A training dataset photogrammetric simulation framework for neural networks rockfall identification

Abstract

This dissertation aims to address the critical challenges in monitoring geotechnical stability in quarry environments by leveraging advancements in photogrammetry and neural networks. It seeks to develop a robust photogrammetric simulation framework designed for the generation of synthetic training datasets, employed to train machine learning algorithms to detect and localize rockfall events. The research is set within the broader scope of automating safety measures in dynamic quarry settings, where traditional methods face considerable limitations in terms of labor intensity, data reliability, and adaptability. The focus is on the use of fixed stereo monitoring systems, which present a cost-effective but reliable method for continuous monitoring. However, these systems are not exempt from limitations, particularly regarding their operational reliability and the necessity for careful geometric network design. In order to achieve a reliable machine learning model, the dissertation explores the constraints posed by the availability and quality of training data. Traditional data collection involves intensive manual labeling and bears the potential for human error, thus prompting the need for synthetic datasets. These simulated datasets are generated through a carefully designed photogrammetric simulation framework, which accounts for the complex interplay of factors that contribute to noise in photogrammetric surveys. The simulation framework aims to create synthetic yet realistic datasets that encapsulate the intricacies of real-world conditions, from the noise induced by atmospheric conditions and equipment to the complexities introduced by camera calibration errors and occlusions. This synthesized data is intended to train neural networks in recognizing and predicting rockfall events, ultimately contributing to more comprehensive and automated risk assessment protocols. The research is conducted in collaboration with the Centre for Geotechnical Science and Engineering at the University of Newcastle, and it has broader applicability in the domain of civil structural monitoring. The methodology and findings could inform the development of automated monitoring systems not just in quarries but also in various infrastructural settings. Given its interdisciplinary approach, synthesizing advancements in photogrammetry, computer vision, and machine learning, this dissertation may enhance current practices in geotechnical monitoring, especially in environments that pose significant risks to human safety and operational integrity.