CIVL Alumnus Wins HKIE Ringo Yu Prize for Best PhD Theses in Geotechnical Studies
Dr. SU Zhaoyu, who graduated in 2021 with a PhD from the Department of Civil and Environmental Engineering under the supervision of Prof. Yu-Hsing Wang, has been awarded the annual Ringo Yu Prize for Best PhD Theses in Geotechnical Studies. The prize was established by the HKIE Geotechnical Division (GD) to recognize outstanding academic achievements in geotechnical studies.
Dr. Su’s thesis, titled “Advanced 2D and 3D computer vision for a smarter city: from image to point cloud”, aims to apply the state-of-the-art 2D and 3D computer vision methods to different civil and geotechnical engineering applications in developing smart and resilient cities, such as automated landslide mapping/inventory, and transportation/construction/crowd monitoring and management. The first part of this research introduces LanDCNN, a newly developed 2D semantic segmentation CNN-based method for automated landslide mapping. Given remote sensing data, such as aerial images and digital terrain models, LanDCNN can generate accurate mapping for landslides and nearby infrastructures at the pixel level. In addition, LanDCNN can also return uncertainty maps of the predicted results to facilitate the subsequent quality control and revision tasks that involve domain experts (humans) in the loop, aiming for active and continuous learning. The second part of this research focuses on the state-of-the-art 3D CNN-based interpretation methods for the point cloud data. A novel neural network operator named dynamic voxelization is proposed to address the sparsity and randomness issues of the point cloud data. Based on the dynamic voxelization, two deep learning models for 3D object detection are also developed: (1) P2P-Net: Point-2D-Projection that combines the advantages from both 2D and 3D CNNs for rapid point cloud object detection, suitable to be adopted for edge computing, and (2) DV-Det: a two-stage object detection framework that is entirely based on the 3D dynamic voxelization for point cloud object detection at a large scale with high accuracy.
