Transfer Learning-Based Crack Detection by Autonomous UAVs

Unmanned Aerial Vehicles (UAVs) have recently shown great performance collecting visual data through autonomous exploration and mapping in building inspection. Yet, the number of studies is limited considering the post processing of the data and its integration with autonomous UAVs. These will enable huge steps onward into full automation of building inspection. In this regard, this work presents a decision making tool for revisiting tasks in visual building inspection by autonomous UAVs. The tool is an implementation of fine-tuning a pretrained Convolutional Neural Network (CNN) for surface crack detection. It offers an optional mechanism for task planning of revisiting pinpoint locations during inspection. It is integrated to a quadrotor UAV system that can autonomously navigate in GPS-denied environments. The UAV is equipped with onboard sensors and computers for autonomous localization, mapping and motion planning. The integrated system is tested through simulations and real-world experiments. The results show that the system achieves crack detection and autonomous navigation in GPS-denied environments for building inspection.