Executive Summary
This project develops a high-performance Edge AI application designed to detect and log road anomalies such as potholes, cracks, and obstacles in real-time. Optimized for Raspberry Pi hardware, it serves as a robust tool for automated infrastructure auditing and vehicle safety.
Methodology
The system utilizes the latest YOLOv11n architecture, trained on the RDD2022 dataset. To ensure deployment at the edge, we implemented several optimization techniques:
- PT -> TFLite Conversion: Switched from heavy PyTorch weights to mobile-optimized TFLite.
- INT8 Quantization: Reduced model size from 10MB to 2.5MB with minimal accuracy loss.
- Threaded I/O Pipeline: Decoupled frame capture from inference to maintain consistent frame rates.
- XNNPACK Optimized Kernels: Maximized ARM CPU utilization on Raspberry Pi 4/5.
Performance Results
| Class | Precision | Recall | mAP50 |
|---|---|---|---|
| Alligator Crack | 0.592 | 0.512 | 0.543 |
| Longitudinal Crack | 0.658 | 0.602 | 0.660 |
| Other Corruption | 0.697 | 0.745 | 0.769 |
| Pothole | 0.634 | 0.404 | 0.476 |