Identifying Correspondences in Sparse and Varying 3D Point Clouds using Distinctive Features

In a wide range of applications stereo systems are used to extract geometric information from the scene observed with the stereo cameras. One possible solution to reconstruct the motion of such a system is to establish correspondences between points of the point clouds generated from stereo matching of image features at different epochs. There exists a large variety of approaches to establish correspondences between image or 3D data. A special group of algorithms, mostly inspired by the work of Lowe (2004), is based on the notion of distinctive feature descriptions. These algorithms assume the existence of a dense neighbourhood changing not too much over time. But the prevalence of untextured regions or computational constraints hindering the use of computationally expensive dense stereo matching approaches often result in only sparse point clouds and thus these approaches cannot be used for the registration of sparse 3D data. In our work we present a new approach that uses the basic principles of distinctive feature descriptions and extends them in a way that they can be applied to identify corresponding points between sparse 3D point clouds. Furthermore, an evaluation is given investigating the advantages and limitations of our approach. The results clearly show the effectiveness of the presented distinctive features to establish point matches between sparse 3D point clouds.