Ground Penetrating Radar Image Processing towards Underground Utilities Detection for Robotic Applications Evangelos Skartados 1 , Ioannis Kostavelis 1 , Dimitrios Giakoumis 1 , Alessandro Simi 2 , Guido Manacorda 2 , Dimosthenis Ioannidis 1 and Dimitrios Tzovaras 1 Abstract— During the last decades, the development of powerful novel sensor topologies, namely Ground Penetrat- ing Radar (GPR) antennas, gave a thrust in the modeling of underground space. Subsurface mapping is of particular importance for future robotic applications that aim to operate in underground space. However, the current processing methods of GPR images (B-Scans) towards detecting underground utilities is typically laborious, semi-automatic and prone to errors, es- pecially on cluttered subsurface sceneries that produce unclear signatures, and still require the manual annotation of experts. Due to lack of large scale datasets from such sceneries, the adoption of deep learning and model specific methods with increased repeatability is not feasible yet. In this scope, and working towards subsurface mapping for underground robotic applications, the paper at hand introduces an application specific method tailored to operate with realistic GPR data for the automatic detection of underground utilities, by applying a hyperbola detection framework on the radar images obtained from a surface GPR. To achieve this, a segmentation step on the GPR data is applied to prepare them for hand-designed feature extraction. The extracted representative features are used to train a SVM, while a final fitting step is applied to the detected hyperbola segments. The developed methodology has been evaluated on cluttered radar images that correspond to a subsurface area with dense buried pipes and exhibited remarkable performance. I. INTRODUCTION In the recent years, advances in GPR antennas allowed coverage of large scale distances with limited efforts from the operators [1], [2]. The latter, brought a new era in the non- destructive inspection techniques that significantly reduced the construction costs [3]. Yet, the research conducted so far in the domain of automatic analysis of GPR data did not share the same bloom, with the technological advancement of hardware itself. It is evident that realistic and massive subsurface data can be gathered shortly with synchronous arrays of GPR antennas, when the latter are integrated with autonomous robots. However, the analysis and annotation of such data still remains a time consuming procedure and introduces increased labor costs, since the development of a fully automatic procedure is a challenging task, and the biggest part of analysis is currently performed from human experts [4]. 1 Center for Research and Technology Hellas Information Technologies Institute, Thessaloniki, Greece, 57100 {eskartad,gkostave,dgiakoum}@iti.gr {djoannid, tzovaras}@iti.gr 2 IDS GeoRadar, Pisa, Italy,56121 alessandro.simi@idsgeoradar.com guido.manacorda@idsgeoradar.com At the same time, underground utilities detection through surface GPRs is highly important, having already a number of applications that share major significance, from non- destructive testing and understanding of the underground prior to constructions [5], through to safe mines recovery [6]. By giving a look into the future, the mapping of the underground space can be considered as a basic pre-requisite for next generation robots that will target underground operations [7]. In order for such robots to operate in the underground space, either autonomously or not, they need to be provided with a map of the underground that will enable their safe navigation. The latter demands the existence of recognition methods of traces, such as hyperbolic signatures, that indicate structured patterns of the existing infrastructure. Such methods enable automatic underground utilities detec- tion from GPR data. The formulation of hyperbolic patterns on the obtained GPR radar-gram (B-Scan) (see Fig. 1(a)) are strongly related to the propagation of the antenna pulse into a medium with certain electrical properties, as well as the size, the type of material, the shape and the orientation of the underground object with respect to scanning direction of the antenna. The great majority of existing works tackle the issue of hy- perbola detection as a classical computer vision problem [8], [9],[10]. The utilization of off-the-shelf solutions typically introduce tuning parameters that are not intuitively connected to the nature of the given task. Moreover, in the case of dense buried utilities, the performance of hyperbola detection in the respective radar-gram significantly degrades. It is evident that due to the nature of the problem, that requires great effort for the formulation of a proper test field infrastructure and given the rarity of such powerful GPR antennas, the construction of large datasets accompanied with consistent ground truth is a challenging task. Even if there are some of them, their disclosure to the research community is limited since there are not any public available large scale data for this problem. From a different point of view, GPR antennas integrated with robotic platforms produce large amount of data, the manual processing of which is not feasible. Given these factors, the developing of model based deep learning solutions as well as the adoption of global optimization techniques that require large datasets is not applicable in this problem. The paper at hand introduces a method that can be integrated in a robotic system for the automatic detection of hyperbolas by incorporating processing steps with in-depth physical meaning suitable to operate in cluttered realistic radar images. More precisely, initial zero mean value reduc- XXX-XX-XXX/XX/ $XX.XX (c) 2018 IEEE