Pixel Club: Approaches for Efficient Incremental Structure from Motion and Consistent Vision-based Single- and Multi-robot Navigation

Speaker:
Vadim Indelman (Georgia Tech)
Date:
Tuesday, 19.3.2013, 11:30
Place:
EE Meyer Building 1061

This talk will focus on efficient methods for single- and multi-robot localization and structure from motion (SfM) related problems such as mobile vision, augmented reality and 3D reconstruction. High-rate performance and high accuracy are a challenge, in particular when operating in large scale environments, over long time periods and in presence of loop closure observations. This challenge is further enhanced in multi-robot configurations, where communication and computation budgets are limited and consistent information fusion should be enforced. In this talk, I will describe approaches that address these challenges.

First, I will present an incremental and computationally efficient method for bundle adjustment that substantially reduces the involved computational cost, compared to state-of-the-art bundle adjustment techniques. The method, incremental light bundle adjustment (iLBA), incorporates two key components: the observed 3D points are algebraically eliminated, leading to a cost function that is formulated in terms of multi-view constraints instead of the projection equations, thereby reducing the number of variables in the optimization. While only the pose variables (or navigation states) are optimized, if required, the observed 3D points, or any part of them, can be reconstructed based on the optimized poses. The second component is the recently developed incremental smoothing approach, which uses graphical models to adaptively identify the variables that need to be recomputed at each step. The described method will be demonstrated in SfM and robot navigation scenarios. Next, I will continue by overviewing an approach to maintain high-rate performance also in the presence of loop closure observations, since these cannot be guaranteed, in the general case, to be processed at a sufficiently high rate. The approach parallelizes computations by partitioning the underlying graphical structure of the problem at hand.

The second part of the talk will focus on distributed multi-agent localization and navigation. I will present approaches that exploit commonly observed 3D points to both perform cooperative localization and to extend sensing horizon of the robots in the group. A special consideration will be given to consistent information fusion, while minimizing communication between the robots and the computational burden required for incorporating the information obtained from other robots. Two methods will be described: sharing the estimated distributions of observed 3D points, and sharing the actual (image) observations of these points, thereby extending the iLBA approach.

Back to the index of events