The International Association for Automation and Robotics in Construction
Publications / 2025 Proceedings of the 42nd ISARC, Montreal, Canada
Drone-Assisted Progress and Quality Tracking for Surface Finishing in Construction
In the construction industry, achieving high standards in surface finishing is essential for structural integrity and aesthetic quality, yet traditional inspection methods can be slow and miss subtle defects. This paper introduces a drone-based system that utilizes advanced sensor processing for automated surface quality assessment in construction finishing tasks. The system integrates Light Detection and Ranging (LiDAR) and high-resolution RGB cameras to collect detailed surface data. The LiDAR module applies Gaussian filtering for noise reduction, followed by bicubic interpolation to generate a smooth and continuous elevation map, allowing for precise flatness estimation across the scanned surface. For visual defect detection, a Convolutional Neural Network based de-blurring algorithm is first applied to captured images to remove motion blur caused by drone movement. This process enhances image sharpness by passing the blurred image through multiple convolutional layers, where each layer reconstructs finer details, resulting in a deblurred image. Subsequently, a zero-shot transformer-based segmentation model uses multi-head self-attention to analyze spatial relationships within the deblurred image, allowing for effective identification and segmentation of defect regions without requiring extensive training data. Testing was conducted in both a ROS-based simulation and on real-life surface work. The LiDAR-based flatness estimation achieved a Root Mean Square Error (RMSE) of 0.2 mm in simulation and 0.8 mm in real-life tests, with Mean Absolute Error (MAE) values of 0.4 mm and 1.1 mm, respectively. The camera-based defect detection showed substantial improvement with the de-blurring module, increasing Intersection over Union (IoU) for defect segmentation to over 95% with de-blurring.
