Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Real-Time Point-Cloud Summarization
The goal of this project is to develop a method for real-time summarization of point-clouds using geometric primitives or repetitive structures, such as planes, cylinders, and more complex manifolds. This project has as input point-clouds generated either by ground robots or Unmanned Aerial Vehicles
Keywords: Fitting, Clustering, Point Cloud, Optimization, Robotics,Drones, Computer Vision,UAV
With the release of affordable 3D (or 2.5D) sensors, such as structured light sensors, like the Microsoft Kinect, a variety of real-time reconstruction algorithms appeared, for example, KinectFusion. Most of these algorithms are able to create high-fidelity reconstructions of potentially large environments, e.g., Kintinuous, and CHISEL. But most of these algorithms are lacking in understanding what they are reconstructing. Therefore, we would like to take a first step into this direction by creating a reconstruction pipeline that forms reconstructions based on geometric primitives or repetitive structures, similar to the work 'Dense Planar SLAM' [Salas-Moreno et al. ISMAR 2014].
In inspection tasks, such a system would allow performing reconstructions and simultaneously checking if the environment is, as expected, built from locally smooth shapes. If large deviations from such smooth surfaces are encountered, specific regions could be autonomously further inspected.
This project has as input point-clouds generated either by ground robots or Unmanned Aerial Vehicles (UAVs). The ground robots will be inspecting indoor areas while the UAVs will be inspecting outdoor scenarios, such as facades and wind turbines.
The goal of this project is to develop a method for real-time summarization of point-clouds using high-level structures, such as planes, cylinders, and more complex manifolds. Drawing inspiration from similar methods in the literature, the objective here is to develop an algorithm for extracting surfaces from a point-cloud.
The outlook for the results of this project is that the developed algorithms are employed on our robots. As these robots can be extremely fast and agile consistent image registration and accurate point cloud generation is not always possible. The aim is that by approximating these point clouds with different shapes where appropriate, we can reduce complexity of the map maintenance and boost the robustness to outliers.
With the release of affordable 3D (or 2.5D) sensors, such as structured light sensors, like the Microsoft Kinect, a variety of real-time reconstruction algorithms appeared, for example, KinectFusion. Most of these algorithms are able to create high-fidelity reconstructions of potentially large environments, e.g., Kintinuous, and CHISEL. But most of these algorithms are lacking in understanding what they are reconstructing. Therefore, we would like to take a first step into this direction by creating a reconstruction pipeline that forms reconstructions based on geometric primitives or repetitive structures, similar to the work 'Dense Planar SLAM' [Salas-Moreno et al. ISMAR 2014].
In inspection tasks, such a system would allow performing reconstructions and simultaneously checking if the environment is, as expected, built from locally smooth shapes. If large deviations from such smooth surfaces are encountered, specific regions could be autonomously further inspected.
This project has as input point-clouds generated either by ground robots or Unmanned Aerial Vehicles (UAVs). The ground robots will be inspecting indoor areas while the UAVs will be inspecting outdoor scenarios, such as facades and wind turbines.
The goal of this project is to develop a method for real-time summarization of point-clouds using high-level structures, such as planes, cylinders, and more complex manifolds. Drawing inspiration from similar methods in the literature, the objective here is to develop an algorithm for extracting surfaces from a point-cloud. The outlook for the results of this project is that the developed algorithms are employed on our robots. As these robots can be extremely fast and agile consistent image registration and accurate point cloud generation is not always possible. The aim is that by approximating these point clouds with different shapes where appropriate, we can reduce complexity of the map maintenance and boost the robustness to outliers.
- **WP1**: Research into existing works tackling shape extraction from point-clouds.
- **WP2**: Implementation of a pipeline for fitting of a set of known shapes.
- **WP3**: Experimentation and evaluation of this method in terms of runtime and accuracy of estimation, with respect to the state of the art.
- **WP4**: Further optimization of the pipeline of WP3 to work with challenging real data.
- **WP5**: Final evaluation of the methods and report writing.
- **WP1**: Research into existing works tackling shape extraction from point-clouds. - **WP2**: Implementation of a pipeline for fitting of a set of known shapes. - **WP3**: Experimentation and evaluation of this method in terms of runtime and accuracy of estimation, with respect to the state of the art. - **WP4**: Further optimization of the pipeline of WP3 to work with challenging real data. - **WP5**: Final evaluation of the methods and report writing.
- Highly motivated.
- C++ programming experience.
- Experience with Computer Vision, Linux, ROS are advantageous.
- Highly motivated. - C++ programming experience. - Experience with Computer Vision, Linux, ROS are advantageous.
Interested student please contact Lucas Teixeira (lteixeira@mavt.ethz.ch) with cc to Fadri Furrer ( fadri.furrer@mavt.ethz.ch).
Interested student please contact Lucas Teixeira (lteixeira@mavt.ethz.ch) with cc to Fadri Furrer ( fadri.furrer@mavt.ethz.ch).