Automating construction-plan interpretation with computer vision
Each road construction site produces dense technical boards. To calculate them requires counting each panel by hand, format by format, then re-entering everything. Rather than a set of rules, we trained a vision model on the company's boards, so that it learned to read what its teams read.
The challenge
The signage boards combine traffic plans, standardized pictograms, visual cues and legends. Counting the panels by type and format required careful reading, subject to forgetting, before completely re-entering the order form.
Our intervention
Collection and annotation of the company's site boards, panel by panel, to cover all families and all formats encountered in production.
A detection model is trained on this corpus, then refined until it recognizes standardized pictograms regardless of the density of the board.
Each panel detected is classified by family and by dimension, from 600 × 600 to 900 × 900, and the total is calculated alone.
The order table is populated from the counts, ready to be verified and then sent to the supplier.
The operator validates the total before transmission and corrects if necessary in the table. The corrections feed into the following learning games.
The results
During an order session, 168 panels, including 165 orange, were identified and encrypted without re-entry. The time spent counting switches to verification, and each operator correction enriches the model for the next session.
Let’s discuss your project
Let’s talk about your AI project
Tell us what you need. Our team will contact you to clarify your objectives and next steps.
