Image processing tasks hold an important position in machine learning research today. Photogrammetric pictures of the physical environment contain high-level information about surrounding objects and, therefore, get processed in a variety of applications, like autonomous driving systems or medical examinations. The construction industry as well exhibits high potential of digitization, standardization, and automation of processes. This thesis proposes an approach for automated object detection and segmentation on construction site photos, based on a Convolutional Neural Network (CNN). The presented method allows a pixel accurate identification of objects like worksite elements on images of building sites. It offers an implementation of the Mask R-CNN for the application in construction monitoring that allows almost real-time processing of photos or videos. The gained information can be used for the analysis of construction progresses and for quality assurance, for instance in the course of automated monitoring tools like progressTrack. Based on a dataset of construction site photos captured by unmanned aircraft vehicles (UAVs), the network is trained to segment separate building elements, in the course of this thesis realized for formwork elements. The results contain a classification of the instances, a localization in the form of bounding boxes and a pixel by pixel segmentation of every detected object. It is implemented in Python, using a TensorFlow backend and the ResNet-101 and FPN backbone architecture of Mask R-CNN.     «

Image processing tasks hold an important position in machine learning research today. Photogrammetric pictures of the physical environment contain high-level information about surrounding objects and, therefore, get processed in a variety of applications, like autonomous driving systems or medical examinations. The construction industry as well exhibits high potential of digitization, standardization, and automation of processes. This thesis proposes an approach for automated object detection an...     »