Simultaneous localization an‎d mapping: place recognition, rotation an‎d translation in deep learning based self-driving cars

8/26/2026 12:00:00 AM

In this research, a novel framework for simultaneous localization an‎d mapping (SLAM) based on LiDAR in autonomous vehicles is proposed, leveraging implicit neural representations an‎d signed distance functions (SDF) to generate accurate, stable, an‎d compact 3D maps in complex an‎d dynamic environments. The proposed method is designed to be efficient in terms of memory an‎d computational optimization while maintaining high accuracy. In this framework, hash voxels are employed as an efficient data structure for representing neural points, which not only reduces computational complexity but also enables fast querying of spatial data. Additionally, to enhance loop closure an‎d environmental feature detection, Siamese neural networks are utilized. These networks, by learning structural similarities in input data, enable effective matching of previous observations an‎d reduce cumulative errors. To improve map registration an‎d alignment accuracy, SDF is used to optimize the point cloud registration process. This approach leads to reduced alignment errors, increased stability of generated maps, an‎d improved accuracy in modeling complex environmental structures. Furthermore, to extract multi-scale features an‎d enhance mapping accuracy in challenging environments, wavelet transforms are applied. This technique, by decomposing information across different scales, facilitates a better understan‎ding of environmental details an‎d improves SLAM system performance. The proposed framework was eva‎luated on the KITTI dataset, an‎d experimental results demonstrate that this method not only achieves higher localization accuracy compared to existing approaches but also supports real-time execution. Moreover, the generated maps are not only superior in terms of accuracy an‎d coherence but also more compact in data volume an‎d offer higher efficiency. These characteristics make the proposed method suitable for practical applications in autonomous navigation, robotics, an‎d intelligent transportation systems.

/ , //

/

Mehdi Khoori

Dr. Abdollah Amirkhani