Initial Registration and Range of Convergence

Most registration algorithms use an iterative algorithm that ideally converges to the best multiple data set registration. Many algorithms only converge to a good solution if the initial relative pose estimate is sufficiently close to the optimal registration. So, one issue is how to find a good initial relative pose estimate. Further, different approaches have a wider range of tolerance about the optimal pose within which convergence occurs, so researchers are investigating methods for increasing the range and robustness. The main approaches can be classified as: 1) using special features or points for initial alignment (12,14,73,23,76,40,33,1) 2) special circumstances, such as properties of the registered objects or of the imaging device (74,19,6,59), 3) human assisted registration (54). It is interesting to note that even with good initialisation, algorithms may not find an optimal registration or even may not be able to refine the initial coarse registration. This is due to a large number of local minima due to noise, occlusion, appearance and disappearance of points, and general lack of knowledge about the distribution of points. Moreover, the distribution and characteristics of these local minima are dependent on image data of specific objects rendering it difficult to theoretically characterise these local minima. Consequently, finding an optimal solution is an unresolved issue and future research may have to focus on investigating trully general purpose registration techniques with a large convergence range without requiring good initialisation.