Home Download Credits Publications Netbot Lab
Multilayer Feature Graph based Visual Navigation
Motivation
Modern visual navigation approaches mostly use homogeneous features (e.g. salient points) as landmarks. However, this choice cannot fully exploit the information possessed in the heterogeneous landmarks from man-made environments. As illustrated in the figure, there exist abundant lines in parallel directions and salient building facades in typical urban environments. These heterogeneous landmarks can help improve navigation performance if the geometric constraints between them are modeled and utilized properly.

This is also the source code release for the following paper:

Yan Lu and Dezhen Song, "Visual Navigation Using Heterogeneous Landmarks and Unsupervised Geometric Constraints", IEEE Transactions on Robotics (T-RO), vol. 31, no. 3, June 2015, pp. 736 – 749. [Download from IEEE]

Multilayer Feature Graph (MFG)
We propose Multilayer Feature Graph (MFG), an open data structure, to organize heterogeneous features and their geometric relationships for robot visual navigation. As illustrated in the figure, MFG consists of five types of features:
  • Key point : interesting points such as SIFT, FAST
  • Line segment: finite linear objects
  • Ideal line: infinite linear objects, fusion of collinear line segments
  • Vanishing point: organizing parallel lines
  • Primary plane: salient planar objects like building facades
and four types of geometric relationships:
  • Adjacency: between key points and ideal lines
  • Collinearity: between line segments and ideal lines
  • Parallelism: between ideal lines and vanishing points
  • Coplanarity: between key points/ideal lines and primary planes
Contact
Please contact Dezhen Song (dzsong@cse.tamu.edu) for questions.