Ladar scan preprocessing for robust motion estimation

Two-dimensional laser radars (2D-ladars) are sensors extensively used in mobile robotics for map building, self-localization, and obstacle detection due to their accuracy and reliability. Due to their fast sampling of the environment they also perfectly suit incremental ego-motion estimation, that is, to find how the position of the vehicle changes in short periods of time only by comparing sensor readings. Some of the most accurate methods dealing with this problem rely on a consistent computation of derivatives of range scans provided by the ladar. In practice, edges in the environment and sensor noise lead to inconsistencies in this computation. In this work we introduce an approach in the frequency domain, which robustly detects the continuous contour patches in the scan and then filters out the noise in those sections. In contrast with other methods based on batches of filters and heuristic rules, our approach employs spectral information to automatically select the filter parameters. We validate our method with experimental results on real environments.