Unmanned aerial vehicle navigation in underground structure inspection: A review

摘要

Many years after construction, a number of existing old tunnels and underground structures are deteriorating with time as evidenced by cracks, large deformations, water leakage and so forth, which usually require regular site inspections to record their structural deterioration by taking high‐pixel, high‐overlap images along miles of a tunnel network. For complex underground structures (e.g., long tunnels and large caves), unmanned aerial vehicles (UAVs) may be adaptive in acquiring images at multiple heights and angles with low operational costs. So far, UAV underground structural health monitoring has become mature for open‐air surveying with rapid developments in robotic software and hardware. However, the UAV image acquisition for underground working conditions still faces a number of key challenges. This paper aims to provide an overview of UAV navigation techniques in confined dark spaces for geotechnical engineers, geologists, drone developers and other interdisciplinary researchers & professionals in the structural health monitoring field. It specifies the challenges for UAV application in underground space, mainly including lack of Global Navigation Satellite System (GNSS) signals, poor lighting conditions, weak features and obstacle avoidance and then followed by strategic solutions. For example, in light of poor GNSS signals, the fusion of multi‐sensors (e.g., laser imaging, detection and ranging (LiDAR) and multi‐cameras) can enhance localization accuracy in low‐luminance underground conditions. To address obstacle avoidance, computer vision (CV)‐based navigation algorithms (e.g., deep reinforced learning [DRL]) enable effective navigation in complex 3D spaces, but their adaptability is limited by arithmetic power and pre‐training needs. The review of relevant previous studies concludes that further development for UAVs in underground space inspection may focus on operation in large‐scale geometric inspection environments, obstacle avoidance, features and semantic recognition.