Manual parking sign surveys are expensive, slow, and error-prone. Our digital twin approach automates the entire process — from raw dashcam footage to structured, map-ready sign data — using a pipeline of vision language models and deterministic rules engines.
The Problem
Local councils across Australia manage thousands of parking signs. Keeping an accurate, up-to-date inventory of every sign — its restrictions, time conditions, applicable days, and precise location — has traditionally required teams of field workers manually recording each sign. This process is costly, inconsistent, and quickly outdated as signs change.
Our Approach
iRoadTech's parking sign digital twin platform takes a different approach. Instead of dedicated survey teams, we use the video footage that councils already capture during routine road inspections. Dashcam-equipped vehicles drive through the network, and our AI pipeline handles the rest.
1. Video Ingestion & Frame Extraction
Inspection videos are uploaded in standard formats (MP4, AVI, MOV, MKV) along with GPS trace data. Our system extracts frames at configurable intervals, then synchronises each frame with interpolated GPS coordinates — bridging the gap between 30Hz camera capture and 1Hz GPS logging.
2. Sign Detection
Each frame passes through a vision language model (currently Qwen3-VL) that identifies parking sign candidates. Detected regions are cropped and forwarded for detailed interpretation. The system handles multiple signs per frame, varying lighting conditions, and partial occlusions.
3. Sign Interpretation
Cropped sign images are processed by a configurable multi-VLM pipeline. We support Claude (Anthropic), GPT-4o (OpenAI), Gemini (Google), and Qwen (Alibaba) — selectable per deployment. The VLM output is then normalised and validated by a deterministic rules engine that enforces Australian parking sign standards (VicRoads TEM Vol 2 Part 211, AS1742.11).
The result is a structured JSON record for each sign panel, including:
{
"time_limit": "2P",
"time_range": "8:00am - 6:00pm",
"applicable_days": ["Mon", "Tue", "Wed", "Thu", "Fri"],
"arrow_direction": "left",
"special_conditions": [],
"confidence": 0.94
}4. GPS Mapping & Road Snapping
Each interpreted sign is placed on the map using its interpolated GPS coordinate, then snapped to the nearest road segment using the OSRM Nearest API with OpenStreetMap road network data. This ensures markers sit on actual streets rather than at raw GPS positions that may drift into buildings or footpaths.
Results
In our demonstration with streets around Monash University Clayton campus — including Beddoe Avenue, Marshall Avenue, Stockdale Avenue, and Koonawarra Street — the system successfully detected and interpreted parking signs with high accuracy, producing a complete interactive map of parking restrictions.
What's Next
We're expanding the platform to handle additional sign types beyond parking (speed, directional, regulatory), and working with more councils to validate the system at scale. Our goal is a comprehensive road sign digital twin for every council in Australia.