| Summary: | With the continuous improvement of computer graphics and information technology, three-dimensional images have become a hot research topic in recent years, and point cloud registration is the most basic but important step in many applications such as 3D reconstruction. Point cloud registration can calculate the rotation matrix and translation matrix (rigid transformation) from two different point clouds with different views of same objects or scenes. Among the numerous methods, Iterative Closest Point (ICP) and its variants are the most popular ones, but they have problems such as tracking into local optima or having high computational complexity. The development of deep learning has provided new directions for point cloud registration. This dissertation mainly studies the Deep Closest Point (DCP) algorithm and applies it to the large-scale outdoor dataset KITTI. DCP maps the data into a high dimensional space, extracts the features by Dynamic Graph Convolution Neural Network(DGCNN) as well as the Transformer and finally uses SVD to calculate the rigid transformation. To reduce the computational complexity, a series of data preprocessing operations are performed on the point clouds before feeding them to the DCP model for registration, such as Random Sample Consensus(RANSAC), data clustering and key point extraction. The results of experiments show that the preprocessing steps can improve the accuracy of DCP, but this network is still unsuitable for the panoramic outdoor scenes.
|
|---|