Parallel Tracking and Mapping for small AR Workspaces

Recently I read a paper about how to map the future points and determine the space coordinate system as well as track an object from university of Oxford. Actually since my thesis I have looked for the approaches estimating the pose and feature points-image points correspondence simutaneously. Some good candidates are like SoftPOSIT, RANSAC and Blind PnP. However during the practical implementation, I have found that the convergence ratio and run time are always the stumbling block for the practical applications, especially for the small number of feature points (between 4 to 11). Although this PTAM method requires large number of points, too, this is really an efficient and robust algorithm which could be applied in real time. Here let's have a look at a demo video and feel the power of this approach.



From this video I think most of us can have the intuitive experience of this PTAM approach. This method is mainly used for the hand-held camera pose estimation and it makes use of the dual-core computer to carry out the tracking and feature points mapping simutaneously. What surprises me more is that it requires no prior map before the tracking and a frame with 660 successful observations just used 18 ms. The implementation of this method for mapping and tracking respectively is certainly complecated, especially for the mapping. But if such an amazing application runs on your cellphone, it is still worth while. And here is another video of PTAM applications on IPhone.




And the paper of this approach is here.

Posted in: Augmented Reality,Videos,Virtual Reality