PSI - Issue 78

Simone Felicioni et al. / Procedia Structural Integrity 78 (2026) 1285–1292

1288

particularly important in this early phase to ensure a high degree of accuracy, the reliance on manual inspection represents a scalability limitation for large-scale or frequent assessments. To address this, thanks to the modularity of the pipeline, we plan to integrate automatic damage detection algorithms based on deep learning and computer vision techniques in future extensions. These methods will enable the identification of relevant frames without human intervention, significantly reducing the time and effort required. 2.4. Visual Relocalization via Image-to-Point Cloud Matching To understand which regions of the structure are affected by the structural damage, it is essential to spatially localize the frames containing cracks within the 3D map. To accomplish this, we formulate the relocalization task as an image to-point cloud matching by using Patch-NetVLAD, a place recognition framework introduced by Hausler et al. (2021), which extends NetVLAD (Arandjelovic et al. (2016)) with a geometric consistency check via a RANSAC-based scoring procedure to increase matching robustness. Each frame collected by the MAV (denoted as query image) is matched against a database of images, aiming to retrieve the most similar image. The database is constructed by rendering 3D-to-2D projections of the point cloud from 36 viewpoints distributed throughout the digital replica of the environment. To improve the visual similarity between the RGB query images and the sparse database images, we apply an inpainting technique from OpenCV library and Gaussian blur to fill missing regions, reduce sparsity, and enhance visual uniformity of the images. Since each database image is associated with a known camera pose within the 3D environment, matching a query image effectively allows (i) determining the portion of the point cloud affected by the detected damage and (ii) associating the images of the same structural damage captured over time to a single POI, thus enabling temporal tracking and analysis of damage progression. 2.5. Immersive Visualization in Virtual Environment In the final stage of the proposed framework, the point cloud and the matched images are imported and integrated into an immersive virtual environment. Users can explore the digital twin of the structure through a VR headset, navigating the space as if conducting an in-person inspection. For monitoring purposes, interactive visual markers are placed at locations where cracks have been detected and relocalized within the point cloud. When a user selects one of these markers, a panel displays the images acquired over time from that location, enabling a temporal analysis of damage progression and supporting long-term monitoring. 3. Case study The proposed methodology has been applied to the Church of Sant’Andrea delle Fratte, a historic masonry structure located in the homonymous district of Perugia, Italy. The earliest records of the building date back to the 11th century, with substantial transformations occurring over the centuries, including a documented renovation in 1498, the construction of a detached bell tower in 1923, and comprehensive restoration works carried out between 2006 and 2009. The church features a rectangular floor plan, timber truss roofing, and mixed masonry walls, showing the typical complexity of heritage buildings affected by multiple construction phases. Several signs of cracking and material degradation are currently visible, partially attributed to superficial groundwater and long-term moisture-related issues. This site provides a representative case for testing immersive and automated structural monitoring techniques. The integration of Virtual Reality with drone-based visual inspection and 3D mapping enables non-invasive, remote assessment of damage progression, offering a valuable tool for proactive maintenance and conservation of architectural heritage.