AI-Based On-Site Construction Quality Control: Experiences in Training AI for Constructionspecific Applications

Abstract

This study investigates the applicability of artificial intelligence in the construction industry, specifically focusing on the automation of on-site quality control through the training and fine-tuning of a YOLO-based segmentation model. The research presents the process of creating custom, construction-specific datasets within a Docker-based CVAT environment, highlighting the significance of the Human-in-the-Loop (HITL) methodology. To minimize human annotation time and prevent faulty feedback loops, the workflow was optimized using a custom Python script. By integrating the Douglas-Peucker algorithm and IoU-based filtering, this solution automatically smooths the fragmentation of AI-predicted polygons and eliminates duplicates. The results demonstrate that combining algorithmic cleaning of raw detection data with expert supervision is essential for developing reliable, industrial-grade AI solutions in the construction sector.