Updated: Program (PDF)

Keynotes

• Shiyu Song (Baidu, China)
  High Definition Mapping, Localization and Self-Driving Cars at Baidu

  Shiyu Song, PhD, is senior research scientist and technical head of the mapping and localization team at the Baidu Autonomous Driving Business Unit. Before joining Baidu, he was a research scientist at NEC Labs America where he worked on visual SLAM, a vision based vehicle localization system for autonomous driving. He joined Baidu in 2014. He is one of the founding team members of the Baidu autonomous driving car project. His research interests include map reconstruction, structure from motion (SFM), SLAM and localization.

  In his presentation, Dr. Song will discuss the development history of the Baidu self-driving car, the Baidu Apollo platform for developers, the techniques behind the Baidu HD Map and their multi-sensor fusion based localization system. Dr. Song will present a robust and precise localization system that achieves centimeter-level localization accuracy in varying city scenes. This system adaptively uses information from complementary sensors such as GNSS, LiDAR, and IMU to achieve high localization accuracy and robustness in various challenging scenes, including urban downtown, highway, tunnel and so on. Both the Baidu HD Map products and localization system have been deployed in a large autonomous driving fleet, and make it possible that Baidu vehicles operate in a fully autonomous way in crowded city streets at a daily basis.

• George Vosselman (ITC, University of Twente, Netherlands)
  Deep learning for the extraction of DTMs from point clouds generated by dense matching

  George Vosselman obtained his MSc degree in geodetic engineering at the Delft University of Technology, the Netherlands, in 1986 and his PhD degree at the University of Bonn, Germany, in 1991. From 1987 to 1992 he was Researcher at the University of Stuttgart, Germany. In 1993 he was appointed Professor in Photogrammetry and Remote Sensing at the Delft University of Technology. Since 2004 he is Professor in Geo-Information Extraction with Sensor Systems at the University of Twente, Enschede, the Netherlands. His research interests include information extraction from point clouds and imagery, 3D building and landscape modelling, quality analysis of point clouds, and sensor integration and fusion.

  Existing algorithms for Digital Terrain Model extraction still face difficulties due to data outliers and geometric ambiguities in the scene such as contiguous off-ground areas or sloped environments. We postulate that in such challenging cases, the radiometric information contained in aerial imagery may be leveraged to distinguish between ground and off-ground objects. The proposed method first applies morphological filters to the Digital Surface Model to obtain candidate ground and off-ground training samples. These samples are used to train a Fully Convolution Network (FCN) in the second step, which can then be used to identify ground samples for the entire dataset. The proposed method harnesses the power of state-of-the-art deep learning methods, while showing how they can be adapted to the application of DTM extraction by (1) automatically selecting and labelling dataset-specific samples which can be used to train the network, and (2) adapting the network architecture to consider a larger surface area without unnecessarily increasing the computational burden. The method is successfully tested on three datasets, indicating that the automatic labelling strategy can achieve an accuracy which is comparable to the use of manually labelled training samples. Furthermore, we demonstrate that the proposed method outperforms two reference DTM extraction algorithms in challenging areas.