# Why semantics matters: A deep study on semantic particle-filtering localization in a LiDAR semantic pole-map

**Authors:** Yuming Huang, Yi Gu, Chengzhong Xu, Hui Kong  
**arXiv:** [2305.14038](https://arxiv.org/abs/2305.14038)  
**Date:** 2023-05-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.14038v1/

## Key points

- Pole-like structures serve as important landmarks for autonomous vehicle localization in urban and suburban areas
- This study focuses on creating a precise map, called a pole-map, using pole-like structures as primary localization landmarks
- The research leverages semantic information associated with these structures, not just their geometric properties
- A novel mask-range transformer network is used for semantic segmentation of pole-like structures
- Multi-layer semantic pole-map is generated by aggregating detected poles from all frames
- Semantic particle-filtering localization scheme proposed for vehicle localization using the semantic pole-map
- Theoretical analysis shows how incorporating semantics enhances particle-filter localization performance
- Empirical validation on the SemanticKITTI dataset demonstrates improved performance with semantics in the presence of uncertainties in odometry predictions and online observations
- Considering semantics in mapping and localization tasks involving pole-like structures is crucial for accurate and robust autonomous vehicle localization in urban environments.

## Summary

In urban and suburban areas, pole-like structures such as utility poles or tree trunks are commonly found. These structures serve as important landmarks for the localization of autonomous vehicles, as they provide geometric information that can be used in conjunction with sensor measurements and maps. This study focuses on creating a precise map, known as a pole-map, specifically designed for autonomous vehicles or robots using pole-like structures as the primary localization landmarks. Unlike previous mapping or localization methods that solely rely on poles' geometric properties, this research leverages the semantic information associated with these structures. A novel mask-range transformer network is employed to achieve semantic segmentation in a mask-classification paradigm. By extracting semantics for pole-like structures in each frame, a multi-layer semantic pole-map is generated by aggregating detected poles from all frames. Using the semantic pole-map, the researchers propose a semantic particle-filtering localization scheme for vehicle localization. Theoretical analysis reveals how incorporating semantic information can enhance particle-filter localization performance. Empirical validation on the public SemanticKITTI dataset demonstrates that particle-filtering localization with semantics outperforms its counterpart without semantics when faced with uncertainties in odometry predictions and online observations. Overall, this study highlights the importance of considering semantics in mapping and localization tasks involving pole-like structures. By exploiting both geometric and semantic information, autonomous vehicles can achieve more accurate and robust localization in urban environments where poles are prevalent.