USING MACHINE LEARNING ON DEPTH MAPS AND IMAGES FOR TUNNEL EQUIPMENT SURVEYING

In this contribution, we illustrate how to use high-resolution 3D data and images to model the global shape of a tunnel and to survey its equipment, in a semi-automatic way using pattern recognition and machine learning techniques. We first implement a robust Bspline fitting algorithm, based on a parametric family of M-estimators that allows an efficient deterministic optimization strategy, to accurately model the tunnel lining using range data, despite the presence of acquisition artifacts and significant perturbations related to the equipment and some surface defects. The residual maps from the robust fit can be exploited to segment the equipment in an unsupervised manner using clustering algorithms, but at the cost of post-processing which makes the method rather ineffective for routine use. However, we deploy it to annotate data for the supervised learning of a deep learning model, namely Mask R-CNN. We comment on the first results, obtained on a still limited number of examples, and from image data only, and discuss the possibilities of improving the method, in the immediate or longer term.