Semantic Assisted Loop Closure Detection for Automated Driving

LiDAR-based simultaneous localization and mapping (SLAM) can provide reliable and accurate location information for most automated vehicles. Loop closure detection plays an important role to eliminate accumulation errors in the SLAM system. Most existing methods usually extract descriptors from low-level features such as coordinate, normal, or reflection intensity of raw point clouds to represent scenes. It is difficult for these methods to keep robustness and accuracy in the case of occlusion and viewpoint changes. In this paper, we introduce the object-level information, as semantics to loop closure detection. Benefitting from semantics, our approach can achieve better performance in complex situations. Since the number of points contained in raw point cloud data is huge and redundant, we sample the point cloud with semantic information to reduce the point cloud density without losing effective information. Exhaustive experiments on the KITTI data set show our approach achieves competitive performance compared with the state-of-the-art methods.