TLS point cloud registration for deformation monitoring

Terrestrial laser scanning (TLS) and accompanying point cloud analysis techniques offer excellent potential for deformation monitoring in the context of natural hazards and structural health monitoring. While new point cloud–based deformation analysis methods, such as M3C2 and F2S3, have improved detection capabilities, they rely on accurately registered point clouds as input data. For this, the geodetic community heavily depends upon either target-based registration approaches or (variants of) ICP. In many real-world monitoring applications, none of these approaches works sufficiently well, and poor registration is consequently one of the main limitations of accuracy, sensitivity, and applicability. Simultaneously, the fields of robotics and computer vision have made significant strides in developing registration techniques by leveraging deep-learning approaches, particularly in the context of supporting autonomous driving. While not directly transferable, these advances offer great potential for improvements of registration also in the monitoring application case.

The proposed doctoral research aims to investigate the limitations of the approaches established in the geodetic community and to develop solutions that overcome the domain gap limiting the application of novel deep-learning techniques and allow reducing the monitoring uncertainty related to registration. The research will be divided into four work packages (WP). The first two will focus on evaluating the geodetic approaches, while the others will center on deep-learning approaches. WP1 will establish a benchmark using precise target-based approaches with an end-to-end automatic target detection and center point coordinate estimation workflow and implementation. WP2 will investigate the performance of commonly used registration approaches on large-scale, outdoor, and non-urban scenes. WP3 will introduce deep-learning approaches into the evaluation while investigating domain transfer. WP4 will provide synthetic TLS data for the training of deep-learning modules within WP1 and 3. The investigation results shall be documented in consecutive publications. Additionally, the developed algorithms will be shared as open-source software code, making the published results easily verifiable.