# Online Range Image-based Pole Extractor for Long-term LiDAR Localization in Urban Environments

**Authors:** Hao Dong, Xieyuanli Chen, Cyrill Stachniss  
**arXiv:** [2108.08621](https://arxiv.org/abs/2108.08621)  
**Date:** 2021-08-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.08621v1/

## Key points

- Novel approach for reliable and accurate localization of mobile autonomous systems in urban environments
- Use of pole-like objects (traffic signs, poles, lamps) as landmarks for localization
- Utilization of range images generated from 3D LiDAR scans to extract poles in an online and computationally efficient manner
- Development of a range image-based pole extractor instead of directly processing raw point clouds
- Extraction process involves projecting the raw point cloud into a range image and identifying poles within that image
- Extracted poles are reprojected into a global coordinate system using ground-truth poses to build a global map
- Monte Carlo localization (MCL) employed during localization to update importance weights of particles by matching detected poles with those in the global map
- Proposed approach outperforms other state-of-the-art methods while running online without requiring a GPU
- Approach extracts more reliable poles leading to improved localization performance in various environments compared to baseline methods
- Approach can run efficiently without a GPU
- Open-source implementation and pole dataset provided for evaluation purposes
- Related work section discusses traditional approaches to robot localization using probabilistic state estimation techniques like Monte Carlo localization
- Introduction of innovative range image-based pole extraction method for long-term LiDAR localization in urban environments
- Leveraging range images instead of raw point clouds achieves accurate and fast pole extraction leading to improved localization performance

## Summary

This paper presents a novel approach for reliable and accurate localization of mobile autonomous systems in urban environments. The focus is on using pole-like objects, such as traffic signs, poles, and lamps, as landmarks for localization due to their distinctiveness and long-term stability. The proposed method utilizes range images generated from 3D LiDAR scans to extract poles in an online and computationally efficient manner. The main contribution of this work is the development of a range image-based pole extractor. Instead of directly processing the raw point clouds obtained from LiDAR sensors, the authors investigate the use of range images which are lighter and faster to process. Range images also retain neighborhood information implicitly in their 2D structure allowing for effective segmentation. The extraction process involves projecting the raw point cloud into a range image and then identifying poles within that image. These extracted poles are then reprojected into a global coordinate system using ground-truth poses to build a global map. During localization Monte Carlo localization (MCL) is employed to update the importance weights of particles by matching detected poles from online sensor data with those in the global map. Experimental results demonstrate that the proposed approach outperforms other state-of-the-art methods while running online without requiring a GPU. The paper makes three key claims: (i) the approach extracts more reliable poles compared to baseline methods leading to improved localization performance in various environments; (ii) it can run efficiently without a GPU; and (iii) it provides an open-source implementation along with a pole dataset for evaluating performance. The related work section discusses traditional approaches to robot localization using probabilistic state estimation techniques like Monte Carlo localization. LiDAR-based approaches are specifically focused on in this study. In summary, this paper introduces an innovative range image-based pole extraction method for long-term LiDAR localization in urban environments. By leveraging range images instead of raw point clouds, the proposed approach achieves accurate and fast pole extraction leading to improved localization performance. The availability of the open-source implementation and pole dataset further facilitates evaluation and adoption of the approach.