Our Solution
FROM ROAD
TO DATA.
System Architecture
CAMERA
IMX296 global-shutter dashcam captures 1280×1088 frames at 30 fps
GPS
NEO-M8N receiver records location concurrently with every frame
EDGE PROCESSING
Raspberry Pi 5 + AI HAT+ (26 TOPS) preprocesses frames with OpenCV
DETECTION & SEGMENTATION
YOLO26n-Seg produces a per-pixel mask of damaged pavement
DEPTH ESTIMATION
YOLO26n-Depth estimates pothole depth from the same frame
SEVERITY CLASSIFICATION
Random Forest assigns each pothole a severity tier (Low <3cm, Medium 3–7cm, High >7cm)
WORK ORDER GENERATION
Automatic DPWH/LGU work orders dispatched via SMS or web
DASHBOARD
Pothole Watch surfaces detections, severity, and GPS for road personnel
Computer Vision
How Detection Works
AXIOM uses YOLO26n-Seg, an instance-segmentation model trained on Roboflow-annotated pothole imagery. Unlike plain object detection that draws a bounding box, YOLO26n-Seg produces a per-pixel mask of the damaged surface — so we know the exact area, not just the rough region.
A second model, YOLO26n-Depth, estimates the pothole's depth from the same frame. Together, the segmentation area and depth feed a Random Forest classifier that assigns each pothole a severity tier (Low, Medium, High).
The whole stack runs locally on a Raspberry Pi 5 with the AI HAT+ (26 TOPS) — no cloud round-trip is needed during the drive.
Location Intelligence
Mapping Road Damage
Each detection is tagged with GPS coordinates, allowing the system to plot potholes on an interactive map. This helps identify problem areas and prioritize repairs.
Dashboard
Road Condition Overview
Detected Potholes
—
Avg Severity
—
Recent Detections
—
Last 24 hours
Coverage Area
—
km monitored
Interactive Map
Real pothole locations will appear here