MoLi-PoseNet: Model-based indoor relocalization using deep pose regression from synthetic LiDAR scans

In indoor environments, the low availability of global navigation satellite systems (GNSS) makes the LiDAR pose initialization a challenge. Approximate spatial data, such as 3D models are easy to extract from building information model (BIM) or construct, and can be used for relocalization. This paper presents MoLi-PoseNet: a novel LiDAR relocalization method in indoor environments. The method estimates LiDAR poses using a surface-based 3D model and a convolutional regression network. In the training phase, synthetic LiDAR scans with ground truth poses generated by a LiDAR simulator placed in the 3D model are used to train the regression network. In the inference phase, the trained network is applied to real LiDAR scans to predict LiDAR poses with respect to the model. We test the performance of MoLi-PoseNet with regression tasks in three environments. The results show that MoLi-PoseNet can achieve meter-level relocalization accuracy in large indoor environments with low-level-of-detail models.