Real-Time Road Anomaly Detection (Raspberry Pi Edge AI)

This project implements an edge AI application that processes dashcam footage in real-time to detect and log road anomalies such as potholes, cracks, and obstacles. It is optimized for high-performance inference on Raspberry Pi 4/5 hardware.

---

🛠 Project Components

1. Core Application

• src/pi_inference.py: The main deployment script.
  • Implements a threaded video pipeline for smooth frame capture.
  • Uses TFLite Runtime for low-latency inference on ARM CPUs.
  • Features Automated Logging: Saves detected anomalies to anomaly_log.csv.
  • Features Evidence Capture: Automatically saves Bounding Box snapshots to the detections/ folder.
  • Custom Class Support: Configured for 7 anomaly classes (Cracks, Potholes, Repairs, etc.).

2. Optimization Utilities

• scripts/export_tflite.py: Converts training weights (.pt) to optimized edge formats.
  • Exports to TFLite Float32 (Standard) and includes logic for INT8 Quantization.
  • Scales input resolution to 320x320 for a ~4x speed boost on edge hardware.
  • Automatically manages file organization in the models/ folder.

---

📂 Project Structure

anomaly-detection/
├── data/               # Training datasets and configuration
├── detections/         # (Auto-generated) Snapshots of detected anomalies
├── models/             # Optimized model files
│   ├── best_road_anomaly.pt         # Original training weights
│   └── best_road_anomaly_float32.tflite # DEPLOY THIS to Pi
├── scripts/            # Optimization and data scripts
│   ├── export_tflite.py     # PT -> TFLite conversion script
│   ├── sync_labels.py       # Data preparation utility
│   └── ...                  # Other helper scripts
├── src/                # Source code for inference
│   ├── pi_inference.py      # MAIN INFERENCE SCRIPT
│   └── inference.ipynb      # Development/Testing notebook
├── DEPLOYMENT.md       # Pi setup and optimization guide
├── README.md           # Project overview (this file)
└── anomaly_log.csv     # (Auto-generated) Detection history

---

🚀 Quick Start (on Raspberry Pi)

1. Install Dependencies:

   pip install tflite-runtime opencv-python numpy

2. Run Inference:

   python src/pi_inference.py

3. View Results:

   • Watch the live feed with detection boxes.
   • Check anomaly_log.csv for timestamped entries.
   • Review detections/ for captured JPEG evidence of road damage.

---

📈 Optimization Stats

• Target Resolution: 320x320
• Target FPS: ≥ 5-10 FPS (on Pi 4)
• Model Size: ~10MB (Float32) / ~2.5MB (INT8)
• Classes: Alligator Crack, Box Crack, Longitudinal Crack, Other Corruption, Pothole, Repair, Transverse Crack.