[Paper Review] A Survey on 3D LiDAR Localization for Autonomous Vehicles

Mahdi Elhousni, Xinming Huang|arXiv (Cornell University)|Jun 1, 2020

Robotics and Sensor-Based Localization57 references4 citations

TL;DR

This survey reviews 3D LiDAR-based localization methods for autonomous vehicles, categorizing them into 3D registration, feature-based, and deep learning approaches. It finds that while deep learning methods show strong performance on benchmark datasets, traditional methods like LOAM remain more practical for real-time deployment due to speed and reliability.

ABSTRACT

LiDAR sensors are becoming one of the most essential sensors in achieving full autonomy for self driving cars. LiDARs are able to produce rich, dense and precise spatial data, which can tremendously help in localizing and tracking a moving vehicle. In this paper, we review the latest finding in 3D LiDAR localization for autonomous driving cars, and analyse the results obtained by each method, in an effort to guide the research community towards the path that seems to be the most promising.

Motivation & Objective

• To review and analyze recent advances in 3D LiDAR localization for autonomous driving.
• To compare the performance of 3D registration, feature-based, and deep learning-based methods on the KITTI odometry benchmark.
• To evaluate the real-world deployability of these methods, focusing on efficiency and robustness.
• To guide future research by identifying the most promising pathways for accurate and real-time localization.

Proposed method

• Categorizes 3D LiDAR localization into three main approaches: 3D registration, 3D feature extraction, and deep learning-based methods.
• Reviews key techniques such as Iterative Closest Point (ICP), its variants (point-to-plane, Generalized ICP), and feature extraction using Normal Covariance Filter and Geometric Correspondence Rejector.
• Analyzes deep learning models including L3-Net and DeepICP, which use PointNet++ for feature extraction and learn residual corrections to improve odometry estimates.
• Employs a multi-loss training strategy in deep learning models, combining geometric consistency, odometry regression, and cross-entropy for regularization.
• Uses the KITTI odometry dataset for evaluation, focusing on sequences with pre-processed LiDAR scans and GPS/INS ground truth.
• Compares results on both training and test splits, excluding methods with loop closure to ensure fair comparison of core localization performance.

Experimental results

Research questions

• RQ1 How do 3D registration-based methods perform in terms of accuracy and computational efficiency for LiDAR localization?
• RQ2 What are the strengths and limitations of 3D feature-based methods in real-world autonomous driving scenarios?
• RQ3 To what extent do deep learning-based approaches improve localization accuracy compared to classical methods?
• RQ4 Why do state-of-the-art deep learning models still lag in real-time deployment despite strong benchmark results?
• RQ5 What factors make traditional methods like LOAM more suitable for real-world autonomous vehicle deployment than deep learning alternatives?

Key findings

• DeepICP reported the best average performance on the KITTI training dataset, outperforming other methods in terms of localization accuracy.
• Despite strong training results, DeepICP requires approximately 2 seconds per frame, making it infeasible for real-time autonomous vehicle operation.
• LOAM and its variants remain the most reliable and deployable solution due to their balance of accuracy, speed, and robustness in real-world conditions.
• Deep learning-based methods show high promise but have not yet demonstrated generalization on the KITTI test set, limiting claims of state-of-the-art performance.
• Feature-based methods like LOAM maintain strong performance due to their efficiency and stability, especially in dynamic outdoor environments.
• The lack of test set results for deep learning models prevents definitive conclusions about their real-world applicability, despite strong training results.