3D Surface Registration on Embedded Systems

In this paper we present a fully embedded implementation of a 3D surface registration algorithm for robot navigation and mapping. The target platform is a multi-core digital signal processor, on which we achieved a significant speedup in comparison to the PC-based implementation. The main contribution of this paper is to show the potential of using multi-core DSP platforms for real-time capable implementation of computationally intensive tasks providing surface registration as application example. Surface registration describes the process of finding a rigid transformation between two sets of 3D points describing the same object, captured from different points of view. This is used to generate a complete 3D scan of an object if the transformation between the single scans are not known. One method for solving this problem is to leverage the assumption that both point clouds perfectly overlap. Minimization of the data term (eg distances between matching points) would then lead to the target transformation. Therefore the result heavily depends on the similarity of the datasets, which usually cannot be guaranteed due to (i) noise and outliers in the sensor data,(ii) only small overlap, and (iii) occlusions. Thus a more robust error metric is required with which a rigid transformation can be estimated iteratively. A well-known solution is using the iterative closest point (ICP) approach by Besl and McKay [1] as shown in Fig. 1. A and B represent point clouds of the same surface captured from different locations. At every iteration, for each point of A the nearest point of B is searched. Wrong correspondences can occur as can be seen at location b3. Using the …