ParkScope is a sophisticated computer vision system designed to detect available and occupied street parking spots in real-time. By leveraging YOLO11 and bridging the data scarcity gap with Generative AI (Stable Diffusion), ParkScope achieves high reliability in dynamic urban environments.
The project addresses the "Empty Spot" data deficiency by implementing a Synthetic Data Pipeline:
- GenAI Inpainting: Used Stable Diffusion to remove vehicles from KITTI dataset images, creating realistic "available" parking spot samples.
- Hybrid Decision Logic: The system cross-references direct "available" detections with "vehicle" detections to ensure maximum reliability and prevent false positives.
- Spatial Awareness: Trained with advanced augmentations including Perspective Transformation and Mosaic to handle complex camera angles and varying distances.
- High Accuracy: Achieved a mAP50 of 88.8% for available parking spot detection.
- Real-Time Inference: Optimized for high-speed processing, reaching ~1.0ms inference time on NVIDIA RTX 4070.
- Robustness: Successfully generalizes across different street layouts and lighting conditions.
(Example: Green boxes = FREE | Red boxes = BUSY)
code/: Core implementation scripts (Download, GenAI Data Gen, Training, Inference).data/synthetic_empty_slots/: The proprietary dataset of AI-generated empty parking spots.runs/detect/parkscope_final_model/: Production-ready model weights (best.pt).slides/: Project presentations and interim reports in PDF format.
# Clone the repository
git clone https://github.com/Lioryanwo/ParkScope.git
cd ParkScope
# Install required dependencies
pip install -r requirements.txt
To analyze a parking scene using the GenAI-enhanced model:
python code/final_parking_system.py
- Architecture: YOLO11 (Ultralytics)
- Generative AI: Stable Diffusion (Inpainting)
- Computer Vision: OpenCV, NumPy
- Dataset: KITTI Vision Benchmark + Custom Synthetic Data
To validate the model's reliability, we analyzed the training metrics across 100 epochs:
- Precision-Recall Balance: The model achieved a high area under the curve (AUC), demonstrating robust detection even in complex street scenes.
- Classification Accuracy: As shown in the Confusion Matrix, the system effectively distinguishes between 'Available' (84% accuracy) and 'Occupied' (88% accuracy) spots, minimizing false positives.
| Precision-Recall Curve | Normalized Confusion Matrix |
|---|---|
 |
 |