This research work aims at applying the state-of-the-art 2D and 3D computer vision methods to civil engineering areas, including remote sensing, transportation and security surveillance. In this research, explorations on both applications and theory breakthroughs of convolutional neural networks (CNN) empowered computer vision techniques are made towards accurate, efficient and economic real-world applications that benefits both academia and the society. The first part of this research introduces LanDCNN, a 2D semantic segmentation CNN based method for landslide detection. Given remote sensing data like aerial images and digital terrain models, LanDCNN is able to generate accurate pixel-level detection results for both landslides and infrastructures. Compared with the conventional lands...[ Read more ]

This research work aims at applying the state-of-the-art 2D and 3D computer vision methods to civil engineering areas, including remote sensing, transportation and security surveillance. In this research, explorations on both applications and theory breakthroughs of convolutional neural networks (CNN) empowered computer vision techniques are made towards accurate, efficient and economic real-world applications that benefits both academia and the society. The first part of this research introduces LanDCNN, a 2D semantic segmentation CNN based method for landslide detection. Given remote sensing data like aerial images and digital terrain models, LanDCNN is able to generate accurate pixel-level detection results for both landslides and infrastructures. Compared with the conventional landslide detection methods, the proposed method is able to generate more continuous and semantically meaningful results. Meanwhile, LanDCNN can also return uncertainty maps regarding its output results to benefit the quality checking and revision tasks afterwards, so that a better human-machine interactive collaboration fashion can be achieved. The second part of this research focuses on the state-of-the-art 3D CNN based interpretation methods for point cloud. To address the sparsity and randomness issues of point cloud, a novel neural network operator named dynamic voxelization is proposed, which constructs convolution kernels only at the necessary locations of the point cloud on-the-fly during the network forward propagation, avoiding the computation power wasted on void space. The effectiveness of dynamic voxelization is verified on several benchmark datasets with satisfactory performance with high inference efficiency. Based on the dynamic voxelization, two deep learning models for 3D object detection are also proposed: 1. P2P-Net: Point-2D-Projection which combines the advantages from both 2D and 3D CNNs for rapid point cloud object detection on edge devices; 2. DV-Det: a two-stage object detection framework that is fully based on the 3D dynamic voxelization for point cloud object detection at large scale with high accuracy. Both of the two proposed algorithms have been evaluated on the standard benchmark datasets and deployed in the real-world applications with satisfactory performance.