Automated Detection of Manhole Covers in MLS Point Clouds Using a Deep Learning Approach

Abstract

Road manhole cover works as an important part of road construction. Timely detection can make a great progress in the development of road management. This paper proposes a rapid road manhole detection method using mobile LiDAR with state-of-the-art computer vision and deep learning techniques. Firstly, the road surface data is extracted from mobile laser scanning system(MLS). Then, the 2D geographic reference feature(GRF) images are formed from 3D point cloud. Finally, the object detector using deep learning technology was applied to locate and annotate the road manholes. Also, we adjusted the training model to present the better result with high confidence over 0.90. Compared with the previous method, the proposed method can correctly detect the manhole cover with higher rate of precision and FI-feature at 0.952 and 0.975 respectively, especially in the complex road situation.