Developing an Efficient Real-Time Terrestrial Infrastructure Inspection System Using Autonomous Drones and Deep Learning
Abstract
Unmanned aerial vehicles (UAV), commonly referred to as drones (Dynamic Remotely Operated Navigation Equipment), show promise for deploying regular, automated structural inspections remotely. Deep learning has shown great potential for robustly detecting structural faults from collected images, through convolutional neural networks (CNN). However, running computationally demanding tasks (such as deep learning algorithms) on-board drones is difficult due to on-board memory and processing constraints. Moreover, the potential for fully automating drone navigation for structural data collection while optimizing deep learning models deployed to computationally constrained on-board processing units has yet to be realized for infrastructure inspection.
Thus, an efficient, fully autonomous drone infrastructure inspection system is introduced. Using inertial sensors, mounted time-of-flight (ToF) and optical sensors to calculate distance readings for obstacle avoidance, a drone can autonomously track around structures. The drone can localize and extract faults in real-time on low-power processing units, through pixel-wise segmentation of faults from structural images collected by an on-board digital camera. Furthermore, proposed modifications to a CNN-based U-Net architecture show notable improvements to the baseline U-Net, in terms of pixel-wise segmentation accuracy and efficiency on computationally constrained on-board devices.
After fault segmentation, the fault points corresponding to the predicted fault pixels are passed into a custom fault tracking algorithm; based on a robust line estimation technique, modifications are proposed using a quadtree data structure and a smart sampling approach. Using this approach, the drone is capable of following along faults robustly and efficiently during inspection to better gauge the extent of the spread of the faults.