CNN-Based Sub-Surface Object Detection Using Ground Penetrating Radar

Abstract

Ground Penetrating Radar (GPR) has a wide range of applications such as scanning underground surface, locating utilities and detecting road damages by analysing the radargrams. Detecting sub-surface road damages is of great importance to the road maintenance authorities as it serves for monitoring of construction processes and helps in early detection of the damages leading to reduced repair costs. The road damages are detected by manual processing and require interpretation of domain experts. This often is too uneconomic for large scale application, therefore one way to solve this problem is to use an AI approach. In this work, this problem is addressed by developing a single-stage object detection system based on the YOLO series for detecting various patterns under the road surface including sub-surface damages. Advanced machine learning techniques like data augmentation and transfer learning are used to improve the detection results. We also present a model ensembling technique that can be used to combine multiple models for making better predictions. The ensemble helps in reducing the generalization errors and dispersion of predictions coming from the individual models. Experimental results verify that YOLO combined with model ensembling provides considerable performance improvements in comparison to the classical computer vision methods.