusive for people with disabilities by leveraging cutting-edge technology.

What It Does

ParkGuard is an AI-powered monitoring system that helps property owners find and correct ADA parking violations to ensure their properties remain accessible for all patrons. Our system uses two different AI models for both livestream and long-term monitoring (in practice, these feeds would come from on-premise cameras set up by owners and integrated with our help).

Using YOLOv11, our system detects the International Symbol of Access (ISA) in real-time video streams, identifying accessible vehicles by recognizing the ISA on license plates, windows, or body decals, allowing property owners to model traffic and violations in real time.

We also include a functionality that uses a multimodal model that intelligently detects and identifies obstructed parking spots and other violations (e.g., illegally parked vehicles), so that property managers can swiftly find and respond to these issues and ensure their lot is accessible as it can be for patrons of _ all _ abilities.

Processes at 300 fps (frames per second).
Outputs bounding boxes and confidence scores for ISA detections.

How We Built It

YOLOv11 and PyTorch for real-time object detection.
Used Flask for backend and HTML and CSS for frontend
Propel Auth for authentication
Used OpenCV for video processing.
Trained on a custom dataset, combining open-source data and self-recorded footage.

Challenges

Finding the right dataset was difficult, so we created our own.
Optimizing the model to run faster than 33ms per frame while maintaining accuracy was a significant challenge.

Accomplishments

Achieved 3ms per frame, far exceeding the required 33ms.
Successfully detected ISA symbols in real-time video with high accuracy.

What We Learned

We learned how to fine-tune YOLOv11 for specialized detection tasks and balanced speed, memory, and accuracy to optimize real-time processing.