The use of aerial robotics in the domain of civil engineering and infrastructure assessment has unlocked a very powerful domain in the Architecture, Engineering, and Construction Industry. As these infrastructures start to age, their structural integrity starts to deteriorate. Traditionally, the inspection of infrastructures is carried out manually, which can take days for normal infrastructures like a bridge and weeks for larger infrastructures like high-voltage powerlines. These applications demand methods that are scalable, rapid, and cost-effective solutions that can automate this process by understanding the information available in the scene. The cost effectiveness factor is defined to be the type of sensor and the expense of operation of an Unmanned Aerial Vehicle (UAV)-based inspection task. In comparison, LiDAR and depth sensors are readily available but have operational complexities like the payload and cost of the sensor. Current levels of autonomy involve complete control over the dynamics of an infrastructure inspection mission. Another implication is that traditional path planning methods require detailed information and insights of the infrastructure to be inspected. In time critical tasks like disaster response, such requirements can be problematic as the requirement of a predefined model is not guaranteed.
This research provides a solution to these problems by autonomously exploring a large 3D infrastructure object using UAVs equipped with an RGB Sensor. The autonomous exploration task is labelled as the Next Best View (NBV) problem. The autonomous exploration workflow estimates the NBVs required to cover an infrastructure object. A state-of-the-art NBV called MACARONS is selected in this research. The MACARONS-NBV is a self-supervised NBV prediction method that uses a monocular RGB camera to establish a 3D occupancy of an environment with infrastructure objects. The MACARONS-NBV architecture is designed in such a way to select a set of candidate camera views (NBVs) that maximize the surface coverage of an object. This trajectory is then assigned to a collaborative system of UAVs that coordinate with one another in a centralized communication channel to optimize their trajectories and avoid obstacles if encountered. The goal of the collaborative multi-UAV is to collect the infrastructure image data in an optimized manner. We utilize a set of methodologies to dynamically assign the workload of a UAV with awareness of their surroundings and any other UAV agents. The outcome of this is a network of UAVs that can simultaneously collect data for a large infrastructure object and can balance one another’s workload depending on the remaining energy level. We then use an offline image-based 3D Reconstruction workflow of the infrastructure object and assess the completion of the object using Cloud-to-mesh (C2M) and Cloud-to-Cloud (C2C) distances.
We were able to optimize the path of UAVs to efficiently traverse large scenes as well as demonstrate dynamic Task Reassignment for the collaborative aspect of UAVs. Additionally, we also conduct a detailed comparison of an autonomous UAV mission using LiDAR-based inspection compared to a monocular RGB-based with reference to the ground truth.