Combination of Infrared Thermography and Structure from Motion for detection of heat related building’s defects. A case study and an algorithm for automatic anomalies detection

Structural health monitoring is a field of study that embeds several scientific areas that are combined to properly inspect structures and evaluate their integrity. In this scenario, nondestructive testing (NDT) has become an efficient and reliable inspection tool able to detect flaws and anomalies, as well as material properties. Its nature of non-harassing the object under analysis, led it to be used in a broad variety of applications. At the same time, developments in computer vision made possible to reconstruct three-dimensional models of structures starting from a suitable set of images. The aim of this research is to present a novel inspection technique that exploits both nondestructive testing and computer vision for detecting heat-related defects in structures and infrastructures. In practice, the related work concerns the combination of infrared thermography (IRT) and structure from motion (SfM) in order to detect flaws and anomalies in a laboratory-scaled structure. IRT is a well-known nondestructive technique able to highlight and detect regions within the structure with different thermal properties by measuring electromagnetic radiations in the infrared spectrum, whereas structure from motion is a photogrammetric technique that allows to reconstruct 3D models starting from a suitable set of images. Not much work has been done towards this direction, therefore the aim of the study is to first, validate the theoretical reasonings of the presented technique, and then evaluate the obtainable accuracy of the results. For what concerns the experimental part, two different infrared thermography tests were performed; first using a hand-held thermal imager and then using an unmanned aerial vehicle (UAV) platform, embedding a dual-spectrum camera. The novelty introduced by the technique is the direct usage of thermal images to reconstruct three-dimensional virtual models, so that the output is a 3D thermal map that can be interrogated for remote inspections. In both cases, the set of images acquired are suitably post-processed and adjusted in a way that the reconstruction is as accurate as possible. To this end, it is possible to perform quantitative and qualitative analysis of possible damages and anomalies remotely, exploiting the reconstruction rather than the actual structure. In parallel with the aforementioned work, a computer vision algorithm for automatic damage detection was developed. First, temperature values are mapped in intensity pixels’ values of each image, then different image processing techniques are exploited to automatically detect heat-related defects starting from thermal images. By applying a cascade of different image processing techniques, the aim is to identify the contours of damaged areas and separate them from sound part of the image. Then, the detected contours are overlapped to the corresponding visible image, providing a straightforward result of the detection. The algorithm was tested both in a controlled environment application and in a real-world scenario, obtaining quite accurate results in both cases.