TL;DR
This paper introduces a novel Lidar-based localization system for autonomous vehicles that significantly improves accuracy and robustness in challenging urban environments. By leveraging semantic segmentation for feature extraction and a multi-stage optimization approach, the system achieves centimeter-level accuracy at a high frequency, outperforming existing methods while reducing reliance on expensive RTK-GNSS systems.
What Problem Does This Paper Address?
Precise and reliable localization is a cornerstone for autonomous driving, enabling critical functions like path planning and control. Current high-accuracy solutions often rely on expensive RTK-GNSS receivers and IMUs, which can be prohibitively costly for widespread adoption. Furthermore, these systems can struggle in urban canyons or areas with poor satellite reception, leading to accuracy degradation or complete failure. While Lidar offers a promising alternative for dense 3D mapping, achieving robust and accurate localization solely with Lidar, especially in complex and dynamic urban settings, remains a significant challenge. Existing Lidar-based methods can suffer from low output frequencies, insufficient accuracy, or poor performance in adverse weather or textureless environments. This paper aims to bridge this gap by developing a cost-effective, accurate, and robust Lidar localization system specifically for autonomous urban driving.
Methods
The proposed system integrates several novel components:
- Efficient Low-Level Semantic Segmentation-Based Feature Extraction: Instead of using raw point clouds, the system first extracts meaningful semantic features: ground, road-curb, surface, and edge. This is achieved through a series of efficient algorithms that leverage geometric properties of the Lidar data, reducing the computational burden and focusing on stable features. This segmentation is crucial for improving data association and robustness.
- Prior Information Considered Category Matching: To enhance data association, the system only matches features of the same category (e.g., edge features with edge features). This significantly reduces the search space and the likelihood of incorrect correspondences. Additionally, it incorporates prior knowledge about feature types (e.g., ground is planar, edges are vertical) to simplify the calculation of geometric primitives (lines and planes), improving efficiency and accuracy.
- Multi-Group-Step L-M Optimization: A standard Levenberg-Marquardt (L-M) optimization for pose estimation can be computationally intensive and prone to local optima. This paper proposes a multi-group-step L-M approach. It iteratively optimizes different pose variables (e.g., only translation and yaw for edge features, then adding others) based on the characteristics of the extracted features. This staged optimization uses the output of one step as the input for the next, leading to a more accurate final pose with reduced computational cost.
- Lidar and MEMS IMU Fusion: To achieve high-frequency (200Hz) and smooth localization, the system fuses Lidar-derived pose estimates with data from a low-cost MEMS IMU using an error-state Kalman filter. This fusion leverages the high accuracy of Lidar for state correction and the high frequency of IMU for state prediction, resulting in a robust and smooth trajectory.
- Efficient Prior Map Management: For large-scale environments, the system employs a tiled map management strategy with overlap between tiles. This allows for dynamic loading of only the relevant local map around the vehicle, saving computational resources.
Datasets
The paper describes experiments conducted using data collected from a Velodyne VLP-32C Lidar and a MEMS IMU installed on an unmanned vehicle. The test area covered approximately 3x4 km, including various challenging scenes like congested urban roads, narrow tunnels, and textureless highways. The total test mileage exceeded 1000 km. The authors do not mention using publicly available datasets for training or evaluation, but rather their own collected data.
Key Insights and Ideas
- Semantic Segmentation is Key for Robustness: By explicitly identifying and using semantic features like ground, road-curb, edges, and surfaces, the system gains robustness against dynamic objects and noise, which are common issues in urban environments.
- Leveraging Prior Knowledge Dramatically Improves Efficiency: Incorporating geometric priors about feature types (e.g., ground is planar) into the matching and optimization process significantly reduces computational complexity without sacrificing accuracy.
- Staged Optimization for Better Results: The multi-group-step L-M optimization is a clever way to tackle the high-dimensional pose estimation problem. By breaking it down and using intermediate results as priors, it avoids local minima and achieves better accuracy.
- Low-Cost IMU Fusion is Crucial for High Frequency: The fusion with a MEMS IMU is essential for achieving the high output frequency (200Hz) required for real-time autonomous driving, complementing the Lidar's accuracy.
Key Results
The proposed Lidar localization system demonstrates impressive performance:
- Accuracy: Achieves 3-5 cm in position and 0.05-0.1° in orientation RMS accuracy.
- Frequency: Outputs localization information at a high frequency of 200Hz.
- Robustness: Verified over 1000 km of testing in challenging scenes, including congested urban roads, narrow tunnels, textureless highways, and adverse weather conditions like rain and fog.
- Performance Comparison: Outperforms previous state-of-the-art Lidar-based localization systems, with translation errors reduced by approximately half compared to a reference system [41].
- Feature Extraction Speed: The segmentation algorithms are lightweight, with each frame processed in about 35ms.
Limitations and Future Work
- Map Dependency: The system relies on a pre-built prior map. Significant changes in the environment, such as construction or demolition, can degrade localization accuracy.
- Map Update Strategy: The authors identify the need for methods to automatically, efficiently, and cheaply update map datasets as a crucial area for future research.
- Data Association Methods: Exploring model-to-model data association instead of the current point-to-line or point-to-plane methods could further reduce computational complexity.
Summary updated