SMLAD: Simultaneous Matching, Localization, and Detection for Intelligent Vehicle From LiDAR Map With Semantic Likelihood Model

Matching with a semantic map can feed accurate prior information into perception and localization modules in the intelligent vehicle system. This article proposes a simultaneous matching, localization, and detection (SMLAD) method from a semantic LiDAR map for intelligent vehicles, which integrates the separated steps of map matching, vehicle localization, and semantic detection into maximum a posteriori probability (MAP) inference problem in the particle filter framework. Firstly, semantic objects, such as cubes, poles, walls, and ground, are extracted from the query point cloud by LiDAR segmentation method. Then semantic likelihood models (SLMs) are generated from the extracted semantic objects with kernel density estimation (KDE). Finally, the likelihood between the map and the query point cloud is calculated from SLMs that contributes to weight assignment in the particle filter framework. And the proposed point-to-likelihood association allows simultaneous map matching, localization, and detection afterwards. The proposed SMLAD method has been validated with both the public KITTI dataset and the real tests. Experimental results demonstrate that the proposed SMLAD method can achieve up to 10-cm localization accuracy on both test datasets, and the detection results can be derived simultaneously by mapping accurate semantic objects from the map into the query point cloud.