LiDAR-Based Localization on Highways Using Raw Data and Pole-like Object Features

Abstract

Poles on highways provide important cues for how a scan should be localized onto a map. However existing point cloud scan matching algorithms do not fully leverage such cues, leading to suboptimal matching accuracy in highway environments. To improve the ability to match in such scenarios, we include pole-like objects for lateral information and add this information to the current matching algorithm. First, we classify the points from the LiDAR sensor using the Random Forests classifier to find the points that represent poles. Each detected pole point will then generate a residual by the distance to the nearest pole in map. The pole residuals are later optimized along with the point-to-distribution residuals proposed in the normal distributions transform (NDT) using a nonlinear least squares optimization to get the localization result. Compared to the baseline (NDT), our proposed method obtains a 34% improvement in accuracy on highway scenes in the localization problem. In addition, our experiment shows that the convergence area is significantly enlarged, increasing the usability of the self-driving car localization algorithm on highway scenarios.