# Online Pole Segmentation on Range Images for Long-term LiDAR Localization in Urban Environments

**Authors:** Hao Dong, Xieyuanli Chen, Simo Särkkä, Cyrill Stachniss  
**arXiv:** [2208.07364](https://arxiv.org/abs/2208.07364)  
**Date:** 2022-08-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2208.07364v1/

## Key points

- Robust and accurate localization is crucial for mobile autonomous systems, especially in urban environments.
- Pole-like objects, such as traffic signs, poles, and lamps, are commonly used landmarks for localization due to their distinctiveness and long-term stability.
- The authors propose a novel approach for online pole extraction based on geometric features that is accurate and fast with minimal computational demands.
- The method operates directly on range images generated from 3D LiDAR scans, eliminating the need to process 3D point clouds explicitly.
- Extracted poles are used as pseudo labels to train a deep neural network for online range image-based pole segmentation.
- The authors conduct experiments on different datasets with varying LiDAR scanners, routes, and seasonal changes to evaluate the performance of their approach.
- Their methods outperform other state-of-the-art approaches in terms of localization accuracy.
- By leveraging pseudo pole labels extracted from multiple datasets, the learning-based method achieves even better localization results across different datasets compared to the geometry-based method.
- The authors have released their pole datasets to the public to facilitate further research and evaluation of pole extractors as well as implementation of their approach.
- They compare their approach's runtime performance with a baseline method proposed by Schaefer et al., demonstrating that it runs online at the sensor frame rate while significantly reducing processing time.
- This paper introduces an innovative approach for online pole extraction using range images from LiDAR scans which combines geometric features with deep learning to enable accurate and efficient localization in urban environments.

## Summary

Robust and accurate localization is crucial for mobile autonomous systems, especially in urban environments. Pole-like objects, such as traffic signs, poles, and lamps, are commonly used landmarks for localization due to their distinctiveness and long-term stability. In this paper, the authors propose a novel approach for online pole extraction based on geometric features that is both accurate and fast with minimal computational demands. The method operates directly on range images generated from 3D LiDAR scans, eliminating the need to process 3D point clouds explicitly. This enables efficient pole extraction for each scan. Additionally, the extracted poles are used as pseudo labels to train a deep neural network for online range image-based pole segmentation. To evaluate the performance of their approach, the authors conduct experiments on different datasets with varying LiDAR scanners, routes, and seasonal changes. The results demonstrate that their methods outperform other state-of-the-art approaches in terms of localization accuracy. Furthermore, by leveraging pseudo pole labels extracted from multiple datasets, the learning-based method achieves even better localization results across different datasets compared to the geometry-based method. To facilitate further research and evaluation of pole extractors as well as implementation of their approach, the authors have released their pole datasets to the public. In addition to presenting their methodology and experimental results, they compare their approach's runtime performance with a baseline method proposed by Schaefer et al., demonstrating that it runs online at the sensor frame rate while significantly reducing processing time. Overall, this paper introduces an innovative approach for online pole extraction using range images from LiDAR scans which combines geometric features with deep learning to enable accurate and efficient localization in urban environments. The experimental results validate the effectiveness of their methods compared to existing approaches.