|
|
3D Mapping from Sparse LiDAR Data
Veľas, Martin ; Hofierka,, Jaroslav (oponent) ; Kaartinen,, Harri (oponent) ; Herout, Adam (vedoucí práce)
This work deals with the proposal of novel algorithms for sparse 3D LiDAR data processing, including the design of a whole mobile backpack mapping solution. This research was driven by the need for such solutions in the field of geodesy, mobile surveying, and the building construction. Firstly, there is a proposal of the iterative algorithm for reliable point cloud registration and odometry estimation from 3D LiDAR point clouds. The sparsity and the size of these data are overcome using random sampling by Collar Line Segments (CLS). The evaluation, using standard KITTI dataset, showed superior accuracy over the well known General ICP algorithm. Convolutional neural networks play an important role in the second method of odometry estimation, which processes encoded LiDAR data in form of 2D matrices. The method is able to run online, while the accuracy is preserved when only translation motion parameters are required. This can be handy when the online preview of mapping is required and the rotation parameters can be reliably provided by e.g. IMU sensor. Based on the CLS algorithm, mobile backpack mapping solution 4RECON was designed and implemented. Using the calibrated and synchronized pair of Velodyne LiDARS and the deployment of dual antenna GNSS/INS solution, the universal system, providing accurate 3D modeling of both small indoor and large open environments, was developed. Our evaluation proved that the requirements set for this system were fulfilled -- relative accuracy up to $5$~cm and the average error of georeferencing under $12$~cm. The last pages contain the description and the evaluation of another method based on the convolutional neural networks -- designed for ground segmentation of 3D LiDAR point clouds. This method outperformed the current state-of-the-art in this task and represents the way semantics cad be introduced into the 3D laser data.
|
|
|
3D Mapping from Sparse LiDAR Data
Veľas, Martin ; Hofierka,, Jaroslav (oponent) ; Kaartinen,, Harri (oponent) ; Herout, Adam (vedoucí práce)
This work deals with the proposal of novel algorithms for sparse 3D LiDAR data processing, including the design of a whole mobile backpack mapping solution. This research was driven by the need for such solutions in the field of geodesy, mobile surveying, and the building construction. Firstly, there is a proposal of the iterative algorithm for reliable point cloud registration and odometry estimation from 3D LiDAR point clouds. The sparsity and the size of these data are overcome using random sampling by Collar Line Segments (CLS). The evaluation, using standard KITTI dataset, showed superior accuracy over the well known General ICP algorithm. Convolutional neural networks play an important role in the second method of odometry estimation, which processes encoded LiDAR data in form of 2D matrices. The method is able to run online, while the accuracy is preserved when only translation motion parameters are required. This can be handy when the online preview of mapping is required and the rotation parameters can be reliably provided by e.g. IMU sensor. Based on the CLS algorithm, mobile backpack mapping solution 4RECON was designed and implemented. Using the calibrated and synchronized pair of Velodyne LiDARS and the deployment of dual antenna GNSS/INS solution, the universal system, providing accurate 3D modeling of both small indoor and large open environments, was developed. Our evaluation proved that the requirements set for this system were fulfilled -- relative accuracy up to $5$~cm and the average error of georeferencing under $12$~cm. The last pages contain the description and the evaluation of another method based on the convolutional neural networks -- designed for ground segmentation of 3D LiDAR point clouds. This method outperformed the current state-of-the-art in this task and represents the way semantics cad be introduced into the 3D laser data.
|
host ::
RSS.