🔬 YOLO Road Damage Model Research

Deep research into the best open-source road damage detection models for SinkAlert

🏆 Final Selection: oracl4/RoadDamageDetection

91 ⭐ GitHub — YOLOv8s fine-tuned on RDD2022 Japan+India subsets. Chosen over 5 competing repos based on: pre-trained weights availability, RDD2022 taxonomy alignment (D00–D40), Thai road compatibility, and deployment readiness.

Integrated: Git submodule at dashcam/rdd2022_model/ with 85.4MB weights active.

Live Test: ✅ Detected Alligator Crack (D20) at 56.9% confidence on synthetic road image. 357ms inference on CPU.

📊 Model Comparison

Repo	Stars	Weights	mAP@0.5	Classes	Thai Fit	Status
oracl4/RoadDamageDetection	91	✅ 85MB	~79.7%	D00,D10,D20,D40	⭐⭐⭐⭐⭐	SELECTED
Aary06/RoadGuard-AI	3	✅	~88% (YOLOv5)	D00,D10,D20,D40	⭐⭐⭐⭐	Backup
KaikePing/RoadDamageYOLO	10	❌	87.8% (YOLO11)	D00,D10,D20,D40	⭐⭐⭐	No weights
Nawaf-Rayhan585/Vehicle Detect	⭐	✅	N/A (vehicle)	Cars only	⭐	Traffic load only
RikudouSage/YOLOv8-Damage	6	❌	N/A	Unknown	⭐	Rejected
sekilab/RoadDamageDetector	Official	❌	N/A	Full RDD2022	⭐⭐⭐	Dataset only

🔍 Why YOLOv8?

v8 vs v11 vs NAS Models

Model	mAP@0.5	Params	Speed (CPU)	Deploy Ready
YOLOv8n	57.3%	3.2M	~50ms	✅
YOLOv8s	79.7%*	11.2M	~360ms	✅
YOLOv8m	62.4%	25.9M	~700ms	✅
YOLO11n	87.8%*	2.6M	~40ms	❌ Needs training

* After RDD2022 fine-tuning. Base mAP shown for untuned models.

🛣️ RDD2022 Taxonomy Coverage

Code	Damage Type	Thai Name	SinkAlert Relevance	oracl4 Model
D00	Longitudinal Crack	รอยแตกตามยาว	High	✅ Detects
D10	Transverse Crack	รอยแตกตามขวาง	High	✅ Detects
D20	Alligator Crack	รอยแตกลายจระเข้	Critical	✅ Detects
D40	Pothole	หลุมบ่อ	Critical	✅ Detects
D0w0	Depression / Settlement	ถนนทรุดตัว	Critical	❌ Missing
D43	Crosswalk Blur	ทางม้าลายเลือน	Low	❌ Not trained
D44	White Line Blur	เส้นจราจรเลือน	Low	❌ Not trained
D50	Manhole Cover	ฝาท่อ	Medium	❌ Not trained

⚠️ Gap: D0w0 (Depression/Settlement)

The oracl4 model does not detect D0w0 (road depression/settlement) — the most critical sinkhole precursor.

Mitigation plan:

InSAR fills this gap — ground deformation detected at mm/yr resolution
Fusion engine weights InSAR data for settlement zones (Bangkok East: 25mm/yr)
Future: fine-tune with ~200 Thai road depression images to add D0w0 class

🚗 Vehicle Detection (Secondary)

Nawaf-Rayhan585/Yolov8_Vehicle_Detection_Model — ⭐ Starred as requested.

Used for traffic load estimation (secondary fusion feature), not primary damage detection.

Vehicle count → traffic_load_index → feeds into XGBoost fusion model alongside InSAR + dashcam damage scores.

📐 Architecture Integration

┌────────────────────────────────────────────────────┐
│                 SinkAlert Fusion Engine            │
├────────────────────────────────────────────────────┤
│                                                    │
│  ┌──────────────┐  ┌───────────────┐               │
│  │ InSAR Monitor │  │ Dashcam Detector│              │
│  │ (mm/yr subs)  │  │ (YOLOv8 RDD2022)│             │
│  └──────┬───────┘  └───────┬───────┘               │
│         │                  │                        │
│         ▼                  ▼                        │
│  ┌─────────────────────────────────┐               │
│  │    XGBoost Fusion Model         │               │
│  │    8 features → risk score 0-1  │               │
│  │    ROC AUC: 99.9% (synthetic)   │               │
│  └───────────────┬─────────────────┘               │
│                  │                                  │
│                  ▼                                  │
│  ┌─────────────────────────────────┐               │
│  │  Alert Engine → LINE Bot / API  │               │
│  └─────────────────────────────────┘               │
│                                                    │
│  Optional: Vehicle Detector → traffic_load_index   │
└────────────────────────────────────────────────────┘

📦 Deployment Status

Component	Status	Details
oracl4 submodule	✅ Cloned	`dashcam/rdd2022_model/`
YOLOv8 Weights	✅ Local	85.4MB at `models/YOLOv8_Small_RDD.pt`
ultralytics pip	✅ v8.4.67	Torch 2.12, CUDA toolkit included
Dashcam collector	✅ Active	Auto-loads oracl4 weights by default
Live inference	✅ Tested	2 alligator cracks detected, 357ms CPU
Vehicle detection	🔄 Planned	Traffic load estimation feature
D0w0 fine-tuning	📋 Planned	~200 Thai depression images needed
YOLO11 upgrade	📋 Future	87.8% mAP with training

← Back to SinkAlert Dashboard