Ground Penetrating Radar (GPR) is a Non-destructive Testing (NDT) method used to investigate subsurface conditions of civil engineering structures and locate buried objects using wideband electromagnetic (EM) pulse. The adoption of GPR to locate utilities has increased due to its ability to detect both metallic and non-metallic pipes. Further, the technology facilitates localization of the buried pipes with the support of signal processing steps and GPS coordinates. In this process, the presence of a pipe yields hyperbolae signatures on the B-scan. Thus, identification and localization of such hyperbolae is a vital step in the GPR signal processing towards 3D localization. For smaller GPR data sets, the human interpretation is sufficient to identify hyperbolae. However, in large-scale utility survey, precise and fast hyperbolae detection is required to accelerate the processing time and minimize human resource and costs. From the literature, several studies have been conducted previously to develop automatic hyperbola detection models based on physical methods and machine learning techniques. The performance of the models varied depending on the signal preprocessing, annotation strategy and machine learning algorithms. The common drawback of these existing models were higher false positives as any hyperbola formed by multiple reflection or other effects were also detected as true positives. Therefore, considering all pending challenges and advancement of deep learning techniques, we have proposed Faster Region-based Convolutional Neural Network (Faster R-CNN) automatic hyperbola detection models using two annotation strategies. The model has been numerically validated using 2D gprMax based on FDTD model, followed by validation on field data.