MoLi-PoseGAN: Model-based Indoor Relocalization Using GAN and Deep Pose Regression from Synthetic LiDAR Scans

Model-based LiDAR localization systems provide accurate pose estimation but they highly rely on the accuracy of 3D models. The inaccurate parts of 3D models will introduce localization errors. This paper presents a novel LiDAR relocalization method using synthetic LiDAR scans generated from a LiDAR generative adversarial network. Synthetic LiDAR scans are generated in a 3D model using the poses of a set of real LiDAR scans and to train a change detection network together with the corresponding real LiDAR scans to detect differences between the 3D models and the real environments. The synthetic and real data, and the differences are used in a generative adversarial network to correct the difference in synthetic LiDAR scans. A pose regression network is then trained using the corrected synthetic LiDAR scans and tested using new real LiDAR data. Experimental results show the proposed method achieves a higher accuracy than previous model-based pose regression methods.