Urban Scene Extraction

What to extract

  • Urban scene features which include: Road, Lamp posts, Power lines, Fire hydrants, Road signs, and Building facades.

Why is it useful?

  • Urban planning and maintenance, 3D mapping/vision, Autonomous navigation.



Mobile Ground Based LiDAR Data

TITAN® from Terrapoint

  • A mobile platform that consists of: 4 LiDARs + 1 GPS + 1 IMU.
  • Accurate 3D data of the urban scene registered with respect to the GPS coordinate.

3D urban data collected using TITAN

  • A 500 X 500 m^2 data tile has an average of 5 million points.
  • Provides excellent sampling of building facades and some urban structures.



Proposed Methods for Automatic Extraction

Road Extraction

  • Road is modeled as a series of connected 3D planes, and Kalman filtering (KF) is used to predict and measure.
  • A small section of the road is approximated as a plane using RANSAC and Least Squares Fit.
  • Change in road direction and elevation are modeled as discrete 3D transformations between two 3D planes.
  • KF predicts the location of the next 3D plane using the transition matrix of the dynamic road model.
  • Measurement update of the KF is provided by estimating the local plane at the proximity of the predicted plane.

Posts and Power Line Extraction

  • Search for prominent near-road structures. Search is limited to 3D bounding boxes projected from the extracted road network.
  • RANSAC and best fit line are used to locate post structures; according to the thresholds provided by the true dimension of urban structures.



Results
  • Experiments on computer with 3GB of RAM and a quad-core 2.26 GHz processor.
  • 7 data sets were tested, the size for each data tile is 500 X 500 m^2.
  • Multiple seeds were used for road and curb (yellow or cyan) in some data sets.
  • All roads were successfully extracted. However, some lamp posts were mis-identified as power line posts.



Publications
[1]


J. Lam, K. Kusevic, P. Mrstik, R. Harrap, M. Greenspan. “Urban Scene Extraction from Mobile Ground Based LiDAR Data”, 3DPVT 2010: 5th International Symposium on 3D Data, Processing, Visualization & Transmission, Paris, France, May 17 – 20, 2010.
Copyright Queen's University 2009, 2010