Ground Penetrating Radar (GPR) data analytics
Description
Ground Penetrating Radar (GPR) is one of the most effective sensors used in geological surveys. It provides us with a non-destructive way of extracting crucial information about concrete and structures present under the road surface. This information can be used in many applications, including maintenance and monitoring of roads, and in autonomous driving. It uses electromagnetic waves that propagate in to the ground through transmitter in the antenna, and receives back the reflections generated from the underground surface and different materials. We are using GPR sensor in the scope of RADSPOT project for mainly two purposes, sub-surface damage detection and localization. The conventional sensor set is extended by a bi-frequency range GPR to detect damaged areas in the road substructure, and utilize the low frequency range to mark deep infrastructure as fingerprints for robust and high precision localization.
Keywords
GPR data analysis
Sub-surface damage detection
Object detection
Generative adversarial networks
Synthetic data generation
Contact
M.Sc. Rajat Mehta
Gottlieb-Daimler-Str. 65
67663, Kaiserslautern
Phone: +49 631/205-4045
Fax: +49 (0)631/205-4201
rajat.mehta(at)mv.uni-kl.de
Funding
BMVI, Bundesministerium für Verkehr und digitale Infrastruktur
Sub Surface Damage Detection
The research goal is to develop a GPR data analysis pipeline that starts from collecting raw GPR data which is then converted into radargrams as shown in the image above. The collected radargram images are then fed to the machine learning algorithms for detecting damages present under the road surface. Once the damages have been detected, we want our autonomous vehicle to avoid driving on top of these damages to prevent further deterioration of the roads. Over the last many years, deep-learning based methods have been dominating the task of object detection. There exist two main categories of detection algorithms, namely, two-stage and single-stage detectors. As the ability to run such algorithms in real-time is an important aspect of this project, we implemented a single stage detector named You Only Look Once (YOLO). It is an end-to-end trainable deep convolutional neural network consisting of over a hundred convolutional layers that deals with object detection as a regression problem i.e. straight from image pixels to bounding box coordinates and class probabilities. The developed algorithm runs at a speed of 25 frames per second.
Goals
- Develop real-time data collection pipeline
- Expand the dataset by applying artificial data augmentation techniques
- Analyse Ground Penetrating Radar (GPR) data for monitoring, and identification of different structural patterns under the road
- Detection of sub-surface road damages from GPR data using deep-learning based algorithms
- Once the damages have been detected, we want our autonomous vehicle to avoid driving on top of these damages
- Main focus is to develop robust algorithms capable of running in real-time, and evaluate them them in our test vehicle
Digital Twin
Synthetic Radargram Generation using Generative Adversarial Networks
In the last decade, autonomous driving has come into the focus of research, and the range of research topics has expanded enormously since then. The autonomous car of TUK is equipped with sensors, including a Ground Penetrating Radar (GPR) sensor. Radar grams are a non-intrusive way of identifying underground phenomenons. Due to the limitations in the GPR sensor coverage area, many problems arise while training the CNN-based models, which require plenty of training data to detect damages. Augmenting training images for machine learning tasks such as classification and identification has proven to be a reasonable approach to increase training data. In contrast to well-examined methods like linear interpolation and augmentation, generative models expand the output space beyond the given feature space. This work addresses two different problems the GPR sensor faces while recording the data, i.e., lack of unlimited tracks and missing data within the tracks due to maximum velocity of the car.
The first part of the work concerns generating similar radargrams utilizing a Generative Adversarial Network to estimate the real radargrams' probability density function such that unlimited tracks could be generated using a neural network (generator). Fig above depict the realness of the data generated by the developed WGAN model [1]. The second part of the work deals with the missing data within the images dealt with by Boundless GANs and image extrapolation concepts [2]. The Boundless architecture incorporates state of the art convolution operations to achieve seamless images and feature consistency. The experiments demonstrated that generating ground penetrating radargrams with Generative Adversarial Networks renders realistic samples. The Boundless architecture manages to produce seamless image extrapolations not distinguishable from real images. In the future the work will be extended by the condition specific scenarios.
Goals
- Generation of synthetic radargrams using DCGAN and WGAN for unlimited tracks of damages along the road
- Image extension using Boundless GAN and Image Extrapolation concepts for the generation of the missing data in radargrams along the road
- Develop optimized deep generative models for the generation of Ground Penetrating Radar Grams