01
Site Segmentation
A general-purpose SegFormer model, being adapted to construction, segments each site image into the main visible elements - earthworks, structure, facade, scaffolding, finishes. The area share of each is stored per image as a time series.
02
Progress Against Plan
Class areas are mapped to foundation, structure, facade and interior with weights calibrated per project, compared with the planned schedule and drawn as an S-curve with delay in days and a completion forecast. A project manager can override any figure, and the override is kept as such.
03
Safety Violation Detection
YOLOv8 flags missing hardhats and vests, people inside danger zones defined per camera, people close to machinery and lone workers. High-severity events reach the dashboard in real time and Telegram.
04
Worker Counting
Each processed image is counted for people, whether or not a violation was found, so the site gets a daily manpower curve and a peak figure per day.
05
Timelapse & Reports
Pick a camera, a date range and a frame rate and the archive becomes an MP4 timelapse. PDF reports combine progress, safety and site photos, with an optional written summary in Vietnamese or English.
06
Drone Defect Detection
A YOLOv8 segmentation model finds cracks, corrosion, spalling, deformation and missing bolts; ground sample distance turns pixel masks into length, width and area in millimetres.
07
Severity & Health Index
EfficientNet-B2 grades each defect as light, medium, heavy or dangerous. The grades roll up into a 0-100 health index per asset, so a bridge has one number to watch between inspections.
08
3D Model & Defect History (Roadmap)
Next steps: rebuild the structure from the flight as a point cloud and mesh you can orbit in the browser, place each defect on that model, and match defects with those found in earlier inspections, so a crack has a history instead of a new ID every time.
