Pothole_Detection_Integration_Plan.md

Pothole Detection Integration (Research + Architecture)

1) Computer vision approaches for road damage detection

Typical successful families:

• Object detection (bounding boxes): YOLOv5/YOLOv8, Faster R-CNN

  • Pros: simple outputs, fast, easy UI.
  • Cons: boxes are coarse; struggles with thin cracks and complex shapes.

• Instance segmentation: Mask R-CNN, YOLACT

  • Pros: tighter region boundary and size estimation.
  • Cons: heavier models, more training complexity.

• Semantic segmentation: U-Net / DeepLabv3+ / SegFormer

  • Pros: best for pixel-level damage maps, severity estimation.
  • Cons: needs mask labels; inference cost.

• Two-stage pipelines:

  1. road surface ROI extraction (segment road), then
  2. damage detection inside road only
  • Pros: reduces false positives (buildings, shadows, non-road textures).

2) Datasets and pretrained models (starting points)

Common public datasets (road damage + potholes):

• RDD (Road Damage Dataset / Road Damage Detection)
  Includes potholes and other damage classes from multiple countries.

• Pothole-600 / Pothole datasets on Kaggle
  Smaller but useful for prototyping.

• CrackForest / CFD / other crack datasets
  More focused on cracks; can help pretraining for surface defects.

Practical approach:

• Use a model pretrained on COCO, then fine-tune on RDD/pothole datasets.
• For best results, fine-tune on your region-specific imagery (road texture and lighting differs).

3) Feasibility: drone vs satellite vs vehicle camera

• Vehicle camera (recommended):

  • Highest feasibility for potholes.
  • Typical resolution and perspective supports pothole features.

• Drone (good):

  • Works well at low altitude with good GSD (cm/px).
  • Requires flight plan and stable capture.

• Satellite (usually not feasible):

  • Most satellite imagery is too low resolution for potholes.
  • Only very high-res (sub-10cm) commercial imagery could work, and still hard due to shadows and angle.

4) Detection pipeline (integrated with current system)

Implemented integration strategy:

• Add Pothole Detection as another detection type in the menu.
• Route through the existing POST /api/detect using:
  • detection_type=pothole_detection
  • pothole_model=<selected>
• Separate engine module:
  • app/pothole_engine.py

Starter logic implemented (Rule-Based v1)

For fast CPU MVP (vehicle/drone imagery):

• shadow/dark region score (local brightness drop)
• rough texture score (Laplacian roughness)
• edge score (Canny)
• fuse + sensitivity percentile threshold
• region extraction + severity/confidence

5) Model architectures to implement next

• YOLOv8n (boxes) for fast detection and scalable deployment.
• SegFormer-b0 / U-Net for pixel-level damage mapping.
• Optional road ROI segmentation first to reduce false positives.

6) Immediate next steps for dataset preprocessing / feature extraction

1. Define input standard: camera height, FOV, resolution, and capture protocol.
2. Build a labeled dataset:
   • bounding boxes or masks for potholes,
   • metadata: day/night, wet/dry, shadows.
3. Add preprocessing:
   • road ROI extraction,
   • illumination normalization,
   • motion blur handling.
4. Train baseline YOLO model and integrate as pothole_model=YOLOv8.